LCOV - code coverage report
Current view: top level - lib/nvmf - rdma.c (source / functions) Hit Total Coverage
Test: ut_cov_unit.info Lines: 668 2398 27.9 %
Date: 2024-07-12 13:42:51 Functions: 30 119 25.2 %

          Line data    Source code
       1             : /*   SPDX-License-Identifier: BSD-3-Clause
       2             :  *   Copyright (C) 2016 Intel Corporation. All rights reserved.
       3             :  *   Copyright (c) 2019-2021 Mellanox Technologies LTD. All rights reserved.
       4             :  *   Copyright (c) 2021-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
       5             :  */
       6             : 
       7             : #include "spdk/stdinc.h"
       8             : 
       9             : #include "spdk/config.h"
      10             : #include "spdk/thread.h"
      11             : #include "spdk/likely.h"
      12             : #include "spdk/nvmf_transport.h"
      13             : #include "spdk/string.h"
      14             : #include "spdk/trace.h"
      15             : #include "spdk/tree.h"
      16             : #include "spdk/util.h"
      17             : 
      18             : #include "spdk_internal/assert.h"
      19             : #include "spdk/log.h"
      20             : #include "spdk_internal/rdma_provider.h"
      21             : #include "spdk_internal/rdma_utils.h"
      22             : 
      23             : #include "nvmf_internal.h"
      24             : #include "transport.h"
      25             : 
      26             : #include "spdk_internal/trace_defs.h"
      27             : 
      28             : struct spdk_nvme_rdma_hooks g_nvmf_hooks = {};
      29             : const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma;
      30             : 
      31             : /*
      32             :  RDMA Connection Resource Defaults
      33             :  */
      34             : #define NVMF_DEFAULT_MSDBD              16
      35             : #define NVMF_DEFAULT_TX_SGE             SPDK_NVMF_MAX_SGL_ENTRIES
      36             : #define NVMF_DEFAULT_RSP_SGE            1
      37             : #define NVMF_DEFAULT_RX_SGE             2
      38             : 
      39             : #define NVMF_RDMA_MAX_EVENTS_PER_POLL   32
      40             : 
      41             : SPDK_STATIC_ASSERT(NVMF_DEFAULT_MSDBD <= SPDK_NVMF_MAX_SGL_ENTRIES,
      42             :                    "MSDBD must not exceed SPDK_NVMF_MAX_SGL_ENTRIES");
      43             : 
      44             : /* The RDMA completion queue size */
      45             : #define DEFAULT_NVMF_RDMA_CQ_SIZE       4096
      46             : #define MAX_WR_PER_QP(queue_depth)      (queue_depth * 3 + 2)
      47             : 
      48             : enum spdk_nvmf_rdma_request_state {
      49             :         /* The request is not currently in use */
      50             :         RDMA_REQUEST_STATE_FREE = 0,
      51             : 
      52             :         /* Initial state when request first received */
      53             :         RDMA_REQUEST_STATE_NEW,
      54             : 
      55             :         /* The request is queued until a data buffer is available. */
      56             :         RDMA_REQUEST_STATE_NEED_BUFFER,
      57             : 
      58             :         /* The request is waiting on RDMA queue depth availability
      59             :          * to transfer data from the host to the controller.
      60             :          */
      61             :         RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING,
      62             : 
      63             :         /* The request is currently transferring data from the host to the controller. */
      64             :         RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER,
      65             : 
      66             :         /* The request is ready to execute at the block device */
      67             :         RDMA_REQUEST_STATE_READY_TO_EXECUTE,
      68             : 
      69             :         /* The request is currently executing at the block device */
      70             :         RDMA_REQUEST_STATE_EXECUTING,
      71             : 
      72             :         /* The request finished executing at the block device */
      73             :         RDMA_REQUEST_STATE_EXECUTED,
      74             : 
      75             :         /* The request is waiting on RDMA queue depth availability
      76             :          * to transfer data from the controller to the host.
      77             :          */
      78             :         RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING,
      79             : 
      80             :         /* The request is waiting on RDMA queue depth availability
      81             :          * to send response to the host.
      82             :          */
      83             :         RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING,
      84             : 
      85             :         /* The request is ready to send a completion */
      86             :         RDMA_REQUEST_STATE_READY_TO_COMPLETE,
      87             : 
      88             :         /* The request is currently transferring data from the controller to the host. */
      89             :         RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST,
      90             : 
      91             :         /* The request currently has an outstanding completion without an
      92             :          * associated data transfer.
      93             :          */
      94             :         RDMA_REQUEST_STATE_COMPLETING,
      95             : 
      96             :         /* The request completed and can be marked free. */
      97             :         RDMA_REQUEST_STATE_COMPLETED,
      98             : 
      99             :         /* Terminator */
     100             :         RDMA_REQUEST_NUM_STATES,
     101             : };
     102             : 
     103           2 : SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA)
     104             : {
     105           0 :         spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r');
     106             : 
     107           0 :         struct spdk_trace_tpoint_opts opts[] = {
     108             :                 {
     109             :                         "RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW,
     110             :                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 1,
     111             :                         {
     112             :                                 { "qpair", SPDK_TRACE_ARG_TYPE_PTR, 8 },
     113             :                                 { "qd", SPDK_TRACE_ARG_TYPE_INT, 4 }
     114             :                         }
     115             :                 },
     116             :                 {
     117             :                         "RDMA_REQ_COMPLETED", TRACE_RDMA_REQUEST_STATE_COMPLETED,
     118             :                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     119             :                         {
     120             :                                 { "qpair", SPDK_TRACE_ARG_TYPE_PTR, 8 },
     121             :                                 { "qd", SPDK_TRACE_ARG_TYPE_INT, 4 }
     122             :                         }
     123             :                 },
     124             :         };
     125             : 
     126           0 :         spdk_trace_register_description_ext(opts, SPDK_COUNTOF(opts));
     127           0 :         spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER,
     128             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     129             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     130           0 :         spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H",
     131             :                                         TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING,
     132             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     133             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     134           0 :         spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C",
     135             :                                         TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING,
     136             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     137             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     138           0 :         spdk_trace_register_description("RDMA_REQ_TX_H2C",
     139             :                                         TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER,
     140             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     141             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     142           0 :         spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE",
     143             :                                         TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE,
     144             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     145             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     146           0 :         spdk_trace_register_description("RDMA_REQ_EXECUTING",
     147             :                                         TRACE_RDMA_REQUEST_STATE_EXECUTING,
     148             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     149             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     150           0 :         spdk_trace_register_description("RDMA_REQ_EXECUTED",
     151             :                                         TRACE_RDMA_REQUEST_STATE_EXECUTED,
     152             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     153             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     154           0 :         spdk_trace_register_description("RDMA_REQ_RDY2COMPL_PEND",
     155             :                                         TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING,
     156             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     157             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     158           0 :         spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL",
     159             :                                         TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE,
     160             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     161             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     162           0 :         spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H",
     163             :                                         TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST,
     164             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     165             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     166           0 :         spdk_trace_register_description("RDMA_REQ_COMPLETING",
     167             :                                         TRACE_RDMA_REQUEST_STATE_COMPLETING,
     168             :                                         OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
     169             :                                         SPDK_TRACE_ARG_TYPE_PTR, "qpair");
     170             : 
     171           0 :         spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE,
     172             :                                         OWNER_TYPE_NONE, OBJECT_NONE, 0,
     173             :                                         SPDK_TRACE_ARG_TYPE_INT, "");
     174           0 :         spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT,
     175             :                                         OWNER_TYPE_NONE, OBJECT_NONE, 0,
     176             :                                         SPDK_TRACE_ARG_TYPE_INT, "type");
     177           0 :         spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT,
     178             :                                         OWNER_TYPE_NONE, OBJECT_NONE, 0,
     179             :                                         SPDK_TRACE_ARG_TYPE_INT, "type");
     180           0 :         spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT,
     181             :                                         OWNER_TYPE_NONE, OBJECT_NONE, 0,
     182             :                                         SPDK_TRACE_ARG_TYPE_INT, "");
     183           0 :         spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY,
     184             :                                         OWNER_TYPE_NONE, OBJECT_NONE, 0,
     185             :                                         SPDK_TRACE_ARG_TYPE_INT, "");
     186           0 : }
     187             : 
     188             : enum spdk_nvmf_rdma_wr_type {
     189             :         RDMA_WR_TYPE_RECV,
     190             :         RDMA_WR_TYPE_SEND,
     191             :         RDMA_WR_TYPE_DATA,
     192             : };
     193             : 
     194             : struct spdk_nvmf_rdma_wr {
     195             :         /* Uses enum spdk_nvmf_rdma_wr_type */
     196             :         uint8_t type;
     197             : };
     198             : 
     199             : /* This structure holds commands as they are received off the wire.
     200             :  * It must be dynamically paired with a full request object
     201             :  * (spdk_nvmf_rdma_request) to service a request. It is separate
     202             :  * from the request because RDMA does not appear to order
     203             :  * completions, so occasionally we'll get a new incoming
     204             :  * command when there aren't any free request objects.
     205             :  */
     206             : struct spdk_nvmf_rdma_recv {
     207             :         struct ibv_recv_wr                      wr;
     208             :         struct ibv_sge                          sgl[NVMF_DEFAULT_RX_SGE];
     209             : 
     210             :         struct spdk_nvmf_rdma_qpair             *qpair;
     211             : 
     212             :         /* In-capsule data buffer */
     213             :         uint8_t                                 *buf;
     214             : 
     215             :         struct spdk_nvmf_rdma_wr                rdma_wr;
     216             :         uint64_t                                receive_tsc;
     217             : 
     218             :         STAILQ_ENTRY(spdk_nvmf_rdma_recv)       link;
     219             : };
     220             : 
     221             : struct spdk_nvmf_rdma_request_data {
     222             :         struct ibv_send_wr              wr;
     223             :         struct ibv_sge                  sgl[SPDK_NVMF_MAX_SGL_ENTRIES];
     224             : };
     225             : 
     226             : struct spdk_nvmf_rdma_request {
     227             :         struct spdk_nvmf_request                req;
     228             : 
     229             :         bool                                    fused_failed;
     230             : 
     231             :         struct spdk_nvmf_rdma_wr                data_wr;
     232             :         struct spdk_nvmf_rdma_wr                rsp_wr;
     233             : 
     234             :         /* Uses enum spdk_nvmf_rdma_request_state */
     235             :         uint8_t                                 state;
     236             : 
     237             :         /* Data offset in req.iov */
     238             :         uint32_t                                offset;
     239             : 
     240             :         struct spdk_nvmf_rdma_recv              *recv;
     241             : 
     242             :         struct {
     243             :                 struct  ibv_send_wr             wr;
     244             :                 struct  ibv_sge                 sgl[NVMF_DEFAULT_RSP_SGE];
     245             :         } rsp;
     246             : 
     247             :         uint16_t                                iovpos;
     248             :         uint16_t                                num_outstanding_data_wr;
     249             :         /* Used to split Write IO with multi SGL payload */
     250             :         uint16_t                                num_remaining_data_wr;
     251             :         uint64_t                                receive_tsc;
     252             :         struct spdk_nvmf_rdma_request           *fused_pair;
     253             :         STAILQ_ENTRY(spdk_nvmf_rdma_request)    state_link;
     254             :         struct ibv_send_wr                      *remaining_tranfer_in_wrs;
     255             :         struct ibv_send_wr                      *transfer_wr;
     256             :         struct spdk_nvmf_rdma_request_data      data;
     257             : };
     258             : 
     259             : struct spdk_nvmf_rdma_resource_opts {
     260             :         struct spdk_nvmf_rdma_qpair     *qpair;
     261             :         /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */
     262             :         void                            *qp;
     263             :         struct spdk_rdma_utils_mem_map  *map;
     264             :         uint32_t                        max_queue_depth;
     265             :         uint32_t                        in_capsule_data_size;
     266             :         bool                            shared;
     267             : };
     268             : 
     269             : struct spdk_nvmf_rdma_resources {
     270             :         /* Array of size "max_queue_depth" containing RDMA requests. */
     271             :         struct spdk_nvmf_rdma_request           *reqs;
     272             : 
     273             :         /* Array of size "max_queue_depth" containing RDMA recvs. */
     274             :         struct spdk_nvmf_rdma_recv              *recvs;
     275             : 
     276             :         /* Array of size "max_queue_depth" containing 64 byte capsules
     277             :          * used for receive.
     278             :          */
     279             :         union nvmf_h2c_msg                      *cmds;
     280             : 
     281             :         /* Array of size "max_queue_depth" containing 16 byte completions
     282             :          * to be sent back to the user.
     283             :          */
     284             :         union nvmf_c2h_msg                      *cpls;
     285             : 
     286             :         /* Array of size "max_queue_depth * InCapsuleDataSize" containing
     287             :          * buffers to be used for in capsule data.
     288             :          */
     289             :         void                                    *bufs;
     290             : 
     291             :         /* Receives that are waiting for a request object */
     292             :         STAILQ_HEAD(, spdk_nvmf_rdma_recv)      incoming_queue;
     293             : 
     294             :         /* Queue to track free requests */
     295             :         STAILQ_HEAD(, spdk_nvmf_rdma_request)   free_queue;
     296             : };
     297             : 
     298             : typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair);
     299             : 
     300             : typedef void (*spdk_poller_destroy_cb)(void *ctx);
     301             : 
     302             : struct spdk_nvmf_rdma_ibv_event_ctx {
     303             :         struct spdk_nvmf_rdma_qpair                     *rqpair;
     304             :         spdk_nvmf_rdma_qpair_ibv_event                  cb_fn;
     305             :         /* Link to other ibv events associated with this qpair */
     306             :         STAILQ_ENTRY(spdk_nvmf_rdma_ibv_event_ctx)      link;
     307             : };
     308             : 
     309             : struct spdk_nvmf_rdma_qpair {
     310             :         struct spdk_nvmf_qpair                  qpair;
     311             : 
     312             :         struct spdk_nvmf_rdma_device            *device;
     313             :         struct spdk_nvmf_rdma_poller            *poller;
     314             : 
     315             :         struct spdk_rdma_provider_qp            *rdma_qp;
     316             :         struct rdma_cm_id                       *cm_id;
     317             :         struct spdk_rdma_provider_srq           *srq;
     318             :         struct rdma_cm_id                       *listen_id;
     319             : 
     320             :         /* Cache the QP number to improve QP search by RB tree. */
     321             :         uint32_t                                qp_num;
     322             : 
     323             :         /* The maximum number of I/O outstanding on this connection at one time */
     324             :         uint16_t                                max_queue_depth;
     325             : 
     326             :         /* The maximum number of active RDMA READ and ATOMIC operations at one time */
     327             :         uint16_t                                max_read_depth;
     328             : 
     329             :         /* The maximum number of RDMA SEND operations at one time */
     330             :         uint32_t                                max_send_depth;
     331             : 
     332             :         /* The current number of outstanding WRs from this qpair's
     333             :          * recv queue. Should not exceed device->attr.max_queue_depth.
     334             :          */
     335             :         uint16_t                                current_recv_depth;
     336             : 
     337             :         /* The current number of active RDMA READ operations */
     338             :         uint16_t                                current_read_depth;
     339             : 
     340             :         /* The current number of posted WRs from this qpair's
     341             :          * send queue. Should not exceed max_send_depth.
     342             :          */
     343             :         uint32_t                                current_send_depth;
     344             : 
     345             :         /* The maximum number of SGEs per WR on the send queue */
     346             :         uint32_t                                max_send_sge;
     347             : 
     348             :         /* The maximum number of SGEs per WR on the recv queue */
     349             :         uint32_t                                max_recv_sge;
     350             : 
     351             :         struct spdk_nvmf_rdma_resources         *resources;
     352             : 
     353             :         STAILQ_HEAD(, spdk_nvmf_rdma_request)   pending_rdma_read_queue;
     354             : 
     355             :         STAILQ_HEAD(, spdk_nvmf_rdma_request)   pending_rdma_write_queue;
     356             : 
     357             :         STAILQ_HEAD(, spdk_nvmf_rdma_request)   pending_rdma_send_queue;
     358             : 
     359             :         /* Number of requests not in the free state */
     360             :         uint32_t                                qd;
     361             : 
     362             :         bool                                    ibv_in_error_state;
     363             : 
     364             :         RB_ENTRY(spdk_nvmf_rdma_qpair)          node;
     365             : 
     366             :         STAILQ_ENTRY(spdk_nvmf_rdma_qpair)      recv_link;
     367             : 
     368             :         STAILQ_ENTRY(spdk_nvmf_rdma_qpair)      send_link;
     369             : 
     370             :         /* Points to the a request that has fuse bits set to
     371             :          * SPDK_NVME_CMD_FUSE_FIRST, when the qpair is waiting
     372             :          * for the request that has SPDK_NVME_CMD_FUSE_SECOND.
     373             :          */
     374             :         struct spdk_nvmf_rdma_request           *fused_first;
     375             : 
     376             :         /*
     377             :          * io_channel which is used to destroy qpair when it is removed from poll group
     378             :          */
     379             :         struct spdk_io_channel          *destruct_channel;
     380             : 
     381             :         /* List of ibv async events */
     382             :         STAILQ_HEAD(, spdk_nvmf_rdma_ibv_event_ctx)     ibv_events;
     383             : 
     384             :         /* Lets us know that we have received the last_wqe event. */
     385             :         bool                                    last_wqe_reached;
     386             : 
     387             :         /* Indicate that nvmf_rdma_close_qpair is called */
     388             :         bool                                    to_close;
     389             : };
     390             : 
     391             : struct spdk_nvmf_rdma_poller_stat {
     392             :         uint64_t                                completions;
     393             :         uint64_t                                polls;
     394             :         uint64_t                                idle_polls;
     395             :         uint64_t                                requests;
     396             :         uint64_t                                request_latency;
     397             :         uint64_t                                pending_free_request;
     398             :         uint64_t                                pending_rdma_read;
     399             :         uint64_t                                pending_rdma_write;
     400             :         uint64_t                                pending_rdma_send;
     401             :         struct spdk_rdma_provider_qp_stats      qp_stats;
     402             : };
     403             : 
     404             : struct spdk_nvmf_rdma_poller {
     405             :         struct spdk_nvmf_rdma_device            *device;
     406             :         struct spdk_nvmf_rdma_poll_group        *group;
     407             : 
     408             :         int                                     num_cqe;
     409             :         int                                     required_num_wr;
     410             :         struct ibv_cq                           *cq;
     411             : 
     412             :         /* The maximum number of I/O outstanding on the shared receive queue at one time */
     413             :         uint16_t                                max_srq_depth;
     414             :         bool                                    need_destroy;
     415             : 
     416             :         /* Shared receive queue */
     417             :         struct spdk_rdma_provider_srq           *srq;
     418             : 
     419             :         struct spdk_nvmf_rdma_resources         *resources;
     420             :         struct spdk_nvmf_rdma_poller_stat       stat;
     421             : 
     422             :         spdk_poller_destroy_cb                  destroy_cb;
     423             :         void                                    *destroy_cb_ctx;
     424             : 
     425             :         RB_HEAD(qpairs_tree, spdk_nvmf_rdma_qpair) qpairs;
     426             : 
     427             :         STAILQ_HEAD(, spdk_nvmf_rdma_qpair)     qpairs_pending_recv;
     428             : 
     429             :         STAILQ_HEAD(, spdk_nvmf_rdma_qpair)     qpairs_pending_send;
     430             : 
     431             :         TAILQ_ENTRY(spdk_nvmf_rdma_poller)      link;
     432             : };
     433             : 
     434             : struct spdk_nvmf_rdma_poll_group_stat {
     435             :         uint64_t                                pending_data_buffer;
     436             : };
     437             : 
     438             : struct spdk_nvmf_rdma_poll_group {
     439             :         struct spdk_nvmf_transport_poll_group           group;
     440             :         struct spdk_nvmf_rdma_poll_group_stat           stat;
     441             :         TAILQ_HEAD(, spdk_nvmf_rdma_poller)             pollers;
     442             :         TAILQ_ENTRY(spdk_nvmf_rdma_poll_group)          link;
     443             : };
     444             : 
     445             : struct spdk_nvmf_rdma_conn_sched {
     446             :         struct spdk_nvmf_rdma_poll_group *next_admin_pg;
     447             :         struct spdk_nvmf_rdma_poll_group *next_io_pg;
     448             : };
     449             : 
     450             : /* Assuming rdma_cm uses just one protection domain per ibv_context. */
     451             : struct spdk_nvmf_rdma_device {
     452             :         struct ibv_device_attr                  attr;
     453             :         struct ibv_context                      *context;
     454             : 
     455             :         struct spdk_rdma_utils_mem_map          *map;
     456             :         struct ibv_pd                           *pd;
     457             : 
     458             :         int                                     num_srq;
     459             :         bool                                    need_destroy;
     460             :         bool                                    ready_to_destroy;
     461             :         bool                                    is_ready;
     462             : 
     463             :         TAILQ_ENTRY(spdk_nvmf_rdma_device)      link;
     464             : };
     465             : 
     466             : struct spdk_nvmf_rdma_port {
     467             :         const struct spdk_nvme_transport_id     *trid;
     468             :         struct rdma_cm_id                       *id;
     469             :         struct spdk_nvmf_rdma_device            *device;
     470             :         TAILQ_ENTRY(spdk_nvmf_rdma_port)        link;
     471             : };
     472             : 
     473             : struct rdma_transport_opts {
     474             :         int             num_cqe;
     475             :         uint32_t        max_srq_depth;
     476             :         bool            no_srq;
     477             :         bool            no_wr_batching;
     478             :         int             acceptor_backlog;
     479             : };
     480             : 
     481             : struct spdk_nvmf_rdma_transport {
     482             :         struct spdk_nvmf_transport      transport;
     483             :         struct rdma_transport_opts      rdma_opts;
     484             : 
     485             :         struct spdk_nvmf_rdma_conn_sched conn_sched;
     486             : 
     487             :         struct rdma_event_channel       *event_channel;
     488             : 
     489             :         struct spdk_mempool             *data_wr_pool;
     490             : 
     491             :         struct spdk_poller              *accept_poller;
     492             : 
     493             :         /* fields used to poll RDMA/IB events */
     494             :         nfds_t                  npoll_fds;
     495             :         struct pollfd           *poll_fds;
     496             : 
     497             :         TAILQ_HEAD(, spdk_nvmf_rdma_device)     devices;
     498             :         TAILQ_HEAD(, spdk_nvmf_rdma_port)       ports;
     499             :         TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups;
     500             : 
     501             :         /* ports that are removed unexpectedly and need retry listen */
     502             :         TAILQ_HEAD(, spdk_nvmf_rdma_port)               retry_ports;
     503             : };
     504             : 
     505             : struct poller_manage_ctx {
     506             :         struct spdk_nvmf_rdma_transport         *rtransport;
     507             :         struct spdk_nvmf_rdma_poll_group        *rgroup;
     508             :         struct spdk_nvmf_rdma_poller            *rpoller;
     509             :         struct spdk_nvmf_rdma_device            *device;
     510             : 
     511             :         struct spdk_thread                      *thread;
     512             :         volatile int                            *inflight_op_counter;
     513             : };
     514             : 
     515             : static const struct spdk_json_object_decoder rdma_transport_opts_decoder[] = {
     516             :         {
     517             :                 "num_cqe", offsetof(struct rdma_transport_opts, num_cqe),
     518             :                 spdk_json_decode_int32, true
     519             :         },
     520             :         {
     521             :                 "max_srq_depth", offsetof(struct rdma_transport_opts, max_srq_depth),
     522             :                 spdk_json_decode_uint32, true
     523             :         },
     524             :         {
     525             :                 "no_srq", offsetof(struct rdma_transport_opts, no_srq),
     526             :                 spdk_json_decode_bool, true
     527             :         },
     528             :         {
     529             :                 "no_wr_batching", offsetof(struct rdma_transport_opts, no_wr_batching),
     530             :                 spdk_json_decode_bool, true
     531             :         },
     532             :         {
     533             :                 "acceptor_backlog", offsetof(struct rdma_transport_opts, acceptor_backlog),
     534             :                 spdk_json_decode_int32, true
     535             :         },
     536             : };
     537             : 
     538             : static int
     539           2 : nvmf_rdma_qpair_compare(struct spdk_nvmf_rdma_qpair *rqpair1, struct spdk_nvmf_rdma_qpair *rqpair2)
     540             : {
     541           2 :         return rqpair1->qp_num < rqpair2->qp_num ? -1 : rqpair1->qp_num > rqpair2->qp_num;
     542             : }
     543             : 
     544           0 : RB_GENERATE_STATIC(qpairs_tree, spdk_nvmf_rdma_qpair, node, nvmf_rdma_qpair_compare);
     545             : 
     546             : static bool nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport,
     547             :                                       struct spdk_nvmf_rdma_request *rdma_req);
     548             : 
     549             : static void _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport,
     550             :                                  struct spdk_nvmf_rdma_poller *rpoller);
     551             : 
     552             : static void _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport,
     553             :                                  struct spdk_nvmf_rdma_poller *rpoller);
     554             : 
     555             : static void _nvmf_rdma_remove_destroyed_device(void *c);
     556             : 
     557             : static inline enum spdk_nvme_media_error_status_code
     558           0 : nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) {
     559             :         enum spdk_nvme_media_error_status_code result;
     560           0 :         switch (err_type)
     561             :         {
     562           0 :         case SPDK_DIF_REFTAG_ERROR:
     563           0 :                 result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR;
     564           0 :                 break;
     565           0 :         case SPDK_DIF_APPTAG_ERROR:
     566           0 :                 result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR;
     567           0 :                 break;
     568           0 :         case SPDK_DIF_GUARD_ERROR:
     569           0 :                 result = SPDK_NVME_SC_GUARD_CHECK_ERROR;
     570           0 :                 break;
     571           0 :         default:
     572           0 :                 SPDK_UNREACHABLE();
     573             :         }
     574             : 
     575           0 :         return result;
     576             : }
     577             : 
     578             : /*
     579             :  * Return data_wrs to pool starting from \b data_wr
     580             :  * Request's own response and data WR are excluded
     581             :  */
     582             : static void
     583           7 : _nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req,
     584             :                              struct ibv_send_wr *data_wr,
     585             :                              struct spdk_mempool *pool)
     586             : {
     587           7 :         struct spdk_nvmf_rdma_request_data      *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES];
     588             :         struct spdk_nvmf_rdma_request_data      *nvmf_data;
     589             :         struct ibv_send_wr                      *next_send_wr;
     590           7 :         uint64_t                                req_wrid = (uint64_t)&rdma_req->data_wr;
     591           7 :         uint32_t                                num_wrs = 0;
     592             : 
     593          15 :         while (data_wr && data_wr->wr_id == req_wrid) {
     594           8 :                 nvmf_data = SPDK_CONTAINEROF(data_wr, struct spdk_nvmf_rdma_request_data, wr);
     595           8 :                 memset(nvmf_data->sgl, 0, sizeof(data_wr->sg_list[0]) * data_wr->num_sge);
     596           8 :                 data_wr->num_sge = 0;
     597           8 :                 next_send_wr = data_wr->next;
     598           8 :                 if (data_wr != &rdma_req->data.wr) {
     599           1 :                         data_wr->next = NULL;
     600           1 :                         assert(num_wrs < SPDK_NVMF_MAX_SGL_ENTRIES);
     601           1 :                         work_requests[num_wrs] = nvmf_data;
     602           1 :                         num_wrs++;
     603             :                 }
     604           8 :                 data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : next_send_wr;
     605             :         }
     606             : 
     607           7 :         if (num_wrs) {
     608           1 :                 spdk_mempool_put_bulk(pool, (void **) work_requests, num_wrs);
     609             :         }
     610           7 : }
     611             : 
     612             : static void
     613           7 : nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req,
     614             :                             struct spdk_nvmf_rdma_transport *rtransport)
     615             : {
     616           7 :         rdma_req->num_outstanding_data_wr = 0;
     617             : 
     618           7 :         _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool);
     619             : 
     620           7 :         if (rdma_req->remaining_tranfer_in_wrs) {
     621           0 :                 _nvmf_rdma_request_free_data(rdma_req, rdma_req->remaining_tranfer_in_wrs,
     622             :                                              rtransport->data_wr_pool);
     623           0 :                 rdma_req->remaining_tranfer_in_wrs = NULL;
     624             :         }
     625             : 
     626           7 :         rdma_req->data.wr.next = NULL;
     627           7 :         rdma_req->rsp.wr.next = NULL;
     628           7 : }
     629             : 
     630             : static void
     631           0 : nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req)
     632             : {
     633           0 :         SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool);
     634           0 :         if (req->req.cmd) {
     635           0 :                 SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode);
     636             :         }
     637           0 :         if (req->recv) {
     638           0 :                 SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id);
     639             :         }
     640           0 : }
     641             : 
     642             : static void
     643           0 : nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair)
     644             : {
     645             :         int i;
     646             : 
     647           0 :         SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid);
     648           0 :         for (i = 0; i < rqpair->max_queue_depth; i++) {
     649           0 :                 if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) {
     650           0 :                         nvmf_rdma_dump_request(&rqpair->resources->reqs[i]);
     651             :                 }
     652             :         }
     653           0 : }
     654             : 
     655             : static void
     656           1 : nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources)
     657             : {
     658           1 :         spdk_free(resources->cmds);
     659           1 :         spdk_free(resources->cpls);
     660           1 :         spdk_free(resources->bufs);
     661           1 :         spdk_free(resources->reqs);
     662           1 :         spdk_free(resources->recvs);
     663           1 :         free(resources);
     664           1 : }
     665             : 
     666             : 
     667             : static struct spdk_nvmf_rdma_resources *
     668           1 : nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts)
     669             : {
     670             :         struct spdk_nvmf_rdma_resources         *resources;
     671             :         struct spdk_nvmf_rdma_request           *rdma_req;
     672             :         struct spdk_nvmf_rdma_recv              *rdma_recv;
     673           1 :         struct spdk_rdma_provider_qp            *qp = NULL;
     674           1 :         struct spdk_rdma_provider_srq           *srq = NULL;
     675           1 :         struct ibv_recv_wr                      *bad_wr = NULL;
     676           1 :         struct spdk_rdma_utils_memory_translation translation;
     677             :         uint32_t                                i;
     678           1 :         int                                     rc = 0;
     679             : 
     680           1 :         resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources));
     681           1 :         if (!resources) {
     682           0 :                 SPDK_ERRLOG("Unable to allocate resources for receive queue.\n");
     683           0 :                 return NULL;
     684             :         }
     685             : 
     686           1 :         resources->reqs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->reqs),
     687             :                                        0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
     688           1 :         resources->recvs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->recvs),
     689             :                                         0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
     690           1 :         resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds),
     691             :                                        0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
     692           1 :         resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls),
     693             :                                        0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
     694             : 
     695           1 :         if (opts->in_capsule_data_size > 0) {
     696           1 :                 resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size,
     697             :                                                0x1000, NULL, SPDK_ENV_LCORE_ID_ANY,
     698             :                                                SPDK_MALLOC_DMA);
     699             :         }
     700             : 
     701           1 :         if (!resources->reqs || !resources->recvs || !resources->cmds ||
     702           1 :             !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) {
     703           0 :                 SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n");
     704           0 :                 goto cleanup;
     705             :         }
     706             : 
     707           1 :         SPDK_DEBUGLOG(rdma, "Command Array: %p Length: %lx\n",
     708             :                       resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds));
     709           1 :         SPDK_DEBUGLOG(rdma, "Completion Array: %p Length: %lx\n",
     710             :                       resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls));
     711           1 :         if (resources->bufs) {
     712           1 :                 SPDK_DEBUGLOG(rdma, "In Capsule Data Array: %p Length: %x\n",
     713             :                               resources->bufs, opts->max_queue_depth *
     714             :                               opts->in_capsule_data_size);
     715             :         }
     716             : 
     717             :         /* Initialize queues */
     718           1 :         STAILQ_INIT(&resources->incoming_queue);
     719           1 :         STAILQ_INIT(&resources->free_queue);
     720             : 
     721           1 :         if (opts->shared) {
     722           1 :                 srq = (struct spdk_rdma_provider_srq *)opts->qp;
     723             :         } else {
     724           0 :                 qp = (struct spdk_rdma_provider_qp *)opts->qp;
     725             :         }
     726             : 
     727         129 :         for (i = 0; i < opts->max_queue_depth; i++) {
     728         128 :                 rdma_recv = &resources->recvs[i];
     729         128 :                 rdma_recv->qpair = opts->qpair;
     730             : 
     731             :                 /* Set up memory to receive commands */
     732         128 :                 if (resources->bufs) {
     733         128 :                         rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i *
     734         128 :                                                   opts->in_capsule_data_size));
     735             :                 }
     736             : 
     737         128 :                 rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV;
     738             : 
     739         128 :                 rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i];
     740         128 :                 rdma_recv->sgl[0].length = sizeof(resources->cmds[i]);
     741         128 :                 rc = spdk_rdma_utils_get_translation(opts->map, &resources->cmds[i], sizeof(resources->cmds[i]),
     742             :                                                      &translation);
     743         128 :                 if (rc) {
     744           0 :                         goto cleanup;
     745             :                 }
     746         128 :                 rdma_recv->sgl[0].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation);
     747         128 :                 rdma_recv->wr.num_sge = 1;
     748             : 
     749         128 :                 if (rdma_recv->buf) {
     750         128 :                         rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf;
     751         128 :                         rdma_recv->sgl[1].length = opts->in_capsule_data_size;
     752         128 :                         rc = spdk_rdma_utils_get_translation(opts->map, rdma_recv->buf, opts->in_capsule_data_size,
     753             :                                                              &translation);
     754         128 :                         if (rc) {
     755           0 :                                 goto cleanup;
     756             :                         }
     757         128 :                         rdma_recv->sgl[1].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation);
     758         128 :                         rdma_recv->wr.num_sge++;
     759             :                 }
     760             : 
     761         128 :                 rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr;
     762         128 :                 rdma_recv->wr.sg_list = rdma_recv->sgl;
     763         128 :                 if (srq) {
     764           0 :                         spdk_rdma_provider_srq_queue_recv_wrs(srq, &rdma_recv->wr);
     765             :                 } else {
     766         128 :                         spdk_rdma_provider_qp_queue_recv_wrs(qp, &rdma_recv->wr);
     767             :                 }
     768             :         }
     769             : 
     770         129 :         for (i = 0; i < opts->max_queue_depth; i++) {
     771         128 :                 rdma_req = &resources->reqs[i];
     772             : 
     773         128 :                 if (opts->qpair != NULL) {
     774         128 :                         rdma_req->req.qpair = &opts->qpair->qpair;
     775             :                 } else {
     776           0 :                         rdma_req->req.qpair = NULL;
     777             :                 }
     778         128 :                 rdma_req->req.cmd = NULL;
     779         128 :                 rdma_req->req.iovcnt = 0;
     780         128 :                 rdma_req->req.stripped_data = NULL;
     781             : 
     782             :                 /* Set up memory to send responses */
     783         128 :                 rdma_req->req.rsp = &resources->cpls[i];
     784             : 
     785         128 :                 rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i];
     786         128 :                 rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]);
     787         128 :                 rc = spdk_rdma_utils_get_translation(opts->map, &resources->cpls[i], sizeof(resources->cpls[i]),
     788             :                                                      &translation);
     789         128 :                 if (rc) {
     790           0 :                         goto cleanup;
     791             :                 }
     792         128 :                 rdma_req->rsp.sgl[0].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation);
     793             : 
     794         128 :                 rdma_req->rsp_wr.type = RDMA_WR_TYPE_SEND;
     795         128 :                 rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp_wr;
     796         128 :                 rdma_req->rsp.wr.next = NULL;
     797         128 :                 rdma_req->rsp.wr.opcode = IBV_WR_SEND;
     798         128 :                 rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED;
     799         128 :                 rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl;
     800         128 :                 rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl);
     801             : 
     802             :                 /* Set up memory for data buffers */
     803         128 :                 rdma_req->data_wr.type = RDMA_WR_TYPE_DATA;
     804         128 :                 rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data_wr;
     805         128 :                 rdma_req->data.wr.next = NULL;
     806         128 :                 rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED;
     807         128 :                 rdma_req->data.wr.sg_list = rdma_req->data.sgl;
     808         128 :                 rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl);
     809             : 
     810             :                 /* Initialize request state to FREE */
     811         128 :                 rdma_req->state = RDMA_REQUEST_STATE_FREE;
     812         128 :                 STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link);
     813             :         }
     814             : 
     815           1 :         if (srq) {
     816           0 :                 rc = spdk_rdma_provider_srq_flush_recv_wrs(srq, &bad_wr);
     817             :         } else {
     818           1 :                 rc = spdk_rdma_provider_qp_flush_recv_wrs(qp, &bad_wr);
     819             :         }
     820             : 
     821           1 :         if (rc) {
     822           0 :                 goto cleanup;
     823             :         }
     824             : 
     825           1 :         return resources;
     826             : 
     827           0 : cleanup:
     828           0 :         nvmf_rdma_resources_destroy(resources);
     829           0 :         return NULL;
     830             : }
     831             : 
     832             : static void
     833           0 : nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair)
     834             : {
     835             :         struct spdk_nvmf_rdma_ibv_event_ctx *ctx, *tctx;
     836           0 :         STAILQ_FOREACH_SAFE(ctx, &rqpair->ibv_events, link, tctx) {
     837           0 :                 ctx->rqpair = NULL;
     838             :                 /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_ibv_event */
     839           0 :                 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link);
     840             :         }
     841           0 : }
     842             : 
     843             : static void nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller);
     844             : 
     845             : static void
     846           0 : nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair)
     847             : {
     848             :         struct spdk_nvmf_rdma_recv      *rdma_recv, *recv_tmp;
     849           0 :         struct ibv_recv_wr              *bad_recv_wr = NULL;
     850             :         int                             rc;
     851             : 
     852           0 :         spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair);
     853             : 
     854           0 :         if (rqpair->qd != 0) {
     855           0 :                 struct spdk_nvmf_qpair *qpair = &rqpair->qpair;
     856           0 :                 struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport,
     857             :                                 struct spdk_nvmf_rdma_transport, transport);
     858             :                 struct spdk_nvmf_rdma_request *req;
     859           0 :                 uint32_t i, max_req_count = 0;
     860             : 
     861           0 :                 SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd);
     862             : 
     863           0 :                 if (rqpair->srq == NULL) {
     864           0 :                         nvmf_rdma_dump_qpair_contents(rqpair);
     865           0 :                         max_req_count = rqpair->max_queue_depth;
     866           0 :                 } else if (rqpair->poller && rqpair->resources) {
     867           0 :                         max_req_count = rqpair->poller->max_srq_depth;
     868             :                 }
     869             : 
     870           0 :                 SPDK_DEBUGLOG(rdma, "Release incomplete requests\n");
     871           0 :                 for (i = 0; i < max_req_count; i++) {
     872           0 :                         req = &rqpair->resources->reqs[i];
     873           0 :                         if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) {
     874             :                                 /* nvmf_rdma_request_process checks qpair ibv and internal state
     875             :                                  * and completes a request */
     876           0 :                                 nvmf_rdma_request_process(rtransport, req);
     877             :                         }
     878             :                 }
     879           0 :                 assert(rqpair->qd == 0);
     880             :         }
     881             : 
     882           0 :         if (rqpair->poller) {
     883           0 :                 RB_REMOVE(qpairs_tree, &rqpair->poller->qpairs, rqpair);
     884             : 
     885           0 :                 if (rqpair->srq != NULL && rqpair->resources != NULL) {
     886             :                         /* Drop all received but unprocessed commands for this queue and return them to SRQ */
     887           0 :                         STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) {
     888           0 :                                 if (rqpair == rdma_recv->qpair) {
     889           0 :                                         STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link);
     890           0 :                                         spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, &rdma_recv->wr);
     891           0 :                                         rc = spdk_rdma_provider_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr);
     892           0 :                                         if (rc) {
     893           0 :                                                 SPDK_ERRLOG("Unable to re-post rx descriptor\n");
     894             :                                         }
     895             :                                 }
     896             :                         }
     897             :                 }
     898             :         }
     899             : 
     900           0 :         if (rqpair->cm_id) {
     901           0 :                 if (rqpair->rdma_qp != NULL) {
     902           0 :                         spdk_rdma_provider_qp_destroy(rqpair->rdma_qp);
     903           0 :                         rqpair->rdma_qp = NULL;
     904             :                 }
     905             : 
     906           0 :                 if (rqpair->poller != NULL && rqpair->srq == NULL) {
     907           0 :                         rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth);
     908             :                 }
     909             :         }
     910             : 
     911           0 :         if (rqpair->srq == NULL && rqpair->resources != NULL) {
     912           0 :                 nvmf_rdma_resources_destroy(rqpair->resources);
     913             :         }
     914             : 
     915           0 :         nvmf_rdma_qpair_clean_ibv_events(rqpair);
     916             : 
     917           0 :         if (rqpair->destruct_channel) {
     918           0 :                 spdk_put_io_channel(rqpair->destruct_channel);
     919           0 :                 rqpair->destruct_channel = NULL;
     920             :         }
     921             : 
     922           0 :         if (rqpair->poller && rqpair->poller->need_destroy && RB_EMPTY(&rqpair->poller->qpairs)) {
     923           0 :                 nvmf_rdma_poller_destroy(rqpair->poller);
     924             :         }
     925             : 
     926             :         /* destroy cm_id last so cma device will not be freed before we destroy the cq. */
     927           0 :         if (rqpair->cm_id) {
     928           0 :                 rdma_destroy_id(rqpair->cm_id);
     929             :         }
     930             : 
     931           0 :         free(rqpair);
     932           0 : }
     933             : 
     934             : static int
     935           5 : nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device)
     936             : {
     937             :         struct spdk_nvmf_rdma_poller    *rpoller;
     938             :         int                             rc, num_cqe, required_num_wr;
     939             : 
     940             :         /* Enlarge CQ size dynamically */
     941           5 :         rpoller = rqpair->poller;
     942           5 :         required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth);
     943           5 :         num_cqe = rpoller->num_cqe;
     944           5 :         if (num_cqe < required_num_wr) {
     945           4 :                 num_cqe = spdk_max(num_cqe * 2, required_num_wr);
     946           4 :                 num_cqe = spdk_min(num_cqe, device->attr.max_cqe);
     947             :         }
     948             : 
     949           5 :         if (rpoller->num_cqe != num_cqe) {
     950           4 :                 if (device->context->device->transport_type == IBV_TRANSPORT_IWARP) {
     951           1 :                         SPDK_ERRLOG("iWARP doesn't support CQ resize. Current capacity %u, required %u\n"
     952             :                                     "Using CQ of insufficient size may lead to CQ overrun\n", rpoller->num_cqe, num_cqe);
     953           1 :                         return -1;
     954             :                 }
     955           3 :                 if (required_num_wr > device->attr.max_cqe) {
     956           1 :                         SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n",
     957             :                                     required_num_wr, device->attr.max_cqe);
     958           1 :                         return -1;
     959             :                 }
     960             : 
     961           2 :                 SPDK_DEBUGLOG(rdma, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe);
     962           2 :                 rc = ibv_resize_cq(rpoller->cq, num_cqe);
     963           2 :                 if (rc) {
     964           1 :                         SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno));
     965           1 :                         return -1;
     966             :                 }
     967             : 
     968           1 :                 rpoller->num_cqe = num_cqe;
     969             :         }
     970             : 
     971           2 :         rpoller->required_num_wr = required_num_wr;
     972           2 :         return 0;
     973             : }
     974             : 
     975             : static int
     976           0 : nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair)
     977             : {
     978             :         struct spdk_nvmf_rdma_qpair             *rqpair;
     979             :         struct spdk_nvmf_rdma_transport         *rtransport;
     980             :         struct spdk_nvmf_transport              *transport;
     981           0 :         struct spdk_nvmf_rdma_resource_opts     opts;
     982             :         struct spdk_nvmf_rdma_device            *device;
     983           0 :         struct spdk_rdma_provider_qp_init_attr  qp_init_attr = {};
     984             : 
     985           0 :         rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
     986           0 :         device = rqpair->device;
     987             : 
     988           0 :         qp_init_attr.qp_context = rqpair;
     989           0 :         qp_init_attr.pd         = device->pd;
     990           0 :         qp_init_attr.send_cq    = rqpair->poller->cq;
     991           0 :         qp_init_attr.recv_cq    = rqpair->poller->cq;
     992             : 
     993           0 :         if (rqpair->srq) {
     994           0 :                 qp_init_attr.srq                = rqpair->srq->srq;
     995             :         } else {
     996           0 :                 qp_init_attr.cap.max_recv_wr    = rqpair->max_queue_depth;
     997             :         }
     998             : 
     999             :         /* SEND, READ, and WRITE operations */
    1000           0 :         qp_init_attr.cap.max_send_wr    = (uint32_t)rqpair->max_queue_depth * 2;
    1001           0 :         qp_init_attr.cap.max_send_sge   = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE);
    1002           0 :         qp_init_attr.cap.max_recv_sge   = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE);
    1003           0 :         qp_init_attr.stats              = &rqpair->poller->stat.qp_stats;
    1004             : 
    1005           0 :         if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) {
    1006           0 :                 SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n");
    1007           0 :                 goto error;
    1008             :         }
    1009             : 
    1010           0 :         rqpair->rdma_qp = spdk_rdma_provider_qp_create(rqpair->cm_id, &qp_init_attr);
    1011           0 :         if (!rqpair->rdma_qp) {
    1012           0 :                 goto error;
    1013             :         }
    1014             : 
    1015           0 :         rqpair->qp_num = rqpair->rdma_qp->qp->qp_num;
    1016             : 
    1017           0 :         rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2),
    1018             :                                           qp_init_attr.cap.max_send_wr);
    1019           0 :         rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge);
    1020           0 :         rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge);
    1021           0 :         spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair);
    1022           0 :         SPDK_DEBUGLOG(rdma, "New RDMA Connection: %p\n", qpair);
    1023             : 
    1024           0 :         if (rqpair->poller->srq == NULL) {
    1025           0 :                 rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
    1026           0 :                 transport = &rtransport->transport;
    1027             : 
    1028           0 :                 opts.qp = rqpair->rdma_qp;
    1029           0 :                 opts.map = device->map;
    1030           0 :                 opts.qpair = rqpair;
    1031           0 :                 opts.shared = false;
    1032           0 :                 opts.max_queue_depth = rqpair->max_queue_depth;
    1033           0 :                 opts.in_capsule_data_size = transport->opts.in_capsule_data_size;
    1034             : 
    1035           0 :                 rqpair->resources = nvmf_rdma_resources_create(&opts);
    1036             : 
    1037           0 :                 if (!rqpair->resources) {
    1038           0 :                         SPDK_ERRLOG("Unable to allocate resources for receive queue.\n");
    1039           0 :                         rdma_destroy_qp(rqpair->cm_id);
    1040           0 :                         goto error;
    1041             :                 }
    1042             :         } else {
    1043           0 :                 rqpair->resources = rqpair->poller->resources;
    1044             :         }
    1045             : 
    1046           0 :         rqpair->current_recv_depth = 0;
    1047           0 :         STAILQ_INIT(&rqpair->pending_rdma_read_queue);
    1048           0 :         STAILQ_INIT(&rqpair->pending_rdma_write_queue);
    1049           0 :         STAILQ_INIT(&rqpair->pending_rdma_send_queue);
    1050           0 :         rqpair->qpair.queue_depth = 0;
    1051             : 
    1052           0 :         return 0;
    1053             : 
    1054           0 : error:
    1055           0 :         rdma_destroy_id(rqpair->cm_id);
    1056           0 :         rqpair->cm_id = NULL;
    1057           0 :         return -1;
    1058             : }
    1059             : 
    1060             : /* Append the given recv wr structure to the resource structs outstanding recvs list. */
    1061             : /* This function accepts either a single wr or the first wr in a linked list. */
    1062             : static void
    1063           6 : nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first)
    1064             : {
    1065           6 :         struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
    1066             :                         struct spdk_nvmf_rdma_transport, transport);
    1067             : 
    1068           6 :         if (rqpair->srq != NULL) {
    1069           0 :                 spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, first);
    1070             :         } else {
    1071           6 :                 if (spdk_rdma_provider_qp_queue_recv_wrs(rqpair->rdma_qp, first)) {
    1072           6 :                         STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link);
    1073             :                 }
    1074             :         }
    1075             : 
    1076           6 :         if (rtransport->rdma_opts.no_wr_batching) {
    1077           0 :                 _poller_submit_recvs(rtransport, rqpair->poller);
    1078             :         }
    1079           6 : }
    1080             : 
    1081             : static int
    1082           4 : request_transfer_in(struct spdk_nvmf_request *req)
    1083             : {
    1084             :         struct spdk_nvmf_rdma_request   *rdma_req;
    1085             :         struct spdk_nvmf_qpair          *qpair;
    1086             :         struct spdk_nvmf_rdma_qpair     *rqpair;
    1087             :         struct spdk_nvmf_rdma_transport *rtransport;
    1088             : 
    1089           4 :         qpair = req->qpair;
    1090           4 :         rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
    1091           4 :         rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
    1092           4 :         rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
    1093             :                                       struct spdk_nvmf_rdma_transport, transport);
    1094             : 
    1095           4 :         assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
    1096           4 :         assert(rdma_req != NULL);
    1097             : 
    1098           4 :         if (spdk_rdma_provider_qp_queue_send_wrs(rqpair->rdma_qp, rdma_req->transfer_wr)) {
    1099           4 :                 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link);
    1100             :         }
    1101           4 :         if (rtransport->rdma_opts.no_wr_batching) {
    1102           0 :                 _poller_submit_sends(rtransport, rqpair->poller);
    1103             :         }
    1104             : 
    1105           4 :         assert(rqpair->current_read_depth + rdma_req->num_outstanding_data_wr <= rqpair->max_read_depth);
    1106           4 :         rqpair->current_read_depth += rdma_req->num_outstanding_data_wr;
    1107           4 :         assert(rqpair->current_send_depth + rdma_req->num_outstanding_data_wr <= rqpair->max_send_depth);
    1108           4 :         rqpair->current_send_depth += rdma_req->num_outstanding_data_wr;
    1109           4 :         return 0;
    1110             : }
    1111             : 
    1112             : static inline void
    1113           0 : nvmf_rdma_request_reset_transfer_in(struct spdk_nvmf_rdma_request *rdma_req,
    1114             :                                     struct spdk_nvmf_rdma_transport *rtransport)
    1115             : {
    1116             :         /* Put completed WRs back to pool and move transfer_wr pointer */
    1117           0 :         _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool);
    1118           0 :         rdma_req->transfer_wr = rdma_req->remaining_tranfer_in_wrs;
    1119           0 :         rdma_req->remaining_tranfer_in_wrs = NULL;
    1120           0 :         rdma_req->num_outstanding_data_wr = rdma_req->num_remaining_data_wr;
    1121           0 :         rdma_req->num_remaining_data_wr = 0;
    1122           0 : }
    1123             : 
    1124             : static inline int
    1125           0 : request_prepare_transfer_in_part(struct spdk_nvmf_request *req, uint32_t num_reads_available)
    1126             : {
    1127             :         struct spdk_nvmf_rdma_request   *rdma_req;
    1128             :         struct ibv_send_wr              *wr;
    1129             :         uint32_t i;
    1130             : 
    1131           0 :         rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
    1132             : 
    1133           0 :         assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
    1134           0 :         assert(rdma_req != NULL);
    1135           0 :         assert(num_reads_available > 0);
    1136           0 :         assert(rdma_req->num_outstanding_data_wr > num_reads_available);
    1137           0 :         wr = rdma_req->transfer_wr;
    1138             : 
    1139           0 :         for (i = 0; i < num_reads_available - 1; i++) {
    1140           0 :                 wr = wr->next;
    1141             :         }
    1142             : 
    1143           0 :         rdma_req->remaining_tranfer_in_wrs = wr->next;
    1144           0 :         rdma_req->num_remaining_data_wr = rdma_req->num_outstanding_data_wr - num_reads_available;
    1145           0 :         rdma_req->num_outstanding_data_wr = num_reads_available;
    1146             :         /* Break chain of WRs to send only part. Once this portion completes, we continue sending RDMA_READs */
    1147           0 :         wr->next = NULL;
    1148             : 
    1149           0 :         return 0;
    1150             : }
    1151             : 
    1152             : static int
    1153           6 : request_transfer_out(struct spdk_nvmf_request *req, int *data_posted)
    1154             : {
    1155           6 :         int                             num_outstanding_data_wr = 0;
    1156             :         struct spdk_nvmf_rdma_request   *rdma_req;
    1157             :         struct spdk_nvmf_qpair          *qpair;
    1158             :         struct spdk_nvmf_rdma_qpair     *rqpair;
    1159             :         struct spdk_nvme_cpl            *rsp;
    1160           6 :         struct ibv_send_wr              *first = NULL;
    1161             :         struct spdk_nvmf_rdma_transport *rtransport;
    1162             : 
    1163           6 :         *data_posted = 0;
    1164           6 :         qpair = req->qpair;
    1165           6 :         rsp = &req->rsp->nvme_cpl;
    1166           6 :         rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
    1167           6 :         rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
    1168           6 :         rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
    1169             :                                       struct spdk_nvmf_rdma_transport, transport);
    1170             : 
    1171             :         /* Advance our sq_head pointer */
    1172           6 :         if (qpair->sq_head == qpair->sq_head_max) {
    1173           6 :                 qpair->sq_head = 0;
    1174             :         } else {
    1175           0 :                 qpair->sq_head++;
    1176             :         }
    1177           6 :         rsp->sqhd = qpair->sq_head;
    1178             : 
    1179             :         /* queue the capsule for the recv buffer */
    1180           6 :         assert(rdma_req->recv != NULL);
    1181             : 
    1182           6 :         nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr);
    1183             : 
    1184           6 :         rdma_req->recv = NULL;
    1185           6 :         assert(rqpair->current_recv_depth > 0);
    1186           6 :         rqpair->current_recv_depth--;
    1187             : 
    1188             :         /* Build the response which consists of optional
    1189             :          * RDMA WRITEs to transfer data, plus an RDMA SEND
    1190             :          * containing the response.
    1191             :          */
    1192           6 :         first = &rdma_req->rsp.wr;
    1193             : 
    1194           6 :         if (spdk_unlikely(rsp->status.sc != SPDK_NVME_SC_SUCCESS)) {
    1195             :                 /* On failure, data was not read from the controller. So clear the
    1196             :                  * number of outstanding data WRs to zero.
    1197             :                  */
    1198           1 :                 rdma_req->num_outstanding_data_wr = 0;
    1199           5 :         } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
    1200           1 :                 first = rdma_req->transfer_wr;
    1201           1 :                 *data_posted = 1;
    1202           1 :                 num_outstanding_data_wr = rdma_req->num_outstanding_data_wr;
    1203             :         }
    1204           6 :         if (spdk_rdma_provider_qp_queue_send_wrs(rqpair->rdma_qp, first)) {
    1205           6 :                 STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link);
    1206             :         }
    1207           6 :         if (rtransport->rdma_opts.no_wr_batching) {
    1208           0 :                 _poller_submit_sends(rtransport, rqpair->poller);
    1209             :         }
    1210             : 
    1211             :         /* +1 for the rsp wr */
    1212           6 :         assert(rqpair->current_send_depth + num_outstanding_data_wr + 1 <= rqpair->max_send_depth);
    1213           6 :         rqpair->current_send_depth += num_outstanding_data_wr + 1;
    1214             : 
    1215           6 :         return 0;
    1216             : }
    1217             : 
    1218             : static int
    1219           0 : nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair)
    1220             : {
    1221           0 :         struct spdk_nvmf_rdma_accept_private_data       accept_data;
    1222           0 :         struct rdma_conn_param                          ctrlr_event_data = {};
    1223             :         int                                             rc;
    1224             : 
    1225           0 :         accept_data.recfmt = 0;
    1226           0 :         accept_data.crqsize = rqpair->max_queue_depth;
    1227             : 
    1228           0 :         ctrlr_event_data.private_data = &accept_data;
    1229           0 :         ctrlr_event_data.private_data_len = sizeof(accept_data);
    1230           0 :         if (id->ps == RDMA_PS_TCP) {
    1231           0 :                 ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */
    1232           0 :                 ctrlr_event_data.initiator_depth = rqpair->max_read_depth;
    1233             :         }
    1234             : 
    1235             :         /* Configure infinite retries for the initiator side qpair.
    1236             :          * We need to pass this value to the initiator to prevent the
    1237             :          * initiator side NIC from completing SEND requests back to the
    1238             :          * initiator with status rnr_retry_count_exceeded. */
    1239           0 :         ctrlr_event_data.rnr_retry_count = 0x7;
    1240             : 
    1241             :         /* When qpair is created without use of rdma cm API, an additional
    1242             :          * information must be provided to initiator in the connection response:
    1243             :          * whether qpair is using SRQ and its qp_num
    1244             :          * Fields below are ignored by rdma cm if qpair has been
    1245             :          * created using rdma cm API. */
    1246           0 :         ctrlr_event_data.srq = rqpair->srq ? 1 : 0;
    1247           0 :         ctrlr_event_data.qp_num = rqpair->qp_num;
    1248             : 
    1249           0 :         rc = spdk_rdma_provider_qp_accept(rqpair->rdma_qp, &ctrlr_event_data);
    1250           0 :         if (rc) {
    1251           0 :                 SPDK_ERRLOG("Error %d on spdk_rdma_provider_qp_accept\n", errno);
    1252             :         } else {
    1253           0 :                 SPDK_DEBUGLOG(rdma, "Sent back the accept\n");
    1254             :         }
    1255             : 
    1256           0 :         return rc;
    1257             : }
    1258             : 
    1259             : static void
    1260           0 : nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error)
    1261             : {
    1262           0 :         struct spdk_nvmf_rdma_reject_private_data       rej_data;
    1263             : 
    1264           0 :         rej_data.recfmt = 0;
    1265           0 :         rej_data.sts = error;
    1266             : 
    1267           0 :         rdma_reject(id, &rej_data, sizeof(rej_data));
    1268           0 : }
    1269             : 
    1270             : static int
    1271           0 : nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event)
    1272             : {
    1273             :         struct spdk_nvmf_rdma_transport *rtransport;
    1274           0 :         struct spdk_nvmf_rdma_qpair     *rqpair = NULL;
    1275             :         struct spdk_nvmf_rdma_port      *port;
    1276           0 :         struct rdma_conn_param          *rdma_param = NULL;
    1277           0 :         const struct spdk_nvmf_rdma_request_private_data *private_data = NULL;
    1278             :         uint16_t                        max_queue_depth;
    1279             :         uint16_t                        max_read_depth;
    1280             : 
    1281           0 :         rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
    1282             : 
    1283           0 :         assert(event->id != NULL); /* Impossible. Can't even reject the connection. */
    1284           0 :         assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */
    1285             : 
    1286           0 :         rdma_param = &event->param.conn;
    1287           0 :         if (rdma_param->private_data == NULL ||
    1288           0 :             rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) {
    1289           0 :                 SPDK_ERRLOG("connect request: no private data provided\n");
    1290           0 :                 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH);
    1291           0 :                 return -1;
    1292             :         }
    1293             : 
    1294           0 :         private_data = rdma_param->private_data;
    1295           0 :         if (private_data->recfmt != 0) {
    1296           0 :                 SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n");
    1297           0 :                 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT);
    1298           0 :                 return -1;
    1299             :         }
    1300             : 
    1301           0 :         SPDK_DEBUGLOG(rdma, "Connect Recv on fabric intf name %s, dev_name %s\n",
    1302             :                       event->id->verbs->device->name, event->id->verbs->device->dev_name);
    1303             : 
    1304           0 :         port = event->listen_id->context;
    1305           0 :         SPDK_DEBUGLOG(rdma, "Listen Id was %p with verbs %p. ListenAddr: %p\n",
    1306             :                       event->listen_id, event->listen_id->verbs, port);
    1307             : 
    1308             :         /* Figure out the supported queue depth. This is a multi-step process
    1309             :          * that takes into account hardware maximums, host provided values,
    1310             :          * and our target's internal memory limits */
    1311             : 
    1312           0 :         SPDK_DEBUGLOG(rdma, "Calculating Queue Depth\n");
    1313             : 
    1314             :         /* Start with the maximum queue depth allowed by the target */
    1315           0 :         max_queue_depth = rtransport->transport.opts.max_queue_depth;
    1316           0 :         max_read_depth = rtransport->transport.opts.max_queue_depth;
    1317           0 :         SPDK_DEBUGLOG(rdma, "Target Max Queue Depth: %d\n",
    1318             :                       rtransport->transport.opts.max_queue_depth);
    1319             : 
    1320             :         /* Next check the local NIC's hardware limitations */
    1321           0 :         SPDK_DEBUGLOG(rdma,
    1322             :                       "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n",
    1323             :                       port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom);
    1324           0 :         max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr);
    1325           0 :         max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom);
    1326             : 
    1327             :         /* Next check the remote NIC's hardware limitations */
    1328           0 :         SPDK_DEBUGLOG(rdma,
    1329             :                       "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n",
    1330             :                       rdma_param->initiator_depth, rdma_param->responder_resources);
    1331             :         /* from man3 rdma_get_cm_event
    1332             :          * responder_resources - Specifies the number of responder resources that is requested by the recipient.
    1333             :          * The responder_resources field must match the initiator depth specified by the remote node when running
    1334             :          * the rdma_connect and rdma_accept functions. */
    1335           0 :         if (rdma_param->responder_resources != 0) {
    1336           0 :                 if (private_data->qid) {
    1337           0 :                         SPDK_DEBUGLOG(rdma, "Host (Initiator) is not allowed to use RDMA operations,"
    1338             :                                       " responder_resources must be 0 but set to %u\n",
    1339             :                                       rdma_param->responder_resources);
    1340             :                 } else {
    1341           0 :                         SPDK_WARNLOG("Host (Initiator) is not allowed to use RDMA operations,"
    1342             :                                      " responder_resources must be 0 but set to %u\n",
    1343             :                                      rdma_param->responder_resources);
    1344             :                 }
    1345             :         }
    1346             :         /* from man3 rdma_get_cm_event
    1347             :          * initiator_depth - Specifies the maximum number of outstanding RDMA read operations that the recipient holds.
    1348             :          * The initiator_depth field must match the responder resources specified by the remote node when running
    1349             :          * the rdma_connect and rdma_accept functions. */
    1350           0 :         if (rdma_param->initiator_depth == 0) {
    1351           0 :                 SPDK_ERRLOG("Host (Initiator) doesn't support RDMA_READ or atomic operations\n");
    1352           0 :                 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_IRD);
    1353           0 :                 return -1;
    1354             :         }
    1355           0 :         max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth);
    1356             : 
    1357           0 :         SPDK_DEBUGLOG(rdma, "Host Receive Queue Size: %d\n", private_data->hrqsize);
    1358           0 :         SPDK_DEBUGLOG(rdma, "Host Send Queue Size: %d\n", private_data->hsqsize);
    1359           0 :         max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize);
    1360           0 :         max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1);
    1361             : 
    1362           0 :         SPDK_DEBUGLOG(rdma, "Final Negotiated Queue Depth: %d R/W Depth: %d\n",
    1363             :                       max_queue_depth, max_read_depth);
    1364             : 
    1365           0 :         rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair));
    1366           0 :         if (rqpair == NULL) {
    1367           0 :                 SPDK_ERRLOG("Could not allocate new connection.\n");
    1368           0 :                 nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES);
    1369           0 :                 return -1;
    1370             :         }
    1371             : 
    1372           0 :         rqpair->device = port->device;
    1373           0 :         rqpair->max_queue_depth = max_queue_depth;
    1374           0 :         rqpair->max_read_depth = max_read_depth;
    1375           0 :         rqpair->cm_id = event->id;
    1376           0 :         rqpair->listen_id = event->listen_id;
    1377           0 :         rqpair->qpair.transport = transport;
    1378           0 :         STAILQ_INIT(&rqpair->ibv_events);
    1379             :         /* use qid from the private data to determine the qpair type
    1380             :            qid will be set to the appropriate value when the controller is created */
    1381           0 :         rqpair->qpair.qid = private_data->qid;
    1382             : 
    1383           0 :         event->id->context = &rqpair->qpair;
    1384             : 
    1385           0 :         spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair);
    1386             : 
    1387           0 :         return 0;
    1388             : }
    1389             : 
    1390             : static inline void
    1391          28 : nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next,
    1392             :                    enum spdk_nvme_data_transfer xfer)
    1393             : {
    1394          28 :         if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
    1395          24 :                 wr->opcode = IBV_WR_RDMA_WRITE;
    1396          24 :                 wr->send_flags = 0;
    1397          24 :                 wr->next = next;
    1398           4 :         } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) {
    1399           4 :                 wr->opcode = IBV_WR_RDMA_READ;
    1400           4 :                 wr->send_flags = IBV_SEND_SIGNALED;
    1401           4 :                 wr->next = NULL;
    1402             :         } else {
    1403           0 :                 assert(0);
    1404             :         }
    1405          28 : }
    1406             : 
    1407             : static int
    1408           6 : nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport,
    1409             :                        struct spdk_nvmf_rdma_request *rdma_req,
    1410             :                        uint32_t num_sgl_descriptors)
    1411             : {
    1412           6 :         struct spdk_nvmf_rdma_request_data      *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES];
    1413             :         struct spdk_nvmf_rdma_request_data      *current_data_wr;
    1414             :         uint32_t                                i;
    1415             : 
    1416           6 :         if (spdk_unlikely(num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES)) {
    1417           0 :                 SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n",
    1418             :                             num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES);
    1419           0 :                 return -EINVAL;
    1420             :         }
    1421             : 
    1422           6 :         if (spdk_unlikely(spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests,
    1423             :                                                 num_sgl_descriptors))) {
    1424           0 :                 return -ENOMEM;
    1425             :         }
    1426             : 
    1427           6 :         current_data_wr = &rdma_req->data;
    1428             : 
    1429          12 :         for (i = 0; i < num_sgl_descriptors; i++) {
    1430           6 :                 nvmf_rdma_setup_wr(&current_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer);
    1431           6 :                 current_data_wr->wr.next = &work_requests[i]->wr;
    1432           6 :                 current_data_wr = work_requests[i];
    1433           6 :                 current_data_wr->wr.sg_list = current_data_wr->sgl;
    1434           6 :                 current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id;
    1435             :         }
    1436             : 
    1437           6 :         nvmf_rdma_setup_wr(&current_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer);
    1438             : 
    1439           6 :         return 0;
    1440             : }
    1441             : 
    1442             : static inline void
    1443          16 : nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req)
    1444             : {
    1445          16 :         struct ibv_send_wr              *wr = &rdma_req->data.wr;
    1446          16 :         struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1;
    1447             : 
    1448          16 :         wr->wr.rdma.rkey = sgl->keyed.key;
    1449          16 :         wr->wr.rdma.remote_addr = sgl->address;
    1450          16 :         nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer);
    1451          16 : }
    1452             : 
    1453             : static inline void
    1454           1 : nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs)
    1455             : {
    1456           1 :         struct ibv_send_wr              *wr = &rdma_req->data.wr;
    1457           1 :         struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1;
    1458             :         uint32_t                        i;
    1459             :         int                             j;
    1460           1 :         uint64_t                        remote_addr_offset = 0;
    1461             : 
    1462           3 :         for (i = 0; i < num_wrs; ++i) {
    1463           2 :                 wr->wr.rdma.rkey = sgl->keyed.key;
    1464           2 :                 wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset;
    1465          19 :                 for (j = 0; j < wr->num_sge; ++j) {
    1466          17 :                         remote_addr_offset += wr->sg_list[j].length;
    1467             :                 }
    1468           2 :                 wr = wr->next;
    1469             :         }
    1470           1 : }
    1471             : 
    1472             : static int
    1473          15 : nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_device *device,
    1474             :                       struct spdk_nvmf_rdma_request *rdma_req,
    1475             :                       struct ibv_send_wr *wr,
    1476             :                       uint32_t total_length)
    1477             : {
    1478          15 :         struct spdk_rdma_utils_memory_translation mem_translation;
    1479             :         struct ibv_sge  *sg_ele;
    1480             :         struct iovec *iov;
    1481             :         uint32_t lkey, remaining;
    1482             :         int rc;
    1483             : 
    1484          15 :         wr->num_sge = 0;
    1485             : 
    1486          74 :         while (total_length && wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES) {
    1487          59 :                 iov = &rdma_req->req.iov[rdma_req->iovpos];
    1488          59 :                 rc = spdk_rdma_utils_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation);
    1489          59 :                 if (spdk_unlikely(rc)) {
    1490           0 :                         return rc;
    1491             :                 }
    1492             : 
    1493          59 :                 lkey = spdk_rdma_utils_memory_translation_get_lkey(&mem_translation);
    1494          59 :                 sg_ele = &wr->sg_list[wr->num_sge];
    1495          59 :                 remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length);
    1496             : 
    1497          59 :                 sg_ele->lkey = lkey;
    1498          59 :                 sg_ele->addr = (uintptr_t)iov->iov_base + rdma_req->offset;
    1499          59 :                 sg_ele->length = remaining;
    1500          59 :                 SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, sg_ele->addr,
    1501             :                               sg_ele->length);
    1502          59 :                 rdma_req->offset += sg_ele->length;
    1503          59 :                 total_length -= sg_ele->length;
    1504          59 :                 wr->num_sge++;
    1505             : 
    1506          59 :                 if (rdma_req->offset == iov->iov_len) {
    1507          57 :                         rdma_req->offset = 0;
    1508          57 :                         rdma_req->iovpos++;
    1509             :                 }
    1510             :         }
    1511             : 
    1512          15 :         if (spdk_unlikely(total_length)) {
    1513           0 :                 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n");
    1514           0 :                 return -EINVAL;
    1515             :         }
    1516             : 
    1517          15 :         return 0;
    1518             : }
    1519             : 
    1520             : static int
    1521          10 : nvmf_rdma_fill_wr_sgl_with_dif(struct spdk_nvmf_rdma_device *device,
    1522             :                                struct spdk_nvmf_rdma_request *rdma_req,
    1523             :                                struct ibv_send_wr *wr,
    1524             :                                uint32_t total_length,
    1525             :                                uint32_t num_extra_wrs)
    1526             : {
    1527          10 :         struct spdk_rdma_utils_memory_translation mem_translation;
    1528          10 :         struct spdk_dif_ctx *dif_ctx = &rdma_req->req.dif.dif_ctx;
    1529             :         struct ibv_sge *sg_ele;
    1530             :         struct iovec *iov;
    1531             :         struct iovec *rdma_iov;
    1532             :         uint32_t lkey, remaining;
    1533             :         uint32_t remaining_data_block, data_block_size, md_size;
    1534             :         uint32_t sge_len;
    1535             :         int rc;
    1536             : 
    1537          10 :         data_block_size = dif_ctx->block_size - dif_ctx->md_size;
    1538             : 
    1539          10 :         if (spdk_likely(!rdma_req->req.stripped_data)) {
    1540           5 :                 rdma_iov = rdma_req->req.iov;
    1541           5 :                 remaining_data_block = data_block_size;
    1542           5 :                 md_size = dif_ctx->md_size;
    1543             :         } else {
    1544           5 :                 rdma_iov = rdma_req->req.stripped_data->iov;
    1545           5 :                 total_length = total_length / dif_ctx->block_size * data_block_size;
    1546           5 :                 remaining_data_block = total_length;
    1547           5 :                 md_size = 0;
    1548             :         }
    1549             : 
    1550          10 :         wr->num_sge = 0;
    1551             : 
    1552          25 :         while (total_length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) {
    1553          15 :                 iov = rdma_iov + rdma_req->iovpos;
    1554          15 :                 rc = spdk_rdma_utils_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation);
    1555          15 :                 if (spdk_unlikely(rc)) {
    1556           0 :                         return rc;
    1557             :                 }
    1558             : 
    1559          15 :                 lkey = spdk_rdma_utils_memory_translation_get_lkey(&mem_translation);
    1560          15 :                 sg_ele = &wr->sg_list[wr->num_sge];
    1561          15 :                 remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length);
    1562             : 
    1563          53 :                 while (remaining) {
    1564          38 :                         if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) {
    1565           1 :                                 if (num_extra_wrs > 0 && wr->next) {
    1566           1 :                                         wr = wr->next;
    1567           1 :                                         wr->num_sge = 0;
    1568           1 :                                         sg_ele = &wr->sg_list[wr->num_sge];
    1569           1 :                                         num_extra_wrs--;
    1570             :                                 } else {
    1571             :                                         break;
    1572             :                                 }
    1573             :                         }
    1574          38 :                         sg_ele->lkey = lkey;
    1575          38 :                         sg_ele->addr = (uintptr_t)((char *)iov->iov_base + rdma_req->offset);
    1576          38 :                         sge_len = spdk_min(remaining, remaining_data_block);
    1577          38 :                         sg_ele->length = sge_len;
    1578          38 :                         SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele,
    1579             :                                       sg_ele->addr, sg_ele->length);
    1580          38 :                         remaining -= sge_len;
    1581          38 :                         remaining_data_block -= sge_len;
    1582          38 :                         rdma_req->offset += sge_len;
    1583          38 :                         total_length -= sge_len;
    1584             : 
    1585          38 :                         sg_ele++;
    1586          38 :                         wr->num_sge++;
    1587             : 
    1588          38 :                         if (remaining_data_block == 0) {
    1589             :                                 /* skip metadata */
    1590          34 :                                 rdma_req->offset += md_size;
    1591          34 :                                 total_length -= md_size;
    1592             :                                 /* Metadata that do not fit this IO buffer will be included in the next IO buffer */
    1593          34 :                                 remaining -= spdk_min(remaining, md_size);
    1594          34 :                                 remaining_data_block = data_block_size;
    1595             :                         }
    1596             : 
    1597          38 :                         if (remaining == 0) {
    1598             :                                 /* By subtracting the size of the last IOV from the offset, we ensure that we skip
    1599             :                                    the remaining metadata bits at the beginning of the next buffer */
    1600          15 :                                 rdma_req->offset -= spdk_min(iov->iov_len, rdma_req->offset);
    1601          15 :                                 rdma_req->iovpos++;
    1602             :                         }
    1603             :                 }
    1604             :         }
    1605             : 
    1606          10 :         if (spdk_unlikely(total_length)) {
    1607           0 :                 SPDK_ERRLOG("Not enough SG entries to hold data buffer\n");
    1608           0 :                 return -EINVAL;
    1609             :         }
    1610             : 
    1611          10 :         return 0;
    1612             : }
    1613             : 
    1614             : static inline uint32_t
    1615           8 : nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size)
    1616             : {
    1617             :         /* estimate the number of SG entries and WRs needed to process the request */
    1618           8 :         uint32_t num_sge = 0;
    1619             :         uint32_t i;
    1620           8 :         uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size);
    1621             : 
    1622          23 :         for (i = 0; i < num_buffers && length > 0; i++) {
    1623          15 :                 uint32_t buffer_len = spdk_min(length, io_unit_size);
    1624          15 :                 uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size);
    1625             : 
    1626          15 :                 if (num_sge_in_block * block_size > buffer_len) {
    1627          11 :                         ++num_sge_in_block;
    1628             :                 }
    1629          15 :                 num_sge += num_sge_in_block;
    1630          15 :                 length -= buffer_len;
    1631             :         }
    1632           8 :         return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES);
    1633             : }
    1634             : 
    1635             : static int
    1636          16 : nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport,
    1637             :                             struct spdk_nvmf_rdma_device *device,
    1638             :                             struct spdk_nvmf_rdma_request *rdma_req)
    1639             : {
    1640             :         struct spdk_nvmf_rdma_qpair             *rqpair;
    1641             :         struct spdk_nvmf_rdma_poll_group        *rgroup;
    1642          16 :         struct spdk_nvmf_request                *req = &rdma_req->req;
    1643          16 :         struct ibv_send_wr                      *wr = &rdma_req->data.wr;
    1644             :         int                                     rc;
    1645          16 :         uint32_t                                num_wrs = 1;
    1646             :         uint32_t                                length;
    1647             : 
    1648          16 :         rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair);
    1649          16 :         rgroup = rqpair->poller->group;
    1650             : 
    1651             :         /* rdma wr specifics */
    1652          16 :         nvmf_rdma_setup_request(rdma_req);
    1653             : 
    1654          16 :         length = req->length;
    1655          16 :         if (spdk_unlikely(req->dif_enabled)) {
    1656           8 :                 req->dif.orig_length = length;
    1657           8 :                 length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx);
    1658           8 :                 req->dif.elba_length = length;
    1659             :         }
    1660             : 
    1661          16 :         rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport,
    1662             :                                            length);
    1663          16 :         if (spdk_unlikely(rc != 0)) {
    1664           1 :                 return rc;
    1665             :         }
    1666             : 
    1667          15 :         assert(req->iovcnt <= rqpair->max_send_sge);
    1668             : 
    1669             :         /* When dif_insert_or_strip is true and the I/O data length is greater than one block,
    1670             :          * the stripped_buffers are got for DIF stripping. */
    1671          15 :         if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST)
    1672             :                           && (req->dif.elba_length > req->dif.dif_ctx.block_size))) {
    1673           7 :                 rc = nvmf_request_get_stripped_buffers(req, &rgroup->group,
    1674             :                                                        &rtransport->transport, req->dif.orig_length);
    1675           7 :                 if (rc != 0) {
    1676           4 :                         SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc);
    1677             :                 }
    1678             :         }
    1679             : 
    1680          15 :         rdma_req->iovpos = 0;
    1681             : 
    1682          15 :         if (spdk_unlikely(req->dif_enabled)) {
    1683           8 :                 num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size,
    1684             :                                                  req->dif.dif_ctx.block_size);
    1685           8 :                 if (num_wrs > 1) {
    1686           1 :                         rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1);
    1687           1 :                         if (spdk_unlikely(rc != 0)) {
    1688           0 :                                 goto err_exit;
    1689             :                         }
    1690             :                 }
    1691             : 
    1692           8 :                 rc = nvmf_rdma_fill_wr_sgl_with_dif(device, rdma_req, wr, length, num_wrs - 1);
    1693           8 :                 if (spdk_unlikely(rc != 0)) {
    1694           0 :                         goto err_exit;
    1695             :                 }
    1696             : 
    1697           8 :                 if (num_wrs > 1) {
    1698           1 :                         nvmf_rdma_update_remote_addr(rdma_req, num_wrs);
    1699             :                 }
    1700             :         } else {
    1701           7 :                 rc = nvmf_rdma_fill_wr_sgl(device, rdma_req, wr, length);
    1702           7 :                 if (spdk_unlikely(rc != 0)) {
    1703           0 :                         goto err_exit;
    1704             :                 }
    1705             :         }
    1706             : 
    1707             :         /* set the number of outstanding data WRs for this request. */
    1708          15 :         rdma_req->num_outstanding_data_wr = num_wrs;
    1709             : 
    1710          15 :         return rc;
    1711             : 
    1712           0 : err_exit:
    1713           0 :         spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport);
    1714           0 :         nvmf_rdma_request_free_data(rdma_req, rtransport);
    1715           0 :         req->iovcnt = 0;
    1716           0 :         return rc;
    1717             : }
    1718             : 
    1719             : static int
    1720           5 : nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport,
    1721             :                                       struct spdk_nvmf_rdma_device *device,
    1722             :                                       struct spdk_nvmf_rdma_request *rdma_req)
    1723             : {
    1724             :         struct spdk_nvmf_rdma_qpair             *rqpair;
    1725             :         struct spdk_nvmf_rdma_poll_group        *rgroup;
    1726             :         struct ibv_send_wr                      *current_wr;
    1727           5 :         struct spdk_nvmf_request                *req = &rdma_req->req;
    1728             :         struct spdk_nvme_sgl_descriptor         *inline_segment, *desc;
    1729             :         uint32_t                                num_sgl_descriptors;
    1730           5 :         uint32_t                                lengths[SPDK_NVMF_MAX_SGL_ENTRIES], total_length = 0;
    1731             :         uint32_t                                i;
    1732             :         int                                     rc;
    1733             : 
    1734           5 :         rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
    1735           5 :         rgroup = rqpair->poller->group;
    1736             : 
    1737           5 :         inline_segment = &req->cmd->nvme_cmd.dptr.sgl1;
    1738           5 :         assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT);
    1739           5 :         assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET);
    1740             : 
    1741           5 :         num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor);
    1742           5 :         assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES);
    1743             : 
    1744           5 :         desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address;
    1745          15 :         for (i = 0; i < num_sgl_descriptors; i++) {
    1746          10 :                 if (spdk_likely(!req->dif_enabled)) {
    1747           8 :                         lengths[i] = desc->keyed.length;
    1748             :                 } else {
    1749           2 :                         req->dif.orig_length += desc->keyed.length;
    1750           2 :                         lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx);
    1751           2 :                         req->dif.elba_length += lengths[i];
    1752             :                 }
    1753          10 :                 total_length += lengths[i];
    1754          10 :                 desc++;
    1755             :         }
    1756             : 
    1757           5 :         if (spdk_unlikely(total_length > rtransport->transport.opts.max_io_size)) {
    1758           0 :                 SPDK_ERRLOG("Multi SGL length 0x%x exceeds max io size 0x%x\n",
    1759             :                             total_length, rtransport->transport.opts.max_io_size);
    1760           0 :                 req->rsp->nvme_cpl.status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
    1761           0 :                 return -EINVAL;
    1762             :         }
    1763             : 
    1764           5 :         rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1);
    1765           5 :         if (spdk_unlikely(rc != 0)) {
    1766           0 :                 return -ENOMEM;
    1767             :         }
    1768             : 
    1769           5 :         rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, total_length);
    1770           5 :         if (spdk_unlikely(rc != 0)) {
    1771           0 :                 nvmf_rdma_request_free_data(rdma_req, rtransport);
    1772           0 :                 return rc;
    1773             :         }
    1774             : 
    1775             :         /* When dif_insert_or_strip is true and the I/O data length is greater than one block,
    1776             :          * the stripped_buffers are got for DIF stripping. */
    1777           5 :         if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST)
    1778             :                           && (req->dif.elba_length > req->dif.dif_ctx.block_size))) {
    1779           1 :                 rc = nvmf_request_get_stripped_buffers(req, &rgroup->group,
    1780             :                                                        &rtransport->transport, req->dif.orig_length);
    1781           1 :                 if (spdk_unlikely(rc != 0)) {
    1782           0 :                         SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc);
    1783             :                 }
    1784             :         }
    1785             : 
    1786             :         /* The first WR must always be the embedded data WR. This is how we unwind them later. */
    1787           5 :         current_wr = &rdma_req->data.wr;
    1788           5 :         assert(current_wr != NULL);
    1789             : 
    1790           5 :         req->length = 0;
    1791           5 :         rdma_req->iovpos = 0;
    1792           5 :         desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address;
    1793          15 :         for (i = 0; i < num_sgl_descriptors; i++) {
    1794             :                 /* The descriptors must be keyed data block descriptors with an address, not an offset. */
    1795          10 :                 if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK ||
    1796             :                                   desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) {
    1797           0 :                         rc = -EINVAL;
    1798           0 :                         goto err_exit;
    1799             :                 }
    1800             : 
    1801          10 :                 if (spdk_likely(!req->dif_enabled)) {
    1802           8 :                         rc = nvmf_rdma_fill_wr_sgl(device, rdma_req, current_wr, lengths[i]);
    1803             :                 } else {
    1804           2 :                         rc = nvmf_rdma_fill_wr_sgl_with_dif(device, rdma_req, current_wr,
    1805             :                                                             lengths[i], 0);
    1806             :                 }
    1807          10 :                 if (spdk_unlikely(rc != 0)) {
    1808           0 :                         rc = -ENOMEM;
    1809           0 :                         goto err_exit;
    1810             :                 }
    1811             : 
    1812          10 :                 req->length += desc->keyed.length;
    1813          10 :                 current_wr->wr.rdma.rkey = desc->keyed.key;
    1814          10 :                 current_wr->wr.rdma.remote_addr = desc->address;
    1815          10 :                 current_wr = current_wr->next;
    1816          10 :                 desc++;
    1817             :         }
    1818             : 
    1819             : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
    1820             :         /* Go back to the last descriptor in the list. */
    1821           5 :         desc--;
    1822           5 :         if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) {
    1823           0 :                 if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) {
    1824           0 :                         rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV;
    1825           0 :                         rdma_req->rsp.wr.imm_data = desc->keyed.key;
    1826             :                 }
    1827             :         }
    1828             : #endif
    1829             : 
    1830           5 :         rdma_req->num_outstanding_data_wr = num_sgl_descriptors;
    1831             : 
    1832           5 :         return 0;
    1833             : 
    1834           0 : err_exit:
    1835           0 :         spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport);
    1836           0 :         nvmf_rdma_request_free_data(rdma_req, rtransport);
    1837           0 :         return rc;
    1838             : }
    1839             : 
    1840             : static int
    1841          25 : nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport,
    1842             :                             struct spdk_nvmf_rdma_device *device,
    1843             :                             struct spdk_nvmf_rdma_request *rdma_req)
    1844             : {
    1845          25 :         struct spdk_nvmf_request                *req = &rdma_req->req;
    1846             :         struct spdk_nvme_cpl                    *rsp;
    1847             :         struct spdk_nvme_sgl_descriptor         *sgl;
    1848             :         int                                     rc;
    1849             :         uint32_t                                length;
    1850             : 
    1851          25 :         rsp = &req->rsp->nvme_cpl;
    1852          25 :         sgl = &req->cmd->nvme_cmd.dptr.sgl1;
    1853             : 
    1854          25 :         if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK &&
    1855          17 :             (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS ||
    1856           0 :              sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) {
    1857             : 
    1858          17 :                 length = sgl->keyed.length;
    1859          17 :                 if (spdk_unlikely(length > rtransport->transport.opts.max_io_size)) {
    1860           1 :                         SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n",
    1861             :                                     length, rtransport->transport.opts.max_io_size);
    1862           1 :                         rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
    1863           1 :                         return -1;
    1864             :                 }
    1865             : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
    1866          16 :                 if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) {
    1867           0 :                         if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) {
    1868           0 :                                 rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV;
    1869           0 :                                 rdma_req->rsp.wr.imm_data = sgl->keyed.key;
    1870             :                         }
    1871             :                 }
    1872             : #endif
    1873             : 
    1874             :                 /* fill request length and populate iovs */
    1875          16 :                 req->length = length;
    1876             : 
    1877          16 :                 rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req);
    1878          16 :                 if (spdk_unlikely(rc < 0)) {
    1879           1 :                         if (rc == -EINVAL) {
    1880           0 :                                 SPDK_ERRLOG("SGL length exceeds the max I/O size\n");
    1881           0 :                                 rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
    1882           0 :                                 return -1;
    1883             :                         }
    1884             :                         /* No available buffers. Queue this request up. */
    1885           1 :                         SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req);
    1886           1 :                         return 0;
    1887             :                 }
    1888             : 
    1889          15 :                 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req,
    1890             :                               req->iovcnt);
    1891             : 
    1892          15 :                 return 0;
    1893           8 :         } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK &&
    1894           3 :                    sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) {
    1895           3 :                 uint64_t offset = sgl->address;
    1896           3 :                 uint32_t max_len = rtransport->transport.opts.in_capsule_data_size;
    1897             : 
    1898           3 :                 SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n",
    1899             :                               offset, sgl->unkeyed.length);
    1900             : 
    1901           3 :                 if (spdk_unlikely(offset > max_len)) {
    1902           0 :                         SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n",
    1903             :                                     offset, max_len);
    1904           0 :                         rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET;
    1905           0 :                         return -1;
    1906             :                 }
    1907           3 :                 max_len -= (uint32_t)offset;
    1908             : 
    1909           3 :                 if (spdk_unlikely(sgl->unkeyed.length > max_len)) {
    1910           2 :                         SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n",
    1911             :                                     sgl->unkeyed.length, max_len);
    1912           2 :                         rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
    1913           2 :                         return -1;
    1914             :                 }
    1915             : 
    1916           1 :                 rdma_req->num_outstanding_data_wr = 0;
    1917           1 :                 req->data_from_pool = false;
    1918           1 :                 req->length = sgl->unkeyed.length;
    1919             : 
    1920           1 :                 req->iov[0].iov_base = rdma_req->recv->buf + offset;
    1921           1 :                 req->iov[0].iov_len = req->length;
    1922           1 :                 req->iovcnt = 1;
    1923             : 
    1924           1 :                 return 0;
    1925           5 :         } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT &&
    1926           5 :                    sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) {
    1927             : 
    1928           5 :                 rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req);
    1929           5 :                 if (spdk_unlikely(rc == -ENOMEM)) {
    1930           0 :                         SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req);
    1931           0 :                         return 0;
    1932           5 :                 } else if (spdk_unlikely(rc == -EINVAL)) {
    1933           0 :                         SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n");
    1934           0 :                         rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
    1935           0 :                         return -1;
    1936             :                 }
    1937             : 
    1938           5 :                 SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req,
    1939             :                               req->iovcnt);
    1940             : 
    1941           5 :                 return 0;
    1942             :         }
    1943             : 
    1944           0 :         SPDK_ERRLOG("Invalid NVMf I/O Command SGL:  Type 0x%x, Subtype 0x%x\n",
    1945             :                     sgl->generic.type, sgl->generic.subtype);
    1946           0 :         rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID;
    1947           0 :         return -1;
    1948             : }
    1949             : 
    1950             : static void
    1951           6 : _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req,
    1952             :                         struct spdk_nvmf_rdma_transport *rtransport)
    1953             : {
    1954             :         struct spdk_nvmf_rdma_qpair             *rqpair;
    1955             :         struct spdk_nvmf_rdma_poll_group        *rgroup;
    1956             : 
    1957           6 :         rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
    1958           6 :         if (rdma_req->req.data_from_pool) {
    1959           5 :                 rgroup = rqpair->poller->group;
    1960             : 
    1961           5 :                 spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport);
    1962             :         }
    1963           6 :         if (rdma_req->req.stripped_data) {
    1964           0 :                 nvmf_request_free_stripped_buffers(&rdma_req->req,
    1965           0 :                                                    &rqpair->poller->group->group,
    1966             :                                                    &rtransport->transport);
    1967             :         }
    1968           6 :         nvmf_rdma_request_free_data(rdma_req, rtransport);
    1969           6 :         rdma_req->req.length = 0;
    1970           6 :         rdma_req->req.iovcnt = 0;
    1971           6 :         rdma_req->offset = 0;
    1972           6 :         rdma_req->req.dif_enabled = false;
    1973           6 :         rdma_req->fused_failed = false;
    1974           6 :         rdma_req->transfer_wr = NULL;
    1975           6 :         if (rdma_req->fused_pair) {
    1976             :                 /* This req was part of a valid fused pair, but failed before it got to
    1977             :                  * READ_TO_EXECUTE state.  This means we need to fail the other request
    1978             :                  * in the pair, because it is no longer part of a valid pair.  If the pair
    1979             :                  * already reached READY_TO_EXECUTE state, we need to kick it.
    1980             :                  */
    1981           0 :                 rdma_req->fused_pair->fused_failed = true;
    1982           0 :                 if (rdma_req->fused_pair->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) {
    1983           0 :                         nvmf_rdma_request_process(rtransport, rdma_req->fused_pair);
    1984             :                 }
    1985           0 :                 rdma_req->fused_pair = NULL;
    1986             :         }
    1987           6 :         memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif));
    1988           6 :         rqpair->qd--;
    1989             : 
    1990           6 :         STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link);
    1991           6 :         rqpair->qpair.queue_depth--;
    1992           6 :         rdma_req->state = RDMA_REQUEST_STATE_FREE;
    1993           6 : }
    1994             : 
    1995             : static void
    1996           6 : nvmf_rdma_check_fused_ordering(struct spdk_nvmf_rdma_transport *rtransport,
    1997             :                                struct spdk_nvmf_rdma_qpair *rqpair,
    1998             :                                struct spdk_nvmf_rdma_request *rdma_req)
    1999             : {
    2000             :         enum spdk_nvme_cmd_fuse last, next;
    2001             : 
    2002           6 :         last = rqpair->fused_first ? rqpair->fused_first->req.cmd->nvme_cmd.fuse : SPDK_NVME_CMD_FUSE_NONE;
    2003           6 :         next = rdma_req->req.cmd->nvme_cmd.fuse;
    2004             : 
    2005           6 :         assert(last != SPDK_NVME_CMD_FUSE_SECOND);
    2006             : 
    2007           6 :         if (spdk_likely(last == SPDK_NVME_CMD_FUSE_NONE && next == SPDK_NVME_CMD_FUSE_NONE)) {
    2008           6 :                 return;
    2009             :         }
    2010             : 
    2011           0 :         if (last == SPDK_NVME_CMD_FUSE_FIRST) {
    2012           0 :                 if (next == SPDK_NVME_CMD_FUSE_SECOND) {
    2013             :                         /* This is a valid pair of fused commands.  Point them at each other
    2014             :                          * so they can be submitted consecutively once ready to be executed.
    2015             :                          */
    2016           0 :                         rqpair->fused_first->fused_pair = rdma_req;
    2017           0 :                         rdma_req->fused_pair = rqpair->fused_first;
    2018           0 :                         rqpair->fused_first = NULL;
    2019           0 :                         return;
    2020             :                 } else {
    2021             :                         /* Mark the last req as failed since it wasn't followed by a SECOND. */
    2022           0 :                         rqpair->fused_first->fused_failed = true;
    2023             : 
    2024             :                         /* If the last req is in READY_TO_EXECUTE state, then call
    2025             :                          * nvmf_rdma_request_process(), otherwise nothing else will kick it.
    2026             :                          */
    2027           0 :                         if (rqpair->fused_first->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) {
    2028           0 :                                 nvmf_rdma_request_process(rtransport, rqpair->fused_first);
    2029             :                         }
    2030             : 
    2031           0 :                         rqpair->fused_first = NULL;
    2032             :                 }
    2033             :         }
    2034             : 
    2035           0 :         if (next == SPDK_NVME_CMD_FUSE_FIRST) {
    2036             :                 /* Set rqpair->fused_first here so that we know to check that the next request
    2037             :                  * is a SECOND (and to fail this one if it isn't).
    2038             :                  */
    2039           0 :                 rqpair->fused_first = rdma_req;
    2040           0 :         } else if (next == SPDK_NVME_CMD_FUSE_SECOND) {
    2041             :                 /* Mark this req failed since it ia SECOND and the last one was not a FIRST. */
    2042           0 :                 rdma_req->fused_failed = true;
    2043             :         }
    2044             : }
    2045             : 
    2046             : bool
    2047          23 : nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport,
    2048             :                           struct spdk_nvmf_rdma_request *rdma_req)
    2049             : {
    2050             :         struct spdk_nvmf_rdma_qpair     *rqpair;
    2051             :         struct spdk_nvmf_rdma_device    *device;
    2052             :         struct spdk_nvmf_rdma_poll_group *rgroup;
    2053          23 :         struct spdk_nvme_cpl            *rsp = &rdma_req->req.rsp->nvme_cpl;
    2054             :         int                             rc;
    2055             :         struct spdk_nvmf_rdma_recv      *rdma_recv;
    2056             :         enum spdk_nvmf_rdma_request_state prev_state;
    2057          23 :         bool                            progress = false;
    2058          23 :         int                             data_posted;
    2059             :         uint32_t                        num_blocks, num_rdma_reads_available, qdepth;
    2060             : 
    2061          23 :         rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
    2062          23 :         device = rqpair->device;
    2063          23 :         rgroup = rqpair->poller->group;
    2064             : 
    2065          23 :         assert(rdma_req->state != RDMA_REQUEST_STATE_FREE);
    2066             : 
    2067             :         /* If the queue pair is in an error state, force the request to the completed state
    2068             :          * to release resources. */
    2069          23 :         if (spdk_unlikely(rqpair->ibv_in_error_state || !spdk_nvmf_qpair_is_active(&rqpair->qpair))) {
    2070           0 :                 switch (rdma_req->state) {
    2071           0 :                 case RDMA_REQUEST_STATE_NEED_BUFFER:
    2072           0 :                         STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link);
    2073           0 :                         break;
    2074           0 :                 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING:
    2075           0 :                         STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
    2076           0 :                         break;
    2077           0 :                 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING:
    2078           0 :                         STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
    2079           0 :                         break;
    2080           0 :                 case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING:
    2081           0 :                         STAILQ_REMOVE(&rqpair->pending_rdma_send_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
    2082           0 :                         break;
    2083           0 :                 default:
    2084           0 :                         break;
    2085             :                 }
    2086           0 :                 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
    2087             :         }
    2088             : 
    2089             :         /* The loop here is to allow for several back-to-back state changes. */
    2090             :         do {
    2091          66 :                 prev_state = rdma_req->state;
    2092             : 
    2093          66 :                 SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state);
    2094             : 
    2095          66 :                 switch (rdma_req->state) {
    2096           6 :                 case RDMA_REQUEST_STATE_FREE:
    2097             :                         /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW
    2098             :                          * to escape this state. */
    2099           6 :                         break;
    2100           6 :                 case RDMA_REQUEST_STATE_NEW:
    2101           6 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0,
    2102             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair, rqpair->qpair.queue_depth);
    2103           6 :                         rdma_recv = rdma_req->recv;
    2104             : 
    2105             :                         /* The first element of the SGL is the NVMe command */
    2106           6 :                         rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr;
    2107           6 :                         memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp));
    2108           6 :                         rdma_req->transfer_wr = &rdma_req->data.wr;
    2109             : 
    2110           6 :                         if (spdk_unlikely(rqpair->ibv_in_error_state || !spdk_nvmf_qpair_is_active(&rqpair->qpair))) {
    2111           0 :                                 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
    2112           0 :                                 break;
    2113             :                         }
    2114             : 
    2115           6 :                         if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) {
    2116           0 :                                 rdma_req->req.dif_enabled = true;
    2117             :                         }
    2118             : 
    2119           6 :                         nvmf_rdma_check_fused_ordering(rtransport, rqpair, rdma_req);
    2120             : 
    2121             : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
    2122           6 :                         rdma_req->rsp.wr.opcode = IBV_WR_SEND;
    2123           6 :                         rdma_req->rsp.wr.imm_data = 0;
    2124             : #endif
    2125             : 
    2126             :                         /* The next state transition depends on the data transfer needs of this request. */
    2127           6 :                         rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req);
    2128             : 
    2129           6 :                         if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) {
    2130           1 :                                 rsp->status.sct = SPDK_NVME_SCT_GENERIC;
    2131           1 :                                 rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE;
    2132           1 :                                 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
    2133           1 :                                 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
    2134           1 :                                 SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req);
    2135           1 :                                 break;
    2136             :                         }
    2137             : 
    2138             :                         /* If no data to transfer, ready to execute. */
    2139           5 :                         if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) {
    2140           0 :                                 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
    2141           0 :                                 break;
    2142             :                         }
    2143             : 
    2144           5 :                         rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER;
    2145           5 :                         STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link);
    2146           5 :                         break;
    2147           5 :                 case RDMA_REQUEST_STATE_NEED_BUFFER:
    2148           5 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0,
    2149             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2150             : 
    2151           5 :                         assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE);
    2152             : 
    2153           5 :                         if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) {
    2154             :                                 /* This request needs to wait in line to obtain a buffer */
    2155           0 :                                 break;
    2156             :                         }
    2157             : 
    2158             :                         /* Try to get a data buffer */
    2159           5 :                         rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req);
    2160           5 :                         if (spdk_unlikely(rc < 0)) {
    2161           0 :                                 STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link);
    2162           0 :                                 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
    2163           0 :                                 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
    2164           0 :                                 break;
    2165             :                         }
    2166             : 
    2167           5 :                         if (rdma_req->req.iovcnt == 0) {
    2168             :                                 /* No buffers available. */
    2169           0 :                                 rgroup->stat.pending_data_buffer++;
    2170           0 :                                 break;
    2171             :                         }
    2172             : 
    2173           5 :                         STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link);
    2174             : 
    2175             :                         /* If data is transferring from host to controller and the data didn't
    2176             :                          * arrive using in capsule data, we need to do a transfer from the host.
    2177             :                          */
    2178           5 :                         if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER &&
    2179             :                             rdma_req->req.data_from_pool) {
    2180           4 :                                 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link);
    2181           4 :                                 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING;
    2182           4 :                                 break;
    2183             :                         }
    2184             : 
    2185           1 :                         rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
    2186           1 :                         break;
    2187           4 :                 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING:
    2188           4 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0,
    2189             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2190             : 
    2191           4 :                         if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) {
    2192             :                                 /* This request needs to wait in line to perform RDMA */
    2193           0 :                                 break;
    2194             :                         }
    2195           4 :                         assert(rqpair->max_send_depth >= rqpair->current_send_depth);
    2196           4 :                         qdepth = rqpair->max_send_depth - rqpair->current_send_depth;
    2197           4 :                         assert(rqpair->max_read_depth >= rqpair->current_read_depth);
    2198           4 :                         num_rdma_reads_available = rqpair->max_read_depth - rqpair->current_read_depth;
    2199           4 :                         if (rdma_req->num_outstanding_data_wr > qdepth ||
    2200           4 :                             rdma_req->num_outstanding_data_wr > num_rdma_reads_available) {
    2201           0 :                                 if (num_rdma_reads_available && qdepth) {
    2202             :                                         /* Send as much as we can */
    2203           0 :                                         request_prepare_transfer_in_part(&rdma_req->req, spdk_min(num_rdma_reads_available, qdepth));
    2204             :                                 } else {
    2205             :                                         /* We can only have so many WRs outstanding. we have to wait until some finish. */
    2206           0 :                                         rqpair->poller->stat.pending_rdma_read++;
    2207           0 :                                         break;
    2208             :                                 }
    2209             :                         }
    2210             : 
    2211             :                         /* We have already verified that this request is the head of the queue. */
    2212           4 :                         if (rdma_req->num_remaining_data_wr == 0) {
    2213           4 :                                 STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link);
    2214             :                         }
    2215             : 
    2216           4 :                         rc = request_transfer_in(&rdma_req->req);
    2217           4 :                         if (spdk_likely(rc == 0)) {
    2218           4 :                                 rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER;
    2219             :                         } else {
    2220           0 :                                 rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR;
    2221           0 :                                 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
    2222           0 :                                 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
    2223             :                         }
    2224           4 :                         break;
    2225           4 :                 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
    2226           4 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0,
    2227             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2228             :                         /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE
    2229             :                          * to escape this state. */
    2230           4 :                         break;
    2231           5 :                 case RDMA_REQUEST_STATE_READY_TO_EXECUTE:
    2232           5 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0,
    2233             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2234             : 
    2235           5 :                         if (spdk_unlikely(rdma_req->req.dif_enabled)) {
    2236           0 :                                 if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) {
    2237             :                                         /* generate DIF for write operation */
    2238           0 :                                         num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size);
    2239           0 :                                         assert(num_blocks > 0);
    2240             : 
    2241           0 :                                         rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt,
    2242           0 :                                                                num_blocks, &rdma_req->req.dif.dif_ctx);
    2243           0 :                                         if (rc != 0) {
    2244           0 :                                                 SPDK_ERRLOG("DIF generation failed\n");
    2245           0 :                                                 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
    2246           0 :                                                 spdk_nvmf_qpair_disconnect(&rqpair->qpair);
    2247           0 :                                                 break;
    2248             :                                         }
    2249             :                                 }
    2250             : 
    2251           0 :                                 assert(rdma_req->req.dif.elba_length >= rdma_req->req.length);
    2252             :                                 /* set extended length before IO operation */
    2253           0 :                                 rdma_req->req.length = rdma_req->req.dif.elba_length;
    2254             :                         }
    2255             : 
    2256           5 :                         if (rdma_req->req.cmd->nvme_cmd.fuse != SPDK_NVME_CMD_FUSE_NONE) {
    2257           0 :                                 if (rdma_req->fused_failed) {
    2258             :                                         /* This request failed FUSED semantics.  Fail it immediately, without
    2259             :                                          * even sending it to the target layer.
    2260             :                                          */
    2261           0 :                                         rsp->status.sct = SPDK_NVME_SCT_GENERIC;
    2262           0 :                                         rsp->status.sc = SPDK_NVME_SC_ABORTED_MISSING_FUSED;
    2263           0 :                                         STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
    2264           0 :                                         rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
    2265           0 :                                         break;
    2266             :                                 }
    2267             : 
    2268           0 :                                 if (rdma_req->fused_pair == NULL ||
    2269           0 :                                     rdma_req->fused_pair->state != RDMA_REQUEST_STATE_READY_TO_EXECUTE) {
    2270             :                                         /* This request is ready to execute, but either we don't know yet if it's
    2271             :                                          * valid - i.e. this is a FIRST but we haven't received the next
    2272             :                                          * request yet or the other request of this fused pair isn't ready to
    2273             :                                          * execute.  So break here and this request will get processed later either
    2274             :                                          * when the other request is ready or we find that this request isn't valid.
    2275             :                                          */
    2276             :                                         break;
    2277             :                                 }
    2278             :                         }
    2279             : 
    2280             :                         /* If we get to this point, and this request is a fused command, we know that
    2281             :                          * it is part of valid sequence (FIRST followed by a SECOND) and that both
    2282             :                          * requests are READY_TO_EXECUTE. So call spdk_nvmf_request_exec() both on this
    2283             :                          * request, and the other request of the fused pair, in the correct order.
    2284             :                          * Also clear the ->fused_pair pointers on both requests, since after this point
    2285             :                          * we no longer need to maintain the relationship between these two requests.
    2286             :                          */
    2287           5 :                         if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_SECOND) {
    2288           0 :                                 assert(rdma_req->fused_pair != NULL);
    2289           0 :                                 assert(rdma_req->fused_pair->fused_pair != NULL);
    2290           0 :                                 rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING;
    2291           0 :                                 spdk_nvmf_request_exec(&rdma_req->fused_pair->req);
    2292           0 :                                 rdma_req->fused_pair->fused_pair = NULL;
    2293           0 :                                 rdma_req->fused_pair = NULL;
    2294             :                         }
    2295           5 :                         rdma_req->state = RDMA_REQUEST_STATE_EXECUTING;
    2296           5 :                         spdk_nvmf_request_exec(&rdma_req->req);
    2297           5 :                         if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_FIRST) {
    2298           0 :                                 assert(rdma_req->fused_pair != NULL);
    2299           0 :                                 assert(rdma_req->fused_pair->fused_pair != NULL);
    2300           0 :                                 rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING;
    2301           0 :                                 spdk_nvmf_request_exec(&rdma_req->fused_pair->req);
    2302           0 :                                 rdma_req->fused_pair->fused_pair = NULL;
    2303           0 :                                 rdma_req->fused_pair = NULL;
    2304             :                         }
    2305           5 :                         break;
    2306           5 :                 case RDMA_REQUEST_STATE_EXECUTING:
    2307           5 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0,
    2308             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2309             :                         /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED
    2310             :                          * to escape this state. */
    2311           5 :                         break;
    2312           5 :                 case RDMA_REQUEST_STATE_EXECUTED:
    2313           5 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0,
    2314             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2315           5 :                         if (rsp->status.sc == SPDK_NVME_SC_SUCCESS &&
    2316           5 :                             rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
    2317           1 :                                 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link);
    2318           1 :                                 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING;
    2319             :                         } else {
    2320           4 :                                 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
    2321           4 :                                 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
    2322             :                         }
    2323           5 :                         if (spdk_unlikely(rdma_req->req.dif_enabled)) {
    2324             :                                 /* restore the original length */
    2325           0 :                                 rdma_req->req.length = rdma_req->req.dif.orig_length;
    2326             : 
    2327           0 :                                 if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
    2328           0 :                                         struct spdk_dif_error error_blk;
    2329             : 
    2330           0 :                                         num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size);
    2331           0 :                                         if (!rdma_req->req.stripped_data) {
    2332           0 :                                                 rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks,
    2333           0 :                                                                      &rdma_req->req.dif.dif_ctx, &error_blk);
    2334             :                                         } else {
    2335           0 :                                                 rc = spdk_dif_verify_copy(rdma_req->req.stripped_data->iov,
    2336           0 :                                                                           rdma_req->req.stripped_data->iovcnt,
    2337           0 :                                                                           rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks,
    2338           0 :                                                                           &rdma_req->req.dif.dif_ctx, &error_blk);
    2339             :                                         }
    2340           0 :                                         if (rc) {
    2341           0 :                                                 struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl;
    2342             : 
    2343           0 :                                                 SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type,
    2344             :                                                             error_blk.err_offset);
    2345           0 :                                                 rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR;
    2346           0 :                                                 rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type);
    2347           0 :                                                 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
    2348           0 :                                                 STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
    2349           0 :                                                 rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
    2350             :                                         }
    2351             :                                 }
    2352             :                         }
    2353           5 :                         break;
    2354           1 :                 case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING:
    2355           1 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0,
    2356             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2357             : 
    2358           1 :                         if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) {
    2359             :                                 /* This request needs to wait in line to perform RDMA */
    2360           0 :                                 break;
    2361             :                         }
    2362           1 :                         if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) >
    2363           1 :                             rqpair->max_send_depth) {
    2364             :                                 /* We can only have so many WRs outstanding. we have to wait until some finish.
    2365             :                                  * +1 since each request has an additional wr in the resp. */
    2366           0 :                                 rqpair->poller->stat.pending_rdma_write++;
    2367           0 :                                 break;
    2368             :                         }
    2369             : 
    2370             :                         /* We have already verified that this request is the head of the queue. */
    2371           1 :                         STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link);
    2372             : 
    2373             :                         /* The data transfer will be kicked off from
    2374             :                          * RDMA_REQUEST_STATE_READY_TO_COMPLETE state.
    2375             :                          * We verified that data + response fit into send queue, so we can go to the next state directly
    2376             :                          */
    2377           1 :                         rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE;
    2378           1 :                         break;
    2379           7 :                 case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING:
    2380           7 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING, 0, 0,
    2381             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2382             : 
    2383           7 :                         if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_send_queue)) {
    2384             :                                 /* This request needs to wait in line to send the completion */
    2385           0 :                                 break;
    2386             :                         }
    2387             : 
    2388           7 :                         assert(rqpair->current_send_depth <= rqpair->max_send_depth);
    2389           7 :                         if (rqpair->current_send_depth == rqpair->max_send_depth) {
    2390             :                                 /* We can only have so many WRs outstanding. we have to wait until some finish */
    2391           2 :                                 rqpair->poller->stat.pending_rdma_send++;
    2392           2 :                                 break;
    2393             :                         }
    2394             : 
    2395             :                         /* We have already verified that this request is the head of the queue. */
    2396           5 :                         STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_send_queue, state_link);
    2397             : 
    2398             :                         /* The response sending will be kicked off from
    2399             :                          * RDMA_REQUEST_STATE_READY_TO_COMPLETE state.
    2400             :                          */
    2401           5 :                         rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE;
    2402           5 :                         break;
    2403           6 :                 case RDMA_REQUEST_STATE_READY_TO_COMPLETE:
    2404           6 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0,
    2405             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2406           6 :                         rc = request_transfer_out(&rdma_req->req, &data_posted);
    2407           6 :                         assert(rc == 0); /* No good way to handle this currently */
    2408           6 :                         if (spdk_unlikely(rc)) {
    2409           0 :                                 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
    2410             :                         } else {
    2411           6 :                                 rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST :
    2412             :                                                   RDMA_REQUEST_STATE_COMPLETING;
    2413             :                         }
    2414           6 :                         break;
    2415           1 :                 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST:
    2416           1 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0,
    2417             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2418             :                         /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED
    2419             :                          * to escape this state. */
    2420           1 :                         break;
    2421           5 :                 case RDMA_REQUEST_STATE_COMPLETING:
    2422           5 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0,
    2423             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair);
    2424             :                         /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED
    2425             :                          * to escape this state. */
    2426           5 :                         break;
    2427           6 :                 case RDMA_REQUEST_STATE_COMPLETED:
    2428           6 :                         spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0,
    2429             :                                           (uintptr_t)rdma_req, (uintptr_t)rqpair, rqpair->qpair.queue_depth);
    2430             : 
    2431           6 :                         rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc;
    2432           6 :                         _nvmf_rdma_request_free(rdma_req, rtransport);
    2433           6 :                         break;
    2434           0 :                 case RDMA_REQUEST_NUM_STATES:
    2435             :                 default:
    2436           0 :                         assert(0);
    2437             :                         break;
    2438             :                 }
    2439             : 
    2440          66 :                 if (rdma_req->state != prev_state) {
    2441          43 :                         progress = true;
    2442             :                 }
    2443          66 :         } while (rdma_req->state != prev_state);
    2444             : 
    2445          23 :         return progress;
    2446             : }
    2447             : 
    2448             : /* Public API callbacks begin here */
    2449             : 
    2450             : #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128
    2451             : #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128
    2452             : #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096
    2453             : #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128
    2454             : #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096
    2455             : #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072
    2456             : #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES)
    2457             : #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095
    2458             : #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE UINT32_MAX
    2459             : #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false
    2460             : #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false
    2461             : #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100
    2462             : #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1
    2463             : #define SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING false
    2464             : #define SPDK_NVMF_RDMA_DEFAULT_DATA_WR_POOL_SIZE 4095
    2465             : 
    2466             : static void
    2467           1 : nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts)
    2468             : {
    2469           1 :         opts->max_queue_depth =              SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH;
    2470           1 :         opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR;
    2471           1 :         opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE;
    2472           1 :         opts->max_io_size =          SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE;
    2473           1 :         opts->io_unit_size =         SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE;
    2474           1 :         opts->max_aq_depth =         SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH;
    2475           1 :         opts->num_shared_buffers =   SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS;
    2476           1 :         opts->buf_cache_size =               SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE;
    2477           1 :         opts->dif_insert_or_strip =  SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP;
    2478           1 :         opts->abort_timeout_sec =    SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC;
    2479           1 :         opts->transport_specific =      NULL;
    2480           1 :         opts->data_wr_pool_size      =       SPDK_NVMF_RDMA_DEFAULT_DATA_WR_POOL_SIZE;
    2481           1 : }
    2482             : 
    2483             : static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport,
    2484             :                              spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg);
    2485             : 
    2486             : static inline bool
    2487           0 : nvmf_rdma_is_rxe_device(struct spdk_nvmf_rdma_device *device)
    2488             : {
    2489           0 :         return device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_OLD ||
    2490           0 :                device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_NEW;
    2491             : }
    2492             : 
    2493             : static int nvmf_rdma_accept(void *ctx);
    2494             : static bool nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport);
    2495             : static void destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport,
    2496             :                               struct spdk_nvmf_rdma_device *device);
    2497             : 
    2498             : static int
    2499           0 : create_ib_device(struct spdk_nvmf_rdma_transport *rtransport, struct ibv_context *context,
    2500             :                  struct spdk_nvmf_rdma_device **new_device)
    2501             : {
    2502             :         struct spdk_nvmf_rdma_device    *device;
    2503           0 :         int                             flag = 0;
    2504           0 :         int                             rc = 0;
    2505             : 
    2506           0 :         device = calloc(1, sizeof(*device));
    2507           0 :         if (!device) {
    2508           0 :                 SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n");
    2509           0 :                 return -ENOMEM;
    2510             :         }
    2511           0 :         device->context = context;
    2512           0 :         rc = ibv_query_device(device->context, &device->attr);
    2513           0 :         if (rc < 0) {
    2514           0 :                 SPDK_ERRLOG("Failed to query RDMA device attributes.\n");
    2515           0 :                 free(device);
    2516           0 :                 return rc;
    2517             :         }
    2518             : 
    2519             : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
    2520           0 :         if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) {
    2521           0 :                 SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,");
    2522           0 :                 SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id);
    2523             :         }
    2524             : 
    2525             :         /**
    2526             :          * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE.
    2527             :          * The Soft-RoCE RXE driver does not currently support send with invalidate,
    2528             :          * but incorrectly reports that it does. There are changes making their way
    2529             :          * through the kernel now that will enable this feature. When they are merged,
    2530             :          * we can conditionally enable this feature.
    2531             :          *
    2532             :          * TODO: enable this for versions of the kernel rxe driver that support it.
    2533             :          */
    2534           0 :         if (nvmf_rdma_is_rxe_device(device)) {
    2535           0 :                 device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS);
    2536             :         }
    2537             : #endif
    2538             : 
    2539             :         /* set up device context async ev fd as NON_BLOCKING */
    2540           0 :         flag = fcntl(device->context->async_fd, F_GETFL);
    2541           0 :         rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK);
    2542           0 :         if (rc < 0) {
    2543           0 :                 SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n");
    2544           0 :                 free(device);
    2545           0 :                 return rc;
    2546             :         }
    2547             : 
    2548           0 :         TAILQ_INSERT_TAIL(&rtransport->devices, device, link);
    2549           0 :         SPDK_DEBUGLOG(rdma, "New device %p is added to RDMA trasport\n", device);
    2550             : 
    2551           0 :         if (g_nvmf_hooks.get_ibv_pd) {
    2552           0 :                 device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context);
    2553             :         } else {
    2554           0 :                 device->pd = ibv_alloc_pd(device->context);
    2555             :         }
    2556             : 
    2557           0 :         if (!device->pd) {
    2558           0 :                 SPDK_ERRLOG("Unable to allocate protection domain.\n");
    2559           0 :                 destroy_ib_device(rtransport, device);
    2560           0 :                 return -ENOMEM;
    2561             :         }
    2562             : 
    2563           0 :         assert(device->map == NULL);
    2564             : 
    2565           0 :         device->map = spdk_rdma_utils_create_mem_map(device->pd, &g_nvmf_hooks, IBV_ACCESS_LOCAL_WRITE);
    2566           0 :         if (!device->map) {
    2567           0 :                 SPDK_ERRLOG("Unable to allocate memory map for listen address\n");
    2568           0 :                 destroy_ib_device(rtransport, device);
    2569           0 :                 return -ENOMEM;
    2570             :         }
    2571             : 
    2572           0 :         assert(device->map != NULL);
    2573           0 :         assert(device->pd != NULL);
    2574             : 
    2575           0 :         if (new_device) {
    2576           0 :                 *new_device = device;
    2577             :         }
    2578           0 :         SPDK_NOTICELOG("Create IB device %s(%p/%p) succeed.\n", ibv_get_device_name(context->device),
    2579             :                        device, context);
    2580             : 
    2581           0 :         return 0;
    2582             : }
    2583             : 
    2584             : static void
    2585           0 : free_poll_fds(struct spdk_nvmf_rdma_transport *rtransport)
    2586             : {
    2587           0 :         if (rtransport->poll_fds) {
    2588           0 :                 free(rtransport->poll_fds);
    2589           0 :                 rtransport->poll_fds = NULL;
    2590             :         }
    2591           0 :         rtransport->npoll_fds = 0;
    2592           0 : }
    2593             : 
    2594             : static int
    2595           0 : generate_poll_fds(struct spdk_nvmf_rdma_transport *rtransport)
    2596             : {
    2597             :         /* Set up poll descriptor array to monitor events from RDMA and IB
    2598             :          * in a single poll syscall
    2599             :          */
    2600           0 :         int device_count = 0;
    2601           0 :         int i = 0;
    2602             :         struct spdk_nvmf_rdma_device *device, *tmp;
    2603             : 
    2604           0 :         TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) {
    2605           0 :                 device_count++;
    2606             :         }
    2607             : 
    2608           0 :         rtransport->npoll_fds = device_count + 1;
    2609             : 
    2610           0 :         rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd));
    2611           0 :         if (rtransport->poll_fds == NULL) {
    2612           0 :                 SPDK_ERRLOG("poll_fds allocation failed\n");
    2613           0 :                 return -ENOMEM;
    2614             :         }
    2615             : 
    2616           0 :         rtransport->poll_fds[i].fd = rtransport->event_channel->fd;
    2617           0 :         rtransport->poll_fds[i++].events = POLLIN;
    2618             : 
    2619           0 :         TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) {
    2620           0 :                 rtransport->poll_fds[i].fd = device->context->async_fd;
    2621           0 :                 rtransport->poll_fds[i++].events = POLLIN;
    2622             :         }
    2623             : 
    2624           0 :         return 0;
    2625             : }
    2626             : 
    2627             : static struct spdk_nvmf_transport *
    2628           0 : nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts)
    2629             : {
    2630             :         int rc;
    2631             :         struct spdk_nvmf_rdma_transport *rtransport;
    2632           0 :         struct spdk_nvmf_rdma_device    *device;
    2633             :         struct ibv_context              **contexts;
    2634             :         size_t                          data_wr_pool_size;
    2635             :         uint32_t                        i;
    2636             :         int                             flag;
    2637             :         uint32_t                        sge_count;
    2638             :         uint32_t                        min_shared_buffers;
    2639             :         uint32_t                        min_in_capsule_data_size;
    2640           0 :         int                             max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
    2641             : 
    2642           0 :         rtransport = calloc(1, sizeof(*rtransport));
    2643           0 :         if (!rtransport) {
    2644           0 :                 return NULL;
    2645             :         }
    2646             : 
    2647           0 :         TAILQ_INIT(&rtransport->devices);
    2648           0 :         TAILQ_INIT(&rtransport->ports);
    2649           0 :         TAILQ_INIT(&rtransport->poll_groups);
    2650           0 :         TAILQ_INIT(&rtransport->retry_ports);
    2651             : 
    2652           0 :         rtransport->transport.ops = &spdk_nvmf_transport_rdma;
    2653           0 :         rtransport->rdma_opts.num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE;
    2654           0 :         rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH;
    2655           0 :         rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ;
    2656           0 :         rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG;
    2657           0 :         rtransport->rdma_opts.no_wr_batching = SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING;
    2658           0 :         if (opts->transport_specific != NULL &&
    2659           0 :             spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder,
    2660             :                                             SPDK_COUNTOF(rdma_transport_opts_decoder),
    2661           0 :                                             &rtransport->rdma_opts)) {
    2662           0 :                 SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n");
    2663           0 :                 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2664           0 :                 return NULL;
    2665             :         }
    2666             : 
    2667           0 :         SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n"
    2668             :                      "  Transport opts:  max_ioq_depth=%d, max_io_size=%d,\n"
    2669             :                      "  max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n"
    2670             :                      "  in_capsule_data_size=%d, max_aq_depth=%d,\n"
    2671             :                      "  num_shared_buffers=%d, num_cqe=%d, max_srq_depth=%d, no_srq=%d,"
    2672             :                      "  acceptor_backlog=%d, no_wr_batching=%d abort_timeout_sec=%d\n",
    2673             :                      opts->max_queue_depth,
    2674             :                      opts->max_io_size,
    2675             :                      opts->max_qpairs_per_ctrlr - 1,
    2676             :                      opts->io_unit_size,
    2677             :                      opts->in_capsule_data_size,
    2678             :                      opts->max_aq_depth,
    2679             :                      opts->num_shared_buffers,
    2680             :                      rtransport->rdma_opts.num_cqe,
    2681             :                      rtransport->rdma_opts.max_srq_depth,
    2682             :                      rtransport->rdma_opts.no_srq,
    2683             :                      rtransport->rdma_opts.acceptor_backlog,
    2684             :                      rtransport->rdma_opts.no_wr_batching,
    2685             :                      opts->abort_timeout_sec);
    2686             : 
    2687             :         /* I/O unit size cannot be larger than max I/O size */
    2688           0 :         if (opts->io_unit_size > opts->max_io_size) {
    2689           0 :                 opts->io_unit_size = opts->max_io_size;
    2690             :         }
    2691             : 
    2692           0 :         if (rtransport->rdma_opts.acceptor_backlog <= 0) {
    2693           0 :                 SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n",
    2694             :                             SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG);
    2695           0 :                 rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG;
    2696             :         }
    2697             : 
    2698           0 :         if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) {
    2699           0 :                 SPDK_ERRLOG("The number of shared data buffers (%d) is less than"
    2700             :                             "the minimum number required to guarantee that forward progress can be made (%d)\n",
    2701             :                             opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2));
    2702           0 :                 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2703           0 :                 return NULL;
    2704             :         }
    2705             : 
    2706             :         /* If buf_cache_size == UINT32_MAX, we will dynamically pick a cache size later that we know will fit. */
    2707           0 :         if (opts->buf_cache_size < UINT32_MAX) {
    2708           0 :                 min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size;
    2709           0 :                 if (min_shared_buffers > opts->num_shared_buffers) {
    2710           0 :                         SPDK_ERRLOG("There are not enough buffers to satisfy"
    2711             :                                     "per-poll group caches for each thread. (%" PRIu32 ")"
    2712             :                                     "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers);
    2713           0 :                         SPDK_ERRLOG("Please specify a larger number of shared buffers\n");
    2714           0 :                         nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2715           0 :                         return NULL;
    2716             :                 }
    2717             :         }
    2718             : 
    2719           0 :         sge_count = opts->max_io_size / opts->io_unit_size;
    2720           0 :         if (sge_count > NVMF_DEFAULT_TX_SGE) {
    2721           0 :                 SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size);
    2722           0 :                 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2723           0 :                 return NULL;
    2724             :         }
    2725             : 
    2726           0 :         min_in_capsule_data_size = sizeof(struct spdk_nvme_sgl_descriptor) * SPDK_NVMF_MAX_SGL_ENTRIES;
    2727           0 :         if (opts->in_capsule_data_size < min_in_capsule_data_size) {
    2728           0 :                 SPDK_WARNLOG("In capsule data size is set to %u, this is minimum size required to support msdbd=16\n",
    2729             :                              min_in_capsule_data_size);
    2730           0 :                 opts->in_capsule_data_size = min_in_capsule_data_size;
    2731             :         }
    2732             : 
    2733           0 :         rtransport->event_channel = rdma_create_event_channel();
    2734           0 :         if (rtransport->event_channel == NULL) {
    2735           0 :                 SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno));
    2736           0 :                 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2737           0 :                 return NULL;
    2738             :         }
    2739             : 
    2740           0 :         flag = fcntl(rtransport->event_channel->fd, F_GETFL);
    2741           0 :         if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) {
    2742           0 :                 SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n",
    2743             :                             rtransport->event_channel->fd, spdk_strerror(errno));
    2744           0 :                 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2745           0 :                 return NULL;
    2746             :         }
    2747             : 
    2748           0 :         data_wr_pool_size = opts->data_wr_pool_size;
    2749           0 :         if (data_wr_pool_size < SPDK_NVMF_MAX_SGL_ENTRIES * 2 * spdk_env_get_core_count()) {
    2750           0 :                 data_wr_pool_size = SPDK_NVMF_MAX_SGL_ENTRIES * 2 * spdk_env_get_core_count();
    2751           0 :                 SPDK_NOTICELOG("data_wr_pool_size is changed to %zu to guarantee enough cache for handling "
    2752             :                                "at least one IO in each core\n", data_wr_pool_size);
    2753             :         }
    2754           0 :         rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", data_wr_pool_size,
    2755             :                                    sizeof(struct spdk_nvmf_rdma_request_data), SPDK_MEMPOOL_DEFAULT_CACHE_SIZE,
    2756             :                                    SPDK_ENV_SOCKET_ID_ANY);
    2757           0 :         if (!rtransport->data_wr_pool) {
    2758           0 :                 if (spdk_mempool_lookup("spdk_nvmf_rdma_wr_data") != NULL) {
    2759           0 :                         SPDK_ERRLOG("Unable to allocate work request pool for poll group: already exists\n");
    2760           0 :                         SPDK_ERRLOG("Probably running in multiprocess environment, which is "
    2761             :                                     "unsupported by the nvmf library\n");
    2762             :                 } else {
    2763           0 :                         SPDK_ERRLOG("Unable to allocate work request pool for poll group\n");
    2764             :                 }
    2765           0 :                 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2766           0 :                 return NULL;
    2767             :         }
    2768             : 
    2769           0 :         contexts = rdma_get_devices(NULL);
    2770           0 :         if (contexts == NULL) {
    2771           0 :                 SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno);
    2772           0 :                 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2773           0 :                 return NULL;
    2774             :         }
    2775             : 
    2776           0 :         i = 0;
    2777           0 :         rc = 0;
    2778           0 :         while (contexts[i] != NULL) {
    2779           0 :                 rc = create_ib_device(rtransport, contexts[i], &device);
    2780           0 :                 if (rc < 0) {
    2781           0 :                         break;
    2782             :                 }
    2783           0 :                 i++;
    2784           0 :                 max_device_sge = spdk_min(max_device_sge, device->attr.max_sge);
    2785           0 :                 device->is_ready = true;
    2786             :         }
    2787           0 :         rdma_free_devices(contexts);
    2788             : 
    2789           0 :         if (opts->io_unit_size * max_device_sge < opts->max_io_size) {
    2790             :                 /* divide and round up. */
    2791           0 :                 opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge;
    2792             : 
    2793             :                 /* round up to the nearest 4k. */
    2794           0 :                 opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK;
    2795             : 
    2796           0 :                 opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE);
    2797           0 :                 SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n",
    2798             :                                opts->io_unit_size);
    2799             :         }
    2800             : 
    2801           0 :         if (rc < 0) {
    2802           0 :                 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2803           0 :                 return NULL;
    2804             :         }
    2805             : 
    2806           0 :         rc = generate_poll_fds(rtransport);
    2807           0 :         if (rc < 0) {
    2808           0 :                 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2809           0 :                 return NULL;
    2810             :         }
    2811             : 
    2812           0 :         rtransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_rdma_accept, &rtransport->transport,
    2813             :                                     opts->acceptor_poll_rate);
    2814           0 :         if (!rtransport->accept_poller) {
    2815           0 :                 nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
    2816           0 :                 return NULL;
    2817             :         }
    2818             : 
    2819           0 :         return &rtransport->transport;
    2820             : }
    2821             : 
    2822             : static void
    2823           0 : destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport,
    2824             :                   struct spdk_nvmf_rdma_device *device)
    2825             : {
    2826           0 :         TAILQ_REMOVE(&rtransport->devices, device, link);
    2827           0 :         spdk_rdma_utils_free_mem_map(&device->map);
    2828           0 :         if (device->pd) {
    2829           0 :                 if (!g_nvmf_hooks.get_ibv_pd) {
    2830           0 :                         ibv_dealloc_pd(device->pd);
    2831             :                 }
    2832             :         }
    2833           0 :         SPDK_DEBUGLOG(rdma, "IB device [%p] is destroyed.\n", device);
    2834           0 :         free(device);
    2835           0 : }
    2836             : 
    2837             : static void
    2838           0 : nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w)
    2839             : {
    2840             :         struct spdk_nvmf_rdma_transport *rtransport;
    2841           0 :         assert(w != NULL);
    2842             : 
    2843           0 :         rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
    2844           0 :         spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth);
    2845           0 :         spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq);
    2846           0 :         if (rtransport->rdma_opts.no_srq == true) {
    2847           0 :                 spdk_json_write_named_int32(w, "num_cqe", rtransport->rdma_opts.num_cqe);
    2848             :         }
    2849           0 :         spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog);
    2850           0 :         spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching);
    2851           0 : }
    2852             : 
    2853             : static int
    2854           0 : nvmf_rdma_destroy(struct spdk_nvmf_transport *transport,
    2855             :                   spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg)
    2856             : {
    2857             :         struct spdk_nvmf_rdma_transport *rtransport;
    2858             :         struct spdk_nvmf_rdma_port      *port, *port_tmp;
    2859             :         struct spdk_nvmf_rdma_device    *device, *device_tmp;
    2860             : 
    2861           0 :         rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
    2862             : 
    2863           0 :         TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, port_tmp) {
    2864           0 :                 TAILQ_REMOVE(&rtransport->retry_ports, port, link);
    2865           0 :                 free(port);
    2866             :         }
    2867             : 
    2868           0 :         TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) {
    2869           0 :                 TAILQ_REMOVE(&rtransport->ports, port, link);
    2870           0 :                 rdma_destroy_id(port->id);
    2871           0 :                 free(port);
    2872             :         }
    2873             : 
    2874           0 :         free_poll_fds(rtransport);
    2875             : 
    2876           0 :         if (rtransport->event_channel != NULL) {
    2877           0 :                 rdma_destroy_event_channel(rtransport->event_channel);
    2878             :         }
    2879             : 
    2880           0 :         TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) {
    2881           0 :                 destroy_ib_device(rtransport, device);
    2882             :         }
    2883             : 
    2884           0 :         if (rtransport->data_wr_pool != NULL) {
    2885           0 :                 if (spdk_mempool_count(rtransport->data_wr_pool) != transport->opts.data_wr_pool_size) {
    2886           0 :                         SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n",
    2887             :                                     spdk_mempool_count(rtransport->data_wr_pool),
    2888             :                                     transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES);
    2889             :                 }
    2890             :         }
    2891             : 
    2892           0 :         spdk_mempool_free(rtransport->data_wr_pool);
    2893             : 
    2894           0 :         spdk_poller_unregister(&rtransport->accept_poller);
    2895           0 :         free(rtransport);
    2896             : 
    2897           0 :         if (cb_fn) {
    2898           0 :                 cb_fn(cb_arg);
    2899             :         }
    2900           0 :         return 0;
    2901             : }
    2902             : 
    2903             : static int nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id,
    2904             :                                      struct spdk_nvme_transport_id *trid,
    2905             :                                      bool peer);
    2906             : 
    2907             : static bool nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport);
    2908             : 
    2909             : static int
    2910           0 : nvmf_rdma_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid,
    2911             :                  struct spdk_nvmf_listen_opts *listen_opts)
    2912             : {
    2913             :         struct spdk_nvmf_rdma_transport *rtransport;
    2914             :         struct spdk_nvmf_rdma_device    *device;
    2915             :         struct spdk_nvmf_rdma_port      *port, *tmp_port;
    2916           0 :         struct addrinfo                 *res;
    2917           0 :         struct addrinfo                 hints;
    2918             :         int                             family;
    2919             :         int                             rc;
    2920             :         long int                        port_val;
    2921           0 :         bool                            is_retry = false;
    2922             : 
    2923           0 :         if (!strlen(trid->trsvcid)) {
    2924           0 :                 SPDK_ERRLOG("Service id is required\n");
    2925           0 :                 return -EINVAL;
    2926             :         }
    2927             : 
    2928           0 :         rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
    2929           0 :         assert(rtransport->event_channel != NULL);
    2930             : 
    2931           0 :         port = calloc(1, sizeof(*port));
    2932           0 :         if (!port) {
    2933           0 :                 SPDK_ERRLOG("Port allocation failed\n");
    2934           0 :                 return -ENOMEM;
    2935             :         }
    2936             : 
    2937           0 :         port->trid = trid;
    2938             : 
    2939           0 :         switch (trid->adrfam) {
    2940           0 :         case SPDK_NVMF_ADRFAM_IPV4:
    2941           0 :                 family = AF_INET;
    2942           0 :                 break;
    2943           0 :         case SPDK_NVMF_ADRFAM_IPV6:
    2944           0 :                 family = AF_INET6;
    2945           0 :                 break;
    2946           0 :         default:
    2947           0 :                 SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam);
    2948           0 :                 free(port);
    2949           0 :                 return -EINVAL;
    2950             :         }
    2951             : 
    2952           0 :         memset(&hints, 0, sizeof(hints));
    2953           0 :         hints.ai_family = family;
    2954           0 :         hints.ai_flags = AI_NUMERICSERV;
    2955           0 :         hints.ai_socktype = SOCK_STREAM;
    2956           0 :         hints.ai_protocol = 0;
    2957             : 
    2958             :         /* Range check the trsvcid. Fail in 3 cases:
    2959             :          * < 0: means that spdk_strtol hit an error
    2960             :          * 0: this results in ephemeral port which we don't want
    2961             :          * > 65535: port too high
    2962             :          */
    2963           0 :         port_val = spdk_strtol(trid->trsvcid, 10);
    2964           0 :         if (port_val <= 0 || port_val > 65535) {
    2965           0 :                 SPDK_ERRLOG("invalid trsvcid %s\n", trid->trsvcid);
    2966           0 :                 free(port);
    2967           0 :                 return -EINVAL;
    2968             :         }
    2969             : 
    2970           0 :         rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res);
    2971           0 :         if (rc) {
    2972           0 :                 SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc);
    2973           0 :                 free(port);
    2974           0 :                 return -(abs(rc));
    2975             :         }
    2976             : 
    2977           0 :         rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP);
    2978           0 :         if (rc < 0) {
    2979           0 :                 SPDK_ERRLOG("rdma_create_id() failed\n");
    2980           0 :                 freeaddrinfo(res);
    2981           0 :                 free(port);
    2982           0 :                 return rc;
    2983             :         }
    2984             : 
    2985           0 :         rc = rdma_bind_addr(port->id, res->ai_addr);
    2986           0 :         freeaddrinfo(res);
    2987             : 
    2988           0 :         if (rc < 0) {
    2989           0 :                 TAILQ_FOREACH(tmp_port, &rtransport->retry_ports, link) {
    2990           0 :                         if (spdk_nvme_transport_id_compare(tmp_port->trid, trid) == 0) {
    2991           0 :                                 is_retry = true;
    2992           0 :                                 break;
    2993             :                         }
    2994             :                 }
    2995           0 :                 if (!is_retry) {
    2996           0 :                         SPDK_ERRLOG("rdma_bind_addr() failed\n");
    2997             :                 }
    2998           0 :                 rdma_destroy_id(port->id);
    2999           0 :                 free(port);
    3000           0 :                 return rc;
    3001             :         }
    3002             : 
    3003           0 :         if (!port->id->verbs) {
    3004           0 :                 SPDK_ERRLOG("ibv_context is null\n");
    3005           0 :                 rdma_destroy_id(port->id);
    3006           0 :                 free(port);
    3007           0 :                 return -1;
    3008             :         }
    3009             : 
    3010           0 :         rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog);
    3011           0 :         if (rc < 0) {
    3012           0 :                 SPDK_ERRLOG("rdma_listen() failed\n");
    3013           0 :                 rdma_destroy_id(port->id);
    3014           0 :                 free(port);
    3015           0 :                 return rc;
    3016             :         }
    3017             : 
    3018           0 :         TAILQ_FOREACH(device, &rtransport->devices, link) {
    3019           0 :                 if (device->context == port->id->verbs && device->is_ready) {
    3020           0 :                         port->device = device;
    3021           0 :                         break;
    3022             :                 }
    3023             :         }
    3024           0 :         if (!port->device) {
    3025           0 :                 SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n",
    3026             :                             port->id->verbs);
    3027           0 :                 rdma_destroy_id(port->id);
    3028           0 :                 free(port);
    3029           0 :                 nvmf_rdma_rescan_devices(rtransport);
    3030           0 :                 return -EINVAL;
    3031             :         }
    3032             : 
    3033           0 :         SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n",
    3034             :                        trid->traddr, trid->trsvcid);
    3035             : 
    3036           0 :         TAILQ_INSERT_TAIL(&rtransport->ports, port, link);
    3037           0 :         return 0;
    3038             : }
    3039             : 
    3040             : static void
    3041           0 : nvmf_rdma_stop_listen_ex(struct spdk_nvmf_transport *transport,
    3042             :                          const struct spdk_nvme_transport_id *trid, bool need_retry)
    3043             : {
    3044             :         struct spdk_nvmf_rdma_transport *rtransport;
    3045             :         struct spdk_nvmf_rdma_port      *port, *tmp;
    3046             : 
    3047           0 :         rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
    3048             : 
    3049           0 :         if (!need_retry) {
    3050           0 :                 TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp) {
    3051           0 :                         if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) {
    3052           0 :                                 TAILQ_REMOVE(&rtransport->retry_ports, port, link);
    3053           0 :                                 free(port);
    3054             :                         }
    3055             :                 }
    3056             :         }
    3057             : 
    3058           0 :         TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) {
    3059           0 :                 if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) {
    3060           0 :                         SPDK_DEBUGLOG(rdma, "Port %s:%s removed. need retry: %d\n",
    3061             :                                       port->trid->traddr, port->trid->trsvcid, need_retry);
    3062           0 :                         TAILQ_REMOVE(&rtransport->ports, port, link);
    3063           0 :                         rdma_destroy_id(port->id);
    3064           0 :                         port->id = NULL;
    3065           0 :                         port->device = NULL;
    3066           0 :                         if (need_retry) {
    3067           0 :                                 TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link);
    3068             :                         } else {
    3069           0 :                                 free(port);
    3070             :                         }
    3071           0 :                         break;
    3072             :                 }
    3073             :         }
    3074           0 : }
    3075             : 
    3076             : static void
    3077           0 : nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport,
    3078             :                       const struct spdk_nvme_transport_id *trid)
    3079             : {
    3080           0 :         nvmf_rdma_stop_listen_ex(transport, trid, false);
    3081           0 : }
    3082             : 
    3083             : static void _nvmf_rdma_register_poller_in_group(void *c);
    3084             : static void _nvmf_rdma_remove_poller_in_group(void *c);
    3085             : 
    3086             : static bool
    3087           0 : nvmf_rdma_all_pollers_management_done(void *c)
    3088             : {
    3089           0 :         struct poller_manage_ctx        *ctx = c;
    3090             :         int                             counter;
    3091             : 
    3092           0 :         counter = __atomic_sub_fetch(ctx->inflight_op_counter, 1, __ATOMIC_SEQ_CST);
    3093           0 :         SPDK_DEBUGLOG(rdma, "nvmf_rdma_all_pollers_management_done called. counter: %d, poller: %p\n",
    3094             :                       counter, ctx->rpoller);
    3095             : 
    3096           0 :         if (counter == 0) {
    3097           0 :                 free((void *)ctx->inflight_op_counter);
    3098             :         }
    3099           0 :         free(ctx);
    3100             : 
    3101           0 :         return counter == 0;
    3102             : }
    3103             : 
    3104             : static int
    3105           0 : nvmf_rdma_manage_poller(struct spdk_nvmf_rdma_transport *rtransport,
    3106             :                         struct spdk_nvmf_rdma_device *device, bool *has_inflight, bool is_add)
    3107             : {
    3108             :         struct spdk_nvmf_rdma_poll_group        *rgroup;
    3109             :         struct spdk_nvmf_rdma_poller            *rpoller;
    3110             :         struct spdk_nvmf_poll_group             *poll_group;
    3111             :         struct poller_manage_ctx                *ctx;
    3112             :         bool                                    found;
    3113             :         int                                     *inflight_counter;
    3114             :         spdk_msg_fn                             do_fn;
    3115             : 
    3116           0 :         *has_inflight = false;
    3117           0 :         do_fn = is_add ? _nvmf_rdma_register_poller_in_group : _nvmf_rdma_remove_poller_in_group;
    3118           0 :         inflight_counter = calloc(1, sizeof(int));
    3119           0 :         if (!inflight_counter) {
    3120           0 :                 SPDK_ERRLOG("Failed to allocate inflight counter when removing pollers\n");
    3121           0 :                 return -ENOMEM;
    3122             :         }
    3123             : 
    3124           0 :         TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) {
    3125           0 :                 (*inflight_counter)++;
    3126             :         }
    3127             : 
    3128           0 :         TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) {
    3129           0 :                 found = false;
    3130           0 :                 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
    3131           0 :                         if (rpoller->device == device) {
    3132           0 :                                 found = true;
    3133           0 :                                 break;
    3134             :                         }
    3135             :                 }
    3136           0 :                 if (found == is_add) {
    3137           0 :                         __atomic_fetch_sub(inflight_counter, 1, __ATOMIC_SEQ_CST);
    3138           0 :                         continue;
    3139             :                 }
    3140             : 
    3141           0 :                 ctx = calloc(1, sizeof(struct poller_manage_ctx));
    3142           0 :                 if (!ctx) {
    3143           0 :                         SPDK_ERRLOG("Failed to allocate poller_manage_ctx when removing pollers\n");
    3144           0 :                         if (!*has_inflight) {
    3145           0 :                                 free(inflight_counter);
    3146             :                         }
    3147           0 :                         return -ENOMEM;
    3148             :                 }
    3149             : 
    3150           0 :                 ctx->rtransport = rtransport;
    3151           0 :                 ctx->rgroup = rgroup;
    3152           0 :                 ctx->rpoller = rpoller;
    3153           0 :                 ctx->device = device;
    3154           0 :                 ctx->thread = spdk_get_thread();
    3155           0 :                 ctx->inflight_op_counter = inflight_counter;
    3156           0 :                 *has_inflight = true;
    3157             : 
    3158           0 :                 poll_group = rgroup->group.group;
    3159           0 :                 if (poll_group->thread != spdk_get_thread()) {
    3160           0 :                         spdk_thread_send_msg(poll_group->thread, do_fn, ctx);
    3161             :                 } else {
    3162           0 :                         do_fn(ctx);
    3163             :                 }
    3164             :         }
    3165             : 
    3166           0 :         if (!*has_inflight) {
    3167           0 :                 free(inflight_counter);
    3168             :         }
    3169             : 
    3170           0 :         return 0;
    3171             : }
    3172             : 
    3173             : static void nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport,
    3174             :                 struct spdk_nvmf_rdma_device *device);
    3175             : 
    3176             : static struct spdk_nvmf_rdma_device *
    3177           0 : nvmf_rdma_find_ib_device(struct spdk_nvmf_rdma_transport *rtransport,
    3178             :                          struct ibv_context *context)
    3179             : {
    3180             :         struct spdk_nvmf_rdma_device    *device, *tmp_device;
    3181             : 
    3182           0 :         TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp_device) {
    3183           0 :                 if (device->need_destroy) {
    3184           0 :                         continue;
    3185             :                 }
    3186             : 
    3187           0 :                 if (strcmp(device->context->device->dev_name, context->device->dev_name) == 0) {
    3188           0 :                         return device;
    3189             :                 }
    3190             :         }
    3191             : 
    3192           0 :         return NULL;
    3193             : }
    3194             : 
    3195             : static bool
    3196           0 : nvmf_rdma_check_devices_context(struct spdk_nvmf_rdma_transport *rtransport,
    3197             :                                 struct ibv_context *context)
    3198             : {
    3199           0 :         struct spdk_nvmf_rdma_device    *old_device, *new_device;
    3200           0 :         int                             rc = 0;
    3201           0 :         bool                            has_inflight;
    3202             : 
    3203           0 :         old_device = nvmf_rdma_find_ib_device(rtransport, context);
    3204             : 
    3205           0 :         if (old_device) {
    3206           0 :                 if (old_device->context != context && !old_device->need_destroy && old_device->is_ready) {
    3207             :                         /* context may not have time to be cleaned when rescan. exactly one context
    3208             :                          * is valid for a device so this context must be invalid and just remove it. */
    3209           0 :                         SPDK_WARNLOG("Device %p has a invalid context %p\n", old_device, old_device->context);
    3210           0 :                         old_device->need_destroy = true;
    3211           0 :                         nvmf_rdma_handle_device_removal(rtransport, old_device);
    3212             :                 }
    3213           0 :                 return false;
    3214             :         }
    3215             : 
    3216           0 :         rc = create_ib_device(rtransport, context, &new_device);
    3217             :         /* TODO: update transport opts. */
    3218           0 :         if (rc < 0) {
    3219           0 :                 SPDK_ERRLOG("Failed to create ib device for context: %s(%p)\n",
    3220             :                             ibv_get_device_name(context->device), context);
    3221           0 :                 return false;
    3222             :         }
    3223             : 
    3224           0 :         rc = nvmf_rdma_manage_poller(rtransport, new_device, &has_inflight, true);
    3225           0 :         if (rc < 0) {
    3226           0 :                 SPDK_ERRLOG("Failed to add poller for device context: %s(%p)\n",
    3227             :                             ibv_get_device_name(context->device), context);
    3228           0 :                 return false;
    3229             :         }
    3230             : 
    3231           0 :         if (has_inflight) {
    3232           0 :                 new_device->is_ready = true;
    3233             :         }
    3234             : 
    3235           0 :         return true;
    3236             : }
    3237             : 
    3238             : static bool
    3239           0 : nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport)
    3240             : {
    3241             :         struct spdk_nvmf_rdma_device    *device;
    3242           0 :         struct ibv_device               **ibv_device_list = NULL;
    3243           0 :         struct ibv_context              **contexts = NULL;
    3244           0 :         int                             i = 0;
    3245           0 :         int                             num_dev = 0;
    3246           0 :         bool                            new_create = false, has_new_device = false;
    3247           0 :         struct ibv_context              *tmp_verbs = NULL;
    3248             : 
    3249             :         /* do not rescan when any device is destroying, or context may be freed when
    3250             :          * regenerating the poll fds.
    3251             :          */
    3252           0 :         TAILQ_FOREACH(device, &rtransport->devices, link) {
    3253           0 :                 if (device->need_destroy) {
    3254           0 :                         return false;
    3255             :                 }
    3256             :         }
    3257             : 
    3258           0 :         ibv_device_list = ibv_get_device_list(&num_dev);
    3259             : 
    3260             :         /* There is a bug in librdmacm. If verbs init failed in rdma_get_devices, it'll be
    3261             :          * marked as dead verbs and never be init again. So we need to make sure the
    3262             :          * verbs is available before we call rdma_get_devices. */
    3263           0 :         if (num_dev >= 0) {
    3264           0 :                 for (i = 0; i < num_dev; i++) {
    3265           0 :                         tmp_verbs = ibv_open_device(ibv_device_list[i]);
    3266           0 :                         if (!tmp_verbs) {
    3267           0 :                                 SPDK_WARNLOG("Failed to init ibv device %p, err %d. Skip rescan.\n", ibv_device_list[i], errno);
    3268           0 :                                 break;
    3269             :                         }
    3270           0 :                         if (nvmf_rdma_find_ib_device(rtransport, tmp_verbs) == NULL) {
    3271           0 :                                 SPDK_DEBUGLOG(rdma, "Find new verbs init ibv device %p(%s).\n", ibv_device_list[i],
    3272             :                                               tmp_verbs->device->dev_name);
    3273           0 :                                 has_new_device = true;
    3274             :                         }
    3275           0 :                         ibv_close_device(tmp_verbs);
    3276             :                 }
    3277           0 :                 ibv_free_device_list(ibv_device_list);
    3278           0 :                 if (!tmp_verbs || !has_new_device) {
    3279           0 :                         return false;
    3280             :                 }
    3281             :         }
    3282             : 
    3283           0 :         contexts = rdma_get_devices(NULL);
    3284             : 
    3285           0 :         for (i = 0; contexts && contexts[i] != NULL; i++) {
    3286           0 :                 new_create |= nvmf_rdma_check_devices_context(rtransport, contexts[i]);
    3287             :         }
    3288             : 
    3289           0 :         if (new_create) {
    3290           0 :                 free_poll_fds(rtransport);
    3291           0 :                 generate_poll_fds(rtransport);
    3292             :         }
    3293             : 
    3294           0 :         if (contexts) {
    3295           0 :                 rdma_free_devices(contexts);
    3296             :         }
    3297             : 
    3298           0 :         return new_create;
    3299             : }
    3300             : 
    3301             : static bool
    3302           0 : nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport)
    3303             : {
    3304             :         struct spdk_nvmf_rdma_port      *port, *tmp_port;
    3305           0 :         int                             rc = 0;
    3306           0 :         bool                            new_create = false;
    3307             : 
    3308           0 :         if (TAILQ_EMPTY(&rtransport->retry_ports)) {
    3309           0 :                 return false;
    3310             :         }
    3311             : 
    3312           0 :         new_create = nvmf_rdma_rescan_devices(rtransport);
    3313             : 
    3314           0 :         TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp_port) {
    3315           0 :                 rc = nvmf_rdma_listen(&rtransport->transport, port->trid, NULL);
    3316             : 
    3317           0 :                 TAILQ_REMOVE(&rtransport->retry_ports, port, link);
    3318           0 :                 if (rc) {
    3319           0 :                         if (new_create) {
    3320           0 :                                 SPDK_ERRLOG("Found new IB device but port %s:%s is still failed(%d) to listen.\n",
    3321             :                                             port->trid->traddr, port->trid->trsvcid, rc);
    3322             :                         }
    3323           0 :                         TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link);
    3324           0 :                         break;
    3325             :                 } else {
    3326           0 :                         SPDK_NOTICELOG("Port %s:%s come back\n", port->trid->traddr, port->trid->trsvcid);
    3327           0 :                         free(port);
    3328             :                 }
    3329             :         }
    3330             : 
    3331           0 :         return true;
    3332             : }
    3333             : 
    3334             : static void
    3335           0 : nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport,
    3336             :                                 struct spdk_nvmf_rdma_qpair *rqpair, bool drain)
    3337             : {
    3338             :         struct spdk_nvmf_request *req, *tmp;
    3339             :         struct spdk_nvmf_rdma_request   *rdma_req, *req_tmp;
    3340             :         struct spdk_nvmf_rdma_resources *resources;
    3341             : 
    3342             :         /* First process requests which are waiting for response to be sent */
    3343           0 :         STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_send_queue, state_link, req_tmp) {
    3344           0 :                 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
    3345           0 :                         break;
    3346             :                 }
    3347             :         }
    3348             : 
    3349             :         /* We process I/O in the data transfer pending queue at the highest priority. */
    3350           0 :         STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) {
    3351           0 :                 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
    3352           0 :                         break;
    3353             :                 }
    3354             :         }
    3355             : 
    3356             :         /* Then RDMA writes since reads have stronger restrictions than writes */
    3357           0 :         STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) {
    3358           0 :                 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
    3359           0 :                         break;
    3360             :                 }
    3361             :         }
    3362             : 
    3363             :         /* Then we handle request waiting on memory buffers. */
    3364           0 :         STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) {
    3365           0 :                 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
    3366           0 :                 if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
    3367           0 :                         break;
    3368             :                 }
    3369             :         }
    3370             : 
    3371           0 :         resources = rqpair->resources;
    3372           0 :         while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) {
    3373           0 :                 rdma_req = STAILQ_FIRST(&resources->free_queue);
    3374           0 :                 STAILQ_REMOVE_HEAD(&resources->free_queue, state_link);
    3375           0 :                 rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue);
    3376           0 :                 STAILQ_REMOVE_HEAD(&resources->incoming_queue, link);
    3377             : 
    3378           0 :                 if (rqpair->srq != NULL) {
    3379           0 :                         rdma_req->req.qpair = &rdma_req->recv->qpair->qpair;
    3380           0 :                         rdma_req->recv->qpair->qd++;
    3381             :                 } else {
    3382           0 :                         rqpair->qd++;
    3383             :                 }
    3384             : 
    3385           0 :                 rdma_req->receive_tsc = rdma_req->recv->receive_tsc;
    3386           0 :                 rdma_req->state = RDMA_REQUEST_STATE_NEW;
    3387           0 :                 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) {
    3388           0 :                         break;
    3389             :                 }
    3390             :         }
    3391           0 :         if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) {
    3392           0 :                 rqpair->poller->stat.pending_free_request++;
    3393             :         }
    3394           0 : }
    3395             : 
    3396             : static void
    3397           0 : nvmf_rdma_poller_process_pending_buf_queue(struct spdk_nvmf_rdma_transport *rtransport,
    3398             :                 struct spdk_nvmf_rdma_poller *rpoller)
    3399             : {
    3400             :         struct spdk_nvmf_request *req, *tmp;
    3401             :         struct spdk_nvmf_rdma_request *rdma_req;
    3402             : 
    3403           0 :         STAILQ_FOREACH_SAFE(req, &rpoller->group->group.pending_buf_queue, buf_link, tmp) {
    3404           0 :                 rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
    3405           0 :                 if (nvmf_rdma_request_process(rtransport, rdma_req) == false) {
    3406           0 :                         break;
    3407             :                 }
    3408             :         }
    3409           0 : }
    3410             : 
    3411             : static inline bool
    3412           0 : nvmf_rdma_can_ignore_last_wqe_reached(struct spdk_nvmf_rdma_device *device)
    3413             : {
    3414             :         /* iWARP transport and SoftRoCE driver don't support LAST_WQE_REACHED ibv async event */
    3415           0 :         return nvmf_rdma_is_rxe_device(device) ||
    3416           0 :                device->context->device->transport_type == IBV_TRANSPORT_IWARP;
    3417             : }
    3418             : 
    3419             : static void
    3420           0 : nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair)
    3421             : {
    3422           0 :         struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
    3423             :                         struct spdk_nvmf_rdma_transport, transport);
    3424             : 
    3425           0 :         nvmf_rdma_qpair_process_pending(rtransport, rqpair, true);
    3426             : 
    3427             :         /* nvmf_rdma_close_qpair is not called */
    3428           0 :         if (!rqpair->to_close) {
    3429           0 :                 return;
    3430             :         }
    3431             : 
    3432             :         /* device is already destroyed and we should force destroy this qpair. */
    3433           0 :         if (rqpair->poller && rqpair->poller->need_destroy) {
    3434           0 :                 nvmf_rdma_qpair_destroy(rqpair);
    3435           0 :                 return;
    3436             :         }
    3437             : 
    3438             :         /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */
    3439           0 :         if (rqpair->current_send_depth != 0) {
    3440           0 :                 return;
    3441             :         }
    3442             : 
    3443           0 :         if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) {
    3444           0 :                 return;
    3445             :         }
    3446             : 
    3447           0 :         if (rqpair->srq != NULL && rqpair->last_wqe_reached == false &&
    3448           0 :             !nvmf_rdma_can_ignore_last_wqe_reached(rqpair->device)) {
    3449           0 :                 return;
    3450             :         }
    3451             : 
    3452           0 :         assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR);
    3453             : 
    3454           0 :         nvmf_rdma_qpair_destroy(rqpair);
    3455             : }
    3456             : 
    3457             : static int
    3458           0 : nvmf_rdma_disconnect(struct rdma_cm_event *evt, bool *event_acked)
    3459             : {
    3460             :         struct spdk_nvmf_qpair          *qpair;
    3461             :         struct spdk_nvmf_rdma_qpair     *rqpair;
    3462             : 
    3463           0 :         if (evt->id == NULL) {
    3464           0 :                 SPDK_ERRLOG("disconnect request: missing cm_id\n");
    3465           0 :                 return -1;
    3466             :         }
    3467             : 
    3468           0 :         qpair = evt->id->context;
    3469           0 :         if (qpair == NULL) {
    3470           0 :                 SPDK_ERRLOG("disconnect request: no active connection\n");
    3471           0 :                 return -1;
    3472             :         }
    3473             : 
    3474           0 :         rdma_ack_cm_event(evt);
    3475           0 :         *event_acked = true;
    3476             : 
    3477           0 :         rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
    3478             : 
    3479           0 :         spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair);
    3480             : 
    3481           0 :         spdk_nvmf_qpair_disconnect(&rqpair->qpair);
    3482             : 
    3483           0 :         return 0;
    3484             : }
    3485             : 
    3486             : #ifdef DEBUG
    3487             : static const char *CM_EVENT_STR[] = {
    3488             :         "RDMA_CM_EVENT_ADDR_RESOLVED",
    3489             :         "RDMA_CM_EVENT_ADDR_ERROR",
    3490             :         "RDMA_CM_EVENT_ROUTE_RESOLVED",
    3491             :         "RDMA_CM_EVENT_ROUTE_ERROR",
    3492             :         "RDMA_CM_EVENT_CONNECT_REQUEST",
    3493             :         "RDMA_CM_EVENT_CONNECT_RESPONSE",
    3494             :         "RDMA_CM_EVENT_CONNECT_ERROR",
    3495             :         "RDMA_CM_EVENT_UNREACHABLE",
    3496             :         "RDMA_CM_EVENT_REJECTED",
    3497             :         "RDMA_CM_EVENT_ESTABLISHED",
    3498             :         "RDMA_CM_EVENT_DISCONNECTED",
    3499             :         "RDMA_CM_EVENT_DEVICE_REMOVAL",
    3500             :         "RDMA_CM_EVENT_MULTICAST_JOIN",
    3501             :         "RDMA_CM_EVENT_MULTICAST_ERROR",
    3502             :         "RDMA_CM_EVENT_ADDR_CHANGE",
    3503             :         "RDMA_CM_EVENT_TIMEWAIT_EXIT"
    3504             : };
    3505             : #endif /* DEBUG */
    3506             : 
    3507             : static void
    3508           0 : nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport,
    3509             :                                     struct spdk_nvmf_rdma_port *port)
    3510             : {
    3511             :         struct spdk_nvmf_rdma_poll_group        *rgroup;
    3512             :         struct spdk_nvmf_rdma_poller            *rpoller;
    3513             :         struct spdk_nvmf_rdma_qpair             *rqpair;
    3514             : 
    3515           0 :         TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) {
    3516           0 :                 TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
    3517           0 :                         RB_FOREACH(rqpair, qpairs_tree, &rpoller->qpairs) {
    3518           0 :                                 if (rqpair->listen_id == port->id) {
    3519           0 :                                         spdk_nvmf_qpair_disconnect(&rqpair->qpair);
    3520             :                                 }
    3521             :                         }
    3522             :                 }
    3523             :         }
    3524           0 : }
    3525             : 
    3526             : static bool
    3527           0 : nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport,
    3528             :                                       struct rdma_cm_event *event)
    3529             : {
    3530             :         const struct spdk_nvme_transport_id     *trid;
    3531             :         struct spdk_nvmf_rdma_port              *port;
    3532             :         struct spdk_nvmf_rdma_transport         *rtransport;
    3533           0 :         bool                                    event_acked = false;
    3534             : 
    3535           0 :         rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
    3536           0 :         TAILQ_FOREACH(port, &rtransport->ports, link) {
    3537           0 :                 if (port->id == event->id) {
    3538           0 :                         SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid);
    3539           0 :                         rdma_ack_cm_event(event);
    3540           0 :                         event_acked = true;
    3541           0 :                         trid = port->trid;
    3542           0 :                         break;
    3543             :                 }
    3544             :         }
    3545             : 
    3546           0 :         if (event_acked) {
    3547           0 :                 nvmf_rdma_disconnect_qpairs_on_port(rtransport, port);
    3548             : 
    3549           0 :                 nvmf_rdma_stop_listen(transport, trid);
    3550           0 :                 nvmf_rdma_listen(transport, trid, NULL);
    3551             :         }
    3552             : 
    3553           0 :         return event_acked;
    3554             : }
    3555             : 
    3556             : static void
    3557           0 : nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport,
    3558             :                                 struct spdk_nvmf_rdma_device *device)
    3559             : {
    3560             :         struct spdk_nvmf_rdma_port      *port, *port_tmp;
    3561             :         int                             rc;
    3562           0 :         bool                            has_inflight;
    3563             : 
    3564           0 :         rc = nvmf_rdma_manage_poller(rtransport, device, &has_inflight, false);
    3565           0 :         if (rc) {
    3566           0 :                 SPDK_ERRLOG("Failed to handle device removal, rc %d\n", rc);
    3567           0 :                 return;
    3568             :         }
    3569             : 
    3570           0 :         if (!has_inflight) {
    3571             :                 /* no pollers, destroy the device */
    3572           0 :                 device->ready_to_destroy = true;
    3573           0 :                 spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_remove_destroyed_device, rtransport);
    3574             :         }
    3575             : 
    3576           0 :         TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) {
    3577           0 :                 if (port->device == device) {
    3578           0 :                         SPDK_NOTICELOG("Port %s:%s on device %s is being removed.\n",
    3579             :                                        port->trid->traddr,
    3580             :                                        port->trid->trsvcid,
    3581             :                                        ibv_get_device_name(port->device->context->device));
    3582             : 
    3583             :                         /* keep NVMF listener and only destroy structures of the
    3584             :                          * RDMA transport. when the device comes back we can retry listening
    3585             :                          * and the application's workflow will not be interrupted.
    3586             :                          */
    3587           0 :                         nvmf_rdma_stop_listen_ex(&rtransport->transport, port->trid, true);
    3588             :                 }
    3589             :         }
    3590             : }
    3591             : 
    3592             : static void
    3593           0 : nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport,
    3594             :                                        struct rdma_cm_event *event)
    3595             : {
    3596             :         struct spdk_nvmf_rdma_port              *port, *tmp_port;
    3597             :         struct spdk_nvmf_rdma_transport         *rtransport;
    3598             : 
    3599           0 :         port = event->id->context;
    3600           0 :         rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
    3601             : 
    3602           0 :         rdma_ack_cm_event(event);
    3603             : 
    3604             :         /* if device removal happens during ctrl qpair disconnecting, it's possible that we receive
    3605             :          * an DEVICE_REMOVAL event on qpair but the id->qp is just NULL. So we should make sure that
    3606             :          * we are handling a port event here.
    3607             :          */
    3608           0 :         TAILQ_FOREACH(tmp_port, &rtransport->ports, link) {
    3609           0 :                 if (port == tmp_port && port->device && !port->device->need_destroy) {
    3610           0 :                         port->device->need_destroy = true;
    3611           0 :                         nvmf_rdma_handle_device_removal(rtransport, port->device);
    3612             :                 }
    3613             :         }
    3614           0 : }
    3615             : 
    3616             : static void
    3617           0 : nvmf_process_cm_events(struct spdk_nvmf_transport *transport, uint32_t max_events)
    3618             : {
    3619             :         struct spdk_nvmf_rdma_transport *rtransport;
    3620           0 :         struct rdma_cm_event            *event;
    3621             :         uint32_t                        i;
    3622             :         int                             rc;
    3623           0 :         bool                            event_acked;
    3624             : 
    3625           0 :         rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
    3626             : 
    3627           0 :         if (rtransport->event_channel == NULL) {
    3628           0 :                 return;
    3629             :         }
    3630             : 
    3631           0 :         for (i = 0; i < max_events; i++) {
    3632           0 :                 event_acked = false;
    3633           0 :                 rc = rdma_get_cm_event(rtransport->event_channel, &event);
    3634           0 :                 if (rc) {
    3635           0 :                         if (errno != EAGAIN && errno != EWOULDBLOCK) {
    3636           0 :                                 SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno));
    3637             :                         }
    3638           0 :                         break;
    3639             :                 }
    3640             : 
    3641           0 :                 SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]);
    3642             : 
    3643           0 :                 spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event);
    3644             : 
    3645           0 :                 switch (event->event) {
    3646           0 :                 case RDMA_CM_EVENT_ADDR_RESOLVED:
    3647             :                 case RDMA_CM_EVENT_ADDR_ERROR:
    3648             :                 case RDMA_CM_EVENT_ROUTE_RESOLVED:
    3649             :                 case RDMA_CM_EVENT_ROUTE_ERROR:
    3650             :                         /* No action required. The target never attempts to resolve routes. */
    3651           0 :                         break;
    3652           0 :                 case RDMA_CM_EVENT_CONNECT_REQUEST:
    3653           0 :                         rc = nvmf_rdma_connect(transport, event);
    3654           0 :                         if (rc < 0) {
    3655           0 :                                 SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc);
    3656           0 :                                 break;
    3657             :                         }
    3658           0 :                         break;
    3659           0 :                 case RDMA_CM_EVENT_CONNECT_RESPONSE:
    3660             :                         /* The target never initiates a new connection. So this will not occur. */
    3661           0 :                         break;
    3662           0 :                 case RDMA_CM_EVENT_CONNECT_ERROR:
    3663             :                         /* Can this happen? The docs say it can, but not sure what causes it. */
    3664           0 :                         break;
    3665           0 :                 case RDMA_CM_EVENT_UNREACHABLE:
    3666             :                 case RDMA_CM_EVENT_REJECTED:
    3667             :                         /* These only occur on the client side. */
    3668           0 :                         break;
    3669           0 :                 case RDMA_CM_EVENT_ESTABLISHED:
    3670             :                         /* TODO: Should we be waiting for this event anywhere? */
    3671           0 :                         break;
    3672           0 :                 case RDMA_CM_EVENT_DISCONNECTED:
    3673           0 :                         rc = nvmf_rdma_disconnect(event, &event_acked);
    3674           0 :                         if (rc < 0) {
    3675           0 :                                 SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc);
    3676           0 :                                 break;
    3677             :                         }
    3678           0 :                         break;
    3679           0 :                 case RDMA_CM_EVENT_DEVICE_REMOVAL:
    3680             :                         /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL
    3681             :                          * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s.
    3682             :                          * Once these events are sent to SPDK, we should release all IB resources and
    3683             :                          * don't make attempts to call any ibv_query/modify/create functions. We can only call
    3684             :                          * ibv_destroy* functions to release user space memory allocated by IB. All kernel
    3685             :                          * resources are already cleaned. */
    3686           0 :                         if (event->id->qp) {
    3687             :                                 /* If rdma_cm event has a valid `qp` pointer then the event refers to the
    3688             :                                  * corresponding qpair. Otherwise the event refers to a listening device. */
    3689           0 :                                 rc = nvmf_rdma_disconnect(event, &event_acked);
    3690           0 :                                 if (rc < 0) {
    3691           0 :                                         SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc);
    3692           0 :                                         break;
    3693             :                                 }
    3694             :                         } else {
    3695           0 :                                 nvmf_rdma_handle_cm_event_port_removal(transport, event);
    3696           0 :                                 event_acked = true;
    3697             :                         }
    3698           0 :                         break;
    3699           0 :                 case RDMA_CM_EVENT_MULTICAST_JOIN:
    3700             :                 case RDMA_CM_EVENT_MULTICAST_ERROR:
    3701             :                         /* Multicast is not used */
    3702           0 :                         break;
    3703           0 :                 case RDMA_CM_EVENT_ADDR_CHANGE:
    3704           0 :                         event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event);
    3705           0 :                         break;
    3706           0 :                 case RDMA_CM_EVENT_TIMEWAIT_EXIT:
    3707             :                         /* For now, do nothing. The target never re-uses queue pairs. */
    3708           0 :                         break;
    3709           0 :                 default:
    3710           0 :                         SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event);
    3711           0 :                         break;
    3712             :                 }
    3713           0 :                 if (!event_acked) {
    3714           0 :                         rdma_ack_cm_event(event);
    3715             :                 }
    3716             :         }
    3717             : }
    3718             : 
    3719             : static void
    3720           0 : nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair)
    3721             : {
    3722           0 :         rqpair->last_wqe_reached = true;
    3723           0 :         nvmf_rdma_destroy_drained_qpair(rqpair);
    3724           0 : }
    3725             : 
    3726             : static void
    3727           0 : nvmf_rdma_qpair_process_ibv_event(void *ctx)
    3728             : {
    3729           0 :         struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx;
    3730             : 
    3731           0 :         if (event_ctx->rqpair) {
    3732           0 :                 STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link);
    3733           0 :                 if (event_ctx->cb_fn) {
    3734           0 :                         event_ctx->cb_fn(event_ctx->rqpair);
    3735             :                 }
    3736             :         }
    3737           0 :         free(event_ctx);
    3738           0 : }
    3739             : 
    3740             : static int
    3741           0 : nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair,
    3742             :                                  spdk_nvmf_rdma_qpair_ibv_event fn)
    3743             : {
    3744             :         struct spdk_nvmf_rdma_ibv_event_ctx *ctx;
    3745           0 :         struct spdk_thread *thr = NULL;
    3746             :         int rc;
    3747             : 
    3748           0 :         if (rqpair->qpair.group) {
    3749           0 :                 thr = rqpair->qpair.group->thread;
    3750           0 :         } else if (rqpair->destruct_channel) {
    3751           0 :                 thr = spdk_io_channel_get_thread(rqpair->destruct_channel);
    3752             :         }
    3753             : 
    3754           0 :         if (!thr) {
    3755           0 :                 SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair);
    3756           0 :                 return -EINVAL;
    3757             :         }
    3758             : 
    3759           0 :         ctx = calloc(1, sizeof(*ctx));
    3760           0 :         if (!ctx) {
    3761           0 :                 return -ENOMEM;
    3762             :         }
    3763             : 
    3764           0 :         ctx->rqpair = rqpair;
    3765           0 :         ctx->cb_fn = fn;
    3766           0 :         STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link);
    3767             : 
    3768           0 :         rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx);
    3769           0 :         if (rc) {
    3770           0 :                 STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link);
    3771           0 :                 free(ctx);
    3772             :         }
    3773             : 
    3774           0 :         return rc;
    3775             : }
    3776             : 
    3777             : static int
    3778           0 : nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device)
    3779             : {
    3780             :         int                             rc;
    3781           0 :         struct spdk_nvmf_rdma_qpair     *rqpair = NULL;
    3782           0 :         struct ibv_async_event          event;
    3783             : 
    3784           0 :         rc = ibv_get_async_event(device->context, &event);
    3785             : 
    3786           0 :         if (rc) {
    3787             :                 /* In non-blocking mode -1 means there are no events available */
    3788           0 :                 return rc;
    3789             :         }
    3790             : 
    3791           0 :         switch (event.event_type) {
    3792           0 :         case IBV_EVENT_QP_FATAL:
    3793             :         case IBV_EVENT_QP_LAST_WQE_REACHED:
    3794             :         case IBV_EVENT_QP_REQ_ERR:
    3795             :         case IBV_EVENT_QP_ACCESS_ERR:
    3796             :         case IBV_EVENT_COMM_EST:
    3797             :         case IBV_EVENT_PATH_MIG:
    3798             :         case IBV_EVENT_PATH_MIG_ERR:
    3799           0 :                 rqpair = event.element.qp->qp_context;
    3800           0 :                 if (!rqpair) {
    3801             :                         /* Any QP event for NVMe-RDMA initiator may be returned. */
    3802           0 :                         SPDK_NOTICELOG("Async QP event for unknown QP: %s\n",
    3803             :                                        ibv_event_type_str(event.event_type));
    3804           0 :                         break;
    3805             :                 }
    3806             : 
    3807           0 :                 switch (event.event_type) {
    3808           0 :                 case IBV_EVENT_QP_FATAL:
    3809           0 :                         SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair);
    3810           0 :                         spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0,
    3811             :                                           (uintptr_t)rqpair, event.event_type);
    3812           0 :                         rqpair->ibv_in_error_state = true;
    3813           0 :                         spdk_nvmf_qpair_disconnect(&rqpair->qpair);
    3814           0 :                         break;
    3815           0 :                 case IBV_EVENT_QP_LAST_WQE_REACHED:
    3816             :                         /* This event only occurs for shared receive queues. */
    3817           0 :                         SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair);
    3818           0 :                         rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached);
    3819           0 :                         if (rc) {
    3820           0 :                                 SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc);
    3821           0 :                                 rqpair->last_wqe_reached = true;
    3822             :                         }
    3823           0 :                         break;
    3824           0 :                 case IBV_EVENT_QP_REQ_ERR:
    3825             :                 case IBV_EVENT_QP_ACCESS_ERR:
    3826             :                 case IBV_EVENT_COMM_EST:
    3827             :                 case IBV_EVENT_PATH_MIG:
    3828             :                 case IBV_EVENT_PATH_MIG_ERR:
    3829           0 :                         SPDK_NOTICELOG("Async QP event: %s\n",
    3830             :                                        ibv_event_type_str(event.event_type));
    3831           0 :                         spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0,
    3832             :                                           (uintptr_t)rqpair, event.event_type);
    3833           0 :                         rqpair->ibv_in_error_state = true;
    3834           0 :                         break;
    3835           0 :                 default:
    3836           0 :                         break;
    3837             :                 }
    3838           0 :                 break;
    3839           0 :         case IBV_EVENT_DEVICE_FATAL:
    3840           0 :                 SPDK_ERRLOG("Device Fatal event[%s] received on %s. device: %p\n",
    3841             :                             ibv_event_type_str(event.event_type), ibv_get_device_name(device->context->device), device);
    3842           0 :                 device->need_destroy = true;
    3843           0 :                 break;
    3844           0 :         case IBV_EVENT_CQ_ERR:
    3845             :         case IBV_EVENT_PORT_ACTIVE:
    3846             :         case IBV_EVENT_PORT_ERR:
    3847             :         case IBV_EVENT_LID_CHANGE:
    3848             :         case IBV_EVENT_PKEY_CHANGE:
    3849             :         case IBV_EVENT_SM_CHANGE:
    3850             :         case IBV_EVENT_SRQ_ERR:
    3851             :         case IBV_EVENT_SRQ_LIMIT_REACHED:
    3852             :         case IBV_EVENT_CLIENT_REREGISTER:
    3853             :         case IBV_EVENT_GID_CHANGE:
    3854             :         case IBV_EVENT_SQ_DRAINED:
    3855             :         default:
    3856           0 :                 SPDK_NOTICELOG("Async event: %s\n",
    3857             :                                ibv_event_type_str(event.event_type));
    3858           0 :                 spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type);
    3859           0 :                 break;
    3860             :         }
    3861           0 :         ibv_ack_async_event(&event);
    3862             : 
    3863           0 :         return 0;
    3864             : }
    3865             : 
    3866             : static void
    3867           0 : nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events)
    3868             : {
    3869           0 :         int rc = 0;
    3870           0 :         uint32_t i = 0;
    3871             : 
    3872           0 :         for (i = 0; i < max_events; i++) {
    3873           0 :                 rc = nvmf_process_ib_event(device);
    3874           0 :                 if (rc) {
    3875           0 :                         break;
    3876             :                 }
    3877             :         }
    3878             : 
    3879           0 :         SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i);
    3880           0 : }
    3881             : 
    3882             : static int
    3883           0 : nvmf_rdma_accept(void *ctx)
    3884             : {
    3885           0 :         int     nfds, i = 0;
    3886           0 :         struct spdk_nvmf_transport *transport = ctx;
    3887             :         struct spdk_nvmf_rdma_transport *rtransport;
    3888             :         struct spdk_nvmf_rdma_device *device, *tmp;
    3889             :         uint32_t count;
    3890             :         short revents;
    3891             :         bool do_retry;
    3892             : 
    3893           0 :         rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
    3894           0 :         do_retry = nvmf_rdma_retry_listen_port(rtransport);
    3895             : 
    3896           0 :         count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0);
    3897             : 
    3898           0 :         if (nfds <= 0) {
    3899           0 :                 return do_retry ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE;
    3900             :         }
    3901             : 
    3902             :         /* The first poll descriptor is RDMA CM event */
    3903           0 :         if (rtransport->poll_fds[i++].revents & POLLIN) {
    3904           0 :                 nvmf_process_cm_events(transport, NVMF_RDMA_MAX_EVENTS_PER_POLL);
    3905           0 :                 nfds--;
    3906             :         }
    3907             : 
    3908           0 :         if (nfds == 0) {
    3909           0 :                 return SPDK_POLLER_BUSY;
    3910             :         }
    3911             : 
    3912             :         /* Second and subsequent poll descriptors are IB async events */
    3913           0 :         TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) {
    3914           0 :                 revents = rtransport->poll_fds[i++].revents;
    3915           0 :                 if (revents & POLLIN) {
    3916           0 :                         if (spdk_likely(!device->need_destroy)) {
    3917           0 :                                 nvmf_process_ib_events(device, NVMF_RDMA_MAX_EVENTS_PER_POLL);
    3918           0 :                                 if (spdk_unlikely(device->need_destroy)) {
    3919           0 :                                         nvmf_rdma_handle_device_removal(rtransport, device);
    3920             :                                 }
    3921             :                         }
    3922           0 :                         nfds--;
    3923           0 :                 } else if (revents & POLLNVAL || revents & POLLHUP) {
    3924           0 :                         SPDK_ERRLOG("Receive unknown revent %x on device %p\n", (int)revents, device);
    3925           0 :                         nfds--;
    3926             :                 }
    3927             :         }
    3928             :         /* check all flagged fd's have been served */
    3929           0 :         assert(nfds == 0);
    3930             : 
    3931           0 :         return count > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE;
    3932             : }
    3933             : 
    3934             : static void
    3935           0 : nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem,
    3936             :                      struct spdk_nvmf_ctrlr_data *cdata)
    3937             : {
    3938           0 :         cdata->nvmf_specific.msdbd = NVMF_DEFAULT_MSDBD;
    3939             : 
    3940             :         /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled
    3941             :         since in-capsule data only works with NVME drives that support SGL memory layout */
    3942           0 :         if (transport->opts.dif_insert_or_strip) {
    3943           0 :                 cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16;
    3944             :         }
    3945             : 
    3946           0 :         if (cdata->nvmf_specific.ioccsz > ((sizeof(struct spdk_nvme_cmd) + 0x1000) / 16)) {
    3947           0 :                 SPDK_WARNLOG("RDMA is configured to support up to 16 SGL entries while in capsule"
    3948             :                              " data is greater than 4KiB.\n");
    3949           0 :                 SPDK_WARNLOG("When used in conjunction with the NVMe-oF initiator from the Linux "
    3950             :                              "kernel between versions 5.4 and 5.12 data corruption may occur for "
    3951             :                              "writes that are not a multiple of 4KiB in size.\n");
    3952             :         }
    3953           0 : }
    3954             : 
    3955             : static void
    3956           0 : nvmf_rdma_discover(struct spdk_nvmf_transport *transport,
    3957             :                    struct spdk_nvme_transport_id *trid,
    3958             :                    struct spdk_nvmf_discovery_log_page_entry *entry)
    3959             : {
    3960           0 :         entry->trtype = SPDK_NVMF_TRTYPE_RDMA;
    3961           0 :         entry->adrfam = trid->adrfam;
    3962           0 :         entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED;
    3963             : 
    3964           0 :         spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' ');
    3965           0 :         spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' ');
    3966             : 
    3967           0 :         entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED;
    3968           0 :         entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE;
    3969           0 :         entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM;
    3970           0 : }
    3971             : 
    3972             : static int
    3973           0 : nvmf_rdma_poller_create(struct spdk_nvmf_rdma_transport *rtransport,
    3974             :                         struct spdk_nvmf_rdma_poll_group *rgroup, struct spdk_nvmf_rdma_device *device,
    3975             :                         struct spdk_nvmf_rdma_poller **out_poller)
    3976             : {
    3977             :         struct spdk_nvmf_rdma_poller            *poller;
    3978           0 :         struct spdk_rdma_provider_srq_init_attr srq_init_attr;
    3979           0 :         struct spdk_nvmf_rdma_resource_opts     opts;
    3980             :         int                                     num_cqe;
    3981             : 
    3982           0 :         poller = calloc(1, sizeof(*poller));
    3983           0 :         if (!poller) {
    3984           0 :                 SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n");
    3985           0 :                 return -1;
    3986             :         }
    3987             : 
    3988           0 :         poller->device = device;
    3989           0 :         poller->group = rgroup;
    3990           0 :         *out_poller = poller;
    3991             : 
    3992           0 :         RB_INIT(&poller->qpairs);
    3993           0 :         STAILQ_INIT(&poller->qpairs_pending_send);
    3994           0 :         STAILQ_INIT(&poller->qpairs_pending_recv);
    3995             : 
    3996           0 :         TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link);
    3997           0 :         SPDK_DEBUGLOG(rdma, "Create poller %p on device %p in poll group %p.\n", poller, device, rgroup);
    3998           0 :         if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) {
    3999           0 :                 if ((int)rtransport->rdma_opts.max_srq_depth > device->attr.max_srq_wr) {
    4000           0 :                         SPDK_WARNLOG("Requested SRQ depth %u, max supported by dev %s is %d\n",
    4001             :                                      rtransport->rdma_opts.max_srq_depth, device->context->device->name, device->attr.max_srq_wr);
    4002             :                 }
    4003           0 :                 poller->max_srq_depth = spdk_min((int)rtransport->rdma_opts.max_srq_depth, device->attr.max_srq_wr);
    4004             : 
    4005           0 :                 device->num_srq++;
    4006           0 :                 memset(&srq_init_attr, 0, sizeof(srq_init_attr));
    4007           0 :                 srq_init_attr.pd = device->pd;
    4008           0 :                 srq_init_attr.stats = &poller->stat.qp_stats.recv;
    4009           0 :                 srq_init_attr.srq_init_attr.attr.max_wr = poller->max_srq_depth;
    4010           0 :                 srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE);
    4011           0 :                 poller->srq = spdk_rdma_provider_srq_create(&srq_init_attr);
    4012           0 :                 if (!poller->srq) {
    4013           0 :                         SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno);
    4014           0 :                         return -1;
    4015             :                 }
    4016             : 
    4017           0 :                 opts.qp = poller->srq;
    4018           0 :                 opts.map = device->map;
    4019           0 :                 opts.qpair = NULL;
    4020           0 :                 opts.shared = true;
    4021           0 :                 opts.max_queue_depth = poller->max_srq_depth;
    4022           0 :                 opts.in_capsule_data_size = rtransport->transport.opts.in_capsule_data_size;
    4023             : 
    4024           0 :                 poller->resources = nvmf_rdma_resources_create(&opts);
    4025           0 :                 if (!poller->resources) {
    4026           0 :                         SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n");
    4027           0 :                         return -1;
    4028             :                 }
    4029             :         }
    4030             : 
    4031             :         /*
    4032             :          * When using an srq, we can limit the completion queue at startup.
    4033             :          * The following formula represents the calculation:
    4034             :          * num_cqe = num_recv + num_data_wr + num_send_wr.
    4035             :          * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth
    4036             :          */
    4037           0 :         if (poller->srq) {
    4038           0 :                 num_cqe = poller->max_srq_depth * 3;
    4039             :         } else {
    4040           0 :                 num_cqe = rtransport->rdma_opts.num_cqe;
    4041             :         }
    4042             : 
    4043           0 :         poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0);
    4044           0 :         if (!poller->cq) {
    4045           0 :                 SPDK_ERRLOG("Unable to create completion queue\n");
    4046           0 :                 return -1;
    4047             :         }
    4048           0 :         poller->num_cqe = num_cqe;
    4049           0 :         return 0;
    4050             : }
    4051             : 
    4052             : static void
    4053           0 : _nvmf_rdma_register_poller_in_group(void *c)
    4054             : {
    4055           0 :         struct spdk_nvmf_rdma_poller    *poller;
    4056           0 :         struct poller_manage_ctx        *ctx = c;
    4057             :         struct spdk_nvmf_rdma_device    *device;
    4058             :         int                             rc;
    4059             : 
    4060           0 :         rc = nvmf_rdma_poller_create(ctx->rtransport, ctx->rgroup, ctx->device, &poller);
    4061           0 :         if (rc < 0 && poller) {
    4062           0 :                 nvmf_rdma_poller_destroy(poller);
    4063             :         }
    4064             : 
    4065           0 :         device = ctx->device;
    4066           0 :         if (nvmf_rdma_all_pollers_management_done(ctx)) {
    4067           0 :                 device->is_ready = true;
    4068             :         }
    4069           0 : }
    4070             : 
    4071             : static void nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group);
    4072             : 
    4073             : static struct spdk_nvmf_transport_poll_group *
    4074           5 : nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport,
    4075             :                             struct spdk_nvmf_poll_group *group)
    4076             : {
    4077             :         struct spdk_nvmf_rdma_transport         *rtransport;
    4078             :         struct spdk_nvmf_rdma_poll_group        *rgroup;
    4079           5 :         struct spdk_nvmf_rdma_poller            *poller;
    4080             :         struct spdk_nvmf_rdma_device            *device;
    4081             :         int                                     rc;
    4082             : 
    4083           5 :         if (spdk_interrupt_mode_is_enabled()) {
    4084           0 :                 SPDK_ERRLOG("RDMA transport does not support interrupt mode\n");
    4085           0 :                 return NULL;
    4086             :         }
    4087             : 
    4088           5 :         rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
    4089             : 
    4090           5 :         rgroup = calloc(1, sizeof(*rgroup));
    4091           5 :         if (!rgroup) {
    4092           0 :                 return NULL;
    4093             :         }
    4094             : 
    4095           5 :         TAILQ_INIT(&rgroup->pollers);
    4096             : 
    4097           5 :         TAILQ_FOREACH(device, &rtransport->devices, link) {
    4098           0 :                 rc = nvmf_rdma_poller_create(rtransport, rgroup, device, &poller);
    4099           0 :                 if (rc < 0) {
    4100           0 :                         nvmf_rdma_poll_group_destroy(&rgroup->group);
    4101           0 :                         return NULL;
    4102             :                 }
    4103             :         }
    4104             : 
    4105           5 :         TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link);
    4106           5 :         if (rtransport->conn_sched.next_admin_pg == NULL) {
    4107           1 :                 rtransport->conn_sched.next_admin_pg = rgroup;
    4108           1 :                 rtransport->conn_sched.next_io_pg = rgroup;
    4109             :         }
    4110             : 
    4111           5 :         return &rgroup->group;
    4112             : }
    4113             : 
    4114             : static uint32_t
    4115          12 : nvmf_poll_group_get_io_qpair_count(struct spdk_nvmf_poll_group *pg)
    4116             : {
    4117             :         uint32_t count;
    4118             : 
    4119             :         /* Just assume that unassociated qpairs will eventually be io
    4120             :          * qpairs.  This is close enough for the use cases for this
    4121             :          * function.
    4122             :          */
    4123          12 :         pthread_mutex_lock(&pg->mutex);
    4124          12 :         count = pg->stat.current_io_qpairs + pg->current_unassociated_qpairs;
    4125          12 :         pthread_mutex_unlock(&pg->mutex);
    4126             : 
    4127          12 :         return count;
    4128             : }
    4129             : 
    4130             : static struct spdk_nvmf_transport_poll_group *
    4131          14 : nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair)
    4132             : {
    4133             :         struct spdk_nvmf_rdma_transport *rtransport;
    4134             :         struct spdk_nvmf_rdma_poll_group **pg;
    4135             :         struct spdk_nvmf_transport_poll_group *result;
    4136             :         uint32_t count;
    4137             : 
    4138          14 :         rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
    4139             : 
    4140          14 :         if (TAILQ_EMPTY(&rtransport->poll_groups)) {
    4141           2 :                 return NULL;
    4142             :         }
    4143             : 
    4144          12 :         if (qpair->qid == 0) {
    4145           6 :                 pg = &rtransport->conn_sched.next_admin_pg;
    4146             :         } else {
    4147             :                 struct spdk_nvmf_rdma_poll_group *pg_min, *pg_start, *pg_current;
    4148             :                 uint32_t min_value;
    4149             : 
    4150           6 :                 pg = &rtransport->conn_sched.next_io_pg;
    4151           6 :                 pg_min = *pg;
    4152           6 :                 pg_start = *pg;
    4153           6 :                 pg_current = *pg;
    4154           6 :                 min_value = nvmf_poll_group_get_io_qpair_count(pg_current->group.group);
    4155             : 
    4156             :                 while (1) {
    4157           6 :                         count = nvmf_poll_group_get_io_qpair_count(pg_current->group.group);
    4158             : 
    4159           6 :                         if (count < min_value) {
    4160           0 :                                 min_value = count;
    4161           0 :                                 pg_min = pg_current;
    4162             :                         }
    4163             : 
    4164           6 :                         pg_current = TAILQ_NEXT(pg_current, link);
    4165           6 :                         if (pg_current == NULL) {
    4166           2 :                                 pg_current = TAILQ_FIRST(&rtransport->poll_groups);
    4167             :                         }
    4168             : 
    4169           6 :                         if (pg_current == pg_start || min_value == 0) {
    4170             :                                 break;
    4171             :                         }
    4172             :                 }
    4173           6 :                 *pg = pg_min;
    4174             :         }
    4175             : 
    4176          12 :         assert(*pg != NULL);
    4177             : 
    4178          12 :         result = &(*pg)->group;
    4179             : 
    4180          12 :         *pg = TAILQ_NEXT(*pg, link);
    4181          12 :         if (*pg == NULL) {
    4182           4 :                 *pg = TAILQ_FIRST(&rtransport->poll_groups);
    4183             :         }
    4184             : 
    4185          12 :         return result;
    4186             : }
    4187             : 
    4188             : static void
    4189           0 : nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller)
    4190             : {
    4191             :         struct spdk_nvmf_rdma_qpair     *qpair, *tmp_qpair;
    4192             :         int                             rc;
    4193             : 
    4194           0 :         TAILQ_REMOVE(&poller->group->pollers, poller, link);
    4195           0 :         RB_FOREACH_SAFE(qpair, qpairs_tree, &poller->qpairs, tmp_qpair) {
    4196           0 :                 nvmf_rdma_qpair_destroy(qpair);
    4197             :         }
    4198             : 
    4199           0 :         if (poller->srq) {
    4200           0 :                 if (poller->resources) {
    4201           0 :                         nvmf_rdma_resources_destroy(poller->resources);
    4202             :                 }
    4203           0 :                 spdk_rdma_provider_srq_destroy(poller->srq);
    4204           0 :                 SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq);
    4205             :         }
    4206             : 
    4207           0 :         if (poller->cq) {
    4208           0 :                 rc = ibv_destroy_cq(poller->cq);
    4209           0 :                 if (rc != 0) {
    4210           0 :                         SPDK_ERRLOG("Destroy cq return %d, error: %s\n", rc, strerror(errno));
    4211             :                 }
    4212             :         }
    4213             : 
    4214           0 :         if (poller->destroy_cb) {
    4215           0 :                 poller->destroy_cb(poller->destroy_cb_ctx);
    4216           0 :                 poller->destroy_cb = NULL;
    4217             :         }
    4218             : 
    4219           0 :         free(poller);
    4220           0 : }
    4221             : 
    4222             : static void
    4223           5 : nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group)
    4224             : {
    4225             :         struct spdk_nvmf_rdma_poll_group        *rgroup, *next_rgroup;
    4226             :         struct spdk_nvmf_rdma_poller            *poller, *tmp;
    4227             :         struct spdk_nvmf_rdma_transport         *rtransport;
    4228             : 
    4229           5 :         rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
    4230           5 :         if (!rgroup) {
    4231           0 :                 return;
    4232             :         }
    4233             : 
    4234           5 :         TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) {
    4235           0 :                 nvmf_rdma_poller_destroy(poller);
    4236             :         }
    4237             : 
    4238           5 :         if (rgroup->group.transport == NULL) {
    4239             :                 /* Transport can be NULL when nvmf_rdma_poll_group_create()
    4240             :                  * calls this function directly in a failure path. */
    4241           0 :                 free(rgroup);
    4242           0 :                 return;
    4243             :         }
    4244             : 
    4245           5 :         rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport);
    4246             : 
    4247           5 :         next_rgroup = TAILQ_NEXT(rgroup, link);
    4248           5 :         TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link);
    4249           5 :         if (next_rgroup == NULL) {
    4250           1 :                 next_rgroup = TAILQ_FIRST(&rtransport->poll_groups);
    4251             :         }
    4252           5 :         if (rtransport->conn_sched.next_admin_pg == rgroup) {
    4253           5 :                 rtransport->conn_sched.next_admin_pg = next_rgroup;
    4254             :         }
    4255           5 :         if (rtransport->conn_sched.next_io_pg == rgroup) {
    4256           5 :                 rtransport->conn_sched.next_io_pg = next_rgroup;
    4257             :         }
    4258             : 
    4259           5 :         free(rgroup);
    4260             : }
    4261             : 
    4262             : static void
    4263           0 : nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair)
    4264             : {
    4265           0 :         if (rqpair->cm_id != NULL) {
    4266           0 :                 nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES);
    4267             :         }
    4268           0 : }
    4269             : 
    4270             : static int
    4271           0 : nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group,
    4272             :                          struct spdk_nvmf_qpair *qpair)
    4273             : {
    4274             :         struct spdk_nvmf_rdma_poll_group        *rgroup;
    4275             :         struct spdk_nvmf_rdma_qpair             *rqpair;
    4276             :         struct spdk_nvmf_rdma_device            *device;
    4277             :         struct spdk_nvmf_rdma_poller            *poller;
    4278             :         int                                     rc;
    4279             : 
    4280           0 :         rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
    4281           0 :         rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
    4282             : 
    4283           0 :         device = rqpair->device;
    4284             : 
    4285           0 :         TAILQ_FOREACH(poller, &rgroup->pollers, link) {
    4286           0 :                 if (poller->device == device) {
    4287           0 :                         break;
    4288             :                 }
    4289             :         }
    4290             : 
    4291           0 :         if (!poller) {
    4292           0 :                 SPDK_ERRLOG("No poller found for device.\n");
    4293           0 :                 return -1;
    4294             :         }
    4295             : 
    4296           0 :         if (poller->need_destroy) {
    4297           0 :                 SPDK_ERRLOG("Poller is destroying.\n");
    4298           0 :                 return -1;
    4299             :         }
    4300             : 
    4301           0 :         rqpair->poller = poller;
    4302           0 :         rqpair->srq = rqpair->poller->srq;
    4303             : 
    4304           0 :         rc = nvmf_rdma_qpair_initialize(qpair);
    4305           0 :         if (rc < 0) {
    4306           0 :                 SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair);
    4307           0 :                 rqpair->poller = NULL;
    4308           0 :                 rqpair->srq = NULL;
    4309           0 :                 return -1;
    4310             :         }
    4311             : 
    4312           0 :         RB_INSERT(qpairs_tree, &poller->qpairs, rqpair);
    4313             : 
    4314           0 :         rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair);
    4315           0 :         if (rc) {
    4316             :                 /* Try to reject, but we probably can't */
    4317           0 :                 nvmf_rdma_qpair_reject_connection(rqpair);
    4318           0 :                 return -1;
    4319             :         }
    4320             : 
    4321           0 :         return 0;
    4322             : }
    4323             : 
    4324             : static int
    4325           0 : nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group,
    4326             :                             struct spdk_nvmf_qpair *qpair)
    4327             : {
    4328             :         struct spdk_nvmf_rdma_qpair             *rqpair;
    4329             : 
    4330           0 :         rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
    4331           0 :         assert(group->transport->tgt != NULL);
    4332             : 
    4333           0 :         rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt);
    4334             : 
    4335           0 :         if (!rqpair->destruct_channel) {
    4336           0 :                 SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair);
    4337           0 :                 return 0;
    4338             :         }
    4339             : 
    4340             :         /* Sanity check that we get io_channel on the correct thread */
    4341           0 :         if (qpair->group) {
    4342           0 :                 assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel));
    4343             :         }
    4344             : 
    4345           0 :         return 0;
    4346             : }
    4347             : 
    4348             : static int
    4349           0 : nvmf_rdma_request_free(struct spdk_nvmf_request *req)
    4350             : {
    4351           0 :         struct spdk_nvmf_rdma_request   *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
    4352           0 :         struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport,
    4353             :                         struct spdk_nvmf_rdma_transport, transport);
    4354           0 :         struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair,
    4355             :                                               struct spdk_nvmf_rdma_qpair, qpair);
    4356             : 
    4357             :         /*
    4358             :          * AER requests are freed when a qpair is destroyed. The recv corresponding to that request
    4359             :          * needs to be returned to the shared receive queue or the poll group will eventually be
    4360             :          * starved of RECV structures.
    4361             :          */
    4362           0 :         if (rqpair->srq && rdma_req->recv) {
    4363             :                 int rc;
    4364           0 :                 struct ibv_recv_wr *bad_recv_wr;
    4365             : 
    4366           0 :                 spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, &rdma_req->recv->wr);
    4367           0 :                 rc = spdk_rdma_provider_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr);
    4368           0 :                 if (rc) {
    4369           0 :                         SPDK_ERRLOG("Unable to re-post rx descriptor\n");
    4370             :                 }
    4371             :         }
    4372             : 
    4373           0 :         _nvmf_rdma_request_free(rdma_req, rtransport);
    4374           0 :         return 0;
    4375             : }
    4376             : 
    4377             : static int
    4378           0 : nvmf_rdma_request_complete(struct spdk_nvmf_request *req)
    4379             : {
    4380           0 :         struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport,
    4381             :                         struct spdk_nvmf_rdma_transport, transport);
    4382           0 :         struct spdk_nvmf_rdma_request   *rdma_req = SPDK_CONTAINEROF(req,
    4383             :                         struct spdk_nvmf_rdma_request, req);
    4384           0 :         struct spdk_nvmf_rdma_qpair     *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair,
    4385             :                         struct spdk_nvmf_rdma_qpair, qpair);
    4386             : 
    4387           0 :         if (spdk_unlikely(rqpair->ibv_in_error_state)) {
    4388             :                 /* The connection is dead. Move the request directly to the completed state. */
    4389           0 :                 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
    4390             :         } else {
    4391             :                 /* The connection is alive, so process the request as normal */
    4392           0 :                 rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
    4393             :         }
    4394             : 
    4395           0 :         nvmf_rdma_request_process(rtransport, rdma_req);
    4396             : 
    4397           0 :         return 0;
    4398             : }
    4399             : 
    4400             : static void
    4401           0 : nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair,
    4402             :                       spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg)
    4403             : {
    4404           0 :         struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
    4405             : 
    4406           0 :         rqpair->to_close = true;
    4407             : 
    4408             :         /* This happens only when the qpair is disconnected before
    4409             :          * it is added to the poll group. Since there is no poll group,
    4410             :          * the RDMA qp has not been initialized yet and the RDMA CM
    4411             :          * event has not yet been acknowledged, so we need to reject it.
    4412             :          */
    4413           0 :         if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) {
    4414           0 :                 nvmf_rdma_qpair_reject_connection(rqpair);
    4415           0 :                 nvmf_rdma_qpair_destroy(rqpair);
    4416           0 :                 return;
    4417             :         }
    4418             : 
    4419           0 :         if (rqpair->rdma_qp) {
    4420           0 :                 spdk_rdma_provider_qp_disconnect(rqpair->rdma_qp);
    4421             :         }
    4422             : 
    4423           0 :         nvmf_rdma_destroy_drained_qpair(rqpair);
    4424             : 
    4425           0 :         if (cb_fn) {
    4426           0 :                 cb_fn(cb_arg);
    4427             :         }
    4428             : }
    4429             : 
    4430             : static struct spdk_nvmf_rdma_qpair *
    4431           0 : get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc)
    4432             : {
    4433           0 :         struct spdk_nvmf_rdma_qpair find;
    4434             : 
    4435           0 :         find.qp_num = wc->qp_num;
    4436             : 
    4437           0 :         return RB_FIND(qpairs_tree, &rpoller->qpairs, &find);
    4438             : }
    4439             : 
    4440             : #ifdef DEBUG
    4441             : static int
    4442           0 : nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req)
    4443             : {
    4444           0 :         return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST ||
    4445           0 :                rdma_req->state == RDMA_REQUEST_STATE_COMPLETING;
    4446             : }
    4447             : #endif
    4448             : 
    4449             : static void
    4450           0 : _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr,
    4451             :                            int rc)
    4452             : {
    4453             :         struct spdk_nvmf_rdma_recv      *rdma_recv;
    4454             :         struct spdk_nvmf_rdma_wr        *bad_rdma_wr;
    4455             : 
    4456           0 :         SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc);
    4457           0 :         while (bad_recv_wr != NULL) {
    4458           0 :                 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id;
    4459           0 :                 rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr);
    4460             : 
    4461           0 :                 rdma_recv->qpair->current_recv_depth++;
    4462           0 :                 bad_recv_wr = bad_recv_wr->next;
    4463           0 :                 SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc);
    4464           0 :                 spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair);
    4465             :         }
    4466           0 : }
    4467             : 
    4468             : static void
    4469           0 : _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc)
    4470             : {
    4471           0 :         SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc);
    4472           0 :         while (bad_recv_wr != NULL) {
    4473           0 :                 bad_recv_wr = bad_recv_wr->next;
    4474           0 :                 rqpair->current_recv_depth++;
    4475             :         }
    4476           0 :         spdk_nvmf_qpair_disconnect(&rqpair->qpair);
    4477           0 : }
    4478             : 
    4479             : static void
    4480           0 : _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport,
    4481             :                      struct spdk_nvmf_rdma_poller *rpoller)
    4482             : {
    4483             :         struct spdk_nvmf_rdma_qpair     *rqpair;
    4484           0 :         struct ibv_recv_wr              *bad_recv_wr;
    4485             :         int                             rc;
    4486             : 
    4487           0 :         if (rpoller->srq) {
    4488           0 :                 rc = spdk_rdma_provider_srq_flush_recv_wrs(rpoller->srq, &bad_recv_wr);
    4489           0 :                 if (spdk_unlikely(rc)) {
    4490           0 :                         _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc);
    4491             :                 }
    4492             :         } else {
    4493           0 :                 while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) {
    4494           0 :                         rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv);
    4495           0 :                         rc = spdk_rdma_provider_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr);
    4496           0 :                         if (spdk_unlikely(rc)) {
    4497           0 :                                 _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc);
    4498             :                         }
    4499           0 :                         STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link);
    4500             :                 }
    4501             :         }
    4502           0 : }
    4503             : 
    4504             : static void
    4505           0 : _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport,
    4506             :                        struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc)
    4507             : {
    4508             :         struct spdk_nvmf_rdma_wr        *bad_rdma_wr;
    4509           0 :         struct spdk_nvmf_rdma_request   *prev_rdma_req = NULL, *cur_rdma_req = NULL;
    4510             : 
    4511           0 :         SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc);
    4512           0 :         for (; bad_wr != NULL; bad_wr = bad_wr->next) {
    4513           0 :                 bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id;
    4514           0 :                 assert(rqpair->current_send_depth > 0);
    4515           0 :                 rqpair->current_send_depth--;
    4516           0 :                 switch (bad_rdma_wr->type) {
    4517           0 :                 case RDMA_WR_TYPE_DATA:
    4518           0 :                         cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data_wr);
    4519           0 :                         if (bad_wr->opcode == IBV_WR_RDMA_READ) {
    4520           0 :                                 assert(rqpair->current_read_depth > 0);
    4521           0 :                                 rqpair->current_read_depth--;
    4522             :                         }
    4523           0 :                         break;
    4524           0 :                 case RDMA_WR_TYPE_SEND:
    4525           0 :                         cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr);
    4526           0 :                         break;
    4527           0 :                 default:
    4528           0 :                         SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair);
    4529           0 :                         prev_rdma_req = cur_rdma_req;
    4530           0 :                         continue;
    4531             :                 }
    4532             : 
    4533           0 :                 if (prev_rdma_req == cur_rdma_req) {
    4534             :                         /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */
    4535             :                         /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */
    4536           0 :                         continue;
    4537             :                 }
    4538             : 
    4539           0 :                 switch (cur_rdma_req->state) {
    4540           0 :                 case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
    4541           0 :                         cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR;
    4542           0 :                         STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, cur_rdma_req, state_link);
    4543           0 :                         cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
    4544           0 :                         break;
    4545           0 :                 case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST:
    4546             :                 case RDMA_REQUEST_STATE_COMPLETING:
    4547           0 :                         cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
    4548           0 :                         break;
    4549           0 :                 default:
    4550           0 :                         SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n",
    4551             :                                     cur_rdma_req->state, rqpair);
    4552           0 :                         continue;
    4553             :                 }
    4554             : 
    4555           0 :                 nvmf_rdma_request_process(rtransport, cur_rdma_req);
    4556           0 :                 prev_rdma_req = cur_rdma_req;
    4557             :         }
    4558             : 
    4559           0 :         if (spdk_nvmf_qpair_is_active(&rqpair->qpair)) {
    4560             :                 /* Disconnect the connection. */
    4561           0 :                 spdk_nvmf_qpair_disconnect(&rqpair->qpair);
    4562             :         }
    4563             : 
    4564           0 : }
    4565             : 
    4566             : static void
    4567           0 : _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport,
    4568             :                      struct spdk_nvmf_rdma_poller *rpoller)
    4569             : {
    4570             :         struct spdk_nvmf_rdma_qpair     *rqpair;
    4571           0 :         struct ibv_send_wr              *bad_wr = NULL;
    4572             :         int                             rc;
    4573             : 
    4574           0 :         while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) {
    4575           0 :                 rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send);
    4576           0 :                 rc = spdk_rdma_provider_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr);
    4577             : 
    4578             :                 /* bad wr always points to the first wr that failed. */
    4579           0 :                 if (spdk_unlikely(rc)) {
    4580           0 :                         _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc);
    4581             :                 }
    4582           0 :                 STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link);
    4583             :         }
    4584           0 : }
    4585             : 
    4586             : static const char *
    4587           0 : nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type)
    4588             : {
    4589           0 :         switch (wr_type) {
    4590           0 :         case RDMA_WR_TYPE_RECV:
    4591           0 :                 return "RECV";
    4592           0 :         case RDMA_WR_TYPE_SEND:
    4593           0 :                 return "SEND";
    4594           0 :         case RDMA_WR_TYPE_DATA:
    4595           0 :                 return "DATA";
    4596           0 :         default:
    4597           0 :                 SPDK_ERRLOG("Unknown WR type %d\n", wr_type);
    4598           0 :                 SPDK_UNREACHABLE();
    4599             :         }
    4600             : }
    4601             : 
    4602             : static inline void
    4603           0 : nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc)
    4604             : {
    4605           0 :         enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type;
    4606             : 
    4607           0 :         if (wc->status == IBV_WC_WR_FLUSH_ERR) {
    4608             :                 /* If qpair is in ERR state, we will receive completions for all posted and not completed
    4609             :                  * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */
    4610           0 :                 SPDK_DEBUGLOG(rdma,
    4611             :                               "Error on CQ %p, (qp state %d, in_error %d) request 0x%lu, type %s, status: (%d): %s\n",
    4612             :                               rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_in_error_state, wc->wr_id,
    4613             :                               nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status));
    4614             :         } else {
    4615           0 :                 SPDK_ERRLOG("Error on CQ %p, (qp state %d, in_error %d) request 0x%lu, type %s, status: (%d): %s\n",
    4616             :                             rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_in_error_state, wc->wr_id,
    4617             :                             nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status));
    4618             :         }
    4619           0 : }
    4620             : 
    4621             : static int
    4622           0 : nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport,
    4623             :                       struct spdk_nvmf_rdma_poller *rpoller)
    4624             : {
    4625           0 :         struct ibv_wc wc[32];
    4626             :         struct spdk_nvmf_rdma_wr        *rdma_wr;
    4627             :         struct spdk_nvmf_rdma_request   *rdma_req;
    4628             :         struct spdk_nvmf_rdma_recv      *rdma_recv;
    4629             :         struct spdk_nvmf_rdma_qpair     *rqpair, *tmp_rqpair;
    4630             :         int reaped, i;
    4631           0 :         int count = 0;
    4632             :         int rc;
    4633           0 :         bool error = false;
    4634           0 :         uint64_t poll_tsc = spdk_get_ticks();
    4635             : 
    4636           0 :         if (spdk_unlikely(rpoller->need_destroy)) {
    4637             :                 /* If qpair is closed before poller destroy, nvmf_rdma_destroy_drained_qpair may not
    4638             :                  * be called because we cannot poll anything from cq. So we call that here to force
    4639             :                  * destroy the qpair after to_close turning true.
    4640             :                  */
    4641           0 :                 RB_FOREACH_SAFE(rqpair, qpairs_tree, &rpoller->qpairs, tmp_rqpair) {
    4642           0 :                         nvmf_rdma_destroy_drained_qpair(rqpair);
    4643             :                 }
    4644           0 :                 return 0;
    4645             :         }
    4646             : 
    4647             :         /* Poll for completing operations. */
    4648           0 :         reaped = ibv_poll_cq(rpoller->cq, 32, wc);
    4649           0 :         if (spdk_unlikely(reaped < 0)) {
    4650           0 :                 SPDK_ERRLOG("Error polling CQ! (%d): %s\n",
    4651             :                             errno, spdk_strerror(errno));
    4652           0 :                 return -1;
    4653           0 :         } else if (reaped == 0) {
    4654           0 :                 rpoller->stat.idle_polls++;
    4655             :         }
    4656             : 
    4657           0 :         rpoller->stat.polls++;
    4658           0 :         rpoller->stat.completions += reaped;
    4659             : 
    4660           0 :         for (i = 0; i < reaped; i++) {
    4661             : 
    4662           0 :                 rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id;
    4663             : 
    4664           0 :                 switch (rdma_wr->type) {
    4665           0 :                 case RDMA_WR_TYPE_SEND:
    4666           0 :                         rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr);
    4667           0 :                         rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
    4668             : 
    4669           0 :                         if (spdk_likely(!wc[i].status)) {
    4670           0 :                                 count++;
    4671           0 :                                 assert(wc[i].opcode == IBV_WC_SEND);
    4672           0 :                                 assert(nvmf_rdma_req_is_completing(rdma_req));
    4673             :                         }
    4674             : 
    4675           0 :                         rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
    4676             :                         /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */
    4677           0 :                         assert(rqpair->current_send_depth >= (uint32_t)rdma_req->num_outstanding_data_wr + 1);
    4678           0 :                         rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1;
    4679           0 :                         rdma_req->num_outstanding_data_wr = 0;
    4680             : 
    4681           0 :                         nvmf_rdma_request_process(rtransport, rdma_req);
    4682           0 :                         break;
    4683           0 :                 case RDMA_WR_TYPE_RECV:
    4684             :                         /* rdma_recv->qpair will be invalid if using an SRQ.  In that case we have to get the qpair from the wc. */
    4685           0 :                         rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr);
    4686           0 :                         if (rpoller->srq != NULL) {
    4687           0 :                                 rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]);
    4688             :                                 /* It is possible that there are still some completions for destroyed QP
    4689             :                                  * associated with SRQ. We just ignore these late completions and re-post
    4690             :                                  * receive WRs back to SRQ.
    4691             :                                  */
    4692           0 :                                 if (spdk_unlikely(NULL == rdma_recv->qpair)) {
    4693           0 :                                         struct ibv_recv_wr *bad_wr;
    4694             : 
    4695           0 :                                         rdma_recv->wr.next = NULL;
    4696           0 :                                         spdk_rdma_provider_srq_queue_recv_wrs(rpoller->srq, &rdma_recv->wr);
    4697           0 :                                         rc = spdk_rdma_provider_srq_flush_recv_wrs(rpoller->srq, &bad_wr);
    4698           0 :                                         if (rc) {
    4699           0 :                                                 SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc);
    4700             :                                         }
    4701           0 :                                         continue;
    4702             :                                 }
    4703             :                         }
    4704           0 :                         rqpair = rdma_recv->qpair;
    4705             : 
    4706           0 :                         assert(rqpair != NULL);
    4707           0 :                         if (spdk_likely(!wc[i].status)) {
    4708           0 :                                 assert(wc[i].opcode == IBV_WC_RECV);
    4709           0 :                                 if (rqpair->current_recv_depth >= rqpair->max_queue_depth) {
    4710           0 :                                         spdk_nvmf_qpair_disconnect(&rqpair->qpair);
    4711           0 :                                         break;
    4712             :                                 }
    4713             :                         }
    4714             : 
    4715           0 :                         rdma_recv->wr.next = NULL;
    4716           0 :                         rqpair->current_recv_depth++;
    4717           0 :                         rdma_recv->receive_tsc = poll_tsc;
    4718           0 :                         rpoller->stat.requests++;
    4719           0 :                         STAILQ_INSERT_HEAD(&rqpair->resources->incoming_queue, rdma_recv, link);
    4720           0 :                         rqpair->qpair.queue_depth++;
    4721           0 :                         break;
    4722           0 :                 case RDMA_WR_TYPE_DATA:
    4723           0 :                         rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data_wr);
    4724           0 :                         rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
    4725             : 
    4726           0 :                         assert(rdma_req->num_outstanding_data_wr > 0);
    4727             : 
    4728           0 :                         rqpair->current_send_depth--;
    4729           0 :                         rdma_req->num_outstanding_data_wr--;
    4730           0 :                         if (spdk_likely(!wc[i].status)) {
    4731           0 :                                 assert(wc[i].opcode == IBV_WC_RDMA_READ);
    4732           0 :                                 rqpair->current_read_depth--;
    4733             :                                 /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */
    4734           0 :                                 if (rdma_req->num_outstanding_data_wr == 0) {
    4735           0 :                                         if (rdma_req->num_remaining_data_wr) {
    4736             :                                                 /* Only part of RDMA_READ operations was submitted, process the rest */
    4737           0 :                                                 nvmf_rdma_request_reset_transfer_in(rdma_req, rtransport);
    4738           0 :                                                 rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING;
    4739           0 :                                                 nvmf_rdma_request_process(rtransport, rdma_req);
    4740           0 :                                                 break;
    4741             :                                         }
    4742           0 :                                         rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
    4743           0 :                                         nvmf_rdma_request_process(rtransport, rdma_req);
    4744             :                                 }
    4745             :                         } else {
    4746             :                                 /* If the data transfer fails still force the queue into the error state,
    4747             :                                  * if we were performing an RDMA_READ, we need to force the request into a
    4748             :                                  * completed state since it wasn't linked to a send. However, in the RDMA_WRITE
    4749             :                                  * case, we should wait for the SEND to complete. */
    4750           0 :                                 if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) {
    4751           0 :                                         rqpair->current_read_depth--;
    4752           0 :                                         if (rdma_req->num_outstanding_data_wr == 0) {
    4753           0 :                                                 rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
    4754             :                                         }
    4755             :                                 }
    4756             :                         }
    4757           0 :                         break;
    4758           0 :                 default:
    4759           0 :                         SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode);
    4760           0 :                         continue;
    4761             :                 }
    4762             : 
    4763             :                 /* Handle error conditions */
    4764           0 :                 if (spdk_unlikely(wc[i].status)) {
    4765           0 :                         rqpair->ibv_in_error_state = true;
    4766           0 :                         nvmf_rdma_log_wc_status(rqpair, &wc[i]);
    4767             : 
    4768           0 :                         error = true;
    4769             : 
    4770           0 :                         if (spdk_nvmf_qpair_is_active(&rqpair->qpair)) {
    4771             :                                 /* Disconnect the connection. */
    4772           0 :                                 spdk_nvmf_qpair_disconnect(&rqpair->qpair);
    4773             :                         } else {
    4774           0 :                                 nvmf_rdma_destroy_drained_qpair(rqpair);
    4775             :                         }
    4776           0 :                         continue;
    4777             :                 }
    4778             : 
    4779           0 :                 nvmf_rdma_qpair_process_pending(rtransport, rqpair, false);
    4780             : 
    4781           0 :                 if (spdk_unlikely(!spdk_nvmf_qpair_is_active(&rqpair->qpair))) {
    4782           0 :                         nvmf_rdma_destroy_drained_qpair(rqpair);
    4783             :                 }
    4784             :         }
    4785             : 
    4786           0 :         if (spdk_unlikely(error == true)) {
    4787           0 :                 return -1;
    4788             :         }
    4789             : 
    4790           0 :         if (reaped == 0) {
    4791             :                 /* In some cases we may not receive any CQE but we still may have pending IO requests waiting for
    4792             :                  * a resource (e.g. a WR from the data_wr_pool).
    4793             :                  * We need to start processing of such requests if no CQE reaped */
    4794           0 :                 nvmf_rdma_poller_process_pending_buf_queue(rtransport, rpoller);
    4795             :         }
    4796             : 
    4797             :         /* submit outstanding work requests. */
    4798           0 :         _poller_submit_recvs(rtransport, rpoller);
    4799           0 :         _poller_submit_sends(rtransport, rpoller);
    4800             : 
    4801           0 :         return count;
    4802             : }
    4803             : 
    4804             : static void
    4805           0 : _nvmf_rdma_remove_destroyed_device(void *c)
    4806             : {
    4807           0 :         struct spdk_nvmf_rdma_transport *rtransport = c;
    4808             :         struct spdk_nvmf_rdma_device    *device, *device_tmp;
    4809             :         int                             rc;
    4810             : 
    4811           0 :         TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) {
    4812           0 :                 if (device->ready_to_destroy) {
    4813           0 :                         destroy_ib_device(rtransport, device);
    4814             :                 }
    4815             :         }
    4816             : 
    4817           0 :         free_poll_fds(rtransport);
    4818           0 :         rc = generate_poll_fds(rtransport);
    4819             :         /* cannot handle fd allocation error here */
    4820           0 :         if (rc != 0) {
    4821           0 :                 SPDK_ERRLOG("Failed to generate poll fds after remove ib device.\n");
    4822             :         }
    4823           0 : }
    4824             : 
    4825             : static void
    4826           0 : _nvmf_rdma_remove_poller_in_group_cb(void *c)
    4827             : {
    4828           0 :         struct poller_manage_ctx        *ctx = c;
    4829           0 :         struct spdk_nvmf_rdma_transport *rtransport = ctx->rtransport;
    4830           0 :         struct spdk_nvmf_rdma_device    *device = ctx->device;
    4831           0 :         struct spdk_thread              *thread = ctx->thread;
    4832             : 
    4833           0 :         if (nvmf_rdma_all_pollers_management_done(c)) {
    4834             :                 /* destroy device when last poller is destroyed */
    4835           0 :                 device->ready_to_destroy = true;
    4836           0 :                 spdk_thread_send_msg(thread, _nvmf_rdma_remove_destroyed_device, rtransport);
    4837             :         }
    4838           0 : }
    4839             : 
    4840             : static void
    4841           0 : _nvmf_rdma_remove_poller_in_group(void *c)
    4842             : {
    4843           0 :         struct poller_manage_ctx                *ctx = c;
    4844             : 
    4845           0 :         ctx->rpoller->need_destroy = true;
    4846           0 :         ctx->rpoller->destroy_cb_ctx = ctx;
    4847           0 :         ctx->rpoller->destroy_cb = _nvmf_rdma_remove_poller_in_group_cb;
    4848             : 
    4849             :         /* qp will be disconnected after receiving a RDMA_CM_EVENT_DEVICE_REMOVAL event. */
    4850           0 :         if (RB_EMPTY(&ctx->rpoller->qpairs)) {
    4851           0 :                 nvmf_rdma_poller_destroy(ctx->rpoller);
    4852             :         }
    4853           0 : }
    4854             : 
    4855             : static int
    4856           0 : nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group)
    4857             : {
    4858             :         struct spdk_nvmf_rdma_transport *rtransport;
    4859             :         struct spdk_nvmf_rdma_poll_group *rgroup;
    4860             :         struct spdk_nvmf_rdma_poller    *rpoller, *tmp;
    4861           0 :         int                             count = 0, rc, rc2 = 0;
    4862             : 
    4863           0 :         rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport);
    4864           0 :         rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
    4865             : 
    4866           0 :         TAILQ_FOREACH_SAFE(rpoller, &rgroup->pollers, link, tmp) {
    4867           0 :                 rc = nvmf_rdma_poller_poll(rtransport, rpoller);
    4868           0 :                 if (spdk_unlikely(rc < 0)) {
    4869           0 :                         if (rc2 == 0) {
    4870           0 :                                 rc2 = rc;
    4871             :                         }
    4872           0 :                         continue;
    4873             :                 }
    4874           0 :                 count += rc;
    4875             :         }
    4876             : 
    4877           0 :         return rc2 ? rc2 : count;
    4878             : }
    4879             : 
    4880             : static int
    4881           0 : nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id,
    4882             :                           struct spdk_nvme_transport_id *trid,
    4883             :                           bool peer)
    4884             : {
    4885             :         struct sockaddr *saddr;
    4886             :         uint16_t port;
    4887             : 
    4888           0 :         spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA);
    4889             : 
    4890           0 :         if (peer) {
    4891           0 :                 saddr = rdma_get_peer_addr(id);
    4892             :         } else {
    4893           0 :                 saddr = rdma_get_local_addr(id);
    4894             :         }
    4895           0 :         switch (saddr->sa_family) {
    4896           0 :         case AF_INET: {
    4897           0 :                 struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr;
    4898             : 
    4899           0 :                 trid->adrfam = SPDK_NVMF_ADRFAM_IPV4;
    4900           0 :                 inet_ntop(AF_INET, &saddr_in->sin_addr,
    4901           0 :                           trid->traddr, sizeof(trid->traddr));
    4902           0 :                 if (peer) {
    4903           0 :                         port = ntohs(rdma_get_dst_port(id));
    4904             :                 } else {
    4905           0 :                         port = ntohs(rdma_get_src_port(id));
    4906             :                 }
    4907           0 :                 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port);
    4908           0 :                 break;
    4909             :         }
    4910           0 :         case AF_INET6: {
    4911           0 :                 struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr;
    4912           0 :                 trid->adrfam = SPDK_NVMF_ADRFAM_IPV6;
    4913           0 :                 inet_ntop(AF_INET6, &saddr_in->sin6_addr,
    4914           0 :                           trid->traddr, sizeof(trid->traddr));
    4915           0 :                 if (peer) {
    4916           0 :                         port = ntohs(rdma_get_dst_port(id));
    4917             :                 } else {
    4918           0 :                         port = ntohs(rdma_get_src_port(id));
    4919             :                 }
    4920           0 :                 snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port);
    4921           0 :                 break;
    4922             :         }
    4923           0 :         default:
    4924           0 :                 return -1;
    4925             : 
    4926             :         }
    4927             : 
    4928           0 :         return 0;
    4929             : }
    4930             : 
    4931             : static int
    4932           0 : nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair,
    4933             :                               struct spdk_nvme_transport_id *trid)
    4934             : {
    4935             :         struct spdk_nvmf_rdma_qpair     *rqpair;
    4936             : 
    4937           0 :         rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
    4938             : 
    4939           0 :         return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true);
    4940             : }
    4941             : 
    4942             : static int
    4943           0 : nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair,
    4944             :                                struct spdk_nvme_transport_id *trid)
    4945             : {
    4946             :         struct spdk_nvmf_rdma_qpair     *rqpair;
    4947             : 
    4948           0 :         rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
    4949             : 
    4950           0 :         return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false);
    4951             : }
    4952             : 
    4953             : static int
    4954           0 : nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair,
    4955             :                                 struct spdk_nvme_transport_id *trid)
    4956             : {
    4957             :         struct spdk_nvmf_rdma_qpair     *rqpair;
    4958             : 
    4959           0 :         rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
    4960             : 
    4961           0 :         return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false);
    4962             : }
    4963             : 
    4964             : void
    4965           0 : spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks)
    4966             : {
    4967           0 :         g_nvmf_hooks = *hooks;
    4968           0 : }
    4969             : 
    4970             : static void
    4971           0 : nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req,
    4972             :                                    struct spdk_nvmf_rdma_request *rdma_req_to_abort,
    4973             :                                    struct spdk_nvmf_rdma_qpair *rqpair)
    4974             : {
    4975           0 :         rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC;
    4976           0 :         rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST;
    4977             : 
    4978           0 :         STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req_to_abort, state_link);
    4979           0 :         rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
    4980             : 
    4981           0 :         req->rsp->nvme_cpl.cdw0 &= ~1U;       /* Command was successfully aborted. */
    4982           0 : }
    4983             : 
    4984             : static int
    4985           0 : _nvmf_rdma_qpair_abort_request(void *ctx)
    4986             : {
    4987           0 :         struct spdk_nvmf_request *req = ctx;
    4988           0 :         struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF(
    4989             :                                 req->req_to_abort, struct spdk_nvmf_rdma_request, req);
    4990           0 :         struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair,
    4991             :                                               struct spdk_nvmf_rdma_qpair, qpair);
    4992             :         int rc;
    4993             : 
    4994           0 :         spdk_poller_unregister(&req->poller);
    4995             : 
    4996           0 :         switch (rdma_req_to_abort->state) {
    4997           0 :         case RDMA_REQUEST_STATE_EXECUTING:
    4998           0 :                 rc = nvmf_ctrlr_abort_request(req);
    4999           0 :                 if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) {
    5000           0 :                         return SPDK_POLLER_BUSY;
    5001             :                 }
    5002           0 :                 break;
    5003             : 
    5004           0 :         case RDMA_REQUEST_STATE_NEED_BUFFER:
    5005           0 :                 STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue,
    5006             :                               &rdma_req_to_abort->req, spdk_nvmf_request, buf_link);
    5007             : 
    5008           0 :                 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
    5009           0 :                 break;
    5010             : 
    5011           0 :         case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING:
    5012           0 :                 STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort,
    5013             :                               spdk_nvmf_rdma_request, state_link);
    5014             : 
    5015           0 :                 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
    5016           0 :                 break;
    5017             : 
    5018           0 :         case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING:
    5019           0 :                 STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort,
    5020             :                               spdk_nvmf_rdma_request, state_link);
    5021             : 
    5022           0 :                 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
    5023           0 :                 break;
    5024             : 
    5025           0 :         case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING:
    5026             :                 /* Remove req from the list here to re-use common function */
    5027           0 :                 STAILQ_REMOVE(&rqpair->pending_rdma_send_queue, rdma_req_to_abort,
    5028             :                               spdk_nvmf_rdma_request, state_link);
    5029             : 
    5030           0 :                 nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
    5031           0 :                 break;
    5032             : 
    5033           0 :         case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
    5034           0 :                 if (spdk_get_ticks() < req->timeout_tsc) {
    5035           0 :                         req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0);
    5036           0 :                         return SPDK_POLLER_BUSY;
    5037             :                 }
    5038           0 :                 break;
    5039             : 
    5040           0 :         default:
    5041           0 :                 break;
    5042             :         }
    5043             : 
    5044           0 :         spdk_nvmf_request_complete(req);
    5045           0 :         return SPDK_POLLER_BUSY;
    5046             : }
    5047             : 
    5048             : static void
    5049           0 : nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair,
    5050             :                               struct spdk_nvmf_request *req)
    5051             : {
    5052             :         struct spdk_nvmf_rdma_qpair *rqpair;
    5053             :         struct spdk_nvmf_rdma_transport *rtransport;
    5054             :         struct spdk_nvmf_transport *transport;
    5055             :         uint16_t cid;
    5056             :         uint32_t i, max_req_count;
    5057           0 :         struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL, *rdma_req;
    5058             : 
    5059           0 :         rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
    5060           0 :         rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
    5061           0 :         transport = &rtransport->transport;
    5062             : 
    5063           0 :         cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid;
    5064           0 :         max_req_count = rqpair->srq == NULL ? rqpair->max_queue_depth : rqpair->poller->max_srq_depth;
    5065             : 
    5066           0 :         for (i = 0; i < max_req_count; i++) {
    5067           0 :                 rdma_req = &rqpair->resources->reqs[i];
    5068             :                 /* When SRQ == NULL, rqpair has its own requests and req.qpair pointer always points to the qpair
    5069             :                  * When SRQ != NULL all rqpairs share common requests and qpair pointer is assigned when we start to
    5070             :                  * process a request. So in both cases all requests which are not in FREE state have valid qpair ptr */
    5071           0 :                 if (rdma_req->state != RDMA_REQUEST_STATE_FREE && rdma_req->req.cmd->nvme_cmd.cid == cid &&
    5072           0 :                     rdma_req->req.qpair == qpair) {
    5073           0 :                         rdma_req_to_abort = rdma_req;
    5074           0 :                         break;
    5075             :                 }
    5076             :         }
    5077             : 
    5078           0 :         if (rdma_req_to_abort == NULL) {
    5079           0 :                 spdk_nvmf_request_complete(req);
    5080           0 :                 return;
    5081             :         }
    5082             : 
    5083           0 :         req->req_to_abort = &rdma_req_to_abort->req;
    5084           0 :         req->timeout_tsc = spdk_get_ticks() +
    5085           0 :                            transport->opts.abort_timeout_sec * spdk_get_ticks_hz();
    5086           0 :         req->poller = NULL;
    5087             : 
    5088           0 :         _nvmf_rdma_qpair_abort_request(req);
    5089             : }
    5090             : 
    5091             : static void
    5092           0 : nvmf_rdma_poll_group_dump_stat(struct spdk_nvmf_transport_poll_group *group,
    5093             :                                struct spdk_json_write_ctx *w)
    5094             : {
    5095             :         struct spdk_nvmf_rdma_poll_group *rgroup;
    5096             :         struct spdk_nvmf_rdma_poller *rpoller;
    5097             : 
    5098           0 :         assert(w != NULL);
    5099             : 
    5100           0 :         rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
    5101             : 
    5102           0 :         spdk_json_write_named_uint64(w, "pending_data_buffer", rgroup->stat.pending_data_buffer);
    5103             : 
    5104           0 :         spdk_json_write_named_array_begin(w, "devices");
    5105             : 
    5106           0 :         TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
    5107           0 :                 spdk_json_write_object_begin(w);
    5108           0 :                 spdk_json_write_named_string(w, "name",
    5109           0 :                                              ibv_get_device_name(rpoller->device->context->device));
    5110           0 :                 spdk_json_write_named_uint64(w, "polls",
    5111             :                                              rpoller->stat.polls);
    5112           0 :                 spdk_json_write_named_uint64(w, "idle_polls",
    5113             :                                              rpoller->stat.idle_polls);
    5114           0 :                 spdk_json_write_named_uint64(w, "completions",
    5115             :                                              rpoller->stat.completions);
    5116           0 :                 spdk_json_write_named_uint64(w, "requests",
    5117             :                                              rpoller->stat.requests);
    5118           0 :                 spdk_json_write_named_uint64(w, "request_latency",
    5119             :                                              rpoller->stat.request_latency);
    5120           0 :                 spdk_json_write_named_uint64(w, "pending_free_request",
    5121             :                                              rpoller->stat.pending_free_request);
    5122           0 :                 spdk_json_write_named_uint64(w, "pending_rdma_read",
    5123             :                                              rpoller->stat.pending_rdma_read);
    5124           0 :                 spdk_json_write_named_uint64(w, "pending_rdma_write",
    5125             :                                              rpoller->stat.pending_rdma_write);
    5126           0 :                 spdk_json_write_named_uint64(w, "pending_rdma_send",
    5127             :                                              rpoller->stat.pending_rdma_send);
    5128           0 :                 spdk_json_write_named_uint64(w, "total_send_wrs",
    5129             :                                              rpoller->stat.qp_stats.send.num_submitted_wrs);
    5130           0 :                 spdk_json_write_named_uint64(w, "send_doorbell_updates",
    5131             :                                              rpoller->stat.qp_stats.send.doorbell_updates);
    5132           0 :                 spdk_json_write_named_uint64(w, "total_recv_wrs",
    5133             :                                              rpoller->stat.qp_stats.recv.num_submitted_wrs);
    5134           0 :                 spdk_json_write_named_uint64(w, "recv_doorbell_updates",
    5135             :                                              rpoller->stat.qp_stats.recv.doorbell_updates);
    5136           0 :                 spdk_json_write_object_end(w);
    5137             :         }
    5138             : 
    5139           0 :         spdk_json_write_array_end(w);
    5140           0 : }
    5141             : 
    5142             : const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = {
    5143             :         .name = "RDMA",
    5144             :         .type = SPDK_NVME_TRANSPORT_RDMA,
    5145             :         .opts_init = nvmf_rdma_opts_init,
    5146             :         .create = nvmf_rdma_create,
    5147             :         .dump_opts = nvmf_rdma_dump_opts,
    5148             :         .destroy = nvmf_rdma_destroy,
    5149             : 
    5150             :         .listen = nvmf_rdma_listen,
    5151             :         .stop_listen = nvmf_rdma_stop_listen,
    5152             :         .cdata_init = nvmf_rdma_cdata_init,
    5153             : 
    5154             :         .listener_discover = nvmf_rdma_discover,
    5155             : 
    5156             :         .poll_group_create = nvmf_rdma_poll_group_create,
    5157             :         .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group,
    5158             :         .poll_group_destroy = nvmf_rdma_poll_group_destroy,
    5159             :         .poll_group_add = nvmf_rdma_poll_group_add,
    5160             :         .poll_group_remove = nvmf_rdma_poll_group_remove,
    5161             :         .poll_group_poll = nvmf_rdma_poll_group_poll,
    5162             : 
    5163             :         .req_free = nvmf_rdma_request_free,
    5164             :         .req_complete = nvmf_rdma_request_complete,
    5165             : 
    5166             :         .qpair_fini = nvmf_rdma_close_qpair,
    5167             :         .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid,
    5168             :         .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid,
    5169             :         .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid,
    5170             :         .qpair_abort_request = nvmf_rdma_qpair_abort_request,
    5171             : 
    5172             :         .poll_group_dump_stat = nvmf_rdma_poll_group_dump_stat,
    5173             : };
    5174             : 
    5175           2 : SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma);
    5176           2 : SPDK_LOG_REGISTER_COMPONENT(rdma)

Generated by: LCOV version 1.15