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 : static void
104 0 : nvmf_trace(void)
105 : {
106 0 : spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r');
107 :
108 0 : struct spdk_trace_tpoint_opts opts[] = {
109 : {
110 : "RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW,
111 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 1,
112 : {
113 : { "qpair", SPDK_TRACE_ARG_TYPE_PTR, 8 },
114 : { "qd", SPDK_TRACE_ARG_TYPE_INT, 4 }
115 : }
116 : },
117 : {
118 : "RDMA_REQ_COMPLETED", TRACE_RDMA_REQUEST_STATE_COMPLETED,
119 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
120 : {
121 : { "qpair", SPDK_TRACE_ARG_TYPE_PTR, 8 },
122 : { "qd", SPDK_TRACE_ARG_TYPE_INT, 4 }
123 : }
124 : },
125 : };
126 :
127 0 : spdk_trace_register_description_ext(opts, SPDK_COUNTOF(opts));
128 0 : spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER,
129 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
130 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
131 0 : spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H",
132 : TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING,
133 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
134 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
135 0 : spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C",
136 : TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING,
137 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
138 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
139 0 : spdk_trace_register_description("RDMA_REQ_TX_H2C",
140 : TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER,
141 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
142 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
143 0 : spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE",
144 : TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE,
145 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
146 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
147 0 : spdk_trace_register_description("RDMA_REQ_EXECUTING",
148 : TRACE_RDMA_REQUEST_STATE_EXECUTING,
149 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
150 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
151 0 : spdk_trace_register_description("RDMA_REQ_EXECUTED",
152 : TRACE_RDMA_REQUEST_STATE_EXECUTED,
153 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
154 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
155 0 : spdk_trace_register_description("RDMA_REQ_RDY2COMPL_PEND",
156 : TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING,
157 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
158 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
159 0 : spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL",
160 : TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE,
161 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
162 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
163 0 : spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H",
164 : TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST,
165 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
166 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
167 0 : spdk_trace_register_description("RDMA_REQ_COMPLETING",
168 : TRACE_RDMA_REQUEST_STATE_COMPLETING,
169 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
170 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
171 :
172 0 : spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE,
173 : OWNER_TYPE_NONE, OBJECT_NONE, 0,
174 : SPDK_TRACE_ARG_TYPE_INT, "");
175 0 : spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT,
176 : OWNER_TYPE_NONE, OBJECT_NONE, 0,
177 : SPDK_TRACE_ARG_TYPE_INT, "type");
178 0 : spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT,
179 : OWNER_TYPE_NONE, OBJECT_NONE, 0,
180 : SPDK_TRACE_ARG_TYPE_INT, "type");
181 0 : spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT,
182 : OWNER_TYPE_NONE, OBJECT_NONE, 0,
183 : SPDK_TRACE_ARG_TYPE_INT, "");
184 0 : spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY,
185 : OWNER_TYPE_NONE, OBJECT_NONE, 0,
186 : SPDK_TRACE_ARG_TYPE_INT, "");
187 :
188 0 : spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_START, OBJECT_NVMF_RDMA_IO, 1);
189 0 : spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_DONE, OBJECT_NVMF_RDMA_IO, 0);
190 0 : }
191 2 : SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA)
192 :
193 : enum spdk_nvmf_rdma_wr_type {
194 : RDMA_WR_TYPE_RECV,
195 : RDMA_WR_TYPE_SEND,
196 : RDMA_WR_TYPE_DATA,
197 : };
198 :
199 : struct spdk_nvmf_rdma_wr {
200 : /* Uses enum spdk_nvmf_rdma_wr_type */
201 : uint8_t type;
202 : };
203 :
204 : /* This structure holds commands as they are received off the wire.
205 : * It must be dynamically paired with a full request object
206 : * (spdk_nvmf_rdma_request) to service a request. It is separate
207 : * from the request because RDMA does not appear to order
208 : * completions, so occasionally we'll get a new incoming
209 : * command when there aren't any free request objects.
210 : */
211 : struct spdk_nvmf_rdma_recv {
212 : struct ibv_recv_wr wr;
213 : struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE];
214 :
215 : struct spdk_nvmf_rdma_qpair *qpair;
216 :
217 : /* In-capsule data buffer */
218 : uint8_t *buf;
219 :
220 : struct spdk_nvmf_rdma_wr rdma_wr;
221 : uint64_t receive_tsc;
222 :
223 : STAILQ_ENTRY(spdk_nvmf_rdma_recv) link;
224 : };
225 :
226 : struct spdk_nvmf_rdma_request_data {
227 : struct ibv_send_wr wr;
228 : struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES];
229 : };
230 :
231 : struct spdk_nvmf_rdma_request {
232 : struct spdk_nvmf_request req;
233 :
234 : bool fused_failed;
235 :
236 : struct spdk_nvmf_rdma_wr data_wr;
237 : struct spdk_nvmf_rdma_wr rsp_wr;
238 :
239 : /* Uses enum spdk_nvmf_rdma_request_state */
240 : uint8_t state;
241 :
242 : /* Data offset in req.iov */
243 : uint32_t offset;
244 :
245 : struct spdk_nvmf_rdma_recv *recv;
246 :
247 : struct {
248 : struct ibv_send_wr wr;
249 : struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE];
250 : } rsp;
251 :
252 : uint16_t iovpos;
253 : uint16_t num_outstanding_data_wr;
254 : /* Used to split Write IO with multi SGL payload */
255 : uint16_t num_remaining_data_wr;
256 : uint64_t receive_tsc;
257 : struct spdk_nvmf_rdma_request *fused_pair;
258 : STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link;
259 : struct ibv_send_wr *remaining_tranfer_in_wrs;
260 : struct ibv_send_wr *transfer_wr;
261 : struct spdk_nvmf_rdma_request_data data;
262 : };
263 :
264 : struct spdk_nvmf_rdma_resource_opts {
265 : struct spdk_nvmf_rdma_qpair *qpair;
266 : /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */
267 : void *qp;
268 : struct spdk_rdma_utils_mem_map *map;
269 : uint32_t max_queue_depth;
270 : uint32_t in_capsule_data_size;
271 : bool shared;
272 : };
273 :
274 : struct spdk_nvmf_rdma_resources {
275 : /* Array of size "max_queue_depth" containing RDMA requests. */
276 : struct spdk_nvmf_rdma_request *reqs;
277 :
278 : /* Array of size "max_queue_depth" containing RDMA recvs. */
279 : struct spdk_nvmf_rdma_recv *recvs;
280 :
281 : /* Array of size "max_queue_depth" containing 64 byte capsules
282 : * used for receive.
283 : */
284 : union nvmf_h2c_msg *cmds;
285 :
286 : /* Array of size "max_queue_depth" containing 16 byte completions
287 : * to be sent back to the user.
288 : */
289 : union nvmf_c2h_msg *cpls;
290 :
291 : /* Array of size "max_queue_depth * InCapsuleDataSize" containing
292 : * buffers to be used for in capsule data.
293 : */
294 : void *bufs;
295 :
296 : /* Receives that are waiting for a request object */
297 : STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue;
298 :
299 : /* Queue to track free requests */
300 : STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue;
301 : };
302 :
303 : typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair);
304 :
305 : typedef void (*spdk_poller_destroy_cb)(void *ctx);
306 :
307 : struct spdk_nvmf_rdma_ibv_event_ctx {
308 : struct spdk_nvmf_rdma_qpair *rqpair;
309 : };
310 :
311 : struct spdk_nvmf_rdma_qpair {
312 : struct spdk_nvmf_qpair qpair;
313 :
314 : struct spdk_nvmf_rdma_device *device;
315 : struct spdk_nvmf_rdma_poller *poller;
316 :
317 : struct spdk_rdma_provider_qp *rdma_qp;
318 : struct rdma_cm_id *cm_id;
319 : struct spdk_rdma_provider_srq *srq;
320 : struct rdma_cm_id *listen_id;
321 :
322 : /* Cache the QP number to improve QP search by RB tree. */
323 : uint32_t qp_num;
324 :
325 : /* The maximum number of I/O outstanding on this connection at one time */
326 : uint16_t max_queue_depth;
327 :
328 : /* The maximum number of active RDMA READ and ATOMIC operations at one time */
329 : uint16_t max_read_depth;
330 :
331 : /* The maximum number of RDMA SEND operations at one time */
332 : uint32_t max_send_depth;
333 :
334 : /* The current number of outstanding WRs from this qpair's
335 : * recv queue. Should not exceed device->attr.max_queue_depth.
336 : */
337 : uint16_t current_recv_depth;
338 :
339 : /* The current number of active RDMA READ operations */
340 : uint16_t current_read_depth;
341 :
342 : /* The current number of posted WRs from this qpair's
343 : * send queue. Should not exceed max_send_depth.
344 : */
345 : uint32_t current_send_depth;
346 :
347 : /* The maximum number of SGEs per WR on the send queue */
348 : uint32_t max_send_sge;
349 :
350 : /* The maximum number of SGEs per WR on the recv queue */
351 : uint32_t max_recv_sge;
352 :
353 : struct spdk_nvmf_rdma_resources *resources;
354 :
355 : STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue;
356 :
357 : STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue;
358 :
359 : STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_send_queue;
360 :
361 : /* Number of requests not in the free state */
362 : uint32_t qd;
363 :
364 : bool ibv_in_error_state;
365 :
366 : RB_ENTRY(spdk_nvmf_rdma_qpair) node;
367 :
368 : STAILQ_ENTRY(spdk_nvmf_rdma_qpair) recv_link;
369 :
370 : STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link;
371 :
372 : /* Points to the a request that has fuse bits set to
373 : * SPDK_NVME_CMD_FUSE_FIRST, when the qpair is waiting
374 : * for the request that has SPDK_NVME_CMD_FUSE_SECOND.
375 : */
376 : struct spdk_nvmf_rdma_request *fused_first;
377 :
378 : /*
379 : * io_channel which is used to destroy qpair when it is removed from poll group
380 : */
381 : struct spdk_io_channel *destruct_channel;
382 :
383 : /* ctx for async processing of last_wqe_reached event */
384 : struct spdk_nvmf_rdma_ibv_event_ctx *last_wqe_reached_ctx;
385 :
386 : /* Lets us know that we have received the last_wqe event. */
387 : bool last_wqe_reached;
388 :
389 : /* Indicate that nvmf_rdma_close_qpair is called */
390 : bool to_close;
391 : };
392 :
393 : struct spdk_nvmf_rdma_poller_stat {
394 : uint64_t completions;
395 : uint64_t polls;
396 : uint64_t idle_polls;
397 : uint64_t requests;
398 : uint64_t request_latency;
399 : uint64_t pending_free_request;
400 : uint64_t pending_rdma_read;
401 : uint64_t pending_rdma_write;
402 : uint64_t pending_rdma_send;
403 : struct spdk_rdma_provider_qp_stats qp_stats;
404 : };
405 :
406 : struct spdk_nvmf_rdma_poller {
407 : struct spdk_nvmf_rdma_device *device;
408 : struct spdk_nvmf_rdma_poll_group *group;
409 :
410 : int num_cqe;
411 : int required_num_wr;
412 : struct ibv_cq *cq;
413 :
414 : /* The maximum number of I/O outstanding on the shared receive queue at one time */
415 : uint16_t max_srq_depth;
416 : bool need_destroy;
417 :
418 : /* Shared receive queue */
419 : struct spdk_rdma_provider_srq *srq;
420 :
421 : struct spdk_nvmf_rdma_resources *resources;
422 : struct spdk_nvmf_rdma_poller_stat stat;
423 :
424 : spdk_poller_destroy_cb destroy_cb;
425 : void *destroy_cb_ctx;
426 :
427 : RB_HEAD(qpairs_tree, spdk_nvmf_rdma_qpair) qpairs;
428 :
429 : STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv;
430 :
431 : STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send;
432 :
433 : TAILQ_ENTRY(spdk_nvmf_rdma_poller) link;
434 : };
435 :
436 : struct spdk_nvmf_rdma_poll_group_stat {
437 : uint64_t pending_data_buffer;
438 : };
439 :
440 : struct spdk_nvmf_rdma_poll_group {
441 : struct spdk_nvmf_transport_poll_group group;
442 : struct spdk_nvmf_rdma_poll_group_stat stat;
443 : TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers;
444 : TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link;
445 : };
446 :
447 : struct spdk_nvmf_rdma_conn_sched {
448 : struct spdk_nvmf_rdma_poll_group *next_admin_pg;
449 : struct spdk_nvmf_rdma_poll_group *next_io_pg;
450 : };
451 :
452 : /* Assuming rdma_cm uses just one protection domain per ibv_context. */
453 : struct spdk_nvmf_rdma_device {
454 : struct ibv_device_attr attr;
455 : struct ibv_context *context;
456 :
457 : struct spdk_rdma_utils_mem_map *map;
458 : struct ibv_pd *pd;
459 :
460 : int num_srq;
461 : bool need_destroy;
462 : bool ready_to_destroy;
463 : bool is_ready;
464 :
465 : TAILQ_ENTRY(spdk_nvmf_rdma_device) link;
466 : };
467 :
468 : struct spdk_nvmf_rdma_port {
469 : const struct spdk_nvme_transport_id *trid;
470 : struct rdma_cm_id *id;
471 : struct spdk_nvmf_rdma_device *device;
472 : TAILQ_ENTRY(spdk_nvmf_rdma_port) link;
473 : };
474 :
475 : struct rdma_transport_opts {
476 : int num_cqe;
477 : uint32_t max_srq_depth;
478 : bool no_srq;
479 : bool no_wr_batching;
480 : int acceptor_backlog;
481 : };
482 :
483 : struct spdk_nvmf_rdma_transport {
484 : struct spdk_nvmf_transport transport;
485 : struct rdma_transport_opts rdma_opts;
486 :
487 : struct spdk_nvmf_rdma_conn_sched conn_sched;
488 :
489 : struct rdma_event_channel *event_channel;
490 :
491 : struct spdk_mempool *data_wr_pool;
492 :
493 : struct spdk_poller *accept_poller;
494 :
495 : /* fields used to poll RDMA/IB events */
496 : nfds_t npoll_fds;
497 : struct pollfd *poll_fds;
498 :
499 : TAILQ_HEAD(, spdk_nvmf_rdma_device) devices;
500 : TAILQ_HEAD(, spdk_nvmf_rdma_port) ports;
501 : TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups;
502 :
503 : /* ports that are removed unexpectedly and need retry listen */
504 : TAILQ_HEAD(, spdk_nvmf_rdma_port) retry_ports;
505 : };
506 :
507 : struct poller_manage_ctx {
508 : struct spdk_nvmf_rdma_transport *rtransport;
509 : struct spdk_nvmf_rdma_poll_group *rgroup;
510 : struct spdk_nvmf_rdma_poller *rpoller;
511 : struct spdk_nvmf_rdma_device *device;
512 :
513 : struct spdk_thread *thread;
514 : volatile int *inflight_op_counter;
515 : };
516 :
517 : static const struct spdk_json_object_decoder rdma_transport_opts_decoder[] = {
518 : {
519 : "num_cqe", offsetof(struct rdma_transport_opts, num_cqe),
520 : spdk_json_decode_int32, true
521 : },
522 : {
523 : "max_srq_depth", offsetof(struct rdma_transport_opts, max_srq_depth),
524 : spdk_json_decode_uint32, true
525 : },
526 : {
527 : "no_srq", offsetof(struct rdma_transport_opts, no_srq),
528 : spdk_json_decode_bool, true
529 : },
530 : {
531 : "no_wr_batching", offsetof(struct rdma_transport_opts, no_wr_batching),
532 : spdk_json_decode_bool, true
533 : },
534 : {
535 : "acceptor_backlog", offsetof(struct rdma_transport_opts, acceptor_backlog),
536 : spdk_json_decode_int32, true
537 : },
538 : };
539 :
540 : static int
541 2 : nvmf_rdma_qpair_compare(struct spdk_nvmf_rdma_qpair *rqpair1, struct spdk_nvmf_rdma_qpair *rqpair2)
542 : {
543 2 : return rqpair1->qp_num < rqpair2->qp_num ? -1 : rqpair1->qp_num > rqpair2->qp_num;
544 : }
545 :
546 0 : RB_GENERATE_STATIC(qpairs_tree, spdk_nvmf_rdma_qpair, node, nvmf_rdma_qpair_compare);
547 :
548 : static bool nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport,
549 : struct spdk_nvmf_rdma_request *rdma_req);
550 :
551 : static void _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport,
552 : struct spdk_nvmf_rdma_poller *rpoller);
553 :
554 : static void _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport,
555 : struct spdk_nvmf_rdma_poller *rpoller);
556 :
557 : static void _nvmf_rdma_remove_destroyed_device(void *c);
558 :
559 : static inline enum spdk_nvme_media_error_status_code
560 0 : nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) {
561 0 : enum spdk_nvme_media_error_status_code result;
562 0 : switch (err_type)
563 : {
564 : case SPDK_DIF_REFTAG_ERROR:
565 0 : result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR;
566 0 : break;
567 : case SPDK_DIF_APPTAG_ERROR:
568 0 : result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR;
569 0 : break;
570 : case SPDK_DIF_GUARD_ERROR:
571 0 : result = SPDK_NVME_SC_GUARD_CHECK_ERROR;
572 0 : break;
573 : default:
574 0 : SPDK_UNREACHABLE();
575 : }
576 :
577 0 : return result;
578 0 : }
579 :
580 : /*
581 : * Return data_wrs to pool starting from \b data_wr
582 : * Request's own response and data WR are excluded
583 : */
584 : static void
585 7 : _nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req,
586 : struct ibv_send_wr *data_wr,
587 : struct spdk_mempool *pool)
588 : {
589 7 : struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES];
590 7 : struct spdk_nvmf_rdma_request_data *nvmf_data;
591 7 : struct ibv_send_wr *next_send_wr;
592 7 : uint64_t req_wrid = (uint64_t)&rdma_req->data_wr;
593 7 : uint32_t num_wrs = 0;
594 :
595 15 : while (data_wr && data_wr->wr_id == req_wrid) {
596 8 : nvmf_data = SPDK_CONTAINEROF(data_wr, struct spdk_nvmf_rdma_request_data, wr);
597 8 : memset(nvmf_data->sgl, 0, sizeof(data_wr->sg_list[0]) * data_wr->num_sge);
598 8 : data_wr->num_sge = 0;
599 8 : next_send_wr = data_wr->next;
600 8 : if (data_wr != &rdma_req->data.wr) {
601 1 : data_wr->next = NULL;
602 1 : assert(num_wrs < SPDK_NVMF_MAX_SGL_ENTRIES);
603 1 : work_requests[num_wrs] = nvmf_data;
604 1 : num_wrs++;
605 1 : }
606 8 : data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : next_send_wr;
607 : }
608 :
609 7 : if (num_wrs) {
610 1 : spdk_mempool_put_bulk(pool, (void **) work_requests, num_wrs);
611 1 : }
612 7 : }
613 :
614 : static void
615 7 : nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req,
616 : struct spdk_nvmf_rdma_transport *rtransport)
617 : {
618 7 : rdma_req->num_outstanding_data_wr = 0;
619 :
620 7 : _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool);
621 :
622 7 : if (rdma_req->remaining_tranfer_in_wrs) {
623 0 : _nvmf_rdma_request_free_data(rdma_req, rdma_req->remaining_tranfer_in_wrs,
624 0 : rtransport->data_wr_pool);
625 0 : rdma_req->remaining_tranfer_in_wrs = NULL;
626 0 : }
627 :
628 7 : rdma_req->data.wr.next = NULL;
629 7 : rdma_req->rsp.wr.next = NULL;
630 7 : }
631 :
632 : static void
633 0 : nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req)
634 : {
635 0 : SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool);
636 0 : if (req->req.cmd) {
637 0 : SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode);
638 0 : }
639 0 : if (req->recv) {
640 0 : SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id);
641 0 : }
642 0 : }
643 :
644 : static void
645 0 : nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair)
646 : {
647 0 : int i;
648 :
649 0 : SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid);
650 0 : for (i = 0; i < rqpair->max_queue_depth; i++) {
651 0 : if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) {
652 0 : nvmf_rdma_dump_request(&rqpair->resources->reqs[i]);
653 0 : }
654 0 : }
655 0 : }
656 :
657 : static void
658 1 : nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources)
659 : {
660 1 : spdk_free(resources->cmds);
661 1 : spdk_free(resources->cpls);
662 1 : spdk_free(resources->bufs);
663 1 : spdk_free(resources->reqs);
664 1 : spdk_free(resources->recvs);
665 1 : free(resources);
666 1 : }
667 :
668 :
669 : static struct spdk_nvmf_rdma_resources *
670 1 : nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts)
671 : {
672 1 : struct spdk_nvmf_rdma_resources *resources;
673 1 : struct spdk_nvmf_rdma_request *rdma_req;
674 1 : struct spdk_nvmf_rdma_recv *rdma_recv;
675 1 : struct spdk_rdma_provider_qp *qp = NULL;
676 1 : struct spdk_rdma_provider_srq *srq = NULL;
677 1 : struct ibv_recv_wr *bad_wr = NULL;
678 1 : struct spdk_rdma_utils_memory_translation translation;
679 1 : uint32_t i;
680 1 : int rc = 0;
681 :
682 1 : resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources));
683 1 : if (!resources) {
684 0 : SPDK_ERRLOG("Unable to allocate resources for receive queue.\n");
685 0 : return NULL;
686 : }
687 :
688 1 : resources->reqs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->reqs),
689 : 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
690 1 : resources->recvs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->recvs),
691 : 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
692 1 : resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds),
693 : 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
694 1 : resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls),
695 : 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
696 :
697 1 : if (opts->in_capsule_data_size > 0) {
698 1 : resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size,
699 : 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY,
700 : SPDK_MALLOC_DMA);
701 1 : }
702 :
703 2 : if (!resources->reqs || !resources->recvs || !resources->cmds ||
704 1 : !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) {
705 0 : SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n");
706 0 : goto cleanup;
707 : }
708 :
709 1 : SPDK_DEBUGLOG(rdma, "Command Array: %p Length: %lx\n",
710 : resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds));
711 1 : SPDK_DEBUGLOG(rdma, "Completion Array: %p Length: %lx\n",
712 : resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls));
713 1 : if (resources->bufs) {
714 1 : SPDK_DEBUGLOG(rdma, "In Capsule Data Array: %p Length: %x\n",
715 : resources->bufs, opts->max_queue_depth *
716 : opts->in_capsule_data_size);
717 1 : }
718 :
719 : /* Initialize queues */
720 1 : STAILQ_INIT(&resources->incoming_queue);
721 1 : STAILQ_INIT(&resources->free_queue);
722 :
723 1 : if (opts->shared) {
724 1 : srq = (struct spdk_rdma_provider_srq *)opts->qp;
725 1 : } else {
726 0 : qp = (struct spdk_rdma_provider_qp *)opts->qp;
727 : }
728 :
729 129 : for (i = 0; i < opts->max_queue_depth; i++) {
730 128 : rdma_recv = &resources->recvs[i];
731 128 : rdma_recv->qpair = opts->qpair;
732 :
733 : /* Set up memory to receive commands */
734 128 : if (resources->bufs) {
735 256 : rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i *
736 128 : opts->in_capsule_data_size));
737 128 : }
738 :
739 128 : rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV;
740 :
741 128 : rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i];
742 128 : rdma_recv->sgl[0].length = sizeof(resources->cmds[i]);
743 128 : rc = spdk_rdma_utils_get_translation(opts->map, &resources->cmds[i], sizeof(resources->cmds[i]),
744 : &translation);
745 128 : if (rc) {
746 0 : goto cleanup;
747 : }
748 128 : rdma_recv->sgl[0].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation);
749 128 : rdma_recv->wr.num_sge = 1;
750 :
751 128 : if (rdma_recv->buf) {
752 128 : rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf;
753 128 : rdma_recv->sgl[1].length = opts->in_capsule_data_size;
754 128 : rc = spdk_rdma_utils_get_translation(opts->map, rdma_recv->buf, opts->in_capsule_data_size,
755 : &translation);
756 128 : if (rc) {
757 0 : goto cleanup;
758 : }
759 128 : rdma_recv->sgl[1].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation);
760 128 : rdma_recv->wr.num_sge++;
761 128 : }
762 :
763 128 : rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr;
764 128 : rdma_recv->wr.sg_list = rdma_recv->sgl;
765 128 : if (srq) {
766 0 : spdk_rdma_provider_srq_queue_recv_wrs(srq, &rdma_recv->wr);
767 0 : } else {
768 128 : spdk_rdma_provider_qp_queue_recv_wrs(qp, &rdma_recv->wr);
769 : }
770 128 : }
771 :
772 129 : for (i = 0; i < opts->max_queue_depth; i++) {
773 128 : rdma_req = &resources->reqs[i];
774 :
775 128 : if (opts->qpair != NULL) {
776 128 : rdma_req->req.qpair = &opts->qpair->qpair;
777 128 : } else {
778 0 : rdma_req->req.qpair = NULL;
779 : }
780 128 : rdma_req->req.cmd = NULL;
781 128 : rdma_req->req.iovcnt = 0;
782 128 : rdma_req->req.stripped_data = NULL;
783 :
784 : /* Set up memory to send responses */
785 128 : rdma_req->req.rsp = &resources->cpls[i];
786 :
787 128 : rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i];
788 128 : rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]);
789 128 : rc = spdk_rdma_utils_get_translation(opts->map, &resources->cpls[i], sizeof(resources->cpls[i]),
790 : &translation);
791 128 : if (rc) {
792 0 : goto cleanup;
793 : }
794 128 : rdma_req->rsp.sgl[0].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation);
795 :
796 128 : rdma_req->rsp_wr.type = RDMA_WR_TYPE_SEND;
797 128 : rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp_wr;
798 128 : rdma_req->rsp.wr.next = NULL;
799 128 : rdma_req->rsp.wr.opcode = IBV_WR_SEND;
800 128 : rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED;
801 128 : rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl;
802 128 : rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl);
803 :
804 : /* Set up memory for data buffers */
805 128 : rdma_req->data_wr.type = RDMA_WR_TYPE_DATA;
806 128 : rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data_wr;
807 128 : rdma_req->data.wr.next = NULL;
808 128 : rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED;
809 128 : rdma_req->data.wr.sg_list = rdma_req->data.sgl;
810 128 : rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl);
811 :
812 : /* Initialize request state to FREE */
813 128 : rdma_req->state = RDMA_REQUEST_STATE_FREE;
814 128 : STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link);
815 128 : }
816 :
817 1 : if (srq) {
818 0 : rc = spdk_rdma_provider_srq_flush_recv_wrs(srq, &bad_wr);
819 0 : } else {
820 1 : rc = spdk_rdma_provider_qp_flush_recv_wrs(qp, &bad_wr);
821 : }
822 :
823 1 : if (rc) {
824 0 : goto cleanup;
825 : }
826 :
827 1 : return resources;
828 :
829 : cleanup:
830 0 : nvmf_rdma_resources_destroy(resources);
831 0 : return NULL;
832 1 : }
833 :
834 : static void
835 0 : nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair)
836 : {
837 0 : struct spdk_nvmf_rdma_ibv_event_ctx *ctx;
838 :
839 0 : ctx = rqpair->last_wqe_reached_ctx;
840 0 : if (ctx) {
841 0 : ctx->rqpair = NULL;
842 : /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_last_wqe_event */
843 0 : rqpair->last_wqe_reached_ctx = NULL;
844 0 : }
845 0 : }
846 :
847 : static void nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller);
848 :
849 : static void
850 0 : nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair)
851 : {
852 0 : struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp;
853 0 : struct ibv_recv_wr *bad_recv_wr = NULL;
854 0 : int rc;
855 :
856 0 : spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair);
857 :
858 0 : if (rqpair->qd != 0) {
859 0 : struct spdk_nvmf_qpair *qpair = &rqpair->qpair;
860 0 : struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport,
861 : struct spdk_nvmf_rdma_transport, transport);
862 0 : struct spdk_nvmf_rdma_request *req;
863 0 : uint32_t i, max_req_count = 0;
864 :
865 0 : SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd);
866 :
867 0 : if (rqpair->srq == NULL) {
868 0 : nvmf_rdma_dump_qpair_contents(rqpair);
869 0 : max_req_count = rqpair->max_queue_depth;
870 0 : } else if (rqpair->poller && rqpair->resources) {
871 0 : max_req_count = rqpair->poller->max_srq_depth;
872 0 : }
873 :
874 0 : SPDK_DEBUGLOG(rdma, "Release incomplete requests\n");
875 0 : for (i = 0; i < max_req_count; i++) {
876 0 : req = &rqpair->resources->reqs[i];
877 0 : if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) {
878 : /* nvmf_rdma_request_process checks qpair ibv and internal state
879 : * and completes a request */
880 0 : nvmf_rdma_request_process(rtransport, req);
881 0 : }
882 0 : }
883 0 : assert(rqpair->qd == 0);
884 0 : }
885 :
886 0 : if (rqpair->poller) {
887 0 : RB_REMOVE(qpairs_tree, &rqpair->poller->qpairs, rqpair);
888 :
889 0 : if (rqpair->srq != NULL && rqpair->resources != NULL) {
890 : /* Drop all received but unprocessed commands for this queue and return them to SRQ */
891 0 : STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) {
892 0 : if (rqpair == rdma_recv->qpair) {
893 0 : STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link);
894 0 : spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, &rdma_recv->wr);
895 0 : rc = spdk_rdma_provider_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr);
896 0 : if (rc) {
897 0 : SPDK_ERRLOG("Unable to re-post rx descriptor\n");
898 0 : }
899 0 : }
900 0 : }
901 0 : }
902 0 : }
903 :
904 0 : if (rqpair->cm_id) {
905 0 : if (rqpair->rdma_qp != NULL) {
906 0 : spdk_rdma_provider_qp_destroy(rqpair->rdma_qp);
907 0 : rqpair->rdma_qp = NULL;
908 0 : }
909 :
910 0 : if (rqpair->poller != NULL && rqpair->srq == NULL) {
911 0 : rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth);
912 0 : }
913 0 : }
914 :
915 0 : if (rqpair->srq == NULL && rqpair->resources != NULL) {
916 0 : nvmf_rdma_resources_destroy(rqpair->resources);
917 0 : }
918 :
919 0 : nvmf_rdma_qpair_clean_ibv_events(rqpair);
920 :
921 0 : if (rqpair->destruct_channel) {
922 0 : spdk_put_io_channel(rqpair->destruct_channel);
923 0 : rqpair->destruct_channel = NULL;
924 0 : }
925 :
926 0 : if (rqpair->poller && rqpair->poller->need_destroy && RB_EMPTY(&rqpair->poller->qpairs)) {
927 0 : nvmf_rdma_poller_destroy(rqpair->poller);
928 0 : }
929 :
930 : /* destroy cm_id last so cma device will not be freed before we destroy the cq. */
931 0 : if (rqpair->cm_id) {
932 0 : rdma_destroy_id(rqpair->cm_id);
933 0 : }
934 :
935 0 : free(rqpair);
936 0 : }
937 :
938 : static int
939 5 : nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device)
940 : {
941 5 : struct spdk_nvmf_rdma_poller *rpoller;
942 5 : int rc, num_cqe, required_num_wr;
943 :
944 : /* Enlarge CQ size dynamically */
945 5 : rpoller = rqpair->poller;
946 5 : required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth);
947 5 : num_cqe = rpoller->num_cqe;
948 5 : if (num_cqe < required_num_wr) {
949 4 : num_cqe = spdk_max(num_cqe * 2, required_num_wr);
950 4 : num_cqe = spdk_min(num_cqe, device->attr.max_cqe);
951 4 : }
952 :
953 5 : if (rpoller->num_cqe != num_cqe) {
954 4 : if (device->context->device->transport_type == IBV_TRANSPORT_IWARP) {
955 1 : SPDK_ERRLOG("iWARP doesn't support CQ resize. Current capacity %u, required %u\n"
956 : "Using CQ of insufficient size may lead to CQ overrun\n", rpoller->num_cqe, num_cqe);
957 1 : return -1;
958 : }
959 3 : if (required_num_wr > device->attr.max_cqe) {
960 1 : SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n",
961 : required_num_wr, device->attr.max_cqe);
962 1 : return -1;
963 : }
964 :
965 2 : SPDK_DEBUGLOG(rdma, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe);
966 2 : rc = ibv_resize_cq(rpoller->cq, num_cqe);
967 2 : if (rc) {
968 1 : SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno));
969 1 : return -1;
970 : }
971 :
972 1 : rpoller->num_cqe = num_cqe;
973 1 : }
974 :
975 2 : rpoller->required_num_wr = required_num_wr;
976 2 : return 0;
977 5 : }
978 :
979 : static int
980 0 : nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair)
981 : {
982 0 : struct spdk_nvmf_rdma_qpair *rqpair;
983 0 : struct spdk_nvmf_rdma_transport *rtransport;
984 0 : struct spdk_nvmf_transport *transport;
985 0 : struct spdk_nvmf_rdma_resource_opts opts;
986 0 : struct spdk_nvmf_rdma_device *device;
987 0 : struct spdk_rdma_provider_qp_init_attr qp_init_attr = {};
988 :
989 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
990 0 : device = rqpair->device;
991 :
992 0 : qp_init_attr.qp_context = rqpair;
993 0 : qp_init_attr.pd = device->pd;
994 0 : qp_init_attr.send_cq = rqpair->poller->cq;
995 0 : qp_init_attr.recv_cq = rqpair->poller->cq;
996 :
997 0 : if (rqpair->srq) {
998 0 : qp_init_attr.srq = rqpair->srq->srq;
999 0 : } else {
1000 0 : qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth;
1001 : }
1002 :
1003 : /* SEND, READ, and WRITE operations */
1004 0 : qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2;
1005 0 : qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE);
1006 0 : qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE);
1007 0 : qp_init_attr.stats = &rqpair->poller->stat.qp_stats;
1008 :
1009 0 : if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) {
1010 0 : SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n");
1011 0 : goto error;
1012 : }
1013 :
1014 0 : rqpair->rdma_qp = spdk_rdma_provider_qp_create(rqpair->cm_id, &qp_init_attr);
1015 0 : if (!rqpair->rdma_qp) {
1016 0 : goto error;
1017 : }
1018 :
1019 0 : rqpair->qp_num = rqpair->rdma_qp->qp->qp_num;
1020 :
1021 0 : rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2),
1022 : qp_init_attr.cap.max_send_wr);
1023 0 : rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge);
1024 0 : rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge);
1025 0 : spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair);
1026 0 : SPDK_DEBUGLOG(rdma, "New RDMA Connection: %p\n", qpair);
1027 :
1028 0 : if (rqpair->poller->srq == NULL) {
1029 0 : rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
1030 0 : transport = &rtransport->transport;
1031 :
1032 0 : opts.qp = rqpair->rdma_qp;
1033 0 : opts.map = device->map;
1034 0 : opts.qpair = rqpair;
1035 0 : opts.shared = false;
1036 0 : opts.max_queue_depth = rqpair->max_queue_depth;
1037 0 : opts.in_capsule_data_size = transport->opts.in_capsule_data_size;
1038 :
1039 0 : rqpair->resources = nvmf_rdma_resources_create(&opts);
1040 :
1041 0 : if (!rqpair->resources) {
1042 0 : SPDK_ERRLOG("Unable to allocate resources for receive queue.\n");
1043 0 : rdma_destroy_qp(rqpair->cm_id);
1044 0 : goto error;
1045 : }
1046 0 : } else {
1047 0 : rqpair->resources = rqpair->poller->resources;
1048 : }
1049 :
1050 0 : rqpair->current_recv_depth = 0;
1051 0 : STAILQ_INIT(&rqpair->pending_rdma_read_queue);
1052 0 : STAILQ_INIT(&rqpair->pending_rdma_write_queue);
1053 0 : STAILQ_INIT(&rqpair->pending_rdma_send_queue);
1054 0 : rqpair->qpair.queue_depth = 0;
1055 :
1056 0 : return 0;
1057 :
1058 : error:
1059 0 : rdma_destroy_id(rqpair->cm_id);
1060 0 : rqpair->cm_id = NULL;
1061 0 : return -1;
1062 0 : }
1063 :
1064 : /* Append the given recv wr structure to the resource structs outstanding recvs list. */
1065 : /* This function accepts either a single wr or the first wr in a linked list. */
1066 : static void
1067 6 : nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first)
1068 : {
1069 6 : struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
1070 : struct spdk_nvmf_rdma_transport, transport);
1071 :
1072 6 : if (rqpair->srq != NULL) {
1073 0 : spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, first);
1074 0 : } else {
1075 6 : if (spdk_rdma_provider_qp_queue_recv_wrs(rqpair->rdma_qp, first)) {
1076 6 : STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link);
1077 6 : }
1078 : }
1079 :
1080 6 : if (rtransport->rdma_opts.no_wr_batching) {
1081 0 : _poller_submit_recvs(rtransport, rqpair->poller);
1082 0 : }
1083 6 : }
1084 :
1085 : static inline void
1086 4 : request_transfer_in(struct spdk_nvmf_request *req)
1087 : {
1088 4 : struct spdk_nvmf_rdma_request *rdma_req;
1089 4 : struct spdk_nvmf_qpair *qpair;
1090 4 : struct spdk_nvmf_rdma_qpair *rqpair;
1091 4 : struct spdk_nvmf_rdma_transport *rtransport;
1092 :
1093 4 : qpair = req->qpair;
1094 4 : rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
1095 4 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
1096 4 : rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
1097 : struct spdk_nvmf_rdma_transport, transport);
1098 :
1099 4 : assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
1100 4 : assert(rdma_req != NULL);
1101 :
1102 4 : if (spdk_rdma_provider_qp_queue_send_wrs(rqpair->rdma_qp, rdma_req->transfer_wr)) {
1103 4 : STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link);
1104 4 : }
1105 4 : if (rtransport->rdma_opts.no_wr_batching) {
1106 0 : _poller_submit_sends(rtransport, rqpair->poller);
1107 0 : }
1108 :
1109 4 : assert(rqpair->current_read_depth + rdma_req->num_outstanding_data_wr <= rqpair->max_read_depth);
1110 4 : rqpair->current_read_depth += rdma_req->num_outstanding_data_wr;
1111 4 : assert(rqpair->current_send_depth + rdma_req->num_outstanding_data_wr <= rqpair->max_send_depth);
1112 4 : rqpair->current_send_depth += rdma_req->num_outstanding_data_wr;
1113 4 : }
1114 :
1115 : static inline void
1116 0 : nvmf_rdma_request_reset_transfer_in(struct spdk_nvmf_rdma_request *rdma_req,
1117 : struct spdk_nvmf_rdma_transport *rtransport)
1118 : {
1119 : /* Put completed WRs back to pool and move transfer_wr pointer */
1120 0 : _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool);
1121 0 : rdma_req->transfer_wr = rdma_req->remaining_tranfer_in_wrs;
1122 0 : rdma_req->remaining_tranfer_in_wrs = NULL;
1123 0 : rdma_req->num_outstanding_data_wr = rdma_req->num_remaining_data_wr;
1124 0 : rdma_req->num_remaining_data_wr = 0;
1125 0 : }
1126 :
1127 : static inline int
1128 0 : request_prepare_transfer_in_part(struct spdk_nvmf_request *req, uint32_t num_reads_available)
1129 : {
1130 0 : struct spdk_nvmf_rdma_request *rdma_req;
1131 0 : struct ibv_send_wr *wr;
1132 0 : uint32_t i;
1133 :
1134 0 : rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
1135 :
1136 0 : assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
1137 0 : assert(rdma_req != NULL);
1138 0 : assert(num_reads_available > 0);
1139 0 : assert(rdma_req->num_outstanding_data_wr > num_reads_available);
1140 0 : wr = rdma_req->transfer_wr;
1141 :
1142 0 : for (i = 0; i < num_reads_available - 1; i++) {
1143 0 : wr = wr->next;
1144 0 : }
1145 :
1146 0 : rdma_req->remaining_tranfer_in_wrs = wr->next;
1147 0 : rdma_req->num_remaining_data_wr = rdma_req->num_outstanding_data_wr - num_reads_available;
1148 0 : rdma_req->num_outstanding_data_wr = num_reads_available;
1149 : /* Break chain of WRs to send only part. Once this portion completes, we continue sending RDMA_READs */
1150 0 : wr->next = NULL;
1151 :
1152 0 : return 0;
1153 0 : }
1154 :
1155 : static int
1156 6 : request_transfer_out(struct spdk_nvmf_request *req, int *data_posted)
1157 : {
1158 6 : int num_outstanding_data_wr = 0;
1159 6 : struct spdk_nvmf_rdma_request *rdma_req;
1160 6 : struct spdk_nvmf_qpair *qpair;
1161 6 : struct spdk_nvmf_rdma_qpair *rqpair;
1162 6 : struct spdk_nvme_cpl *rsp;
1163 6 : struct ibv_send_wr *first = NULL;
1164 6 : struct spdk_nvmf_rdma_transport *rtransport;
1165 :
1166 6 : *data_posted = 0;
1167 6 : qpair = req->qpair;
1168 6 : rsp = &req->rsp->nvme_cpl;
1169 6 : rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
1170 6 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
1171 6 : rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
1172 : struct spdk_nvmf_rdma_transport, transport);
1173 :
1174 : /* Advance our sq_head pointer */
1175 6 : if (qpair->sq_head == qpair->sq_head_max) {
1176 6 : qpair->sq_head = 0;
1177 6 : } else {
1178 0 : qpair->sq_head++;
1179 : }
1180 6 : rsp->sqhd = qpair->sq_head;
1181 :
1182 : /* queue the capsule for the recv buffer */
1183 6 : assert(rdma_req->recv != NULL);
1184 :
1185 6 : nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr);
1186 :
1187 6 : rdma_req->recv = NULL;
1188 6 : assert(rqpair->current_recv_depth > 0);
1189 6 : rqpair->current_recv_depth--;
1190 :
1191 : /* Build the response which consists of optional
1192 : * RDMA WRITEs to transfer data, plus an RDMA SEND
1193 : * containing the response.
1194 : */
1195 6 : first = &rdma_req->rsp.wr;
1196 :
1197 6 : if (spdk_unlikely(rsp->status.sc != SPDK_NVME_SC_SUCCESS)) {
1198 : /* On failure, data was not read from the controller. So clear the
1199 : * number of outstanding data WRs to zero.
1200 : */
1201 1 : rdma_req->num_outstanding_data_wr = 0;
1202 6 : } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
1203 1 : first = rdma_req->transfer_wr;
1204 1 : *data_posted = 1;
1205 1 : num_outstanding_data_wr = rdma_req->num_outstanding_data_wr;
1206 1 : }
1207 6 : if (spdk_rdma_provider_qp_queue_send_wrs(rqpair->rdma_qp, first)) {
1208 6 : STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link);
1209 6 : }
1210 6 : if (rtransport->rdma_opts.no_wr_batching) {
1211 0 : _poller_submit_sends(rtransport, rqpair->poller);
1212 0 : }
1213 :
1214 : /* +1 for the rsp wr */
1215 6 : assert(rqpair->current_send_depth + num_outstanding_data_wr + 1 <= rqpair->max_send_depth);
1216 6 : rqpair->current_send_depth += num_outstanding_data_wr + 1;
1217 :
1218 6 : return 0;
1219 6 : }
1220 :
1221 : static int
1222 0 : nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair)
1223 : {
1224 0 : struct spdk_nvmf_rdma_accept_private_data accept_data;
1225 0 : struct rdma_conn_param ctrlr_event_data = {};
1226 0 : int rc;
1227 :
1228 0 : accept_data.recfmt = 0;
1229 0 : accept_data.crqsize = rqpair->max_queue_depth;
1230 :
1231 0 : ctrlr_event_data.private_data = &accept_data;
1232 0 : ctrlr_event_data.private_data_len = sizeof(accept_data);
1233 0 : if (id->ps == RDMA_PS_TCP) {
1234 0 : ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */
1235 0 : ctrlr_event_data.initiator_depth = rqpair->max_read_depth;
1236 0 : }
1237 :
1238 : /* Configure infinite retries for the initiator side qpair.
1239 : * We need to pass this value to the initiator to prevent the
1240 : * initiator side NIC from completing SEND requests back to the
1241 : * initiator with status rnr_retry_count_exceeded. */
1242 0 : ctrlr_event_data.rnr_retry_count = 0x7;
1243 :
1244 : /* When qpair is created without use of rdma cm API, an additional
1245 : * information must be provided to initiator in the connection response:
1246 : * whether qpair is using SRQ and its qp_num
1247 : * Fields below are ignored by rdma cm if qpair has been
1248 : * created using rdma cm API. */
1249 0 : ctrlr_event_data.srq = rqpair->srq ? 1 : 0;
1250 0 : ctrlr_event_data.qp_num = rqpair->qp_num;
1251 :
1252 0 : rc = spdk_rdma_provider_qp_accept(rqpair->rdma_qp, &ctrlr_event_data);
1253 0 : if (rc) {
1254 0 : SPDK_ERRLOG("Error %d on spdk_rdma_provider_qp_accept\n", errno);
1255 0 : } else {
1256 0 : SPDK_DEBUGLOG(rdma, "Sent back the accept\n");
1257 : }
1258 :
1259 0 : return rc;
1260 0 : }
1261 :
1262 : static void
1263 0 : nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error)
1264 : {
1265 0 : struct spdk_nvmf_rdma_reject_private_data rej_data;
1266 :
1267 0 : rej_data.recfmt = 0;
1268 0 : rej_data.sts = error;
1269 :
1270 0 : rdma_reject(id, &rej_data, sizeof(rej_data));
1271 0 : }
1272 :
1273 : static int
1274 0 : nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event)
1275 : {
1276 0 : struct spdk_nvmf_rdma_transport *rtransport;
1277 0 : struct spdk_nvmf_rdma_qpair *rqpair = NULL;
1278 0 : struct spdk_nvmf_rdma_port *port;
1279 0 : struct rdma_conn_param *rdma_param = NULL;
1280 0 : const struct spdk_nvmf_rdma_request_private_data *private_data = NULL;
1281 0 : uint16_t max_queue_depth;
1282 0 : uint16_t max_read_depth;
1283 :
1284 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
1285 :
1286 0 : assert(event->id != NULL); /* Impossible. Can't even reject the connection. */
1287 0 : assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */
1288 :
1289 0 : rdma_param = &event->param.conn;
1290 0 : if (rdma_param->private_data == NULL ||
1291 0 : rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) {
1292 0 : SPDK_ERRLOG("connect request: no private data provided\n");
1293 0 : nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH);
1294 0 : return -1;
1295 : }
1296 :
1297 0 : private_data = rdma_param->private_data;
1298 0 : if (private_data->recfmt != 0) {
1299 0 : SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n");
1300 0 : nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT);
1301 0 : return -1;
1302 : }
1303 :
1304 0 : SPDK_DEBUGLOG(rdma, "Connect Recv on fabric intf name %s, dev_name %s\n",
1305 : event->id->verbs->device->name, event->id->verbs->device->dev_name);
1306 :
1307 0 : port = event->listen_id->context;
1308 0 : SPDK_DEBUGLOG(rdma, "Listen Id was %p with verbs %p. ListenAddr: %p\n",
1309 : event->listen_id, event->listen_id->verbs, port);
1310 :
1311 : /* Figure out the supported queue depth. This is a multi-step process
1312 : * that takes into account hardware maximums, host provided values,
1313 : * and our target's internal memory limits */
1314 :
1315 0 : SPDK_DEBUGLOG(rdma, "Calculating Queue Depth\n");
1316 :
1317 : /* Start with the maximum queue depth allowed by the target */
1318 0 : max_queue_depth = rtransport->transport.opts.max_queue_depth;
1319 0 : max_read_depth = rtransport->transport.opts.max_queue_depth;
1320 0 : SPDK_DEBUGLOG(rdma, "Target Max Queue Depth: %d\n",
1321 : rtransport->transport.opts.max_queue_depth);
1322 :
1323 : /* Next check the local NIC's hardware limitations */
1324 0 : SPDK_DEBUGLOG(rdma,
1325 : "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n",
1326 : port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom);
1327 0 : max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr);
1328 0 : max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom);
1329 :
1330 : /* Next check the remote NIC's hardware limitations */
1331 0 : SPDK_DEBUGLOG(rdma,
1332 : "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n",
1333 : rdma_param->initiator_depth, rdma_param->responder_resources);
1334 : /* from man3 rdma_get_cm_event
1335 : * responder_resources - Specifies the number of responder resources that is requested by the recipient.
1336 : * The responder_resources field must match the initiator depth specified by the remote node when running
1337 : * the rdma_connect and rdma_accept functions. */
1338 0 : if (rdma_param->responder_resources != 0) {
1339 0 : if (private_data->qid) {
1340 0 : SPDK_DEBUGLOG(rdma, "Host (Initiator) is not allowed to use RDMA operations,"
1341 : " responder_resources must be 0 but set to %u\n",
1342 : rdma_param->responder_resources);
1343 0 : } else {
1344 0 : SPDK_WARNLOG("Host (Initiator) is not allowed to use RDMA operations,"
1345 : " responder_resources must be 0 but set to %u\n",
1346 : rdma_param->responder_resources);
1347 : }
1348 0 : }
1349 : /* from man3 rdma_get_cm_event
1350 : * initiator_depth - Specifies the maximum number of outstanding RDMA read operations that the recipient holds.
1351 : * The initiator_depth field must match the responder resources specified by the remote node when running
1352 : * the rdma_connect and rdma_accept functions. */
1353 0 : if (rdma_param->initiator_depth == 0) {
1354 0 : SPDK_ERRLOG("Host (Initiator) doesn't support RDMA_READ or atomic operations\n");
1355 0 : nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_IRD);
1356 0 : return -1;
1357 : }
1358 0 : max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth);
1359 :
1360 0 : SPDK_DEBUGLOG(rdma, "Host Receive Queue Size: %d\n", private_data->hrqsize);
1361 0 : SPDK_DEBUGLOG(rdma, "Host Send Queue Size: %d\n", private_data->hsqsize);
1362 0 : max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize);
1363 0 : max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1);
1364 :
1365 0 : SPDK_DEBUGLOG(rdma, "Final Negotiated Queue Depth: %d R/W Depth: %d\n",
1366 : max_queue_depth, max_read_depth);
1367 :
1368 0 : rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair));
1369 0 : if (rqpair == NULL) {
1370 0 : SPDK_ERRLOG("Could not allocate new connection.\n");
1371 0 : nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES);
1372 0 : return -1;
1373 : }
1374 :
1375 0 : rqpair->device = port->device;
1376 0 : rqpair->max_queue_depth = max_queue_depth;
1377 0 : rqpair->max_read_depth = max_read_depth;
1378 0 : rqpair->cm_id = event->id;
1379 0 : rqpair->listen_id = event->listen_id;
1380 0 : rqpair->qpair.transport = transport;
1381 : /* use qid from the private data to determine the qpair type
1382 : qid will be set to the appropriate value when the controller is created */
1383 0 : rqpair->qpair.qid = private_data->qid;
1384 0 : rqpair->qpair.numa.id_valid = 1;
1385 0 : rqpair->qpair.numa.id = spdk_rdma_cm_id_get_numa_id(rqpair->cm_id);
1386 :
1387 0 : event->id->context = &rqpair->qpair;
1388 :
1389 0 : spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair);
1390 :
1391 0 : return 0;
1392 0 : }
1393 :
1394 : static inline void
1395 28 : nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next,
1396 : enum spdk_nvme_data_transfer xfer)
1397 : {
1398 28 : if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
1399 24 : wr->opcode = IBV_WR_RDMA_WRITE;
1400 24 : wr->send_flags = 0;
1401 24 : wr->next = next;
1402 28 : } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) {
1403 4 : wr->opcode = IBV_WR_RDMA_READ;
1404 4 : wr->send_flags = IBV_SEND_SIGNALED;
1405 4 : wr->next = NULL;
1406 4 : } else {
1407 0 : assert(0);
1408 : }
1409 28 : }
1410 :
1411 : static int
1412 6 : nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport,
1413 : struct spdk_nvmf_rdma_request *rdma_req,
1414 : uint32_t num_sgl_descriptors)
1415 : {
1416 6 : struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES];
1417 6 : struct spdk_nvmf_rdma_request_data *current_data_wr;
1418 6 : uint32_t i;
1419 :
1420 6 : if (spdk_unlikely(num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES)) {
1421 0 : SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n",
1422 : num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES);
1423 0 : return -EINVAL;
1424 : }
1425 :
1426 6 : if (spdk_unlikely(spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests,
1427 : num_sgl_descriptors))) {
1428 0 : return -ENOMEM;
1429 : }
1430 :
1431 6 : current_data_wr = &rdma_req->data;
1432 :
1433 12 : for (i = 0; i < num_sgl_descriptors; i++) {
1434 6 : nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer);
1435 6 : current_data_wr->wr.next = &work_requests[i]->wr;
1436 6 : current_data_wr = work_requests[i];
1437 6 : current_data_wr->wr.sg_list = current_data_wr->sgl;
1438 6 : current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id;
1439 6 : }
1440 :
1441 6 : nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer);
1442 :
1443 6 : return 0;
1444 6 : }
1445 :
1446 : static inline void
1447 16 : nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req)
1448 : {
1449 16 : struct ibv_send_wr *wr = &rdma_req->data.wr;
1450 16 : struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1;
1451 :
1452 16 : wr->wr.rdma.rkey = sgl->keyed.key;
1453 16 : wr->wr.rdma.remote_addr = sgl->address;
1454 16 : nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer);
1455 16 : }
1456 :
1457 : static inline void
1458 1 : nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs)
1459 : {
1460 1 : struct ibv_send_wr *wr = &rdma_req->data.wr;
1461 1 : struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1;
1462 1 : uint32_t i;
1463 1 : int j;
1464 1 : uint64_t remote_addr_offset = 0;
1465 :
1466 3 : for (i = 0; i < num_wrs; ++i) {
1467 2 : wr->wr.rdma.rkey = sgl->keyed.key;
1468 2 : wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset;
1469 19 : for (j = 0; j < wr->num_sge; ++j) {
1470 17 : remote_addr_offset += wr->sg_list[j].length;
1471 17 : }
1472 2 : wr = wr->next;
1473 2 : }
1474 1 : }
1475 :
1476 : static int
1477 15 : nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_device *device,
1478 : struct spdk_nvmf_rdma_request *rdma_req,
1479 : struct ibv_send_wr *wr,
1480 : uint32_t total_length)
1481 : {
1482 15 : struct spdk_rdma_utils_memory_translation mem_translation;
1483 15 : struct ibv_sge *sg_ele;
1484 15 : struct iovec *iov;
1485 15 : uint32_t lkey, remaining;
1486 15 : int rc;
1487 :
1488 15 : wr->num_sge = 0;
1489 :
1490 74 : while (total_length && wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES) {
1491 59 : iov = &rdma_req->req.iov[rdma_req->iovpos];
1492 59 : rc = spdk_rdma_utils_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation);
1493 59 : if (spdk_unlikely(rc)) {
1494 0 : return rc;
1495 : }
1496 :
1497 59 : lkey = spdk_rdma_utils_memory_translation_get_lkey(&mem_translation);
1498 59 : sg_ele = &wr->sg_list[wr->num_sge];
1499 59 : remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length);
1500 :
1501 59 : sg_ele->lkey = lkey;
1502 59 : sg_ele->addr = (uintptr_t)iov->iov_base + rdma_req->offset;
1503 59 : sg_ele->length = remaining;
1504 59 : SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, sg_ele->addr,
1505 : sg_ele->length);
1506 59 : rdma_req->offset += sg_ele->length;
1507 59 : total_length -= sg_ele->length;
1508 59 : wr->num_sge++;
1509 :
1510 59 : if (rdma_req->offset == iov->iov_len) {
1511 57 : rdma_req->offset = 0;
1512 57 : rdma_req->iovpos++;
1513 57 : }
1514 : }
1515 :
1516 15 : if (spdk_unlikely(total_length)) {
1517 0 : SPDK_ERRLOG("Not enough SG entries to hold data buffer\n");
1518 0 : return -EINVAL;
1519 : }
1520 :
1521 15 : return 0;
1522 15 : }
1523 :
1524 : static int
1525 10 : nvmf_rdma_fill_wr_sgl_with_dif(struct spdk_nvmf_rdma_device *device,
1526 : struct spdk_nvmf_rdma_request *rdma_req,
1527 : struct ibv_send_wr *wr,
1528 : uint32_t total_length,
1529 : uint32_t num_extra_wrs)
1530 : {
1531 10 : struct spdk_rdma_utils_memory_translation mem_translation;
1532 10 : struct spdk_dif_ctx *dif_ctx = &rdma_req->req.dif.dif_ctx;
1533 10 : struct ibv_sge *sg_ele;
1534 10 : struct iovec *iov;
1535 10 : struct iovec *rdma_iov;
1536 10 : uint32_t lkey, remaining;
1537 10 : uint32_t remaining_data_block, data_block_size, md_size;
1538 10 : uint32_t sge_len;
1539 10 : int rc;
1540 :
1541 10 : data_block_size = dif_ctx->block_size - dif_ctx->md_size;
1542 :
1543 10 : if (spdk_likely(!rdma_req->req.stripped_data)) {
1544 5 : rdma_iov = rdma_req->req.iov;
1545 5 : remaining_data_block = data_block_size;
1546 5 : md_size = dif_ctx->md_size;
1547 5 : } else {
1548 5 : rdma_iov = rdma_req->req.stripped_data->iov;
1549 5 : total_length = total_length / dif_ctx->block_size * data_block_size;
1550 5 : remaining_data_block = total_length;
1551 5 : md_size = 0;
1552 : }
1553 :
1554 10 : wr->num_sge = 0;
1555 :
1556 40 : while (total_length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) {
1557 15 : iov = rdma_iov + rdma_req->iovpos;
1558 15 : rc = spdk_rdma_utils_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation);
1559 15 : if (spdk_unlikely(rc)) {
1560 0 : return rc;
1561 : }
1562 :
1563 15 : lkey = spdk_rdma_utils_memory_translation_get_lkey(&mem_translation);
1564 15 : sg_ele = &wr->sg_list[wr->num_sge];
1565 15 : remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length);
1566 :
1567 53 : while (remaining) {
1568 38 : if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) {
1569 1 : if (num_extra_wrs > 0 && wr->next) {
1570 1 : wr = wr->next;
1571 1 : wr->num_sge = 0;
1572 1 : sg_ele = &wr->sg_list[wr->num_sge];
1573 1 : num_extra_wrs--;
1574 1 : } else {
1575 0 : break;
1576 : }
1577 1 : }
1578 38 : sg_ele->lkey = lkey;
1579 38 : sg_ele->addr = (uintptr_t)((char *)iov->iov_base + rdma_req->offset);
1580 38 : sge_len = spdk_min(remaining, remaining_data_block);
1581 38 : sg_ele->length = sge_len;
1582 38 : SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele,
1583 : sg_ele->addr, sg_ele->length);
1584 38 : remaining -= sge_len;
1585 38 : remaining_data_block -= sge_len;
1586 38 : rdma_req->offset += sge_len;
1587 38 : total_length -= sge_len;
1588 :
1589 38 : sg_ele++;
1590 38 : wr->num_sge++;
1591 :
1592 38 : if (remaining_data_block == 0) {
1593 : /* skip metadata */
1594 34 : rdma_req->offset += md_size;
1595 34 : total_length -= md_size;
1596 : /* Metadata that do not fit this IO buffer will be included in the next IO buffer */
1597 34 : remaining -= spdk_min(remaining, md_size);
1598 34 : remaining_data_block = data_block_size;
1599 34 : }
1600 :
1601 38 : if (remaining == 0) {
1602 : /* By subtracting the size of the last IOV from the offset, we ensure that we skip
1603 : the remaining metadata bits at the beginning of the next buffer */
1604 15 : rdma_req->offset -= spdk_min(iov->iov_len, rdma_req->offset);
1605 15 : rdma_req->iovpos++;
1606 15 : }
1607 : }
1608 : }
1609 :
1610 10 : if (spdk_unlikely(total_length)) {
1611 0 : SPDK_ERRLOG("Not enough SG entries to hold data buffer\n");
1612 0 : return -EINVAL;
1613 : }
1614 :
1615 10 : return 0;
1616 10 : }
1617 :
1618 : static inline uint32_t
1619 8 : nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size)
1620 : {
1621 : /* estimate the number of SG entries and WRs needed to process the request */
1622 8 : uint32_t num_sge = 0;
1623 8 : uint32_t i;
1624 8 : uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size);
1625 :
1626 23 : for (i = 0; i < num_buffers && length > 0; i++) {
1627 15 : uint32_t buffer_len = spdk_min(length, io_unit_size);
1628 15 : uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size);
1629 :
1630 15 : if (num_sge_in_block * block_size > buffer_len) {
1631 11 : ++num_sge_in_block;
1632 11 : }
1633 15 : num_sge += num_sge_in_block;
1634 15 : length -= buffer_len;
1635 15 : }
1636 8 : return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES);
1637 8 : }
1638 :
1639 : static int
1640 16 : nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport,
1641 : struct spdk_nvmf_rdma_device *device,
1642 : struct spdk_nvmf_rdma_request *rdma_req)
1643 : {
1644 16 : struct spdk_nvmf_rdma_qpair *rqpair;
1645 16 : struct spdk_nvmf_rdma_poll_group *rgroup;
1646 16 : struct spdk_nvmf_request *req = &rdma_req->req;
1647 16 : struct ibv_send_wr *wr = &rdma_req->data.wr;
1648 16 : int rc;
1649 16 : uint32_t num_wrs = 1;
1650 16 : uint32_t length;
1651 :
1652 16 : rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair);
1653 16 : rgroup = rqpair->poller->group;
1654 :
1655 : /* rdma wr specifics */
1656 16 : nvmf_rdma_setup_request(rdma_req);
1657 :
1658 16 : length = req->length;
1659 16 : if (spdk_unlikely(req->dif_enabled)) {
1660 8 : req->dif.orig_length = length;
1661 8 : length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx);
1662 8 : req->dif.elba_length = length;
1663 8 : }
1664 :
1665 16 : rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport,
1666 16 : length);
1667 16 : if (spdk_unlikely(rc != 0)) {
1668 1 : return rc;
1669 : }
1670 :
1671 15 : assert(req->iovcnt <= rqpair->max_send_sge);
1672 :
1673 : /* When dif_insert_or_strip is true and the I/O data length is greater than one block,
1674 : * the stripped_buffers are got for DIF stripping. */
1675 15 : if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST)
1676 : && (req->dif.elba_length > req->dif.dif_ctx.block_size))) {
1677 14 : rc = nvmf_request_get_stripped_buffers(req, &rgroup->group,
1678 7 : &rtransport->transport, req->dif.orig_length);
1679 7 : if (rc != 0) {
1680 4 : SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc);
1681 4 : }
1682 7 : }
1683 :
1684 15 : rdma_req->iovpos = 0;
1685 :
1686 15 : if (spdk_unlikely(req->dif_enabled)) {
1687 16 : num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size,
1688 8 : req->dif.dif_ctx.block_size);
1689 8 : if (num_wrs > 1) {
1690 1 : rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1);
1691 1 : if (spdk_unlikely(rc != 0)) {
1692 0 : goto err_exit;
1693 : }
1694 1 : }
1695 :
1696 8 : rc = nvmf_rdma_fill_wr_sgl_with_dif(device, rdma_req, wr, length, num_wrs - 1);
1697 8 : if (spdk_unlikely(rc != 0)) {
1698 0 : goto err_exit;
1699 : }
1700 :
1701 8 : if (num_wrs > 1) {
1702 1 : nvmf_rdma_update_remote_addr(rdma_req, num_wrs);
1703 1 : }
1704 8 : } else {
1705 7 : rc = nvmf_rdma_fill_wr_sgl(device, rdma_req, wr, length);
1706 7 : if (spdk_unlikely(rc != 0)) {
1707 0 : goto err_exit;
1708 : }
1709 : }
1710 :
1711 : /* set the number of outstanding data WRs for this request. */
1712 15 : rdma_req->num_outstanding_data_wr = num_wrs;
1713 :
1714 15 : return rc;
1715 :
1716 : err_exit:
1717 0 : spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport);
1718 0 : nvmf_rdma_request_free_data(rdma_req, rtransport);
1719 0 : req->iovcnt = 0;
1720 0 : return rc;
1721 16 : }
1722 :
1723 : static int
1724 5 : nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport,
1725 : struct spdk_nvmf_rdma_device *device,
1726 : struct spdk_nvmf_rdma_request *rdma_req)
1727 : {
1728 5 : struct spdk_nvmf_rdma_qpair *rqpair;
1729 5 : struct spdk_nvmf_rdma_poll_group *rgroup;
1730 5 : struct ibv_send_wr *current_wr;
1731 5 : struct spdk_nvmf_request *req = &rdma_req->req;
1732 5 : struct spdk_nvme_sgl_descriptor *inline_segment, *desc;
1733 5 : uint32_t num_sgl_descriptors;
1734 5 : uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES], total_length = 0;
1735 5 : uint32_t i;
1736 5 : int rc;
1737 :
1738 5 : rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
1739 5 : rgroup = rqpair->poller->group;
1740 :
1741 5 : inline_segment = &req->cmd->nvme_cmd.dptr.sgl1;
1742 5 : assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT);
1743 5 : assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET);
1744 :
1745 5 : num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor);
1746 5 : assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES);
1747 :
1748 5 : desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address;
1749 15 : for (i = 0; i < num_sgl_descriptors; i++) {
1750 10 : if (spdk_likely(!req->dif_enabled)) {
1751 8 : lengths[i] = desc->keyed.length;
1752 8 : } else {
1753 2 : req->dif.orig_length += desc->keyed.length;
1754 2 : lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx);
1755 2 : req->dif.elba_length += lengths[i];
1756 : }
1757 10 : total_length += lengths[i];
1758 10 : desc++;
1759 10 : }
1760 :
1761 5 : if (spdk_unlikely(total_length > rtransport->transport.opts.max_io_size)) {
1762 0 : SPDK_ERRLOG("Multi SGL length 0x%x exceeds max io size 0x%x\n",
1763 : total_length, rtransport->transport.opts.max_io_size);
1764 0 : req->rsp->nvme_cpl.status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
1765 0 : return -EINVAL;
1766 : }
1767 :
1768 5 : rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1);
1769 5 : if (spdk_unlikely(rc != 0)) {
1770 0 : return -ENOMEM;
1771 : }
1772 :
1773 5 : rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, total_length);
1774 5 : if (spdk_unlikely(rc != 0)) {
1775 0 : nvmf_rdma_request_free_data(rdma_req, rtransport);
1776 0 : return rc;
1777 : }
1778 :
1779 : /* When dif_insert_or_strip is true and the I/O data length is greater than one block,
1780 : * the stripped_buffers are got for DIF stripping. */
1781 5 : if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST)
1782 : && (req->dif.elba_length > req->dif.dif_ctx.block_size))) {
1783 2 : rc = nvmf_request_get_stripped_buffers(req, &rgroup->group,
1784 1 : &rtransport->transport, req->dif.orig_length);
1785 1 : if (spdk_unlikely(rc != 0)) {
1786 0 : SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc);
1787 0 : }
1788 1 : }
1789 :
1790 : /* The first WR must always be the embedded data WR. This is how we unwind them later. */
1791 5 : current_wr = &rdma_req->data.wr;
1792 5 : assert(current_wr != NULL);
1793 :
1794 5 : req->length = 0;
1795 5 : rdma_req->iovpos = 0;
1796 5 : desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address;
1797 15 : for (i = 0; i < num_sgl_descriptors; i++) {
1798 : /* The descriptors must be keyed data block descriptors with an address, not an offset. */
1799 10 : if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK ||
1800 : desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) {
1801 0 : rc = -EINVAL;
1802 0 : goto err_exit;
1803 : }
1804 :
1805 10 : if (spdk_likely(!req->dif_enabled)) {
1806 8 : rc = nvmf_rdma_fill_wr_sgl(device, rdma_req, current_wr, lengths[i]);
1807 8 : } else {
1808 4 : rc = nvmf_rdma_fill_wr_sgl_with_dif(device, rdma_req, current_wr,
1809 2 : lengths[i], 0);
1810 : }
1811 10 : if (spdk_unlikely(rc != 0)) {
1812 0 : rc = -ENOMEM;
1813 0 : goto err_exit;
1814 : }
1815 :
1816 10 : req->length += desc->keyed.length;
1817 10 : current_wr->wr.rdma.rkey = desc->keyed.key;
1818 10 : current_wr->wr.rdma.remote_addr = desc->address;
1819 10 : current_wr = current_wr->next;
1820 10 : desc++;
1821 10 : }
1822 :
1823 : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
1824 : /* Go back to the last descriptor in the list. */
1825 5 : desc--;
1826 5 : if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) {
1827 0 : if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) {
1828 0 : rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV;
1829 0 : rdma_req->rsp.wr.imm_data = desc->keyed.key;
1830 0 : }
1831 0 : }
1832 : #endif
1833 :
1834 5 : rdma_req->num_outstanding_data_wr = num_sgl_descriptors;
1835 :
1836 5 : return 0;
1837 :
1838 : err_exit:
1839 0 : spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport);
1840 0 : nvmf_rdma_request_free_data(rdma_req, rtransport);
1841 0 : return rc;
1842 5 : }
1843 :
1844 : static int
1845 25 : nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport,
1846 : struct spdk_nvmf_rdma_device *device,
1847 : struct spdk_nvmf_rdma_request *rdma_req)
1848 : {
1849 25 : struct spdk_nvmf_request *req = &rdma_req->req;
1850 25 : struct spdk_nvme_cpl *rsp;
1851 25 : struct spdk_nvme_sgl_descriptor *sgl;
1852 25 : int rc;
1853 25 : uint32_t length;
1854 :
1855 25 : rsp = &req->rsp->nvme_cpl;
1856 25 : sgl = &req->cmd->nvme_cmd.dptr.sgl1;
1857 :
1858 25 : if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK &&
1859 17 : (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS ||
1860 0 : sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) {
1861 :
1862 17 : length = sgl->keyed.length;
1863 17 : if (spdk_unlikely(length > rtransport->transport.opts.max_io_size)) {
1864 1 : SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n",
1865 : length, rtransport->transport.opts.max_io_size);
1866 1 : rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
1867 1 : return -1;
1868 : }
1869 : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
1870 16 : if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) {
1871 0 : if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) {
1872 0 : rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV;
1873 0 : rdma_req->rsp.wr.imm_data = sgl->keyed.key;
1874 0 : }
1875 0 : }
1876 : #endif
1877 :
1878 : /* fill request length and populate iovs */
1879 16 : req->length = length;
1880 :
1881 16 : rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req);
1882 16 : if (spdk_unlikely(rc < 0)) {
1883 1 : if (rc == -EINVAL) {
1884 0 : SPDK_ERRLOG("SGL length exceeds the max I/O size\n");
1885 0 : rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
1886 0 : return -1;
1887 : }
1888 : /* No available buffers. Queue this request up. */
1889 1 : SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req);
1890 1 : return 0;
1891 : }
1892 :
1893 15 : SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req,
1894 : req->iovcnt);
1895 :
1896 15 : return 0;
1897 11 : } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK &&
1898 3 : sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) {
1899 3 : uint64_t offset = sgl->address;
1900 3 : uint32_t max_len = rtransport->transport.opts.in_capsule_data_size;
1901 :
1902 3 : SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n",
1903 : offset, sgl->unkeyed.length);
1904 :
1905 3 : if (spdk_unlikely(offset > max_len)) {
1906 0 : SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n",
1907 : offset, max_len);
1908 0 : rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET;
1909 0 : return -1;
1910 : }
1911 3 : max_len -= (uint32_t)offset;
1912 :
1913 3 : if (spdk_unlikely(sgl->unkeyed.length > max_len)) {
1914 2 : SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n",
1915 : sgl->unkeyed.length, max_len);
1916 2 : rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
1917 2 : return -1;
1918 : }
1919 :
1920 1 : rdma_req->num_outstanding_data_wr = 0;
1921 1 : req->data_from_pool = false;
1922 1 : req->length = sgl->unkeyed.length;
1923 :
1924 1 : req->iov[0].iov_base = rdma_req->recv->buf + offset;
1925 1 : req->iov[0].iov_len = req->length;
1926 1 : req->iovcnt = 1;
1927 :
1928 1 : return 0;
1929 13 : } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT &&
1930 5 : sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) {
1931 :
1932 5 : rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req);
1933 5 : if (spdk_unlikely(rc == -ENOMEM)) {
1934 0 : SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req);
1935 0 : return 0;
1936 5 : } else if (spdk_unlikely(rc == -EINVAL)) {
1937 0 : SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n");
1938 0 : rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
1939 0 : return -1;
1940 : }
1941 :
1942 5 : SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req,
1943 : req->iovcnt);
1944 :
1945 5 : return 0;
1946 : }
1947 :
1948 0 : SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n",
1949 : sgl->generic.type, sgl->generic.subtype);
1950 0 : rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID;
1951 0 : return -1;
1952 25 : }
1953 :
1954 : static void
1955 6 : _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req,
1956 : struct spdk_nvmf_rdma_transport *rtransport)
1957 : {
1958 6 : struct spdk_nvmf_rdma_qpair *rqpair;
1959 6 : struct spdk_nvmf_rdma_poll_group *rgroup;
1960 :
1961 6 : rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
1962 6 : if (rdma_req->req.data_from_pool) {
1963 5 : rgroup = rqpair->poller->group;
1964 :
1965 5 : spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport);
1966 5 : }
1967 6 : if (rdma_req->req.stripped_data) {
1968 0 : nvmf_request_free_stripped_buffers(&rdma_req->req,
1969 0 : &rqpair->poller->group->group,
1970 0 : &rtransport->transport);
1971 0 : }
1972 6 : nvmf_rdma_request_free_data(rdma_req, rtransport);
1973 6 : rdma_req->req.length = 0;
1974 6 : rdma_req->req.iovcnt = 0;
1975 6 : rdma_req->offset = 0;
1976 6 : rdma_req->req.dif_enabled = false;
1977 6 : rdma_req->fused_failed = false;
1978 6 : rdma_req->transfer_wr = NULL;
1979 6 : if (rdma_req->fused_pair) {
1980 : /* This req was part of a valid fused pair, but failed before it got to
1981 : * READ_TO_EXECUTE state. This means we need to fail the other request
1982 : * in the pair, because it is no longer part of a valid pair. If the pair
1983 : * already reached READY_TO_EXECUTE state, we need to kick it.
1984 : */
1985 0 : rdma_req->fused_pair->fused_failed = true;
1986 0 : if (rdma_req->fused_pair->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) {
1987 0 : nvmf_rdma_request_process(rtransport, rdma_req->fused_pair);
1988 0 : }
1989 0 : rdma_req->fused_pair = NULL;
1990 0 : }
1991 6 : memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif));
1992 6 : rqpair->qd--;
1993 :
1994 6 : STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link);
1995 6 : rqpair->qpair.queue_depth--;
1996 6 : rdma_req->state = RDMA_REQUEST_STATE_FREE;
1997 6 : }
1998 :
1999 : static void
2000 6 : nvmf_rdma_check_fused_ordering(struct spdk_nvmf_rdma_transport *rtransport,
2001 : struct spdk_nvmf_rdma_qpair *rqpair,
2002 : struct spdk_nvmf_rdma_request *rdma_req)
2003 : {
2004 6 : enum spdk_nvme_cmd_fuse last, next;
2005 :
2006 6 : last = rqpair->fused_first ? rqpair->fused_first->req.cmd->nvme_cmd.fuse : SPDK_NVME_CMD_FUSE_NONE;
2007 6 : next = rdma_req->req.cmd->nvme_cmd.fuse;
2008 :
2009 6 : assert(last != SPDK_NVME_CMD_FUSE_SECOND);
2010 :
2011 6 : if (spdk_likely(last == SPDK_NVME_CMD_FUSE_NONE && next == SPDK_NVME_CMD_FUSE_NONE)) {
2012 6 : return;
2013 : }
2014 :
2015 0 : if (last == SPDK_NVME_CMD_FUSE_FIRST) {
2016 0 : if (next == SPDK_NVME_CMD_FUSE_SECOND) {
2017 : /* This is a valid pair of fused commands. Point them at each other
2018 : * so they can be submitted consecutively once ready to be executed.
2019 : */
2020 0 : rqpair->fused_first->fused_pair = rdma_req;
2021 0 : rdma_req->fused_pair = rqpair->fused_first;
2022 0 : rqpair->fused_first = NULL;
2023 0 : return;
2024 : } else {
2025 : /* Mark the last req as failed since it wasn't followed by a SECOND. */
2026 0 : rqpair->fused_first->fused_failed = true;
2027 :
2028 : /* If the last req is in READY_TO_EXECUTE state, then call
2029 : * nvmf_rdma_request_process(), otherwise nothing else will kick it.
2030 : */
2031 0 : if (rqpair->fused_first->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) {
2032 0 : nvmf_rdma_request_process(rtransport, rqpair->fused_first);
2033 0 : }
2034 :
2035 0 : rqpair->fused_first = NULL;
2036 : }
2037 0 : }
2038 :
2039 0 : if (next == SPDK_NVME_CMD_FUSE_FIRST) {
2040 : /* Set rqpair->fused_first here so that we know to check that the next request
2041 : * is a SECOND (and to fail this one if it isn't).
2042 : */
2043 0 : rqpair->fused_first = rdma_req;
2044 0 : } else if (next == SPDK_NVME_CMD_FUSE_SECOND) {
2045 : /* Mark this req failed since it ia SECOND and the last one was not a FIRST. */
2046 0 : rdma_req->fused_failed = true;
2047 0 : }
2048 6 : }
2049 :
2050 : static void
2051 5 : nvmf_rdma_poll_group_insert_need_buffer_req(struct spdk_nvmf_rdma_poll_group *rgroup,
2052 : struct spdk_nvmf_rdma_request *rdma_req)
2053 : {
2054 5 : struct spdk_nvmf_request *r;
2055 :
2056 : /* CONNECT commands have a timeout, so we need to avoid a CONNECT command
2057 : * from getting buried behind a long list of other non-FABRIC requests
2058 : * waiting for a buffer. Note that even though the CONNECT command's data is
2059 : * in-capsule, the request still goes to this STAILQ.
2060 : */
2061 5 : if (spdk_likely(rdma_req->req.cmd->nvme_cmd.opc != SPDK_NVME_OPC_FABRIC)) {
2062 : /* This is the most likely case. */
2063 5 : STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link);
2064 5 : return;
2065 : } else {
2066 : /* STAILQ doesn't have INSERT_BEFORE, so we need to either INSERT_HEAD
2067 : * or INSERT_AFTER. Put it after any other FABRIC commands that are
2068 : * already in the queue.
2069 : */
2070 0 : r = STAILQ_FIRST(&rgroup->group.pending_buf_queue);
2071 0 : if (r == NULL || r->cmd->nvme_cmd.opc != SPDK_NVME_OPC_FABRIC) {
2072 0 : STAILQ_INSERT_HEAD(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link);
2073 0 : return;
2074 : }
2075 0 : while (true) {
2076 0 : struct spdk_nvmf_request *next;
2077 :
2078 0 : next = STAILQ_NEXT(r, buf_link);
2079 0 : if (next == NULL || next->cmd->nvme_cmd.opc != SPDK_NVME_OPC_FABRIC) {
2080 0 : STAILQ_INSERT_AFTER(&rgroup->group.pending_buf_queue, r, &rdma_req->req, buf_link);
2081 0 : return;
2082 : }
2083 0 : r = next;
2084 0 : }
2085 : }
2086 5 : }
2087 :
2088 : bool
2089 23 : nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport,
2090 : struct spdk_nvmf_rdma_request *rdma_req)
2091 : {
2092 23 : struct spdk_nvmf_rdma_qpair *rqpair;
2093 23 : struct spdk_nvmf_rdma_device *device;
2094 23 : struct spdk_nvmf_rdma_poll_group *rgroup;
2095 23 : struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl;
2096 23 : int rc;
2097 23 : struct spdk_nvmf_rdma_recv *rdma_recv;
2098 23 : enum spdk_nvmf_rdma_request_state prev_state;
2099 23 : bool progress = false;
2100 23 : int data_posted;
2101 23 : uint32_t num_blocks, num_rdma_reads_available, qdepth;
2102 :
2103 23 : rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
2104 23 : device = rqpair->device;
2105 23 : rgroup = rqpair->poller->group;
2106 :
2107 23 : assert(rdma_req->state != RDMA_REQUEST_STATE_FREE);
2108 :
2109 : /* If the queue pair is in an error state, force the request to the completed state
2110 : * to release resources. */
2111 23 : if (spdk_unlikely(rqpair->ibv_in_error_state || !spdk_nvmf_qpair_is_active(&rqpair->qpair))) {
2112 0 : switch (rdma_req->state) {
2113 : case RDMA_REQUEST_STATE_NEED_BUFFER:
2114 0 : STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link);
2115 0 : break;
2116 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING:
2117 0 : STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
2118 0 : break;
2119 : case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
2120 0 : if (rdma_req->num_remaining_data_wr) {
2121 : /* Partially sent request is still in the pending_rdma_read_queue,
2122 : * remove it before completing */
2123 0 : rdma_req->num_remaining_data_wr = 0;
2124 0 : STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
2125 0 : }
2126 0 : break;
2127 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING:
2128 0 : STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
2129 0 : break;
2130 : case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING:
2131 0 : STAILQ_REMOVE(&rqpair->pending_rdma_send_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
2132 0 : break;
2133 : default:
2134 0 : break;
2135 : }
2136 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
2137 0 : }
2138 :
2139 : /* The loop here is to allow for several back-to-back state changes. */
2140 23 : do {
2141 66 : prev_state = rdma_req->state;
2142 :
2143 66 : SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state);
2144 :
2145 66 : switch (rdma_req->state) {
2146 : case RDMA_REQUEST_STATE_FREE:
2147 : /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW
2148 : * to escape this state. */
2149 6 : break;
2150 : case RDMA_REQUEST_STATE_NEW:
2151 6 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0,
2152 : (uintptr_t)rdma_req, (uintptr_t)rqpair, rqpair->qpair.queue_depth);
2153 6 : rdma_recv = rdma_req->recv;
2154 :
2155 : /* The first element of the SGL is the NVMe command */
2156 6 : rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr;
2157 6 : memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp));
2158 6 : rdma_req->transfer_wr = &rdma_req->data.wr;
2159 :
2160 6 : if (spdk_unlikely(rqpair->ibv_in_error_state || !spdk_nvmf_qpair_is_active(&rqpair->qpair))) {
2161 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
2162 0 : break;
2163 : }
2164 :
2165 6 : if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) {
2166 0 : rdma_req->req.dif_enabled = true;
2167 0 : }
2168 :
2169 6 : nvmf_rdma_check_fused_ordering(rtransport, rqpair, rdma_req);
2170 :
2171 : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
2172 6 : rdma_req->rsp.wr.opcode = IBV_WR_SEND;
2173 6 : rdma_req->rsp.wr.imm_data = 0;
2174 : #endif
2175 :
2176 : /* The next state transition depends on the data transfer needs of this request. */
2177 6 : rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req);
2178 :
2179 6 : if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) {
2180 1 : rsp->status.sct = SPDK_NVME_SCT_GENERIC;
2181 1 : rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE;
2182 1 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
2183 1 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2184 1 : SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req);
2185 1 : break;
2186 : }
2187 :
2188 : /* If no data to transfer, ready to execute. */
2189 5 : if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) {
2190 0 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
2191 0 : break;
2192 : }
2193 :
2194 5 : rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER;
2195 5 : nvmf_rdma_poll_group_insert_need_buffer_req(rgroup, rdma_req);
2196 5 : break;
2197 : case RDMA_REQUEST_STATE_NEED_BUFFER:
2198 5 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0,
2199 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2200 :
2201 5 : assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE);
2202 :
2203 5 : if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) {
2204 : /* This request needs to wait in line to obtain a buffer */
2205 0 : break;
2206 : }
2207 :
2208 : /* Try to get a data buffer */
2209 5 : rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req);
2210 5 : if (spdk_unlikely(rc < 0)) {
2211 0 : STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link);
2212 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
2213 0 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2214 0 : break;
2215 : }
2216 :
2217 5 : if (rdma_req->req.iovcnt == 0) {
2218 : /* No buffers available. */
2219 0 : rgroup->stat.pending_data_buffer++;
2220 0 : break;
2221 : }
2222 :
2223 5 : STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link);
2224 :
2225 : /* If data is transferring from host to controller and the data didn't
2226 : * arrive using in capsule data, we need to do a transfer from the host.
2227 : */
2228 9 : if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER &&
2229 4 : rdma_req->req.data_from_pool) {
2230 4 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link);
2231 4 : rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING;
2232 4 : break;
2233 : }
2234 :
2235 1 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
2236 1 : break;
2237 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING:
2238 4 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0,
2239 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2240 :
2241 4 : if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) {
2242 : /* This request needs to wait in line to perform RDMA */
2243 0 : break;
2244 : }
2245 4 : assert(rqpair->max_send_depth >= rqpair->current_send_depth);
2246 4 : qdepth = rqpair->max_send_depth - rqpair->current_send_depth;
2247 4 : assert(rqpair->max_read_depth >= rqpair->current_read_depth);
2248 4 : num_rdma_reads_available = rqpair->max_read_depth - rqpair->current_read_depth;
2249 8 : if (rdma_req->num_outstanding_data_wr > qdepth ||
2250 4 : rdma_req->num_outstanding_data_wr > num_rdma_reads_available) {
2251 0 : if (num_rdma_reads_available && qdepth) {
2252 : /* Send as much as we can */
2253 0 : request_prepare_transfer_in_part(&rdma_req->req, spdk_min(num_rdma_reads_available, qdepth));
2254 0 : } else {
2255 : /* We can only have so many WRs outstanding. we have to wait until some finish. */
2256 0 : rqpair->poller->stat.pending_rdma_read++;
2257 0 : break;
2258 : }
2259 0 : }
2260 :
2261 : /* We have already verified that this request is the head of the queue. */
2262 4 : if (rdma_req->num_remaining_data_wr == 0) {
2263 4 : STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link);
2264 4 : }
2265 :
2266 4 : request_transfer_in(&rdma_req->req);
2267 4 : rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER;
2268 :
2269 4 : break;
2270 : case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
2271 4 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0,
2272 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2273 : /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE
2274 : * to escape this state. */
2275 4 : break;
2276 : case RDMA_REQUEST_STATE_READY_TO_EXECUTE:
2277 5 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0,
2278 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2279 :
2280 5 : if (spdk_unlikely(rdma_req->req.dif_enabled)) {
2281 0 : if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) {
2282 : /* generate DIF for write operation */
2283 0 : num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size);
2284 0 : assert(num_blocks > 0);
2285 :
2286 0 : rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt,
2287 0 : num_blocks, &rdma_req->req.dif.dif_ctx);
2288 0 : if (rc != 0) {
2289 0 : SPDK_ERRLOG("DIF generation failed\n");
2290 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
2291 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
2292 0 : break;
2293 : }
2294 0 : }
2295 :
2296 0 : assert(rdma_req->req.dif.elba_length >= rdma_req->req.length);
2297 : /* set extended length before IO operation */
2298 0 : rdma_req->req.length = rdma_req->req.dif.elba_length;
2299 0 : }
2300 :
2301 5 : if (rdma_req->req.cmd->nvme_cmd.fuse != SPDK_NVME_CMD_FUSE_NONE) {
2302 0 : if (rdma_req->fused_failed) {
2303 : /* This request failed FUSED semantics. Fail it immediately, without
2304 : * even sending it to the target layer.
2305 : */
2306 0 : rsp->status.sct = SPDK_NVME_SCT_GENERIC;
2307 0 : rsp->status.sc = SPDK_NVME_SC_ABORTED_MISSING_FUSED;
2308 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
2309 0 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2310 0 : break;
2311 : }
2312 :
2313 0 : if (rdma_req->fused_pair == NULL ||
2314 0 : rdma_req->fused_pair->state != RDMA_REQUEST_STATE_READY_TO_EXECUTE) {
2315 : /* This request is ready to execute, but either we don't know yet if it's
2316 : * valid - i.e. this is a FIRST but we haven't received the next
2317 : * request yet or the other request of this fused pair isn't ready to
2318 : * execute. So break here and this request will get processed later either
2319 : * when the other request is ready or we find that this request isn't valid.
2320 : */
2321 0 : break;
2322 : }
2323 0 : }
2324 :
2325 : /* If we get to this point, and this request is a fused command, we know that
2326 : * it is part of valid sequence (FIRST followed by a SECOND) and that both
2327 : * requests are READY_TO_EXECUTE. So call spdk_nvmf_request_exec() both on this
2328 : * request, and the other request of the fused pair, in the correct order.
2329 : * Also clear the ->fused_pair pointers on both requests, since after this point
2330 : * we no longer need to maintain the relationship between these two requests.
2331 : */
2332 5 : if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_SECOND) {
2333 0 : assert(rdma_req->fused_pair != NULL);
2334 0 : assert(rdma_req->fused_pair->fused_pair != NULL);
2335 0 : rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING;
2336 0 : spdk_nvmf_request_exec(&rdma_req->fused_pair->req);
2337 0 : rdma_req->fused_pair->fused_pair = NULL;
2338 0 : rdma_req->fused_pair = NULL;
2339 0 : }
2340 5 : rdma_req->state = RDMA_REQUEST_STATE_EXECUTING;
2341 5 : spdk_nvmf_request_exec(&rdma_req->req);
2342 5 : if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_FIRST) {
2343 0 : assert(rdma_req->fused_pair != NULL);
2344 0 : assert(rdma_req->fused_pair->fused_pair != NULL);
2345 0 : rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING;
2346 0 : spdk_nvmf_request_exec(&rdma_req->fused_pair->req);
2347 0 : rdma_req->fused_pair->fused_pair = NULL;
2348 0 : rdma_req->fused_pair = NULL;
2349 0 : }
2350 5 : break;
2351 : case RDMA_REQUEST_STATE_EXECUTING:
2352 5 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0,
2353 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2354 : /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED
2355 : * to escape this state. */
2356 5 : break;
2357 : case RDMA_REQUEST_STATE_EXECUTED:
2358 5 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0,
2359 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2360 10 : if (rsp->status.sc == SPDK_NVME_SC_SUCCESS &&
2361 5 : rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
2362 1 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link);
2363 1 : rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING;
2364 1 : } else {
2365 4 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
2366 4 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2367 : }
2368 5 : if (spdk_unlikely(rdma_req->req.dif_enabled)) {
2369 : /* restore the original length */
2370 0 : rdma_req->req.length = rdma_req->req.dif.orig_length;
2371 :
2372 0 : if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
2373 0 : struct spdk_dif_error error_blk;
2374 :
2375 0 : num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size);
2376 0 : if (!rdma_req->req.stripped_data) {
2377 0 : rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks,
2378 0 : &rdma_req->req.dif.dif_ctx, &error_blk);
2379 0 : } else {
2380 0 : rc = spdk_dif_verify_copy(rdma_req->req.stripped_data->iov,
2381 0 : rdma_req->req.stripped_data->iovcnt,
2382 0 : rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks,
2383 0 : &rdma_req->req.dif.dif_ctx, &error_blk);
2384 : }
2385 0 : if (rc) {
2386 0 : struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl;
2387 :
2388 0 : SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type,
2389 : error_blk.err_offset);
2390 0 : rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR;
2391 0 : rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type);
2392 0 : STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
2393 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
2394 0 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2395 0 : }
2396 0 : }
2397 0 : }
2398 5 : break;
2399 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING:
2400 1 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0,
2401 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2402 :
2403 1 : if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) {
2404 : /* This request needs to wait in line to perform RDMA */
2405 0 : break;
2406 : }
2407 2 : if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) >
2408 1 : rqpair->max_send_depth) {
2409 : /* We can only have so many WRs outstanding. we have to wait until some finish.
2410 : * +1 since each request has an additional wr in the resp. */
2411 0 : rqpair->poller->stat.pending_rdma_write++;
2412 0 : break;
2413 : }
2414 :
2415 : /* We have already verified that this request is the head of the queue. */
2416 1 : STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link);
2417 :
2418 : /* The data transfer will be kicked off from
2419 : * RDMA_REQUEST_STATE_READY_TO_COMPLETE state.
2420 : * We verified that data + response fit into send queue, so we can go to the next state directly
2421 : */
2422 1 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE;
2423 1 : break;
2424 : case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING:
2425 7 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING, 0, 0,
2426 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2427 :
2428 7 : if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_send_queue)) {
2429 : /* This request needs to wait in line to send the completion */
2430 0 : break;
2431 : }
2432 :
2433 7 : assert(rqpair->current_send_depth <= rqpair->max_send_depth);
2434 7 : if (rqpair->current_send_depth == rqpair->max_send_depth) {
2435 : /* We can only have so many WRs outstanding. we have to wait until some finish */
2436 2 : rqpair->poller->stat.pending_rdma_send++;
2437 2 : break;
2438 : }
2439 :
2440 : /* We have already verified that this request is the head of the queue. */
2441 5 : STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_send_queue, state_link);
2442 :
2443 : /* The response sending will be kicked off from
2444 : * RDMA_REQUEST_STATE_READY_TO_COMPLETE state.
2445 : */
2446 5 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE;
2447 5 : break;
2448 : case RDMA_REQUEST_STATE_READY_TO_COMPLETE:
2449 6 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0,
2450 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2451 6 : rc = request_transfer_out(&rdma_req->req, &data_posted);
2452 6 : assert(rc == 0); /* No good way to handle this currently */
2453 6 : if (spdk_unlikely(rc)) {
2454 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
2455 0 : } else {
2456 6 : rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST :
2457 : RDMA_REQUEST_STATE_COMPLETING;
2458 : }
2459 6 : break;
2460 : case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST:
2461 1 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0,
2462 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2463 : /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED
2464 : * to escape this state. */
2465 1 : break;
2466 : case RDMA_REQUEST_STATE_COMPLETING:
2467 5 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0,
2468 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2469 : /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED
2470 : * to escape this state. */
2471 5 : break;
2472 : case RDMA_REQUEST_STATE_COMPLETED:
2473 6 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0,
2474 : (uintptr_t)rdma_req, (uintptr_t)rqpair, rqpair->qpair.queue_depth);
2475 :
2476 6 : rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc;
2477 6 : _nvmf_rdma_request_free(rdma_req, rtransport);
2478 6 : break;
2479 0 : case RDMA_REQUEST_NUM_STATES:
2480 : default:
2481 0 : assert(0);
2482 : break;
2483 : }
2484 :
2485 66 : if (rdma_req->state != prev_state) {
2486 43 : progress = true;
2487 43 : }
2488 66 : } while (rdma_req->state != prev_state);
2489 :
2490 23 : return progress;
2491 23 : }
2492 :
2493 : /* Public API callbacks begin here */
2494 :
2495 : #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128
2496 : #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128
2497 : #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096
2498 : #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128
2499 : #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096
2500 : #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072
2501 : #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES)
2502 : #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095
2503 : #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE UINT32_MAX
2504 : #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false
2505 : #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false
2506 : #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100
2507 : #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1
2508 : #define SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING false
2509 : #define SPDK_NVMF_RDMA_DEFAULT_DATA_WR_POOL_SIZE 4095
2510 :
2511 : static void
2512 1 : nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts)
2513 : {
2514 1 : opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH;
2515 1 : opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR;
2516 1 : opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE;
2517 1 : opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE;
2518 1 : opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE;
2519 1 : opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH;
2520 1 : opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS;
2521 1 : opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE;
2522 1 : opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP;
2523 1 : opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC;
2524 1 : opts->transport_specific = NULL;
2525 1 : opts->data_wr_pool_size = SPDK_NVMF_RDMA_DEFAULT_DATA_WR_POOL_SIZE;
2526 1 : }
2527 :
2528 : static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport,
2529 : spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg);
2530 :
2531 : static inline bool
2532 0 : nvmf_rdma_is_rxe_device(struct spdk_nvmf_rdma_device *device)
2533 : {
2534 0 : return device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_OLD ||
2535 0 : device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_NEW;
2536 : }
2537 :
2538 : static int nvmf_rdma_accept(void *ctx);
2539 : static bool nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport);
2540 : static void destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport,
2541 : struct spdk_nvmf_rdma_device *device);
2542 :
2543 : static int
2544 0 : create_ib_device(struct spdk_nvmf_rdma_transport *rtransport, struct ibv_context *context,
2545 : struct spdk_nvmf_rdma_device **new_device)
2546 : {
2547 0 : struct spdk_nvmf_rdma_device *device;
2548 0 : int flag = 0;
2549 0 : int rc = 0;
2550 :
2551 0 : device = calloc(1, sizeof(*device));
2552 0 : if (!device) {
2553 0 : SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n");
2554 0 : return -ENOMEM;
2555 : }
2556 0 : device->context = context;
2557 0 : rc = ibv_query_device(device->context, &device->attr);
2558 0 : if (rc < 0) {
2559 0 : SPDK_ERRLOG("Failed to query RDMA device attributes.\n");
2560 0 : free(device);
2561 0 : return rc;
2562 : }
2563 :
2564 : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
2565 0 : if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) {
2566 0 : SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,");
2567 0 : SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id);
2568 0 : }
2569 :
2570 : /**
2571 : * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE.
2572 : * The Soft-RoCE RXE driver does not currently support send with invalidate,
2573 : * but incorrectly reports that it does. There are changes making their way
2574 : * through the kernel now that will enable this feature. When they are merged,
2575 : * we can conditionally enable this feature.
2576 : *
2577 : * TODO: enable this for versions of the kernel rxe driver that support it.
2578 : */
2579 0 : if (nvmf_rdma_is_rxe_device(device)) {
2580 0 : device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS);
2581 0 : }
2582 : #endif
2583 :
2584 : /* set up device context async ev fd as NON_BLOCKING */
2585 0 : flag = fcntl(device->context->async_fd, F_GETFL);
2586 0 : rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK);
2587 0 : if (rc < 0) {
2588 0 : SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n");
2589 0 : free(device);
2590 0 : return rc;
2591 : }
2592 :
2593 0 : TAILQ_INSERT_TAIL(&rtransport->devices, device, link);
2594 0 : SPDK_DEBUGLOG(rdma, "New device %p is added to RDMA transport\n", device);
2595 :
2596 0 : if (g_nvmf_hooks.get_ibv_pd) {
2597 0 : device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context);
2598 0 : } else {
2599 0 : device->pd = ibv_alloc_pd(device->context);
2600 : }
2601 :
2602 0 : if (!device->pd) {
2603 0 : SPDK_ERRLOG("Unable to allocate protection domain.\n");
2604 0 : destroy_ib_device(rtransport, device);
2605 0 : return -ENOMEM;
2606 : }
2607 :
2608 0 : assert(device->map == NULL);
2609 :
2610 0 : device->map = spdk_rdma_utils_create_mem_map(device->pd, &g_nvmf_hooks, IBV_ACCESS_LOCAL_WRITE);
2611 0 : if (!device->map) {
2612 0 : SPDK_ERRLOG("Unable to allocate memory map for listen address\n");
2613 0 : destroy_ib_device(rtransport, device);
2614 0 : return -ENOMEM;
2615 : }
2616 :
2617 0 : assert(device->map != NULL);
2618 0 : assert(device->pd != NULL);
2619 :
2620 0 : if (new_device) {
2621 0 : *new_device = device;
2622 0 : }
2623 0 : SPDK_NOTICELOG("Create IB device %s(%p/%p) succeed.\n", ibv_get_device_name(context->device),
2624 : device, context);
2625 :
2626 0 : return 0;
2627 0 : }
2628 :
2629 : static void
2630 0 : free_poll_fds(struct spdk_nvmf_rdma_transport *rtransport)
2631 : {
2632 0 : if (rtransport->poll_fds) {
2633 0 : free(rtransport->poll_fds);
2634 0 : rtransport->poll_fds = NULL;
2635 0 : }
2636 0 : rtransport->npoll_fds = 0;
2637 0 : }
2638 :
2639 : static int
2640 0 : generate_poll_fds(struct spdk_nvmf_rdma_transport *rtransport)
2641 : {
2642 : /* Set up poll descriptor array to monitor events from RDMA and IB
2643 : * in a single poll syscall
2644 : */
2645 0 : int device_count = 0;
2646 0 : int i = 0;
2647 0 : struct spdk_nvmf_rdma_device *device, *tmp;
2648 :
2649 0 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) {
2650 0 : device_count++;
2651 0 : }
2652 :
2653 0 : rtransport->npoll_fds = device_count + 1;
2654 :
2655 0 : rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd));
2656 0 : if (rtransport->poll_fds == NULL) {
2657 0 : SPDK_ERRLOG("poll_fds allocation failed\n");
2658 0 : return -ENOMEM;
2659 : }
2660 :
2661 0 : rtransport->poll_fds[i].fd = rtransport->event_channel->fd;
2662 0 : rtransport->poll_fds[i++].events = POLLIN;
2663 :
2664 0 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) {
2665 0 : rtransport->poll_fds[i].fd = device->context->async_fd;
2666 0 : rtransport->poll_fds[i++].events = POLLIN;
2667 0 : }
2668 :
2669 0 : return 0;
2670 0 : }
2671 :
2672 : static struct spdk_nvmf_transport *
2673 0 : nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts)
2674 : {
2675 0 : int rc;
2676 0 : struct spdk_nvmf_rdma_transport *rtransport;
2677 0 : struct spdk_nvmf_rdma_device *device;
2678 0 : struct ibv_context **contexts;
2679 0 : size_t data_wr_pool_size;
2680 0 : uint32_t i;
2681 0 : int flag;
2682 0 : uint32_t sge_count;
2683 0 : uint32_t min_shared_buffers;
2684 0 : uint32_t min_in_capsule_data_size;
2685 0 : int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
2686 :
2687 0 : rtransport = calloc(1, sizeof(*rtransport));
2688 0 : if (!rtransport) {
2689 0 : return NULL;
2690 : }
2691 :
2692 0 : TAILQ_INIT(&rtransport->devices);
2693 0 : TAILQ_INIT(&rtransport->ports);
2694 0 : TAILQ_INIT(&rtransport->poll_groups);
2695 0 : TAILQ_INIT(&rtransport->retry_ports);
2696 :
2697 0 : rtransport->transport.ops = &spdk_nvmf_transport_rdma;
2698 0 : rtransport->rdma_opts.num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE;
2699 0 : rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH;
2700 0 : rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ;
2701 0 : rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG;
2702 0 : rtransport->rdma_opts.no_wr_batching = SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING;
2703 0 : if (opts->transport_specific != NULL &&
2704 0 : spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder,
2705 : SPDK_COUNTOF(rdma_transport_opts_decoder),
2706 0 : &rtransport->rdma_opts)) {
2707 0 : SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n");
2708 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2709 0 : return NULL;
2710 : }
2711 :
2712 0 : SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n"
2713 : " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n"
2714 : " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n"
2715 : " in_capsule_data_size=%d, max_aq_depth=%d,\n"
2716 : " num_shared_buffers=%d, num_cqe=%d, max_srq_depth=%d, no_srq=%d,"
2717 : " acceptor_backlog=%d, no_wr_batching=%d abort_timeout_sec=%d\n",
2718 : opts->max_queue_depth,
2719 : opts->max_io_size,
2720 : opts->max_qpairs_per_ctrlr - 1,
2721 : opts->io_unit_size,
2722 : opts->in_capsule_data_size,
2723 : opts->max_aq_depth,
2724 : opts->num_shared_buffers,
2725 : rtransport->rdma_opts.num_cqe,
2726 : rtransport->rdma_opts.max_srq_depth,
2727 : rtransport->rdma_opts.no_srq,
2728 : rtransport->rdma_opts.acceptor_backlog,
2729 : rtransport->rdma_opts.no_wr_batching,
2730 : opts->abort_timeout_sec);
2731 :
2732 : /* I/O unit size cannot be larger than max I/O size */
2733 0 : if (opts->io_unit_size > opts->max_io_size) {
2734 0 : opts->io_unit_size = opts->max_io_size;
2735 0 : }
2736 :
2737 0 : if (rtransport->rdma_opts.acceptor_backlog <= 0) {
2738 0 : SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n",
2739 : SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG);
2740 0 : rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG;
2741 0 : }
2742 :
2743 0 : if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) {
2744 0 : SPDK_ERRLOG("The number of shared data buffers (%d) is less than"
2745 : "the minimum number required to guarantee that forward progress can be made (%d)\n",
2746 : opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2));
2747 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2748 0 : return NULL;
2749 : }
2750 :
2751 : /* If buf_cache_size == UINT32_MAX, we will dynamically pick a cache size later that we know will fit. */
2752 0 : if (opts->buf_cache_size < UINT32_MAX) {
2753 0 : min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size;
2754 0 : if (min_shared_buffers > opts->num_shared_buffers) {
2755 0 : SPDK_ERRLOG("There are not enough buffers to satisfy"
2756 : "per-poll group caches for each thread. (%" PRIu32 ")"
2757 : "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers);
2758 0 : SPDK_ERRLOG("Please specify a larger number of shared buffers\n");
2759 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2760 0 : return NULL;
2761 : }
2762 0 : }
2763 :
2764 0 : sge_count = opts->max_io_size / opts->io_unit_size;
2765 0 : if (sge_count > NVMF_DEFAULT_TX_SGE) {
2766 0 : SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size);
2767 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2768 0 : return NULL;
2769 : }
2770 :
2771 0 : min_in_capsule_data_size = sizeof(struct spdk_nvme_sgl_descriptor) * SPDK_NVMF_MAX_SGL_ENTRIES;
2772 0 : if (opts->in_capsule_data_size < min_in_capsule_data_size) {
2773 0 : SPDK_WARNLOG("In capsule data size is set to %u, this is minimum size required to support msdbd=16\n",
2774 : min_in_capsule_data_size);
2775 0 : opts->in_capsule_data_size = min_in_capsule_data_size;
2776 0 : }
2777 :
2778 0 : rtransport->event_channel = rdma_create_event_channel();
2779 0 : if (rtransport->event_channel == NULL) {
2780 0 : SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno));
2781 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2782 0 : return NULL;
2783 : }
2784 :
2785 0 : flag = fcntl(rtransport->event_channel->fd, F_GETFL);
2786 0 : if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) {
2787 0 : SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n",
2788 : rtransport->event_channel->fd, spdk_strerror(errno));
2789 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2790 0 : return NULL;
2791 : }
2792 :
2793 0 : data_wr_pool_size = opts->data_wr_pool_size;
2794 0 : if (data_wr_pool_size < SPDK_NVMF_MAX_SGL_ENTRIES * 2 * spdk_env_get_core_count()) {
2795 0 : data_wr_pool_size = SPDK_NVMF_MAX_SGL_ENTRIES * 2 * spdk_env_get_core_count();
2796 0 : SPDK_NOTICELOG("data_wr_pool_size is changed to %zu to guarantee enough cache for handling "
2797 : "at least one IO in each core\n", data_wr_pool_size);
2798 0 : }
2799 0 : rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", data_wr_pool_size,
2800 : sizeof(struct spdk_nvmf_rdma_request_data), SPDK_MEMPOOL_DEFAULT_CACHE_SIZE,
2801 : SPDK_ENV_NUMA_ID_ANY);
2802 0 : if (!rtransport->data_wr_pool) {
2803 0 : if (spdk_mempool_lookup("spdk_nvmf_rdma_wr_data") != NULL) {
2804 0 : SPDK_ERRLOG("Unable to allocate work request pool for poll group: already exists\n");
2805 0 : SPDK_ERRLOG("Probably running in multiprocess environment, which is "
2806 : "unsupported by the nvmf library\n");
2807 0 : } else {
2808 0 : SPDK_ERRLOG("Unable to allocate work request pool for poll group\n");
2809 : }
2810 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2811 0 : return NULL;
2812 : }
2813 :
2814 0 : contexts = rdma_get_devices(NULL);
2815 0 : if (contexts == NULL) {
2816 0 : SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno);
2817 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2818 0 : return NULL;
2819 : }
2820 :
2821 0 : i = 0;
2822 0 : rc = 0;
2823 0 : while (contexts[i] != NULL) {
2824 0 : rc = create_ib_device(rtransport, contexts[i], &device);
2825 0 : if (rc < 0) {
2826 0 : break;
2827 : }
2828 0 : i++;
2829 0 : max_device_sge = spdk_min(max_device_sge, device->attr.max_sge);
2830 0 : device->is_ready = true;
2831 : }
2832 0 : rdma_free_devices(contexts);
2833 :
2834 0 : if (opts->io_unit_size * max_device_sge < opts->max_io_size) {
2835 : /* divide and round up. */
2836 0 : opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge;
2837 :
2838 : /* round up to the nearest 4k. */
2839 0 : opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK;
2840 :
2841 0 : opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE);
2842 0 : SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n",
2843 : opts->io_unit_size);
2844 0 : }
2845 :
2846 0 : if (rc < 0) {
2847 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2848 0 : return NULL;
2849 : }
2850 :
2851 0 : rc = generate_poll_fds(rtransport);
2852 0 : if (rc < 0) {
2853 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2854 0 : return NULL;
2855 : }
2856 :
2857 0 : rtransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_rdma_accept, &rtransport->transport,
2858 : opts->acceptor_poll_rate);
2859 0 : if (!rtransport->accept_poller) {
2860 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2861 0 : return NULL;
2862 : }
2863 :
2864 0 : return &rtransport->transport;
2865 0 : }
2866 :
2867 : static void
2868 0 : destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport,
2869 : struct spdk_nvmf_rdma_device *device)
2870 : {
2871 0 : TAILQ_REMOVE(&rtransport->devices, device, link);
2872 0 : spdk_rdma_utils_free_mem_map(&device->map);
2873 0 : if (device->pd) {
2874 0 : if (!g_nvmf_hooks.get_ibv_pd) {
2875 0 : ibv_dealloc_pd(device->pd);
2876 0 : }
2877 0 : }
2878 0 : SPDK_DEBUGLOG(rdma, "IB device [%p] is destroyed.\n", device);
2879 0 : free(device);
2880 0 : }
2881 :
2882 : static void
2883 0 : nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w)
2884 : {
2885 0 : struct spdk_nvmf_rdma_transport *rtransport;
2886 0 : assert(w != NULL);
2887 :
2888 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
2889 0 : spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth);
2890 0 : spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq);
2891 0 : if (rtransport->rdma_opts.no_srq == true) {
2892 0 : spdk_json_write_named_int32(w, "num_cqe", rtransport->rdma_opts.num_cqe);
2893 0 : }
2894 0 : spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog);
2895 0 : spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching);
2896 0 : }
2897 :
2898 : static int
2899 0 : nvmf_rdma_destroy(struct spdk_nvmf_transport *transport,
2900 : spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg)
2901 : {
2902 0 : struct spdk_nvmf_rdma_transport *rtransport;
2903 0 : struct spdk_nvmf_rdma_port *port, *port_tmp;
2904 0 : struct spdk_nvmf_rdma_device *device, *device_tmp;
2905 :
2906 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
2907 :
2908 0 : TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, port_tmp) {
2909 0 : TAILQ_REMOVE(&rtransport->retry_ports, port, link);
2910 0 : free(port);
2911 0 : }
2912 :
2913 0 : TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) {
2914 0 : TAILQ_REMOVE(&rtransport->ports, port, link);
2915 0 : rdma_destroy_id(port->id);
2916 0 : free(port);
2917 0 : }
2918 :
2919 0 : free_poll_fds(rtransport);
2920 :
2921 0 : if (rtransport->event_channel != NULL) {
2922 0 : rdma_destroy_event_channel(rtransport->event_channel);
2923 0 : }
2924 :
2925 0 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) {
2926 0 : destroy_ib_device(rtransport, device);
2927 0 : }
2928 :
2929 0 : if (rtransport->data_wr_pool != NULL) {
2930 0 : if (spdk_mempool_count(rtransport->data_wr_pool) != transport->opts.data_wr_pool_size) {
2931 0 : SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n",
2932 : spdk_mempool_count(rtransport->data_wr_pool),
2933 : transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES);
2934 0 : }
2935 0 : }
2936 :
2937 0 : spdk_mempool_free(rtransport->data_wr_pool);
2938 :
2939 0 : spdk_poller_unregister(&rtransport->accept_poller);
2940 0 : free(rtransport);
2941 :
2942 0 : if (cb_fn) {
2943 0 : cb_fn(cb_arg);
2944 0 : }
2945 0 : return 0;
2946 0 : }
2947 :
2948 : static int nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id,
2949 : struct spdk_nvme_transport_id *trid,
2950 : bool peer);
2951 :
2952 : static bool nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport);
2953 :
2954 : static int
2955 0 : nvmf_rdma_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid,
2956 : struct spdk_nvmf_listen_opts *listen_opts)
2957 : {
2958 0 : struct spdk_nvmf_rdma_transport *rtransport;
2959 0 : struct spdk_nvmf_rdma_device *device;
2960 0 : struct spdk_nvmf_rdma_port *port, *tmp_port;
2961 0 : struct addrinfo *res;
2962 0 : struct addrinfo hints;
2963 0 : int family;
2964 0 : int rc;
2965 0 : long int port_val;
2966 0 : bool is_retry = false;
2967 :
2968 0 : if (!strlen(trid->trsvcid)) {
2969 0 : SPDK_ERRLOG("Service id is required\n");
2970 0 : return -EINVAL;
2971 : }
2972 :
2973 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
2974 0 : assert(rtransport->event_channel != NULL);
2975 :
2976 0 : port = calloc(1, sizeof(*port));
2977 0 : if (!port) {
2978 0 : SPDK_ERRLOG("Port allocation failed\n");
2979 0 : return -ENOMEM;
2980 : }
2981 :
2982 0 : port->trid = trid;
2983 :
2984 0 : switch (trid->adrfam) {
2985 : case SPDK_NVMF_ADRFAM_IPV4:
2986 0 : family = AF_INET;
2987 0 : break;
2988 : case SPDK_NVMF_ADRFAM_IPV6:
2989 0 : family = AF_INET6;
2990 0 : break;
2991 : default:
2992 0 : SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam);
2993 0 : free(port);
2994 0 : return -EINVAL;
2995 : }
2996 :
2997 0 : memset(&hints, 0, sizeof(hints));
2998 0 : hints.ai_family = family;
2999 0 : hints.ai_flags = AI_NUMERICSERV;
3000 0 : hints.ai_socktype = SOCK_STREAM;
3001 0 : hints.ai_protocol = 0;
3002 :
3003 : /* Range check the trsvcid. Fail in 3 cases:
3004 : * < 0: means that spdk_strtol hit an error
3005 : * 0: this results in ephemeral port which we don't want
3006 : * > 65535: port too high
3007 : */
3008 0 : port_val = spdk_strtol(trid->trsvcid, 10);
3009 0 : if (port_val <= 0 || port_val > 65535) {
3010 0 : SPDK_ERRLOG("invalid trsvcid %s\n", trid->trsvcid);
3011 0 : free(port);
3012 0 : return -EINVAL;
3013 : }
3014 :
3015 0 : rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res);
3016 0 : if (rc) {
3017 0 : SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc);
3018 0 : free(port);
3019 0 : return -(abs(rc));
3020 : }
3021 :
3022 0 : rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP);
3023 0 : if (rc < 0) {
3024 0 : SPDK_ERRLOG("rdma_create_id() failed\n");
3025 0 : freeaddrinfo(res);
3026 0 : free(port);
3027 0 : return rc;
3028 : }
3029 :
3030 0 : rc = rdma_bind_addr(port->id, res->ai_addr);
3031 0 : freeaddrinfo(res);
3032 :
3033 0 : if (rc < 0) {
3034 0 : TAILQ_FOREACH(tmp_port, &rtransport->retry_ports, link) {
3035 0 : if (spdk_nvme_transport_id_compare(tmp_port->trid, trid) == 0) {
3036 0 : is_retry = true;
3037 0 : break;
3038 : }
3039 0 : }
3040 0 : if (!is_retry) {
3041 0 : SPDK_ERRLOG("rdma_bind_addr() failed\n");
3042 0 : }
3043 0 : rdma_destroy_id(port->id);
3044 0 : free(port);
3045 0 : return rc;
3046 : }
3047 :
3048 0 : if (!port->id->verbs) {
3049 0 : SPDK_ERRLOG("ibv_context is null\n");
3050 0 : rdma_destroy_id(port->id);
3051 0 : free(port);
3052 0 : return -1;
3053 : }
3054 :
3055 0 : rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog);
3056 0 : if (rc < 0) {
3057 0 : SPDK_ERRLOG("rdma_listen() failed\n");
3058 0 : rdma_destroy_id(port->id);
3059 0 : free(port);
3060 0 : return rc;
3061 : }
3062 :
3063 0 : TAILQ_FOREACH(device, &rtransport->devices, link) {
3064 0 : if (device->context == port->id->verbs && device->is_ready) {
3065 0 : port->device = device;
3066 0 : break;
3067 : }
3068 0 : }
3069 0 : if (!port->device) {
3070 0 : SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n",
3071 : port->id->verbs);
3072 0 : rdma_destroy_id(port->id);
3073 0 : free(port);
3074 0 : nvmf_rdma_rescan_devices(rtransport);
3075 0 : return -EINVAL;
3076 : }
3077 :
3078 0 : SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n",
3079 : trid->traddr, trid->trsvcid);
3080 :
3081 0 : TAILQ_INSERT_TAIL(&rtransport->ports, port, link);
3082 0 : return 0;
3083 0 : }
3084 :
3085 : static void
3086 0 : nvmf_rdma_stop_listen_ex(struct spdk_nvmf_transport *transport,
3087 : const struct spdk_nvme_transport_id *trid, bool need_retry)
3088 : {
3089 0 : struct spdk_nvmf_rdma_transport *rtransport;
3090 0 : struct spdk_nvmf_rdma_port *port, *tmp;
3091 :
3092 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3093 :
3094 0 : if (!need_retry) {
3095 0 : TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp) {
3096 0 : if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) {
3097 0 : TAILQ_REMOVE(&rtransport->retry_ports, port, link);
3098 0 : free(port);
3099 0 : }
3100 0 : }
3101 0 : }
3102 :
3103 0 : TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) {
3104 0 : if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) {
3105 0 : SPDK_DEBUGLOG(rdma, "Port %s:%s removed. need retry: %d\n",
3106 : port->trid->traddr, port->trid->trsvcid, need_retry);
3107 0 : TAILQ_REMOVE(&rtransport->ports, port, link);
3108 0 : rdma_destroy_id(port->id);
3109 0 : port->id = NULL;
3110 0 : port->device = NULL;
3111 0 : if (need_retry) {
3112 0 : TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link);
3113 0 : } else {
3114 0 : free(port);
3115 : }
3116 0 : break;
3117 : }
3118 0 : }
3119 0 : }
3120 :
3121 : static void
3122 0 : nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport,
3123 : const struct spdk_nvme_transport_id *trid)
3124 : {
3125 0 : nvmf_rdma_stop_listen_ex(transport, trid, false);
3126 0 : }
3127 :
3128 : static void _nvmf_rdma_register_poller_in_group(void *c);
3129 : static void _nvmf_rdma_remove_poller_in_group(void *c);
3130 :
3131 : static bool
3132 0 : nvmf_rdma_all_pollers_management_done(void *c)
3133 : {
3134 0 : struct poller_manage_ctx *ctx = c;
3135 0 : int counter;
3136 :
3137 0 : counter = __atomic_sub_fetch(ctx->inflight_op_counter, 1, __ATOMIC_SEQ_CST);
3138 0 : SPDK_DEBUGLOG(rdma, "nvmf_rdma_all_pollers_management_done called. counter: %d, poller: %p\n",
3139 : counter, ctx->rpoller);
3140 :
3141 0 : if (counter == 0) {
3142 0 : free((void *)ctx->inflight_op_counter);
3143 0 : }
3144 0 : free(ctx);
3145 :
3146 0 : return counter == 0;
3147 0 : }
3148 :
3149 : static int
3150 0 : nvmf_rdma_manage_poller(struct spdk_nvmf_rdma_transport *rtransport,
3151 : struct spdk_nvmf_rdma_device *device, bool *has_inflight, bool is_add)
3152 : {
3153 0 : struct spdk_nvmf_rdma_poll_group *rgroup;
3154 0 : struct spdk_nvmf_rdma_poller *rpoller;
3155 0 : struct spdk_nvmf_poll_group *poll_group;
3156 0 : struct poller_manage_ctx *ctx;
3157 0 : bool found;
3158 0 : int *inflight_counter;
3159 0 : spdk_msg_fn do_fn;
3160 :
3161 0 : *has_inflight = false;
3162 0 : do_fn = is_add ? _nvmf_rdma_register_poller_in_group : _nvmf_rdma_remove_poller_in_group;
3163 0 : inflight_counter = calloc(1, sizeof(int));
3164 0 : if (!inflight_counter) {
3165 0 : SPDK_ERRLOG("Failed to allocate inflight counter when removing pollers\n");
3166 0 : return -ENOMEM;
3167 : }
3168 :
3169 0 : TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) {
3170 0 : (*inflight_counter)++;
3171 0 : }
3172 :
3173 0 : TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) {
3174 0 : found = false;
3175 0 : TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
3176 0 : if (rpoller->device == device) {
3177 0 : found = true;
3178 0 : break;
3179 : }
3180 0 : }
3181 0 : if (found == is_add) {
3182 0 : __atomic_fetch_sub(inflight_counter, 1, __ATOMIC_SEQ_CST);
3183 0 : continue;
3184 : }
3185 :
3186 0 : ctx = calloc(1, sizeof(struct poller_manage_ctx));
3187 0 : if (!ctx) {
3188 0 : SPDK_ERRLOG("Failed to allocate poller_manage_ctx when removing pollers\n");
3189 0 : if (!*has_inflight) {
3190 0 : free(inflight_counter);
3191 0 : }
3192 0 : return -ENOMEM;
3193 : }
3194 :
3195 0 : ctx->rtransport = rtransport;
3196 0 : ctx->rgroup = rgroup;
3197 0 : ctx->rpoller = rpoller;
3198 0 : ctx->device = device;
3199 0 : ctx->thread = spdk_get_thread();
3200 0 : ctx->inflight_op_counter = inflight_counter;
3201 0 : *has_inflight = true;
3202 :
3203 0 : poll_group = rgroup->group.group;
3204 0 : if (poll_group->thread != spdk_get_thread()) {
3205 0 : spdk_thread_send_msg(poll_group->thread, do_fn, ctx);
3206 0 : } else {
3207 0 : do_fn(ctx);
3208 : }
3209 0 : }
3210 :
3211 0 : if (!*has_inflight) {
3212 0 : free(inflight_counter);
3213 0 : }
3214 :
3215 0 : return 0;
3216 0 : }
3217 :
3218 : static void nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport,
3219 : struct spdk_nvmf_rdma_device *device);
3220 :
3221 : static struct spdk_nvmf_rdma_device *
3222 0 : nvmf_rdma_find_ib_device(struct spdk_nvmf_rdma_transport *rtransport,
3223 : struct ibv_context *context)
3224 : {
3225 0 : struct spdk_nvmf_rdma_device *device, *tmp_device;
3226 :
3227 0 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp_device) {
3228 0 : if (device->need_destroy) {
3229 0 : continue;
3230 : }
3231 :
3232 0 : if (strcmp(device->context->device->dev_name, context->device->dev_name) == 0) {
3233 0 : return device;
3234 : }
3235 0 : }
3236 :
3237 0 : return NULL;
3238 0 : }
3239 :
3240 : static bool
3241 0 : nvmf_rdma_check_devices_context(struct spdk_nvmf_rdma_transport *rtransport,
3242 : struct ibv_context *context)
3243 : {
3244 0 : struct spdk_nvmf_rdma_device *old_device, *new_device;
3245 0 : int rc = 0;
3246 0 : bool has_inflight;
3247 :
3248 0 : old_device = nvmf_rdma_find_ib_device(rtransport, context);
3249 :
3250 0 : if (old_device) {
3251 0 : if (old_device->context != context && !old_device->need_destroy && old_device->is_ready) {
3252 : /* context may not have time to be cleaned when rescan. exactly one context
3253 : * is valid for a device so this context must be invalid and just remove it. */
3254 0 : SPDK_WARNLOG("Device %p has a invalid context %p\n", old_device, old_device->context);
3255 0 : old_device->need_destroy = true;
3256 0 : nvmf_rdma_handle_device_removal(rtransport, old_device);
3257 0 : }
3258 0 : return false;
3259 : }
3260 :
3261 0 : rc = create_ib_device(rtransport, context, &new_device);
3262 : /* TODO: update transport opts. */
3263 0 : if (rc < 0) {
3264 0 : SPDK_ERRLOG("Failed to create ib device for context: %s(%p)\n",
3265 : ibv_get_device_name(context->device), context);
3266 0 : return false;
3267 : }
3268 :
3269 0 : rc = nvmf_rdma_manage_poller(rtransport, new_device, &has_inflight, true);
3270 0 : if (rc < 0) {
3271 0 : SPDK_ERRLOG("Failed to add poller for device context: %s(%p)\n",
3272 : ibv_get_device_name(context->device), context);
3273 0 : return false;
3274 : }
3275 :
3276 0 : if (has_inflight) {
3277 0 : new_device->is_ready = true;
3278 0 : }
3279 :
3280 0 : return true;
3281 0 : }
3282 :
3283 : static bool
3284 0 : nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport)
3285 : {
3286 0 : struct spdk_nvmf_rdma_device *device;
3287 0 : struct ibv_device **ibv_device_list = NULL;
3288 0 : struct ibv_context **contexts = NULL;
3289 0 : int i = 0;
3290 0 : int num_dev = 0;
3291 0 : bool new_create = false, has_new_device = false;
3292 0 : struct ibv_context *tmp_verbs = NULL;
3293 :
3294 : /* do not rescan when any device is destroying, or context may be freed when
3295 : * regenerating the poll fds.
3296 : */
3297 0 : TAILQ_FOREACH(device, &rtransport->devices, link) {
3298 0 : if (device->need_destroy) {
3299 0 : return false;
3300 : }
3301 0 : }
3302 :
3303 0 : ibv_device_list = ibv_get_device_list(&num_dev);
3304 :
3305 : /* There is a bug in librdmacm. If verbs init failed in rdma_get_devices, it'll be
3306 : * marked as dead verbs and never be init again. So we need to make sure the
3307 : * verbs is available before we call rdma_get_devices. */
3308 0 : if (num_dev >= 0) {
3309 0 : for (i = 0; i < num_dev; i++) {
3310 0 : tmp_verbs = ibv_open_device(ibv_device_list[i]);
3311 0 : if (!tmp_verbs) {
3312 0 : SPDK_WARNLOG("Failed to init ibv device %p, err %d. Skip rescan.\n", ibv_device_list[i], errno);
3313 0 : break;
3314 : }
3315 0 : if (nvmf_rdma_find_ib_device(rtransport, tmp_verbs) == NULL) {
3316 0 : SPDK_DEBUGLOG(rdma, "Find new verbs init ibv device %p(%s).\n", ibv_device_list[i],
3317 : tmp_verbs->device->dev_name);
3318 0 : has_new_device = true;
3319 0 : }
3320 0 : ibv_close_device(tmp_verbs);
3321 0 : }
3322 0 : ibv_free_device_list(ibv_device_list);
3323 0 : if (!tmp_verbs || !has_new_device) {
3324 0 : return false;
3325 : }
3326 0 : }
3327 :
3328 0 : contexts = rdma_get_devices(NULL);
3329 :
3330 0 : for (i = 0; contexts && contexts[i] != NULL; i++) {
3331 0 : new_create |= nvmf_rdma_check_devices_context(rtransport, contexts[i]);
3332 0 : }
3333 :
3334 0 : if (new_create) {
3335 0 : free_poll_fds(rtransport);
3336 0 : generate_poll_fds(rtransport);
3337 0 : }
3338 :
3339 0 : if (contexts) {
3340 0 : rdma_free_devices(contexts);
3341 0 : }
3342 :
3343 0 : return new_create;
3344 0 : }
3345 :
3346 : static bool
3347 0 : nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport)
3348 : {
3349 0 : struct spdk_nvmf_rdma_port *port, *tmp_port;
3350 0 : int rc = 0;
3351 0 : bool new_create = false;
3352 :
3353 0 : if (TAILQ_EMPTY(&rtransport->retry_ports)) {
3354 0 : return false;
3355 : }
3356 :
3357 0 : new_create = nvmf_rdma_rescan_devices(rtransport);
3358 :
3359 0 : TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp_port) {
3360 0 : rc = nvmf_rdma_listen(&rtransport->transport, port->trid, NULL);
3361 :
3362 0 : TAILQ_REMOVE(&rtransport->retry_ports, port, link);
3363 0 : if (rc) {
3364 0 : if (new_create) {
3365 0 : SPDK_ERRLOG("Found new IB device but port %s:%s is still failed(%d) to listen.\n",
3366 : port->trid->traddr, port->trid->trsvcid, rc);
3367 0 : }
3368 0 : TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link);
3369 0 : break;
3370 : } else {
3371 0 : SPDK_NOTICELOG("Port %s:%s come back\n", port->trid->traddr, port->trid->trsvcid);
3372 0 : free(port);
3373 : }
3374 0 : }
3375 :
3376 0 : return true;
3377 0 : }
3378 :
3379 : static void
3380 0 : nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport,
3381 : struct spdk_nvmf_rdma_qpair *rqpair, bool drain)
3382 : {
3383 0 : struct spdk_nvmf_request *req, *tmp;
3384 0 : struct spdk_nvmf_rdma_request *rdma_req, *req_tmp;
3385 0 : struct spdk_nvmf_rdma_resources *resources;
3386 :
3387 : /* First process requests which are waiting for response to be sent */
3388 0 : STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_send_queue, state_link, req_tmp) {
3389 0 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
3390 0 : break;
3391 : }
3392 0 : }
3393 :
3394 : /* We process I/O in the data transfer pending queue at the highest priority. */
3395 0 : STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) {
3396 0 : if (rdma_req->state != RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) {
3397 : /* Requests in this queue might be in state RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER,
3398 : * they are transmitting data over network but we keep them in the list to guarantee
3399 : * fair processing. */
3400 0 : continue;
3401 : }
3402 0 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
3403 0 : break;
3404 : }
3405 0 : }
3406 :
3407 : /* Then RDMA writes since reads have stronger restrictions than writes */
3408 0 : STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) {
3409 0 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
3410 0 : break;
3411 : }
3412 0 : }
3413 :
3414 : /* Then we handle request waiting on memory buffers. */
3415 0 : STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) {
3416 0 : rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
3417 0 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
3418 0 : break;
3419 : }
3420 0 : }
3421 :
3422 0 : resources = rqpair->resources;
3423 0 : while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) {
3424 0 : rdma_req = STAILQ_FIRST(&resources->free_queue);
3425 0 : STAILQ_REMOVE_HEAD(&resources->free_queue, state_link);
3426 0 : rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue);
3427 0 : STAILQ_REMOVE_HEAD(&resources->incoming_queue, link);
3428 :
3429 0 : if (rqpair->srq != NULL) {
3430 0 : rdma_req->req.qpair = &rdma_req->recv->qpair->qpair;
3431 0 : rdma_req->recv->qpair->qd++;
3432 0 : } else {
3433 0 : rqpair->qd++;
3434 : }
3435 :
3436 0 : rdma_req->receive_tsc = rdma_req->recv->receive_tsc;
3437 0 : rdma_req->state = RDMA_REQUEST_STATE_NEW;
3438 0 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false) {
3439 0 : break;
3440 : }
3441 : }
3442 0 : if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) {
3443 0 : rqpair->poller->stat.pending_free_request++;
3444 0 : }
3445 0 : }
3446 :
3447 : static void
3448 0 : nvmf_rdma_poller_process_pending_buf_queue(struct spdk_nvmf_rdma_transport *rtransport,
3449 : struct spdk_nvmf_rdma_poller *rpoller)
3450 : {
3451 0 : struct spdk_nvmf_request *req, *tmp;
3452 0 : struct spdk_nvmf_rdma_request *rdma_req;
3453 :
3454 0 : STAILQ_FOREACH_SAFE(req, &rpoller->group->group.pending_buf_queue, buf_link, tmp) {
3455 0 : rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
3456 0 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false) {
3457 0 : break;
3458 : }
3459 0 : }
3460 0 : }
3461 :
3462 : static inline bool
3463 0 : nvmf_rdma_device_supports_last_wqe_reached(struct spdk_nvmf_rdma_device *device)
3464 : {
3465 : /* iWARP transport and SoftRoCE driver don't support LAST_WQE_REACHED ibv async event */
3466 0 : return !nvmf_rdma_is_rxe_device(device) &&
3467 0 : device->context->device->transport_type != IBV_TRANSPORT_IWARP;
3468 : }
3469 :
3470 : static void
3471 0 : nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair)
3472 : {
3473 0 : struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
3474 : struct spdk_nvmf_rdma_transport, transport);
3475 :
3476 0 : nvmf_rdma_qpair_process_pending(rtransport, rqpair, true);
3477 :
3478 : /* nvmf_rdma_close_qpair is not called */
3479 0 : if (!rqpair->to_close) {
3480 0 : return;
3481 : }
3482 :
3483 : /* device is already destroyed and we should force destroy this qpair. */
3484 0 : if (rqpair->poller && rqpair->poller->need_destroy) {
3485 0 : nvmf_rdma_qpair_destroy(rqpair);
3486 0 : return;
3487 : }
3488 :
3489 : /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */
3490 0 : if (rqpair->current_send_depth != 0) {
3491 0 : return;
3492 : }
3493 :
3494 0 : if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) {
3495 0 : return;
3496 : }
3497 :
3498 : /* For devices that support LAST_WQE_REACHED with srq, we need to
3499 : * wait to destroy the qpair until that event has been received.
3500 : */
3501 0 : if (rqpair->srq != NULL && rqpair->last_wqe_reached == false &&
3502 0 : nvmf_rdma_device_supports_last_wqe_reached(rqpair->device)) {
3503 0 : return;
3504 : }
3505 :
3506 0 : assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED ||
3507 : rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR);
3508 :
3509 0 : nvmf_rdma_qpair_destroy(rqpair);
3510 0 : }
3511 :
3512 : static int
3513 0 : nvmf_rdma_disconnect(struct rdma_cm_event *evt, bool *event_acked)
3514 : {
3515 0 : struct spdk_nvmf_qpair *qpair;
3516 0 : struct spdk_nvmf_rdma_qpair *rqpair;
3517 :
3518 0 : if (evt->id == NULL) {
3519 0 : SPDK_ERRLOG("disconnect request: missing cm_id\n");
3520 0 : return -1;
3521 : }
3522 :
3523 0 : qpair = evt->id->context;
3524 0 : if (qpair == NULL) {
3525 0 : SPDK_ERRLOG("disconnect request: no active connection\n");
3526 0 : return -1;
3527 : }
3528 :
3529 0 : rdma_ack_cm_event(evt);
3530 0 : *event_acked = true;
3531 :
3532 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
3533 :
3534 0 : spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair);
3535 :
3536 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
3537 :
3538 0 : return 0;
3539 0 : }
3540 :
3541 : #ifdef DEBUG
3542 : static const char *CM_EVENT_STR[] = {
3543 : "RDMA_CM_EVENT_ADDR_RESOLVED",
3544 : "RDMA_CM_EVENT_ADDR_ERROR",
3545 : "RDMA_CM_EVENT_ROUTE_RESOLVED",
3546 : "RDMA_CM_EVENT_ROUTE_ERROR",
3547 : "RDMA_CM_EVENT_CONNECT_REQUEST",
3548 : "RDMA_CM_EVENT_CONNECT_RESPONSE",
3549 : "RDMA_CM_EVENT_CONNECT_ERROR",
3550 : "RDMA_CM_EVENT_UNREACHABLE",
3551 : "RDMA_CM_EVENT_REJECTED",
3552 : "RDMA_CM_EVENT_ESTABLISHED",
3553 : "RDMA_CM_EVENT_DISCONNECTED",
3554 : "RDMA_CM_EVENT_DEVICE_REMOVAL",
3555 : "RDMA_CM_EVENT_MULTICAST_JOIN",
3556 : "RDMA_CM_EVENT_MULTICAST_ERROR",
3557 : "RDMA_CM_EVENT_ADDR_CHANGE",
3558 : "RDMA_CM_EVENT_TIMEWAIT_EXIT"
3559 : };
3560 : #endif /* DEBUG */
3561 :
3562 : static void
3563 0 : nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport,
3564 : struct spdk_nvmf_rdma_port *port)
3565 : {
3566 0 : struct spdk_nvmf_rdma_poll_group *rgroup;
3567 0 : struct spdk_nvmf_rdma_poller *rpoller;
3568 0 : struct spdk_nvmf_rdma_qpair *rqpair;
3569 :
3570 0 : TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) {
3571 0 : TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
3572 0 : RB_FOREACH(rqpair, qpairs_tree, &rpoller->qpairs) {
3573 0 : if (rqpair->listen_id == port->id) {
3574 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
3575 0 : }
3576 0 : }
3577 0 : }
3578 0 : }
3579 0 : }
3580 :
3581 : static bool
3582 0 : nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport,
3583 : struct rdma_cm_event *event)
3584 : {
3585 0 : const struct spdk_nvme_transport_id *trid;
3586 0 : struct spdk_nvmf_rdma_port *port;
3587 0 : struct spdk_nvmf_rdma_transport *rtransport;
3588 0 : bool event_acked = false;
3589 :
3590 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3591 0 : TAILQ_FOREACH(port, &rtransport->ports, link) {
3592 0 : if (port->id == event->id) {
3593 0 : SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid);
3594 0 : rdma_ack_cm_event(event);
3595 0 : event_acked = true;
3596 0 : trid = port->trid;
3597 0 : break;
3598 : }
3599 0 : }
3600 :
3601 0 : if (event_acked) {
3602 0 : nvmf_rdma_disconnect_qpairs_on_port(rtransport, port);
3603 :
3604 0 : nvmf_rdma_stop_listen(transport, trid);
3605 0 : nvmf_rdma_listen(transport, trid, NULL);
3606 0 : }
3607 :
3608 0 : return event_acked;
3609 0 : }
3610 :
3611 : static void
3612 0 : nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport,
3613 : struct spdk_nvmf_rdma_device *device)
3614 : {
3615 0 : struct spdk_nvmf_rdma_port *port, *port_tmp;
3616 0 : int rc;
3617 0 : bool has_inflight;
3618 :
3619 0 : rc = nvmf_rdma_manage_poller(rtransport, device, &has_inflight, false);
3620 0 : if (rc) {
3621 0 : SPDK_ERRLOG("Failed to handle device removal, rc %d\n", rc);
3622 0 : return;
3623 : }
3624 :
3625 0 : if (!has_inflight) {
3626 : /* no pollers, destroy the device */
3627 0 : device->ready_to_destroy = true;
3628 0 : spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_remove_destroyed_device, rtransport);
3629 0 : }
3630 :
3631 0 : TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) {
3632 0 : if (port->device == device) {
3633 0 : SPDK_NOTICELOG("Port %s:%s on device %s is being removed.\n",
3634 : port->trid->traddr,
3635 : port->trid->trsvcid,
3636 : ibv_get_device_name(port->device->context->device));
3637 :
3638 : /* keep NVMF listener and only destroy structures of the
3639 : * RDMA transport. when the device comes back we can retry listening
3640 : * and the application's workflow will not be interrupted.
3641 : */
3642 0 : nvmf_rdma_stop_listen_ex(&rtransport->transport, port->trid, true);
3643 0 : }
3644 0 : }
3645 0 : }
3646 :
3647 : static void
3648 0 : nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport,
3649 : struct rdma_cm_event *event)
3650 : {
3651 0 : struct spdk_nvmf_rdma_port *port, *tmp_port;
3652 0 : struct spdk_nvmf_rdma_transport *rtransport;
3653 :
3654 0 : port = event->id->context;
3655 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3656 :
3657 0 : rdma_ack_cm_event(event);
3658 :
3659 : /* if device removal happens during ctrl qpair disconnecting, it's possible that we receive
3660 : * an DEVICE_REMOVAL event on qpair but the id->qp is just NULL. So we should make sure that
3661 : * we are handling a port event here.
3662 : */
3663 0 : TAILQ_FOREACH(tmp_port, &rtransport->ports, link) {
3664 0 : if (port == tmp_port && port->device && !port->device->need_destroy) {
3665 0 : port->device->need_destroy = true;
3666 0 : nvmf_rdma_handle_device_removal(rtransport, port->device);
3667 0 : }
3668 0 : }
3669 0 : }
3670 :
3671 : static void
3672 0 : nvmf_process_cm_events(struct spdk_nvmf_transport *transport, uint32_t max_events)
3673 : {
3674 0 : struct spdk_nvmf_rdma_transport *rtransport;
3675 0 : struct rdma_cm_event *event;
3676 0 : uint32_t i;
3677 0 : int rc;
3678 0 : bool event_acked;
3679 :
3680 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3681 :
3682 0 : if (rtransport->event_channel == NULL) {
3683 0 : return;
3684 : }
3685 :
3686 0 : for (i = 0; i < max_events; i++) {
3687 0 : event_acked = false;
3688 0 : rc = rdma_get_cm_event(rtransport->event_channel, &event);
3689 0 : if (rc) {
3690 0 : if (errno != EAGAIN && errno != EWOULDBLOCK) {
3691 0 : SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno));
3692 0 : }
3693 0 : break;
3694 : }
3695 :
3696 0 : SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]);
3697 :
3698 0 : spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event);
3699 :
3700 0 : switch (event->event) {
3701 : case RDMA_CM_EVENT_ADDR_RESOLVED:
3702 : case RDMA_CM_EVENT_ADDR_ERROR:
3703 : case RDMA_CM_EVENT_ROUTE_RESOLVED:
3704 : case RDMA_CM_EVENT_ROUTE_ERROR:
3705 : /* No action required. The target never attempts to resolve routes. */
3706 0 : break;
3707 : case RDMA_CM_EVENT_CONNECT_REQUEST:
3708 0 : rc = nvmf_rdma_connect(transport, event);
3709 0 : if (rc < 0) {
3710 0 : SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc);
3711 0 : break;
3712 : }
3713 0 : break;
3714 : case RDMA_CM_EVENT_CONNECT_RESPONSE:
3715 : /* The target never initiates a new connection. So this will not occur. */
3716 0 : break;
3717 : case RDMA_CM_EVENT_CONNECT_ERROR:
3718 : /* Can this happen? The docs say it can, but not sure what causes it. */
3719 0 : break;
3720 : case RDMA_CM_EVENT_UNREACHABLE:
3721 : case RDMA_CM_EVENT_REJECTED:
3722 : /* These only occur on the client side. */
3723 0 : break;
3724 : case RDMA_CM_EVENT_ESTABLISHED:
3725 : /* TODO: Should we be waiting for this event anywhere? */
3726 0 : break;
3727 : case RDMA_CM_EVENT_DISCONNECTED:
3728 0 : rc = nvmf_rdma_disconnect(event, &event_acked);
3729 0 : if (rc < 0) {
3730 0 : SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc);
3731 0 : break;
3732 : }
3733 0 : break;
3734 : case RDMA_CM_EVENT_DEVICE_REMOVAL:
3735 : /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL
3736 : * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s.
3737 : * Once these events are sent to SPDK, we should release all IB resources and
3738 : * don't make attempts to call any ibv_query/modify/create functions. We can only call
3739 : * ibv_destroy* functions to release user space memory allocated by IB. All kernel
3740 : * resources are already cleaned. */
3741 0 : if (event->id->qp) {
3742 : /* If rdma_cm event has a valid `qp` pointer then the event refers to the
3743 : * corresponding qpair. Otherwise the event refers to a listening device. */
3744 0 : rc = nvmf_rdma_disconnect(event, &event_acked);
3745 0 : if (rc < 0) {
3746 0 : SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc);
3747 0 : break;
3748 : }
3749 0 : } else {
3750 0 : nvmf_rdma_handle_cm_event_port_removal(transport, event);
3751 0 : event_acked = true;
3752 : }
3753 0 : break;
3754 : case RDMA_CM_EVENT_MULTICAST_JOIN:
3755 : case RDMA_CM_EVENT_MULTICAST_ERROR:
3756 : /* Multicast is not used */
3757 0 : break;
3758 : case RDMA_CM_EVENT_ADDR_CHANGE:
3759 0 : event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event);
3760 0 : break;
3761 : case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3762 : /* For now, do nothing. The target never re-uses queue pairs. */
3763 0 : break;
3764 : default:
3765 0 : SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event);
3766 0 : break;
3767 : }
3768 0 : if (!event_acked) {
3769 0 : rdma_ack_cm_event(event);
3770 0 : }
3771 0 : }
3772 0 : }
3773 :
3774 : static void
3775 0 : nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair)
3776 : {
3777 0 : rqpair->last_wqe_reached = true;
3778 0 : nvmf_rdma_destroy_drained_qpair(rqpair);
3779 0 : }
3780 :
3781 : static void
3782 0 : nvmf_rdma_qpair_process_last_wqe_event(void *ctx)
3783 : {
3784 0 : struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx;
3785 0 : struct spdk_nvmf_rdma_qpair *rqpair;
3786 :
3787 0 : rqpair = event_ctx->rqpair;
3788 :
3789 0 : if (rqpair) {
3790 0 : assert(event_ctx == rqpair->last_wqe_reached_ctx);
3791 0 : rqpair->last_wqe_reached_ctx = NULL;
3792 0 : nvmf_rdma_handle_last_wqe_reached(rqpair);
3793 0 : }
3794 0 : free(event_ctx);
3795 0 : }
3796 :
3797 : static int
3798 0 : nvmf_rdma_send_qpair_last_wqe_event(struct spdk_nvmf_rdma_qpair *rqpair)
3799 : {
3800 0 : struct spdk_nvmf_rdma_ibv_event_ctx *ctx;
3801 0 : struct spdk_thread *thr = NULL;
3802 0 : int rc;
3803 :
3804 0 : if (rqpair->qpair.group) {
3805 0 : thr = rqpair->qpair.group->thread;
3806 0 : } else if (rqpair->destruct_channel) {
3807 0 : thr = spdk_io_channel_get_thread(rqpair->destruct_channel);
3808 0 : }
3809 :
3810 0 : if (!thr) {
3811 0 : SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair);
3812 0 : return -EINVAL;
3813 : }
3814 :
3815 0 : if (rqpair->last_wqe_reached || rqpair->last_wqe_reached_ctx != NULL) {
3816 0 : SPDK_ERRLOG("LAST_WQE_REACHED already received for rqpair %p\n", rqpair);
3817 0 : return -EALREADY;
3818 : }
3819 :
3820 0 : ctx = calloc(1, sizeof(*ctx));
3821 0 : if (!ctx) {
3822 0 : return -ENOMEM;
3823 : }
3824 :
3825 0 : ctx->rqpair = rqpair;
3826 0 : rqpair->last_wqe_reached_ctx = ctx;
3827 :
3828 0 : rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_last_wqe_event, ctx);
3829 0 : if (rc) {
3830 0 : rqpair->last_wqe_reached_ctx = NULL;
3831 0 : free(ctx);
3832 0 : }
3833 :
3834 0 : return rc;
3835 0 : }
3836 :
3837 : static int
3838 0 : nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device)
3839 : {
3840 0 : int rc;
3841 0 : struct spdk_nvmf_rdma_qpair *rqpair = NULL;
3842 0 : struct ibv_async_event event;
3843 :
3844 0 : rc = ibv_get_async_event(device->context, &event);
3845 :
3846 0 : if (rc) {
3847 : /* In non-blocking mode -1 means there are no events available */
3848 0 : return rc;
3849 : }
3850 :
3851 0 : switch (event.event_type) {
3852 : case IBV_EVENT_QP_FATAL:
3853 : case IBV_EVENT_QP_LAST_WQE_REACHED:
3854 : case IBV_EVENT_QP_REQ_ERR:
3855 : case IBV_EVENT_QP_ACCESS_ERR:
3856 : case IBV_EVENT_COMM_EST:
3857 : case IBV_EVENT_PATH_MIG:
3858 : case IBV_EVENT_PATH_MIG_ERR:
3859 0 : rqpair = event.element.qp->qp_context;
3860 0 : if (!rqpair) {
3861 : /* Any QP event for NVMe-RDMA initiator may be returned. */
3862 0 : SPDK_NOTICELOG("Async QP event for unknown QP: %s\n",
3863 : ibv_event_type_str(event.event_type));
3864 0 : break;
3865 : }
3866 :
3867 0 : switch (event.event_type) {
3868 : case IBV_EVENT_QP_FATAL:
3869 0 : SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair);
3870 0 : spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0,
3871 : (uintptr_t)rqpair, event.event_type);
3872 0 : rqpair->ibv_in_error_state = true;
3873 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
3874 0 : break;
3875 : case IBV_EVENT_QP_LAST_WQE_REACHED:
3876 : /* This event only occurs for shared receive queues. */
3877 0 : SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair);
3878 0 : rc = nvmf_rdma_send_qpair_last_wqe_event(rqpair);
3879 0 : if (rc) {
3880 0 : SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc);
3881 0 : rqpair->last_wqe_reached = true;
3882 0 : }
3883 0 : break;
3884 : case IBV_EVENT_QP_REQ_ERR:
3885 : case IBV_EVENT_QP_ACCESS_ERR:
3886 : case IBV_EVENT_COMM_EST:
3887 : case IBV_EVENT_PATH_MIG:
3888 : case IBV_EVENT_PATH_MIG_ERR:
3889 0 : SPDK_NOTICELOG("Async QP event: %s\n",
3890 : ibv_event_type_str(event.event_type));
3891 0 : spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0,
3892 : (uintptr_t)rqpair, event.event_type);
3893 0 : rqpair->ibv_in_error_state = true;
3894 0 : break;
3895 : default:
3896 0 : break;
3897 : }
3898 0 : break;
3899 : case IBV_EVENT_DEVICE_FATAL:
3900 0 : SPDK_ERRLOG("Device Fatal event[%s] received on %s. device: %p\n",
3901 : ibv_event_type_str(event.event_type), ibv_get_device_name(device->context->device), device);
3902 0 : device->need_destroy = true;
3903 0 : break;
3904 : case IBV_EVENT_CQ_ERR:
3905 : case IBV_EVENT_PORT_ACTIVE:
3906 : case IBV_EVENT_PORT_ERR:
3907 : case IBV_EVENT_LID_CHANGE:
3908 : case IBV_EVENT_PKEY_CHANGE:
3909 : case IBV_EVENT_SM_CHANGE:
3910 : case IBV_EVENT_SRQ_ERR:
3911 : case IBV_EVENT_SRQ_LIMIT_REACHED:
3912 : case IBV_EVENT_CLIENT_REREGISTER:
3913 : case IBV_EVENT_GID_CHANGE:
3914 0 : case IBV_EVENT_SQ_DRAINED:
3915 : default:
3916 0 : SPDK_NOTICELOG("Async event: %s\n",
3917 : ibv_event_type_str(event.event_type));
3918 0 : spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type);
3919 0 : break;
3920 : }
3921 0 : ibv_ack_async_event(&event);
3922 :
3923 0 : return 0;
3924 0 : }
3925 :
3926 : static void
3927 0 : nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events)
3928 : {
3929 0 : int rc = 0;
3930 0 : uint32_t i = 0;
3931 :
3932 0 : for (i = 0; i < max_events; i++) {
3933 0 : rc = nvmf_process_ib_event(device);
3934 0 : if (rc) {
3935 0 : break;
3936 : }
3937 0 : }
3938 :
3939 0 : SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i);
3940 0 : }
3941 :
3942 : static int
3943 0 : nvmf_rdma_accept(void *ctx)
3944 : {
3945 0 : int nfds, i = 0;
3946 0 : struct spdk_nvmf_transport *transport = ctx;
3947 0 : struct spdk_nvmf_rdma_transport *rtransport;
3948 0 : struct spdk_nvmf_rdma_device *device, *tmp;
3949 0 : uint32_t count;
3950 0 : short revents;
3951 0 : bool do_retry;
3952 :
3953 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3954 0 : do_retry = nvmf_rdma_retry_listen_port(rtransport);
3955 :
3956 0 : count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0);
3957 :
3958 0 : if (nfds <= 0) {
3959 0 : return do_retry ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE;
3960 : }
3961 :
3962 : /* The first poll descriptor is RDMA CM event */
3963 0 : if (rtransport->poll_fds[i++].revents & POLLIN) {
3964 0 : nvmf_process_cm_events(transport, NVMF_RDMA_MAX_EVENTS_PER_POLL);
3965 0 : nfds--;
3966 0 : }
3967 :
3968 0 : if (nfds == 0) {
3969 0 : return SPDK_POLLER_BUSY;
3970 : }
3971 :
3972 : /* Second and subsequent poll descriptors are IB async events */
3973 0 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) {
3974 0 : revents = rtransport->poll_fds[i++].revents;
3975 0 : if (revents & POLLIN) {
3976 0 : if (spdk_likely(!device->need_destroy)) {
3977 0 : nvmf_process_ib_events(device, NVMF_RDMA_MAX_EVENTS_PER_POLL);
3978 0 : if (spdk_unlikely(device->need_destroy)) {
3979 0 : nvmf_rdma_handle_device_removal(rtransport, device);
3980 0 : }
3981 0 : }
3982 0 : nfds--;
3983 0 : } else if (revents & POLLNVAL || revents & POLLHUP) {
3984 0 : SPDK_ERRLOG("Receive unknown revent %x on device %p\n", (int)revents, device);
3985 0 : nfds--;
3986 0 : }
3987 0 : }
3988 : /* check all flagged fd's have been served */
3989 0 : assert(nfds == 0);
3990 :
3991 0 : return count > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE;
3992 0 : }
3993 :
3994 : static void
3995 0 : nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem,
3996 : struct spdk_nvmf_ctrlr_data *cdata)
3997 : {
3998 0 : cdata->nvmf_specific.msdbd = NVMF_DEFAULT_MSDBD;
3999 :
4000 : /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled
4001 : since in-capsule data only works with NVME drives that support SGL memory layout */
4002 0 : if (transport->opts.dif_insert_or_strip) {
4003 0 : cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16;
4004 0 : }
4005 :
4006 0 : if (cdata->nvmf_specific.ioccsz > ((sizeof(struct spdk_nvme_cmd) + 0x1000) / 16)) {
4007 0 : SPDK_WARNLOG("RDMA is configured to support up to 16 SGL entries while in capsule"
4008 : " data is greater than 4KiB.\n");
4009 0 : SPDK_WARNLOG("When used in conjunction with the NVMe-oF initiator from the Linux "
4010 : "kernel between versions 5.4 and 5.12 data corruption may occur for "
4011 : "writes that are not a multiple of 4KiB in size.\n");
4012 0 : }
4013 0 : }
4014 :
4015 : static void
4016 0 : nvmf_rdma_discover(struct spdk_nvmf_transport *transport,
4017 : struct spdk_nvme_transport_id *trid,
4018 : struct spdk_nvmf_discovery_log_page_entry *entry)
4019 : {
4020 0 : entry->trtype = SPDK_NVMF_TRTYPE_RDMA;
4021 0 : entry->adrfam = trid->adrfam;
4022 0 : entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED;
4023 :
4024 0 : spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' ');
4025 0 : spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' ');
4026 :
4027 0 : entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED;
4028 0 : entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE;
4029 0 : entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM;
4030 0 : }
4031 :
4032 : static int
4033 0 : nvmf_rdma_poller_create(struct spdk_nvmf_rdma_transport *rtransport,
4034 : struct spdk_nvmf_rdma_poll_group *rgroup, struct spdk_nvmf_rdma_device *device,
4035 : struct spdk_nvmf_rdma_poller **out_poller)
4036 : {
4037 0 : struct spdk_nvmf_rdma_poller *poller;
4038 0 : struct spdk_rdma_provider_srq_init_attr srq_init_attr;
4039 0 : struct spdk_nvmf_rdma_resource_opts opts;
4040 0 : int num_cqe;
4041 :
4042 0 : poller = calloc(1, sizeof(*poller));
4043 0 : if (!poller) {
4044 0 : SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n");
4045 0 : return -1;
4046 : }
4047 :
4048 0 : poller->device = device;
4049 0 : poller->group = rgroup;
4050 0 : *out_poller = poller;
4051 :
4052 0 : RB_INIT(&poller->qpairs);
4053 0 : STAILQ_INIT(&poller->qpairs_pending_send);
4054 0 : STAILQ_INIT(&poller->qpairs_pending_recv);
4055 :
4056 0 : TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link);
4057 0 : SPDK_DEBUGLOG(rdma, "Create poller %p on device %p in poll group %p.\n", poller, device, rgroup);
4058 0 : if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) {
4059 0 : if ((int)rtransport->rdma_opts.max_srq_depth > device->attr.max_srq_wr) {
4060 0 : SPDK_WARNLOG("Requested SRQ depth %u, max supported by dev %s is %d\n",
4061 : rtransport->rdma_opts.max_srq_depth, device->context->device->name, device->attr.max_srq_wr);
4062 0 : }
4063 0 : poller->max_srq_depth = spdk_min((int)rtransport->rdma_opts.max_srq_depth, device->attr.max_srq_wr);
4064 :
4065 0 : device->num_srq++;
4066 0 : memset(&srq_init_attr, 0, sizeof(srq_init_attr));
4067 0 : srq_init_attr.pd = device->pd;
4068 0 : srq_init_attr.stats = &poller->stat.qp_stats.recv;
4069 0 : srq_init_attr.srq_init_attr.attr.max_wr = poller->max_srq_depth;
4070 0 : srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE);
4071 0 : poller->srq = spdk_rdma_provider_srq_create(&srq_init_attr);
4072 0 : if (!poller->srq) {
4073 0 : SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno);
4074 0 : return -1;
4075 : }
4076 :
4077 0 : opts.qp = poller->srq;
4078 0 : opts.map = device->map;
4079 0 : opts.qpair = NULL;
4080 0 : opts.shared = true;
4081 0 : opts.max_queue_depth = poller->max_srq_depth;
4082 0 : opts.in_capsule_data_size = rtransport->transport.opts.in_capsule_data_size;
4083 :
4084 0 : poller->resources = nvmf_rdma_resources_create(&opts);
4085 0 : if (!poller->resources) {
4086 0 : SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n");
4087 0 : return -1;
4088 : }
4089 0 : }
4090 :
4091 : /*
4092 : * When using an srq, we can limit the completion queue at startup.
4093 : * The following formula represents the calculation:
4094 : * num_cqe = num_recv + num_data_wr + num_send_wr.
4095 : * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth
4096 : */
4097 0 : if (poller->srq) {
4098 0 : num_cqe = poller->max_srq_depth * 3;
4099 0 : } else {
4100 0 : num_cqe = rtransport->rdma_opts.num_cqe;
4101 : }
4102 :
4103 0 : poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0);
4104 0 : if (!poller->cq) {
4105 0 : SPDK_ERRLOG("Unable to create completion queue\n");
4106 0 : return -1;
4107 : }
4108 0 : poller->num_cqe = num_cqe;
4109 0 : return 0;
4110 0 : }
4111 :
4112 : static void
4113 0 : _nvmf_rdma_register_poller_in_group(void *c)
4114 : {
4115 0 : struct spdk_nvmf_rdma_poller *poller;
4116 0 : struct poller_manage_ctx *ctx = c;
4117 0 : struct spdk_nvmf_rdma_device *device;
4118 0 : int rc;
4119 :
4120 0 : rc = nvmf_rdma_poller_create(ctx->rtransport, ctx->rgroup, ctx->device, &poller);
4121 0 : if (rc < 0 && poller) {
4122 0 : nvmf_rdma_poller_destroy(poller);
4123 0 : }
4124 :
4125 0 : device = ctx->device;
4126 0 : if (nvmf_rdma_all_pollers_management_done(ctx)) {
4127 0 : device->is_ready = true;
4128 0 : }
4129 0 : }
4130 :
4131 : static void nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group);
4132 :
4133 : static struct spdk_nvmf_transport_poll_group *
4134 5 : nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport,
4135 : struct spdk_nvmf_poll_group *group)
4136 : {
4137 5 : struct spdk_nvmf_rdma_transport *rtransport;
4138 5 : struct spdk_nvmf_rdma_poll_group *rgroup;
4139 5 : struct spdk_nvmf_rdma_poller *poller;
4140 5 : struct spdk_nvmf_rdma_device *device;
4141 5 : int rc;
4142 :
4143 5 : if (spdk_interrupt_mode_is_enabled()) {
4144 0 : SPDK_ERRLOG("RDMA transport does not support interrupt mode\n");
4145 0 : return NULL;
4146 : }
4147 :
4148 5 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
4149 :
4150 5 : rgroup = calloc(1, sizeof(*rgroup));
4151 5 : if (!rgroup) {
4152 0 : return NULL;
4153 : }
4154 :
4155 5 : TAILQ_INIT(&rgroup->pollers);
4156 :
4157 5 : TAILQ_FOREACH(device, &rtransport->devices, link) {
4158 0 : rc = nvmf_rdma_poller_create(rtransport, rgroup, device, &poller);
4159 0 : if (rc < 0) {
4160 0 : nvmf_rdma_poll_group_destroy(&rgroup->group);
4161 0 : return NULL;
4162 : }
4163 0 : }
4164 :
4165 5 : TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link);
4166 5 : if (rtransport->conn_sched.next_admin_pg == NULL) {
4167 1 : rtransport->conn_sched.next_admin_pg = rgroup;
4168 1 : rtransport->conn_sched.next_io_pg = rgroup;
4169 1 : }
4170 :
4171 5 : return &rgroup->group;
4172 5 : }
4173 :
4174 : static uint32_t
4175 12 : nvmf_poll_group_get_io_qpair_count(struct spdk_nvmf_poll_group *pg)
4176 : {
4177 12 : uint32_t count;
4178 :
4179 : /* Just assume that unassociated qpairs will eventually be io
4180 : * qpairs. This is close enough for the use cases for this
4181 : * function.
4182 : */
4183 12 : pthread_mutex_lock(&pg->mutex);
4184 12 : count = pg->stat.current_io_qpairs + pg->current_unassociated_qpairs;
4185 12 : pthread_mutex_unlock(&pg->mutex);
4186 :
4187 24 : return count;
4188 12 : }
4189 :
4190 : static struct spdk_nvmf_transport_poll_group *
4191 14 : nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair)
4192 : {
4193 14 : struct spdk_nvmf_rdma_transport *rtransport;
4194 14 : struct spdk_nvmf_rdma_poll_group **pg;
4195 14 : struct spdk_nvmf_transport_poll_group *result;
4196 14 : uint32_t count;
4197 :
4198 14 : rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
4199 :
4200 14 : if (TAILQ_EMPTY(&rtransport->poll_groups)) {
4201 2 : return NULL;
4202 : }
4203 :
4204 12 : if (qpair->qid == 0) {
4205 6 : pg = &rtransport->conn_sched.next_admin_pg;
4206 6 : } else {
4207 6 : struct spdk_nvmf_rdma_poll_group *pg_min, *pg_start, *pg_current;
4208 6 : uint32_t min_value;
4209 :
4210 6 : pg = &rtransport->conn_sched.next_io_pg;
4211 6 : pg_min = *pg;
4212 6 : pg_start = *pg;
4213 6 : pg_current = *pg;
4214 6 : min_value = nvmf_poll_group_get_io_qpair_count(pg_current->group.group);
4215 :
4216 6 : while (1) {
4217 6 : count = nvmf_poll_group_get_io_qpair_count(pg_current->group.group);
4218 :
4219 6 : if (count < min_value) {
4220 0 : min_value = count;
4221 0 : pg_min = pg_current;
4222 0 : }
4223 :
4224 6 : pg_current = TAILQ_NEXT(pg_current, link);
4225 6 : if (pg_current == NULL) {
4226 2 : pg_current = TAILQ_FIRST(&rtransport->poll_groups);
4227 2 : }
4228 :
4229 6 : if (pg_current == pg_start || min_value == 0) {
4230 6 : break;
4231 : }
4232 : }
4233 6 : *pg = pg_min;
4234 6 : }
4235 :
4236 12 : assert(*pg != NULL);
4237 :
4238 12 : result = &(*pg)->group;
4239 :
4240 12 : *pg = TAILQ_NEXT(*pg, link);
4241 12 : if (*pg == NULL) {
4242 4 : *pg = TAILQ_FIRST(&rtransport->poll_groups);
4243 4 : }
4244 :
4245 12 : return result;
4246 14 : }
4247 :
4248 : static void
4249 0 : nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller)
4250 : {
4251 0 : struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair;
4252 0 : int rc;
4253 :
4254 0 : TAILQ_REMOVE(&poller->group->pollers, poller, link);
4255 0 : RB_FOREACH_SAFE(qpair, qpairs_tree, &poller->qpairs, tmp_qpair) {
4256 0 : nvmf_rdma_qpair_destroy(qpair);
4257 0 : }
4258 :
4259 0 : if (poller->srq) {
4260 0 : if (poller->resources) {
4261 0 : nvmf_rdma_resources_destroy(poller->resources);
4262 0 : }
4263 0 : spdk_rdma_provider_srq_destroy(poller->srq);
4264 0 : SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq);
4265 0 : }
4266 :
4267 0 : if (poller->cq) {
4268 0 : rc = ibv_destroy_cq(poller->cq);
4269 0 : if (rc != 0) {
4270 0 : SPDK_ERRLOG("Destroy cq return %d, error: %s\n", rc, strerror(errno));
4271 0 : }
4272 0 : }
4273 :
4274 0 : if (poller->destroy_cb) {
4275 0 : poller->destroy_cb(poller->destroy_cb_ctx);
4276 0 : poller->destroy_cb = NULL;
4277 0 : }
4278 :
4279 0 : free(poller);
4280 0 : }
4281 :
4282 : static void
4283 5 : nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group)
4284 : {
4285 5 : struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup;
4286 5 : struct spdk_nvmf_rdma_poller *poller, *tmp;
4287 5 : struct spdk_nvmf_rdma_transport *rtransport;
4288 :
4289 5 : rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
4290 5 : if (!rgroup) {
4291 0 : return;
4292 : }
4293 :
4294 5 : TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) {
4295 0 : nvmf_rdma_poller_destroy(poller);
4296 0 : }
4297 :
4298 5 : if (rgroup->group.transport == NULL) {
4299 : /* Transport can be NULL when nvmf_rdma_poll_group_create()
4300 : * calls this function directly in a failure path. */
4301 0 : free(rgroup);
4302 0 : return;
4303 : }
4304 :
4305 5 : rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport);
4306 :
4307 5 : next_rgroup = TAILQ_NEXT(rgroup, link);
4308 5 : TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link);
4309 5 : if (next_rgroup == NULL) {
4310 1 : next_rgroup = TAILQ_FIRST(&rtransport->poll_groups);
4311 1 : }
4312 5 : if (rtransport->conn_sched.next_admin_pg == rgroup) {
4313 5 : rtransport->conn_sched.next_admin_pg = next_rgroup;
4314 5 : }
4315 5 : if (rtransport->conn_sched.next_io_pg == rgroup) {
4316 5 : rtransport->conn_sched.next_io_pg = next_rgroup;
4317 5 : }
4318 :
4319 5 : free(rgroup);
4320 5 : }
4321 :
4322 : static void
4323 0 : nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair)
4324 : {
4325 0 : if (rqpair->cm_id != NULL) {
4326 0 : nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES);
4327 0 : }
4328 0 : }
4329 :
4330 : static int
4331 0 : nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group,
4332 : struct spdk_nvmf_qpair *qpair)
4333 : {
4334 0 : struct spdk_nvmf_rdma_poll_group *rgroup;
4335 0 : struct spdk_nvmf_rdma_qpair *rqpair;
4336 0 : struct spdk_nvmf_rdma_device *device;
4337 0 : struct spdk_nvmf_rdma_poller *poller;
4338 0 : int rc;
4339 :
4340 0 : rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
4341 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4342 :
4343 0 : device = rqpair->device;
4344 :
4345 0 : TAILQ_FOREACH(poller, &rgroup->pollers, link) {
4346 0 : if (poller->device == device) {
4347 0 : break;
4348 : }
4349 0 : }
4350 :
4351 0 : if (!poller) {
4352 0 : SPDK_ERRLOG("No poller found for device.\n");
4353 0 : return -1;
4354 : }
4355 :
4356 0 : if (poller->need_destroy) {
4357 0 : SPDK_ERRLOG("Poller is destroying.\n");
4358 0 : return -1;
4359 : }
4360 :
4361 0 : rqpair->poller = poller;
4362 0 : rqpair->srq = rqpair->poller->srq;
4363 :
4364 0 : rc = nvmf_rdma_qpair_initialize(qpair);
4365 0 : if (rc < 0) {
4366 0 : SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair);
4367 0 : rqpair->poller = NULL;
4368 0 : rqpair->srq = NULL;
4369 0 : return -1;
4370 : }
4371 :
4372 0 : RB_INSERT(qpairs_tree, &poller->qpairs, rqpair);
4373 :
4374 0 : rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair);
4375 0 : if (rc) {
4376 : /* Try to reject, but we probably can't */
4377 0 : nvmf_rdma_qpair_reject_connection(rqpair);
4378 0 : return -1;
4379 : }
4380 :
4381 0 : return 0;
4382 0 : }
4383 :
4384 : static int
4385 0 : nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group,
4386 : struct spdk_nvmf_qpair *qpair)
4387 : {
4388 0 : struct spdk_nvmf_rdma_qpair *rqpair;
4389 :
4390 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4391 0 : assert(group->transport->tgt != NULL);
4392 :
4393 0 : rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt);
4394 :
4395 0 : if (!rqpair->destruct_channel) {
4396 0 : SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair);
4397 0 : return 0;
4398 : }
4399 :
4400 : /* Sanity check that we get io_channel on the correct thread */
4401 0 : if (qpair->group) {
4402 0 : assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel));
4403 0 : }
4404 :
4405 0 : return 0;
4406 0 : }
4407 :
4408 : static int
4409 0 : nvmf_rdma_request_free(struct spdk_nvmf_request *req)
4410 : {
4411 0 : struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
4412 0 : struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport,
4413 : struct spdk_nvmf_rdma_transport, transport);
4414 0 : struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair,
4415 : struct spdk_nvmf_rdma_qpair, qpair);
4416 :
4417 : /*
4418 : * AER requests are freed when a qpair is destroyed. The recv corresponding to that request
4419 : * needs to be returned to the shared receive queue or the poll group will eventually be
4420 : * starved of RECV structures.
4421 : */
4422 0 : if (rqpair->srq && rdma_req->recv) {
4423 0 : int rc;
4424 0 : struct ibv_recv_wr *bad_recv_wr;
4425 :
4426 0 : spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, &rdma_req->recv->wr);
4427 0 : rc = spdk_rdma_provider_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr);
4428 0 : if (rc) {
4429 0 : SPDK_ERRLOG("Unable to re-post rx descriptor\n");
4430 0 : }
4431 0 : }
4432 :
4433 0 : _nvmf_rdma_request_free(rdma_req, rtransport);
4434 0 : return 0;
4435 0 : }
4436 :
4437 : static int
4438 0 : nvmf_rdma_request_complete(struct spdk_nvmf_request *req)
4439 : {
4440 0 : struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport,
4441 : struct spdk_nvmf_rdma_transport, transport);
4442 0 : struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req,
4443 : struct spdk_nvmf_rdma_request, req);
4444 0 : struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair,
4445 : struct spdk_nvmf_rdma_qpair, qpair);
4446 :
4447 0 : if (spdk_unlikely(rqpair->ibv_in_error_state)) {
4448 : /* The connection is dead. Move the request directly to the completed state. */
4449 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
4450 0 : } else {
4451 : /* The connection is alive, so process the request as normal */
4452 0 : rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
4453 : }
4454 :
4455 0 : nvmf_rdma_request_process(rtransport, rdma_req);
4456 :
4457 0 : return 0;
4458 0 : }
4459 :
4460 : static void
4461 0 : nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair,
4462 : spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg)
4463 : {
4464 0 : struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4465 :
4466 0 : rqpair->to_close = true;
4467 :
4468 0 : if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) {
4469 0 : nvmf_rdma_qpair_reject_connection(rqpair);
4470 0 : }
4471 0 : if (rqpair->rdma_qp) {
4472 0 : spdk_rdma_provider_qp_disconnect(rqpair->rdma_qp);
4473 0 : }
4474 :
4475 0 : nvmf_rdma_destroy_drained_qpair(rqpair);
4476 :
4477 0 : if (cb_fn) {
4478 0 : cb_fn(cb_arg);
4479 0 : }
4480 0 : }
4481 :
4482 : static struct spdk_nvmf_rdma_qpair *
4483 0 : get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc)
4484 : {
4485 0 : struct spdk_nvmf_rdma_qpair find;
4486 :
4487 0 : find.qp_num = wc->qp_num;
4488 :
4489 0 : return RB_FIND(qpairs_tree, &rpoller->qpairs, &find);
4490 0 : }
4491 :
4492 : #ifdef DEBUG
4493 : static int
4494 0 : nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req)
4495 : {
4496 0 : return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST ||
4497 0 : rdma_req->state == RDMA_REQUEST_STATE_COMPLETING;
4498 : }
4499 : #endif
4500 :
4501 : static void
4502 0 : _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr,
4503 : int rc)
4504 : {
4505 0 : struct spdk_nvmf_rdma_recv *rdma_recv;
4506 0 : struct spdk_nvmf_rdma_wr *bad_rdma_wr;
4507 :
4508 0 : SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc);
4509 0 : while (bad_recv_wr != NULL) {
4510 0 : bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id;
4511 0 : rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr);
4512 :
4513 0 : rdma_recv->qpair->current_recv_depth++;
4514 0 : bad_recv_wr = bad_recv_wr->next;
4515 0 : SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc);
4516 0 : spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair);
4517 : }
4518 0 : }
4519 :
4520 : static void
4521 0 : _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc)
4522 : {
4523 0 : SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc);
4524 0 : while (bad_recv_wr != NULL) {
4525 0 : bad_recv_wr = bad_recv_wr->next;
4526 0 : rqpair->current_recv_depth++;
4527 : }
4528 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
4529 0 : }
4530 :
4531 : static void
4532 0 : _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport,
4533 : struct spdk_nvmf_rdma_poller *rpoller)
4534 : {
4535 0 : struct spdk_nvmf_rdma_qpair *rqpair;
4536 0 : struct ibv_recv_wr *bad_recv_wr;
4537 0 : int rc;
4538 :
4539 0 : if (rpoller->srq) {
4540 0 : rc = spdk_rdma_provider_srq_flush_recv_wrs(rpoller->srq, &bad_recv_wr);
4541 0 : if (spdk_unlikely(rc)) {
4542 0 : _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc);
4543 0 : }
4544 0 : } else {
4545 0 : while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) {
4546 0 : rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv);
4547 0 : rc = spdk_rdma_provider_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr);
4548 0 : if (spdk_unlikely(rc)) {
4549 0 : _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc);
4550 0 : }
4551 0 : STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link);
4552 : }
4553 : }
4554 0 : }
4555 :
4556 : static void
4557 0 : _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport,
4558 : struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc)
4559 : {
4560 0 : struct spdk_nvmf_rdma_wr *bad_rdma_wr;
4561 0 : struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL;
4562 :
4563 0 : SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc);
4564 0 : for (; bad_wr != NULL; bad_wr = bad_wr->next) {
4565 0 : bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id;
4566 0 : assert(rqpair->current_send_depth > 0);
4567 0 : rqpair->current_send_depth--;
4568 0 : switch (bad_rdma_wr->type) {
4569 : case RDMA_WR_TYPE_DATA:
4570 0 : cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data_wr);
4571 0 : if (bad_wr->opcode == IBV_WR_RDMA_READ) {
4572 0 : assert(rqpair->current_read_depth > 0);
4573 0 : rqpair->current_read_depth--;
4574 0 : }
4575 0 : break;
4576 : case RDMA_WR_TYPE_SEND:
4577 0 : cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr);
4578 0 : break;
4579 : default:
4580 0 : SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair);
4581 0 : prev_rdma_req = cur_rdma_req;
4582 0 : continue;
4583 : }
4584 :
4585 0 : if (prev_rdma_req == cur_rdma_req) {
4586 : /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */
4587 : /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */
4588 0 : continue;
4589 : }
4590 :
4591 0 : switch (cur_rdma_req->state) {
4592 : case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
4593 0 : cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR;
4594 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, cur_rdma_req, state_link);
4595 0 : cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
4596 0 : break;
4597 : case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST:
4598 : case RDMA_REQUEST_STATE_COMPLETING:
4599 0 : cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
4600 0 : break;
4601 : default:
4602 0 : SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n",
4603 : cur_rdma_req->state, rqpair);
4604 0 : continue;
4605 : }
4606 :
4607 0 : nvmf_rdma_request_process(rtransport, cur_rdma_req);
4608 0 : prev_rdma_req = cur_rdma_req;
4609 0 : }
4610 :
4611 0 : if (spdk_nvmf_qpair_is_active(&rqpair->qpair)) {
4612 : /* Disconnect the connection. */
4613 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
4614 0 : }
4615 :
4616 0 : }
4617 :
4618 : static void
4619 0 : _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport,
4620 : struct spdk_nvmf_rdma_poller *rpoller)
4621 : {
4622 0 : struct spdk_nvmf_rdma_qpair *rqpair;
4623 0 : struct ibv_send_wr *bad_wr = NULL;
4624 0 : int rc;
4625 :
4626 0 : while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) {
4627 0 : rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send);
4628 0 : rc = spdk_rdma_provider_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr);
4629 :
4630 : /* bad wr always points to the first wr that failed. */
4631 0 : if (spdk_unlikely(rc)) {
4632 0 : _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc);
4633 0 : }
4634 0 : STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link);
4635 : }
4636 0 : }
4637 :
4638 : static const char *
4639 0 : nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type)
4640 : {
4641 0 : switch (wr_type) {
4642 : case RDMA_WR_TYPE_RECV:
4643 0 : return "RECV";
4644 : case RDMA_WR_TYPE_SEND:
4645 0 : return "SEND";
4646 : case RDMA_WR_TYPE_DATA:
4647 0 : return "DATA";
4648 : default:
4649 0 : SPDK_ERRLOG("Unknown WR type %d\n", wr_type);
4650 0 : SPDK_UNREACHABLE();
4651 : }
4652 0 : }
4653 :
4654 : static inline void
4655 0 : nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc)
4656 : {
4657 0 : enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type;
4658 :
4659 0 : if (wc->status == IBV_WC_WR_FLUSH_ERR) {
4660 : /* If qpair is in ERR state, we will receive completions for all posted and not completed
4661 : * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */
4662 0 : SPDK_DEBUGLOG(rdma,
4663 : "Error on CQ %p, (qp state %d, in_error %d) request 0x%lu, type %s, status: (%d): %s\n",
4664 : rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_in_error_state, wc->wr_id,
4665 : nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status));
4666 0 : } else {
4667 0 : SPDK_ERRLOG("Error on CQ %p, (qp state %d, in_error %d) request 0x%lu, type %s, status: (%d): %s\n",
4668 : rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_in_error_state, wc->wr_id,
4669 : nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status));
4670 : }
4671 0 : }
4672 :
4673 : static int
4674 0 : nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport,
4675 : struct spdk_nvmf_rdma_poller *rpoller)
4676 : {
4677 0 : struct ibv_wc wc[32];
4678 0 : struct spdk_nvmf_rdma_wr *rdma_wr;
4679 0 : struct spdk_nvmf_rdma_request *rdma_req;
4680 0 : struct spdk_nvmf_rdma_recv *rdma_recv;
4681 0 : struct spdk_nvmf_rdma_qpair *rqpair, *tmp_rqpair;
4682 0 : int reaped, i;
4683 0 : int count = 0;
4684 0 : int rc;
4685 0 : bool error = false;
4686 0 : uint64_t poll_tsc = spdk_get_ticks();
4687 :
4688 0 : if (spdk_unlikely(rpoller->need_destroy)) {
4689 : /* If qpair is closed before poller destroy, nvmf_rdma_destroy_drained_qpair may not
4690 : * be called because we cannot poll anything from cq. So we call that here to force
4691 : * destroy the qpair after to_close turning true.
4692 : */
4693 0 : RB_FOREACH_SAFE(rqpair, qpairs_tree, &rpoller->qpairs, tmp_rqpair) {
4694 0 : nvmf_rdma_destroy_drained_qpair(rqpair);
4695 0 : }
4696 0 : return 0;
4697 : }
4698 :
4699 : /* Poll for completing operations. */
4700 0 : reaped = ibv_poll_cq(rpoller->cq, 32, wc);
4701 0 : if (spdk_unlikely(reaped < 0)) {
4702 0 : SPDK_ERRLOG("Error polling CQ! (%d): %s\n",
4703 : errno, spdk_strerror(errno));
4704 0 : return -1;
4705 0 : } else if (reaped == 0) {
4706 0 : rpoller->stat.idle_polls++;
4707 0 : }
4708 :
4709 0 : rpoller->stat.polls++;
4710 0 : rpoller->stat.completions += reaped;
4711 :
4712 0 : for (i = 0; i < reaped; i++) {
4713 :
4714 0 : rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id;
4715 :
4716 0 : switch (rdma_wr->type) {
4717 : case RDMA_WR_TYPE_SEND:
4718 0 : rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr);
4719 0 : rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
4720 :
4721 0 : if (spdk_likely(!wc[i].status)) {
4722 0 : count++;
4723 0 : assert(wc[i].opcode == IBV_WC_SEND);
4724 0 : assert(nvmf_rdma_req_is_completing(rdma_req));
4725 0 : }
4726 :
4727 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
4728 : /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */
4729 0 : assert(rqpair->current_send_depth >= (uint32_t)rdma_req->num_outstanding_data_wr + 1);
4730 0 : rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1;
4731 0 : rdma_req->num_outstanding_data_wr = 0;
4732 :
4733 0 : nvmf_rdma_request_process(rtransport, rdma_req);
4734 0 : break;
4735 : case RDMA_WR_TYPE_RECV:
4736 : /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */
4737 0 : rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr);
4738 0 : if (rpoller->srq != NULL) {
4739 0 : rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]);
4740 : /* It is possible that there are still some completions for destroyed QP
4741 : * associated with SRQ. We just ignore these late completions and re-post
4742 : * receive WRs back to SRQ.
4743 : */
4744 0 : if (spdk_unlikely(NULL == rdma_recv->qpair)) {
4745 0 : struct ibv_recv_wr *bad_wr;
4746 :
4747 0 : rdma_recv->wr.next = NULL;
4748 0 : spdk_rdma_provider_srq_queue_recv_wrs(rpoller->srq, &rdma_recv->wr);
4749 0 : rc = spdk_rdma_provider_srq_flush_recv_wrs(rpoller->srq, &bad_wr);
4750 0 : if (rc) {
4751 0 : SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc);
4752 0 : }
4753 : continue;
4754 0 : }
4755 0 : }
4756 0 : rqpair = rdma_recv->qpair;
4757 :
4758 0 : assert(rqpair != NULL);
4759 0 : if (spdk_likely(!wc[i].status)) {
4760 0 : assert(wc[i].opcode == IBV_WC_RECV);
4761 0 : if (rqpair->current_recv_depth >= rqpair->max_queue_depth) {
4762 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
4763 0 : break;
4764 : }
4765 0 : }
4766 :
4767 0 : rdma_recv->wr.next = NULL;
4768 0 : rqpair->current_recv_depth++;
4769 0 : rdma_recv->receive_tsc = poll_tsc;
4770 0 : rpoller->stat.requests++;
4771 0 : STAILQ_INSERT_HEAD(&rqpair->resources->incoming_queue, rdma_recv, link);
4772 0 : rqpair->qpair.queue_depth++;
4773 0 : break;
4774 : case RDMA_WR_TYPE_DATA:
4775 0 : rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data_wr);
4776 0 : rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
4777 :
4778 0 : assert(rdma_req->num_outstanding_data_wr > 0);
4779 :
4780 0 : rqpair->current_send_depth--;
4781 0 : rdma_req->num_outstanding_data_wr--;
4782 0 : if (spdk_likely(!wc[i].status)) {
4783 0 : assert(wc[i].opcode == IBV_WC_RDMA_READ);
4784 0 : rqpair->current_read_depth--;
4785 : /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */
4786 0 : if (rdma_req->num_outstanding_data_wr == 0) {
4787 0 : if (rdma_req->num_remaining_data_wr) {
4788 : /* Only part of RDMA_READ operations was submitted, process the rest */
4789 0 : nvmf_rdma_request_reset_transfer_in(rdma_req, rtransport);
4790 0 : rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING;
4791 0 : nvmf_rdma_request_process(rtransport, rdma_req);
4792 0 : break;
4793 : }
4794 0 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
4795 0 : nvmf_rdma_request_process(rtransport, rdma_req);
4796 0 : }
4797 0 : } else {
4798 : /* If the data transfer fails still force the queue into the error state,
4799 : * if we were performing an RDMA_READ, we need to force the request into a
4800 : * completed state since it wasn't linked to a send. However, in the RDMA_WRITE
4801 : * case, we should wait for the SEND to complete. */
4802 0 : if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) {
4803 0 : rqpair->current_read_depth--;
4804 0 : if (rdma_req->num_outstanding_data_wr == 0) {
4805 0 : if (rdma_req->num_remaining_data_wr) {
4806 : /* Partially sent request is still in the pending_rdma_read_queue,
4807 : * remove it now before completing */
4808 0 : rdma_req->num_remaining_data_wr = 0;
4809 0 : STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
4810 0 : }
4811 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
4812 0 : nvmf_rdma_request_process(rtransport, rdma_req);
4813 0 : }
4814 0 : }
4815 : }
4816 0 : break;
4817 : default:
4818 0 : SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode);
4819 0 : continue;
4820 : }
4821 :
4822 : /* Handle error conditions */
4823 0 : if (spdk_unlikely(wc[i].status)) {
4824 0 : rqpair->ibv_in_error_state = true;
4825 0 : nvmf_rdma_log_wc_status(rqpair, &wc[i]);
4826 :
4827 0 : error = true;
4828 :
4829 0 : if (spdk_nvmf_qpair_is_active(&rqpair->qpair)) {
4830 : /* Disconnect the connection. */
4831 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
4832 0 : } else {
4833 0 : nvmf_rdma_destroy_drained_qpair(rqpair);
4834 : }
4835 0 : continue;
4836 : }
4837 :
4838 0 : nvmf_rdma_qpair_process_pending(rtransport, rqpair, false);
4839 :
4840 0 : if (spdk_unlikely(!spdk_nvmf_qpair_is_active(&rqpair->qpair))) {
4841 0 : nvmf_rdma_destroy_drained_qpair(rqpair);
4842 0 : }
4843 0 : }
4844 :
4845 0 : if (spdk_unlikely(error == true)) {
4846 0 : return -1;
4847 : }
4848 :
4849 0 : if (reaped == 0) {
4850 : /* In some cases we may not receive any CQE but we still may have pending IO requests waiting for
4851 : * a resource (e.g. a WR from the data_wr_pool).
4852 : * We need to start processing of such requests if no CQE reaped */
4853 0 : nvmf_rdma_poller_process_pending_buf_queue(rtransport, rpoller);
4854 0 : }
4855 :
4856 : /* submit outstanding work requests. */
4857 0 : _poller_submit_recvs(rtransport, rpoller);
4858 0 : _poller_submit_sends(rtransport, rpoller);
4859 :
4860 0 : return count;
4861 0 : }
4862 :
4863 : static void
4864 0 : _nvmf_rdma_remove_destroyed_device(void *c)
4865 : {
4866 0 : struct spdk_nvmf_rdma_transport *rtransport = c;
4867 0 : struct spdk_nvmf_rdma_device *device, *device_tmp;
4868 0 : int rc;
4869 :
4870 0 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) {
4871 0 : if (device->ready_to_destroy) {
4872 0 : destroy_ib_device(rtransport, device);
4873 0 : }
4874 0 : }
4875 :
4876 0 : free_poll_fds(rtransport);
4877 0 : rc = generate_poll_fds(rtransport);
4878 : /* cannot handle fd allocation error here */
4879 0 : if (rc != 0) {
4880 0 : SPDK_ERRLOG("Failed to generate poll fds after remove ib device.\n");
4881 0 : }
4882 0 : }
4883 :
4884 : static void
4885 0 : _nvmf_rdma_remove_poller_in_group_cb(void *c)
4886 : {
4887 0 : struct poller_manage_ctx *ctx = c;
4888 0 : struct spdk_nvmf_rdma_transport *rtransport = ctx->rtransport;
4889 0 : struct spdk_nvmf_rdma_device *device = ctx->device;
4890 0 : struct spdk_thread *thread = ctx->thread;
4891 :
4892 0 : if (nvmf_rdma_all_pollers_management_done(c)) {
4893 : /* destroy device when last poller is destroyed */
4894 0 : device->ready_to_destroy = true;
4895 0 : spdk_thread_send_msg(thread, _nvmf_rdma_remove_destroyed_device, rtransport);
4896 0 : }
4897 0 : }
4898 :
4899 : static void
4900 0 : _nvmf_rdma_remove_poller_in_group(void *c)
4901 : {
4902 0 : struct poller_manage_ctx *ctx = c;
4903 :
4904 0 : ctx->rpoller->need_destroy = true;
4905 0 : ctx->rpoller->destroy_cb_ctx = ctx;
4906 0 : ctx->rpoller->destroy_cb = _nvmf_rdma_remove_poller_in_group_cb;
4907 :
4908 : /* qp will be disconnected after receiving a RDMA_CM_EVENT_DEVICE_REMOVAL event. */
4909 0 : if (RB_EMPTY(&ctx->rpoller->qpairs)) {
4910 0 : nvmf_rdma_poller_destroy(ctx->rpoller);
4911 0 : }
4912 0 : }
4913 :
4914 : static int
4915 0 : nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group)
4916 : {
4917 0 : struct spdk_nvmf_rdma_transport *rtransport;
4918 0 : struct spdk_nvmf_rdma_poll_group *rgroup;
4919 0 : struct spdk_nvmf_rdma_poller *rpoller, *tmp;
4920 0 : int count = 0, rc, rc2 = 0;
4921 :
4922 0 : rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport);
4923 0 : rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
4924 :
4925 0 : TAILQ_FOREACH_SAFE(rpoller, &rgroup->pollers, link, tmp) {
4926 0 : rc = nvmf_rdma_poller_poll(rtransport, rpoller);
4927 0 : if (spdk_unlikely(rc < 0)) {
4928 0 : if (rc2 == 0) {
4929 0 : rc2 = rc;
4930 0 : }
4931 0 : continue;
4932 : }
4933 0 : count += rc;
4934 0 : }
4935 :
4936 0 : return rc2 ? rc2 : count;
4937 0 : }
4938 :
4939 : static int
4940 0 : nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id,
4941 : struct spdk_nvme_transport_id *trid,
4942 : bool peer)
4943 : {
4944 0 : struct sockaddr *saddr;
4945 0 : uint16_t port;
4946 :
4947 0 : spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA);
4948 :
4949 0 : if (peer) {
4950 0 : saddr = rdma_get_peer_addr(id);
4951 0 : } else {
4952 0 : saddr = rdma_get_local_addr(id);
4953 : }
4954 0 : switch (saddr->sa_family) {
4955 : case AF_INET: {
4956 0 : struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr;
4957 :
4958 0 : trid->adrfam = SPDK_NVMF_ADRFAM_IPV4;
4959 0 : inet_ntop(AF_INET, &saddr_in->sin_addr,
4960 0 : trid->traddr, sizeof(trid->traddr));
4961 0 : if (peer) {
4962 0 : port = ntohs(rdma_get_dst_port(id));
4963 0 : } else {
4964 0 : port = ntohs(rdma_get_src_port(id));
4965 : }
4966 0 : snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port);
4967 : break;
4968 0 : }
4969 : case AF_INET6: {
4970 0 : struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr;
4971 0 : trid->adrfam = SPDK_NVMF_ADRFAM_IPV6;
4972 0 : inet_ntop(AF_INET6, &saddr_in->sin6_addr,
4973 0 : trid->traddr, sizeof(trid->traddr));
4974 0 : if (peer) {
4975 0 : port = ntohs(rdma_get_dst_port(id));
4976 0 : } else {
4977 0 : port = ntohs(rdma_get_src_port(id));
4978 : }
4979 0 : snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port);
4980 : break;
4981 0 : }
4982 : default:
4983 0 : return -1;
4984 :
4985 : }
4986 :
4987 0 : return 0;
4988 0 : }
4989 :
4990 : static int
4991 0 : nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair,
4992 : struct spdk_nvme_transport_id *trid)
4993 : {
4994 0 : struct spdk_nvmf_rdma_qpair *rqpair;
4995 :
4996 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4997 :
4998 0 : return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true);
4999 0 : }
5000 :
5001 : static int
5002 0 : nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair,
5003 : struct spdk_nvme_transport_id *trid)
5004 : {
5005 0 : struct spdk_nvmf_rdma_qpair *rqpair;
5006 :
5007 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
5008 :
5009 0 : return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false);
5010 0 : }
5011 :
5012 : static int
5013 0 : nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair,
5014 : struct spdk_nvme_transport_id *trid)
5015 : {
5016 0 : struct spdk_nvmf_rdma_qpair *rqpair;
5017 :
5018 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
5019 :
5020 0 : return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false);
5021 0 : }
5022 :
5023 : void
5024 0 : spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks)
5025 : {
5026 0 : g_nvmf_hooks = *hooks;
5027 0 : }
5028 :
5029 : static void
5030 0 : nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req,
5031 : struct spdk_nvmf_rdma_request *rdma_req_to_abort,
5032 : struct spdk_nvmf_rdma_qpair *rqpair)
5033 : {
5034 0 : rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC;
5035 0 : rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST;
5036 :
5037 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req_to_abort, state_link);
5038 0 : rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
5039 :
5040 0 : req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */
5041 0 : }
5042 :
5043 : static int
5044 0 : _nvmf_rdma_qpair_abort_request(void *ctx)
5045 : {
5046 0 : struct spdk_nvmf_request *req = ctx;
5047 0 : struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF(
5048 : req->req_to_abort, struct spdk_nvmf_rdma_request, req);
5049 0 : struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair,
5050 : struct spdk_nvmf_rdma_qpair, qpair);
5051 0 : int rc;
5052 :
5053 0 : spdk_poller_unregister(&req->poller);
5054 :
5055 0 : switch (rdma_req_to_abort->state) {
5056 : case RDMA_REQUEST_STATE_EXECUTING:
5057 0 : rc = nvmf_ctrlr_abort_request(req);
5058 0 : if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) {
5059 0 : return SPDK_POLLER_BUSY;
5060 : }
5061 0 : break;
5062 :
5063 : case RDMA_REQUEST_STATE_NEED_BUFFER:
5064 0 : STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue,
5065 : &rdma_req_to_abort->req, spdk_nvmf_request, buf_link);
5066 :
5067 0 : nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
5068 0 : break;
5069 :
5070 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING:
5071 0 : STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort,
5072 : spdk_nvmf_rdma_request, state_link);
5073 :
5074 0 : nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
5075 0 : break;
5076 :
5077 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING:
5078 0 : STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort,
5079 : spdk_nvmf_rdma_request, state_link);
5080 :
5081 0 : nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
5082 0 : break;
5083 :
5084 : case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING:
5085 : /* Remove req from the list here to re-use common function */
5086 0 : STAILQ_REMOVE(&rqpair->pending_rdma_send_queue, rdma_req_to_abort,
5087 : spdk_nvmf_rdma_request, state_link);
5088 :
5089 0 : nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
5090 0 : break;
5091 :
5092 : case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
5093 0 : if (spdk_get_ticks() < req->timeout_tsc) {
5094 0 : req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0);
5095 0 : return SPDK_POLLER_BUSY;
5096 : }
5097 0 : break;
5098 :
5099 : default:
5100 0 : break;
5101 : }
5102 :
5103 0 : spdk_nvmf_request_complete(req);
5104 0 : return SPDK_POLLER_BUSY;
5105 0 : }
5106 :
5107 : static void
5108 0 : nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair,
5109 : struct spdk_nvmf_request *req)
5110 : {
5111 0 : struct spdk_nvmf_rdma_qpair *rqpair;
5112 0 : struct spdk_nvmf_rdma_transport *rtransport;
5113 0 : struct spdk_nvmf_transport *transport;
5114 0 : uint16_t cid;
5115 0 : uint32_t i, max_req_count;
5116 0 : struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL, *rdma_req;
5117 :
5118 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
5119 0 : rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
5120 0 : transport = &rtransport->transport;
5121 :
5122 0 : cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid;
5123 0 : max_req_count = rqpair->srq == NULL ? rqpair->max_queue_depth : rqpair->poller->max_srq_depth;
5124 :
5125 0 : for (i = 0; i < max_req_count; i++) {
5126 0 : rdma_req = &rqpair->resources->reqs[i];
5127 : /* When SRQ == NULL, rqpair has its own requests and req.qpair pointer always points to the qpair
5128 : * When SRQ != NULL all rqpairs share common requests and qpair pointer is assigned when we start to
5129 : * process a request. So in both cases all requests which are not in FREE state have valid qpair ptr */
5130 0 : if (rdma_req->state != RDMA_REQUEST_STATE_FREE && rdma_req->req.cmd->nvme_cmd.cid == cid &&
5131 0 : rdma_req->req.qpair == qpair) {
5132 0 : rdma_req_to_abort = rdma_req;
5133 0 : break;
5134 : }
5135 0 : }
5136 :
5137 0 : if (rdma_req_to_abort == NULL) {
5138 0 : spdk_nvmf_request_complete(req);
5139 0 : return;
5140 : }
5141 :
5142 0 : req->req_to_abort = &rdma_req_to_abort->req;
5143 0 : req->timeout_tsc = spdk_get_ticks() +
5144 0 : transport->opts.abort_timeout_sec * spdk_get_ticks_hz();
5145 0 : req->poller = NULL;
5146 :
5147 0 : _nvmf_rdma_qpair_abort_request(req);
5148 0 : }
5149 :
5150 : static void
5151 0 : nvmf_rdma_poll_group_dump_stat(struct spdk_nvmf_transport_poll_group *group,
5152 : struct spdk_json_write_ctx *w)
5153 : {
5154 0 : struct spdk_nvmf_rdma_poll_group *rgroup;
5155 0 : struct spdk_nvmf_rdma_poller *rpoller;
5156 :
5157 0 : assert(w != NULL);
5158 :
5159 0 : rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
5160 :
5161 0 : spdk_json_write_named_uint64(w, "pending_data_buffer", rgroup->stat.pending_data_buffer);
5162 :
5163 0 : spdk_json_write_named_array_begin(w, "devices");
5164 :
5165 0 : TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
5166 0 : spdk_json_write_object_begin(w);
5167 0 : spdk_json_write_named_string(w, "name",
5168 0 : ibv_get_device_name(rpoller->device->context->device));
5169 0 : spdk_json_write_named_uint64(w, "polls",
5170 0 : rpoller->stat.polls);
5171 0 : spdk_json_write_named_uint64(w, "idle_polls",
5172 0 : rpoller->stat.idle_polls);
5173 0 : spdk_json_write_named_uint64(w, "completions",
5174 0 : rpoller->stat.completions);
5175 0 : spdk_json_write_named_uint64(w, "requests",
5176 0 : rpoller->stat.requests);
5177 0 : spdk_json_write_named_uint64(w, "request_latency",
5178 0 : rpoller->stat.request_latency);
5179 0 : spdk_json_write_named_uint64(w, "pending_free_request",
5180 0 : rpoller->stat.pending_free_request);
5181 0 : spdk_json_write_named_uint64(w, "pending_rdma_read",
5182 0 : rpoller->stat.pending_rdma_read);
5183 0 : spdk_json_write_named_uint64(w, "pending_rdma_write",
5184 0 : rpoller->stat.pending_rdma_write);
5185 0 : spdk_json_write_named_uint64(w, "pending_rdma_send",
5186 0 : rpoller->stat.pending_rdma_send);
5187 0 : spdk_json_write_named_uint64(w, "total_send_wrs",
5188 0 : rpoller->stat.qp_stats.send.num_submitted_wrs);
5189 0 : spdk_json_write_named_uint64(w, "send_doorbell_updates",
5190 0 : rpoller->stat.qp_stats.send.doorbell_updates);
5191 0 : spdk_json_write_named_uint64(w, "total_recv_wrs",
5192 0 : rpoller->stat.qp_stats.recv.num_submitted_wrs);
5193 0 : spdk_json_write_named_uint64(w, "recv_doorbell_updates",
5194 0 : rpoller->stat.qp_stats.recv.doorbell_updates);
5195 0 : spdk_json_write_object_end(w);
5196 0 : }
5197 :
5198 0 : spdk_json_write_array_end(w);
5199 0 : }
5200 :
5201 : const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = {
5202 : .name = "RDMA",
5203 : .type = SPDK_NVME_TRANSPORT_RDMA,
5204 : .opts_init = nvmf_rdma_opts_init,
5205 : .create = nvmf_rdma_create,
5206 : .dump_opts = nvmf_rdma_dump_opts,
5207 : .destroy = nvmf_rdma_destroy,
5208 :
5209 : .listen = nvmf_rdma_listen,
5210 : .stop_listen = nvmf_rdma_stop_listen,
5211 : .cdata_init = nvmf_rdma_cdata_init,
5212 :
5213 : .listener_discover = nvmf_rdma_discover,
5214 :
5215 : .poll_group_create = nvmf_rdma_poll_group_create,
5216 : .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group,
5217 : .poll_group_destroy = nvmf_rdma_poll_group_destroy,
5218 : .poll_group_add = nvmf_rdma_poll_group_add,
5219 : .poll_group_remove = nvmf_rdma_poll_group_remove,
5220 : .poll_group_poll = nvmf_rdma_poll_group_poll,
5221 :
5222 : .req_free = nvmf_rdma_request_free,
5223 : .req_complete = nvmf_rdma_request_complete,
5224 :
5225 : .qpair_fini = nvmf_rdma_close_qpair,
5226 : .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid,
5227 : .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid,
5228 : .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid,
5229 : .qpair_abort_request = nvmf_rdma_qpair_abort_request,
5230 :
5231 : .poll_group_dump_stat = nvmf_rdma_poll_group_dump_stat,
5232 : };
5233 :
5234 2 : SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma);
5235 2 : SPDK_LOG_REGISTER_COMPONENT(rdma)
|
