Line data Source code
1 : /* SPDX-License-Identifier: BSD-3-Clause
2 : * Copyright (C) 2016 Intel Corporation. All rights reserved.
3 : * Copyright (c) 2019-2021 Mellanox Technologies LTD. All rights reserved.
4 : * Copyright (c) 2021-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
5 : */
6 :
7 : #include "spdk/stdinc.h"
8 :
9 : #include "spdk/config.h"
10 : #include "spdk/thread.h"
11 : #include "spdk/likely.h"
12 : #include "spdk/nvmf_transport.h"
13 : #include "spdk/string.h"
14 : #include "spdk/trace.h"
15 : #include "spdk/tree.h"
16 : #include "spdk/util.h"
17 :
18 : #include "spdk_internal/assert.h"
19 : #include "spdk/log.h"
20 : #include "spdk_internal/rdma_provider.h"
21 : #include "spdk_internal/rdma_utils.h"
22 :
23 : #include "nvmf_internal.h"
24 : #include "transport.h"
25 :
26 : #include "spdk_internal/trace_defs.h"
27 :
28 : struct spdk_nvme_rdma_hooks g_nvmf_hooks = {};
29 : const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma;
30 :
31 : /*
32 : RDMA Connection Resource Defaults
33 : */
34 : #define NVMF_DEFAULT_MSDBD 16
35 : #define NVMF_DEFAULT_TX_SGE SPDK_NVMF_MAX_SGL_ENTRIES
36 : #define NVMF_DEFAULT_RSP_SGE 1
37 : #define NVMF_DEFAULT_RX_SGE 2
38 :
39 : #define NVMF_RDMA_MAX_EVENTS_PER_POLL 32
40 :
41 : SPDK_STATIC_ASSERT(NVMF_DEFAULT_MSDBD <= SPDK_NVMF_MAX_SGL_ENTRIES,
42 : "MSDBD must not exceed SPDK_NVMF_MAX_SGL_ENTRIES");
43 :
44 : /* The RDMA completion queue size */
45 : #define DEFAULT_NVMF_RDMA_CQ_SIZE 4096
46 : #define MAX_WR_PER_QP(queue_depth) (queue_depth * 3 + 2)
47 :
48 : enum spdk_nvmf_rdma_request_state {
49 : /* The request is not currently in use */
50 : RDMA_REQUEST_STATE_FREE = 0,
51 :
52 : /* Initial state when request first received */
53 : RDMA_REQUEST_STATE_NEW,
54 :
55 : /* The request is queued until a data buffer is available. */
56 : RDMA_REQUEST_STATE_NEED_BUFFER,
57 :
58 : /* The request is waiting on RDMA queue depth availability
59 : * to transfer data from the host to the controller.
60 : */
61 : RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING,
62 :
63 : /* The request is currently transferring data from the host to the controller. */
64 : RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER,
65 :
66 : /* The request is ready to execute at the block device */
67 : RDMA_REQUEST_STATE_READY_TO_EXECUTE,
68 :
69 : /* The request is currently executing at the block device */
70 : RDMA_REQUEST_STATE_EXECUTING,
71 :
72 : /* The request finished executing at the block device */
73 : RDMA_REQUEST_STATE_EXECUTED,
74 :
75 : /* The request is waiting on RDMA queue depth availability
76 : * to transfer data from the controller to the host.
77 : */
78 : RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING,
79 :
80 : /* The request is waiting on RDMA queue depth availability
81 : * to send response to the host.
82 : */
83 : RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING,
84 :
85 : /* The request is ready to send a completion */
86 : RDMA_REQUEST_STATE_READY_TO_COMPLETE,
87 :
88 : /* The request is currently transferring data from the controller to the host. */
89 : RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST,
90 :
91 : /* The request currently has an outstanding completion without an
92 : * associated data transfer.
93 : */
94 : RDMA_REQUEST_STATE_COMPLETING,
95 :
96 : /* The request completed and can be marked free. */
97 : RDMA_REQUEST_STATE_COMPLETED,
98 :
99 : /* Terminator */
100 : RDMA_REQUEST_NUM_STATES,
101 : };
102 :
103 2 : SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA)
104 : {
105 0 : spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r');
106 :
107 0 : struct spdk_trace_tpoint_opts opts[] = {
108 : {
109 : "RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW,
110 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 1,
111 : {
112 : { "qpair", SPDK_TRACE_ARG_TYPE_PTR, 8 },
113 : { "qd", SPDK_TRACE_ARG_TYPE_INT, 4 }
114 : }
115 : },
116 : {
117 : "RDMA_REQ_COMPLETED", TRACE_RDMA_REQUEST_STATE_COMPLETED,
118 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
119 : {
120 : { "qpair", SPDK_TRACE_ARG_TYPE_PTR, 8 },
121 : { "qd", SPDK_TRACE_ARG_TYPE_INT, 4 }
122 : }
123 : },
124 : };
125 :
126 0 : spdk_trace_register_description_ext(opts, SPDK_COUNTOF(opts));
127 0 : spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER,
128 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
129 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
130 0 : spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H",
131 : TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING,
132 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
133 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
134 0 : spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C",
135 : TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING,
136 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
137 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
138 0 : spdk_trace_register_description("RDMA_REQ_TX_H2C",
139 : TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER,
140 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
141 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
142 0 : spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE",
143 : TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE,
144 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
145 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
146 0 : spdk_trace_register_description("RDMA_REQ_EXECUTING",
147 : TRACE_RDMA_REQUEST_STATE_EXECUTING,
148 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
149 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
150 0 : spdk_trace_register_description("RDMA_REQ_EXECUTED",
151 : TRACE_RDMA_REQUEST_STATE_EXECUTED,
152 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
153 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
154 0 : spdk_trace_register_description("RDMA_REQ_RDY2COMPL_PEND",
155 : TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING,
156 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
157 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
158 0 : spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL",
159 : TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE,
160 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
161 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
162 0 : spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H",
163 : TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST,
164 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
165 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
166 0 : spdk_trace_register_description("RDMA_REQ_COMPLETING",
167 : TRACE_RDMA_REQUEST_STATE_COMPLETING,
168 : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
169 : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
170 :
171 0 : spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE,
172 : OWNER_TYPE_NONE, OBJECT_NONE, 0,
173 : SPDK_TRACE_ARG_TYPE_INT, "");
174 0 : spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT,
175 : OWNER_TYPE_NONE, OBJECT_NONE, 0,
176 : SPDK_TRACE_ARG_TYPE_INT, "type");
177 0 : spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT,
178 : OWNER_TYPE_NONE, OBJECT_NONE, 0,
179 : SPDK_TRACE_ARG_TYPE_INT, "type");
180 0 : spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT,
181 : OWNER_TYPE_NONE, OBJECT_NONE, 0,
182 : SPDK_TRACE_ARG_TYPE_INT, "");
183 0 : spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY,
184 : OWNER_TYPE_NONE, OBJECT_NONE, 0,
185 : SPDK_TRACE_ARG_TYPE_INT, "");
186 :
187 0 : spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_START, OBJECT_NVMF_RDMA_IO, 1);
188 0 : spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_DONE, OBJECT_NVMF_RDMA_IO, 0);
189 0 : }
190 :
191 : enum spdk_nvmf_rdma_wr_type {
192 : RDMA_WR_TYPE_RECV,
193 : RDMA_WR_TYPE_SEND,
194 : RDMA_WR_TYPE_DATA,
195 : };
196 :
197 : struct spdk_nvmf_rdma_wr {
198 : /* Uses enum spdk_nvmf_rdma_wr_type */
199 : uint8_t type;
200 : };
201 :
202 : /* This structure holds commands as they are received off the wire.
203 : * It must be dynamically paired with a full request object
204 : * (spdk_nvmf_rdma_request) to service a request. It is separate
205 : * from the request because RDMA does not appear to order
206 : * completions, so occasionally we'll get a new incoming
207 : * command when there aren't any free request objects.
208 : */
209 : struct spdk_nvmf_rdma_recv {
210 : struct ibv_recv_wr wr;
211 : struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE];
212 :
213 : struct spdk_nvmf_rdma_qpair *qpair;
214 :
215 : /* In-capsule data buffer */
216 : uint8_t *buf;
217 :
218 : struct spdk_nvmf_rdma_wr rdma_wr;
219 : uint64_t receive_tsc;
220 :
221 : STAILQ_ENTRY(spdk_nvmf_rdma_recv) link;
222 : };
223 :
224 : struct spdk_nvmf_rdma_request_data {
225 : struct ibv_send_wr wr;
226 : struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES];
227 : };
228 :
229 : struct spdk_nvmf_rdma_request {
230 : struct spdk_nvmf_request req;
231 :
232 : bool fused_failed;
233 :
234 : struct spdk_nvmf_rdma_wr data_wr;
235 : struct spdk_nvmf_rdma_wr rsp_wr;
236 :
237 : /* Uses enum spdk_nvmf_rdma_request_state */
238 : uint8_t state;
239 :
240 : /* Data offset in req.iov */
241 : uint32_t offset;
242 :
243 : struct spdk_nvmf_rdma_recv *recv;
244 :
245 : struct {
246 : struct ibv_send_wr wr;
247 : struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE];
248 : } rsp;
249 :
250 : uint16_t iovpos;
251 : uint16_t num_outstanding_data_wr;
252 : /* Used to split Write IO with multi SGL payload */
253 : uint16_t num_remaining_data_wr;
254 : uint64_t receive_tsc;
255 : struct spdk_nvmf_rdma_request *fused_pair;
256 : STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link;
257 : struct ibv_send_wr *remaining_tranfer_in_wrs;
258 : struct ibv_send_wr *transfer_wr;
259 : struct spdk_nvmf_rdma_request_data data;
260 : };
261 :
262 : struct spdk_nvmf_rdma_resource_opts {
263 : struct spdk_nvmf_rdma_qpair *qpair;
264 : /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */
265 : void *qp;
266 : struct spdk_rdma_utils_mem_map *map;
267 : uint32_t max_queue_depth;
268 : uint32_t in_capsule_data_size;
269 : bool shared;
270 : };
271 :
272 : struct spdk_nvmf_rdma_resources {
273 : /* Array of size "max_queue_depth" containing RDMA requests. */
274 : struct spdk_nvmf_rdma_request *reqs;
275 :
276 : /* Array of size "max_queue_depth" containing RDMA recvs. */
277 : struct spdk_nvmf_rdma_recv *recvs;
278 :
279 : /* Array of size "max_queue_depth" containing 64 byte capsules
280 : * used for receive.
281 : */
282 : union nvmf_h2c_msg *cmds;
283 :
284 : /* Array of size "max_queue_depth" containing 16 byte completions
285 : * to be sent back to the user.
286 : */
287 : union nvmf_c2h_msg *cpls;
288 :
289 : /* Array of size "max_queue_depth * InCapsuleDataSize" containing
290 : * buffers to be used for in capsule data.
291 : */
292 : void *bufs;
293 :
294 : /* Receives that are waiting for a request object */
295 : STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue;
296 :
297 : /* Queue to track free requests */
298 : STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue;
299 : };
300 :
301 : typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair);
302 :
303 : typedef void (*spdk_poller_destroy_cb)(void *ctx);
304 :
305 : struct spdk_nvmf_rdma_ibv_event_ctx {
306 : struct spdk_nvmf_rdma_qpair *rqpair;
307 : spdk_nvmf_rdma_qpair_ibv_event cb_fn;
308 : /* Link to other ibv events associated with this qpair */
309 : STAILQ_ENTRY(spdk_nvmf_rdma_ibv_event_ctx) link;
310 : };
311 :
312 : struct spdk_nvmf_rdma_qpair {
313 : struct spdk_nvmf_qpair qpair;
314 :
315 : struct spdk_nvmf_rdma_device *device;
316 : struct spdk_nvmf_rdma_poller *poller;
317 :
318 : struct spdk_rdma_provider_qp *rdma_qp;
319 : struct rdma_cm_id *cm_id;
320 : struct spdk_rdma_provider_srq *srq;
321 : struct rdma_cm_id *listen_id;
322 :
323 : /* Cache the QP number to improve QP search by RB tree. */
324 : uint32_t qp_num;
325 :
326 : /* The maximum number of I/O outstanding on this connection at one time */
327 : uint16_t max_queue_depth;
328 :
329 : /* The maximum number of active RDMA READ and ATOMIC operations at one time */
330 : uint16_t max_read_depth;
331 :
332 : /* The maximum number of RDMA SEND operations at one time */
333 : uint32_t max_send_depth;
334 :
335 : /* The current number of outstanding WRs from this qpair's
336 : * recv queue. Should not exceed device->attr.max_queue_depth.
337 : */
338 : uint16_t current_recv_depth;
339 :
340 : /* The current number of active RDMA READ operations */
341 : uint16_t current_read_depth;
342 :
343 : /* The current number of posted WRs from this qpair's
344 : * send queue. Should not exceed max_send_depth.
345 : */
346 : uint32_t current_send_depth;
347 :
348 : /* The maximum number of SGEs per WR on the send queue */
349 : uint32_t max_send_sge;
350 :
351 : /* The maximum number of SGEs per WR on the recv queue */
352 : uint32_t max_recv_sge;
353 :
354 : struct spdk_nvmf_rdma_resources *resources;
355 :
356 : STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue;
357 :
358 : STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue;
359 :
360 : STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_send_queue;
361 :
362 : /* Number of requests not in the free state */
363 : uint32_t qd;
364 :
365 : bool ibv_in_error_state;
366 :
367 : RB_ENTRY(spdk_nvmf_rdma_qpair) node;
368 :
369 : STAILQ_ENTRY(spdk_nvmf_rdma_qpair) recv_link;
370 :
371 : STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link;
372 :
373 : /* Points to the a request that has fuse bits set to
374 : * SPDK_NVME_CMD_FUSE_FIRST, when the qpair is waiting
375 : * for the request that has SPDK_NVME_CMD_FUSE_SECOND.
376 : */
377 : struct spdk_nvmf_rdma_request *fused_first;
378 :
379 : /*
380 : * io_channel which is used to destroy qpair when it is removed from poll group
381 : */
382 : struct spdk_io_channel *destruct_channel;
383 :
384 : /* List of ibv async events */
385 : STAILQ_HEAD(, spdk_nvmf_rdma_ibv_event_ctx) ibv_events;
386 :
387 : /* Lets us know that we have received the last_wqe event. */
388 : bool last_wqe_reached;
389 :
390 : /* Indicate that nvmf_rdma_close_qpair is called */
391 : bool to_close;
392 : };
393 :
394 : struct spdk_nvmf_rdma_poller_stat {
395 : uint64_t completions;
396 : uint64_t polls;
397 : uint64_t idle_polls;
398 : uint64_t requests;
399 : uint64_t request_latency;
400 : uint64_t pending_free_request;
401 : uint64_t pending_rdma_read;
402 : uint64_t pending_rdma_write;
403 : uint64_t pending_rdma_send;
404 : struct spdk_rdma_provider_qp_stats qp_stats;
405 : };
406 :
407 : struct spdk_nvmf_rdma_poller {
408 : struct spdk_nvmf_rdma_device *device;
409 : struct spdk_nvmf_rdma_poll_group *group;
410 :
411 : int num_cqe;
412 : int required_num_wr;
413 : struct ibv_cq *cq;
414 :
415 : /* The maximum number of I/O outstanding on the shared receive queue at one time */
416 : uint16_t max_srq_depth;
417 : bool need_destroy;
418 :
419 : /* Shared receive queue */
420 : struct spdk_rdma_provider_srq *srq;
421 :
422 : struct spdk_nvmf_rdma_resources *resources;
423 : struct spdk_nvmf_rdma_poller_stat stat;
424 :
425 : spdk_poller_destroy_cb destroy_cb;
426 : void *destroy_cb_ctx;
427 :
428 : RB_HEAD(qpairs_tree, spdk_nvmf_rdma_qpair) qpairs;
429 :
430 : STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv;
431 :
432 : STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send;
433 :
434 : TAILQ_ENTRY(spdk_nvmf_rdma_poller) link;
435 : };
436 :
437 : struct spdk_nvmf_rdma_poll_group_stat {
438 : uint64_t pending_data_buffer;
439 : };
440 :
441 : struct spdk_nvmf_rdma_poll_group {
442 : struct spdk_nvmf_transport_poll_group group;
443 : struct spdk_nvmf_rdma_poll_group_stat stat;
444 : TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers;
445 : TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link;
446 : };
447 :
448 : struct spdk_nvmf_rdma_conn_sched {
449 : struct spdk_nvmf_rdma_poll_group *next_admin_pg;
450 : struct spdk_nvmf_rdma_poll_group *next_io_pg;
451 : };
452 :
453 : /* Assuming rdma_cm uses just one protection domain per ibv_context. */
454 : struct spdk_nvmf_rdma_device {
455 : struct ibv_device_attr attr;
456 : struct ibv_context *context;
457 :
458 : struct spdk_rdma_utils_mem_map *map;
459 : struct ibv_pd *pd;
460 :
461 : int num_srq;
462 : bool need_destroy;
463 : bool ready_to_destroy;
464 : bool is_ready;
465 :
466 : TAILQ_ENTRY(spdk_nvmf_rdma_device) link;
467 : };
468 :
469 : struct spdk_nvmf_rdma_port {
470 : const struct spdk_nvme_transport_id *trid;
471 : struct rdma_cm_id *id;
472 : struct spdk_nvmf_rdma_device *device;
473 : TAILQ_ENTRY(spdk_nvmf_rdma_port) link;
474 : };
475 :
476 : struct rdma_transport_opts {
477 : int num_cqe;
478 : uint32_t max_srq_depth;
479 : bool no_srq;
480 : bool no_wr_batching;
481 : int acceptor_backlog;
482 : };
483 :
484 : struct spdk_nvmf_rdma_transport {
485 : struct spdk_nvmf_transport transport;
486 : struct rdma_transport_opts rdma_opts;
487 :
488 : struct spdk_nvmf_rdma_conn_sched conn_sched;
489 :
490 : struct rdma_event_channel *event_channel;
491 :
492 : struct spdk_mempool *data_wr_pool;
493 :
494 : struct spdk_poller *accept_poller;
495 :
496 : /* fields used to poll RDMA/IB events */
497 : nfds_t npoll_fds;
498 : struct pollfd *poll_fds;
499 :
500 : TAILQ_HEAD(, spdk_nvmf_rdma_device) devices;
501 : TAILQ_HEAD(, spdk_nvmf_rdma_port) ports;
502 : TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups;
503 :
504 : /* ports that are removed unexpectedly and need retry listen */
505 : TAILQ_HEAD(, spdk_nvmf_rdma_port) retry_ports;
506 : };
507 :
508 : struct poller_manage_ctx {
509 : struct spdk_nvmf_rdma_transport *rtransport;
510 : struct spdk_nvmf_rdma_poll_group *rgroup;
511 : struct spdk_nvmf_rdma_poller *rpoller;
512 : struct spdk_nvmf_rdma_device *device;
513 :
514 : struct spdk_thread *thread;
515 : volatile int *inflight_op_counter;
516 : };
517 :
518 : static const struct spdk_json_object_decoder rdma_transport_opts_decoder[] = {
519 : {
520 : "num_cqe", offsetof(struct rdma_transport_opts, num_cqe),
521 : spdk_json_decode_int32, true
522 : },
523 : {
524 : "max_srq_depth", offsetof(struct rdma_transport_opts, max_srq_depth),
525 : spdk_json_decode_uint32, true
526 : },
527 : {
528 : "no_srq", offsetof(struct rdma_transport_opts, no_srq),
529 : spdk_json_decode_bool, true
530 : },
531 : {
532 : "no_wr_batching", offsetof(struct rdma_transport_opts, no_wr_batching),
533 : spdk_json_decode_bool, true
534 : },
535 : {
536 : "acceptor_backlog", offsetof(struct rdma_transport_opts, acceptor_backlog),
537 : spdk_json_decode_int32, true
538 : },
539 : };
540 :
541 : static int
542 2 : nvmf_rdma_qpair_compare(struct spdk_nvmf_rdma_qpair *rqpair1, struct spdk_nvmf_rdma_qpair *rqpair2)
543 : {
544 2 : return rqpair1->qp_num < rqpair2->qp_num ? -1 : rqpair1->qp_num > rqpair2->qp_num;
545 : }
546 :
547 0 : RB_GENERATE_STATIC(qpairs_tree, spdk_nvmf_rdma_qpair, node, nvmf_rdma_qpair_compare);
548 :
549 : static bool nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport,
550 : struct spdk_nvmf_rdma_request *rdma_req);
551 :
552 : static void _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport,
553 : struct spdk_nvmf_rdma_poller *rpoller);
554 :
555 : static void _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport,
556 : struct spdk_nvmf_rdma_poller *rpoller);
557 :
558 : static void _nvmf_rdma_remove_destroyed_device(void *c);
559 :
560 : static inline enum spdk_nvme_media_error_status_code
561 0 : nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) {
562 : enum spdk_nvme_media_error_status_code result;
563 0 : switch (err_type)
564 : {
565 0 : case SPDK_DIF_REFTAG_ERROR:
566 0 : result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR;
567 0 : break;
568 0 : case SPDK_DIF_APPTAG_ERROR:
569 0 : result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR;
570 0 : break;
571 0 : case SPDK_DIF_GUARD_ERROR:
572 0 : result = SPDK_NVME_SC_GUARD_CHECK_ERROR;
573 0 : break;
574 0 : default:
575 0 : SPDK_UNREACHABLE();
576 : }
577 :
578 0 : return result;
579 : }
580 :
581 : /*
582 : * Return data_wrs to pool starting from \b data_wr
583 : * Request's own response and data WR are excluded
584 : */
585 : static void
586 7 : _nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req,
587 : struct ibv_send_wr *data_wr,
588 : struct spdk_mempool *pool)
589 : {
590 7 : struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES];
591 : struct spdk_nvmf_rdma_request_data *nvmf_data;
592 : struct ibv_send_wr *next_send_wr;
593 7 : uint64_t req_wrid = (uint64_t)&rdma_req->data_wr;
594 7 : uint32_t num_wrs = 0;
595 :
596 15 : while (data_wr && data_wr->wr_id == req_wrid) {
597 8 : nvmf_data = SPDK_CONTAINEROF(data_wr, struct spdk_nvmf_rdma_request_data, wr);
598 8 : memset(nvmf_data->sgl, 0, sizeof(data_wr->sg_list[0]) * data_wr->num_sge);
599 8 : data_wr->num_sge = 0;
600 8 : next_send_wr = data_wr->next;
601 8 : if (data_wr != &rdma_req->data.wr) {
602 1 : data_wr->next = NULL;
603 1 : assert(num_wrs < SPDK_NVMF_MAX_SGL_ENTRIES);
604 1 : work_requests[num_wrs] = nvmf_data;
605 1 : num_wrs++;
606 : }
607 8 : data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : next_send_wr;
608 : }
609 :
610 7 : if (num_wrs) {
611 1 : spdk_mempool_put_bulk(pool, (void **) work_requests, num_wrs);
612 : }
613 7 : }
614 :
615 : static void
616 7 : nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req,
617 : struct spdk_nvmf_rdma_transport *rtransport)
618 : {
619 7 : rdma_req->num_outstanding_data_wr = 0;
620 :
621 7 : _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool);
622 :
623 7 : if (rdma_req->remaining_tranfer_in_wrs) {
624 0 : _nvmf_rdma_request_free_data(rdma_req, rdma_req->remaining_tranfer_in_wrs,
625 : rtransport->data_wr_pool);
626 0 : rdma_req->remaining_tranfer_in_wrs = NULL;
627 : }
628 :
629 7 : rdma_req->data.wr.next = NULL;
630 7 : rdma_req->rsp.wr.next = NULL;
631 7 : }
632 :
633 : static void
634 0 : nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req)
635 : {
636 0 : SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool);
637 0 : if (req->req.cmd) {
638 0 : SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode);
639 : }
640 0 : if (req->recv) {
641 0 : SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id);
642 : }
643 0 : }
644 :
645 : static void
646 0 : nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair)
647 : {
648 : int i;
649 :
650 0 : SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid);
651 0 : for (i = 0; i < rqpair->max_queue_depth; i++) {
652 0 : if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) {
653 0 : nvmf_rdma_dump_request(&rqpair->resources->reqs[i]);
654 : }
655 : }
656 0 : }
657 :
658 : static void
659 1 : nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources)
660 : {
661 1 : spdk_free(resources->cmds);
662 1 : spdk_free(resources->cpls);
663 1 : spdk_free(resources->bufs);
664 1 : spdk_free(resources->reqs);
665 1 : spdk_free(resources->recvs);
666 1 : free(resources);
667 1 : }
668 :
669 :
670 : static struct spdk_nvmf_rdma_resources *
671 1 : nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts)
672 : {
673 : struct spdk_nvmf_rdma_resources *resources;
674 : struct spdk_nvmf_rdma_request *rdma_req;
675 : struct spdk_nvmf_rdma_recv *rdma_recv;
676 1 : struct spdk_rdma_provider_qp *qp = NULL;
677 1 : struct spdk_rdma_provider_srq *srq = NULL;
678 1 : struct ibv_recv_wr *bad_wr = NULL;
679 1 : struct spdk_rdma_utils_memory_translation translation;
680 : uint32_t i;
681 1 : int rc = 0;
682 :
683 1 : resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources));
684 1 : if (!resources) {
685 0 : SPDK_ERRLOG("Unable to allocate resources for receive queue.\n");
686 0 : return NULL;
687 : }
688 :
689 1 : resources->reqs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->reqs),
690 : 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
691 1 : resources->recvs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->recvs),
692 : 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
693 1 : resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds),
694 : 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
695 1 : resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls),
696 : 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
697 :
698 1 : if (opts->in_capsule_data_size > 0) {
699 1 : resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size,
700 : 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY,
701 : SPDK_MALLOC_DMA);
702 : }
703 :
704 1 : if (!resources->reqs || !resources->recvs || !resources->cmds ||
705 1 : !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) {
706 0 : SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n");
707 0 : goto cleanup;
708 : }
709 :
710 1 : SPDK_DEBUGLOG(rdma, "Command Array: %p Length: %lx\n",
711 : resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds));
712 1 : SPDK_DEBUGLOG(rdma, "Completion Array: %p Length: %lx\n",
713 : resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls));
714 1 : if (resources->bufs) {
715 1 : SPDK_DEBUGLOG(rdma, "In Capsule Data Array: %p Length: %x\n",
716 : resources->bufs, opts->max_queue_depth *
717 : opts->in_capsule_data_size);
718 : }
719 :
720 : /* Initialize queues */
721 1 : STAILQ_INIT(&resources->incoming_queue);
722 1 : STAILQ_INIT(&resources->free_queue);
723 :
724 1 : if (opts->shared) {
725 1 : srq = (struct spdk_rdma_provider_srq *)opts->qp;
726 : } else {
727 0 : qp = (struct spdk_rdma_provider_qp *)opts->qp;
728 : }
729 :
730 129 : for (i = 0; i < opts->max_queue_depth; i++) {
731 128 : rdma_recv = &resources->recvs[i];
732 128 : rdma_recv->qpair = opts->qpair;
733 :
734 : /* Set up memory to receive commands */
735 128 : if (resources->bufs) {
736 128 : rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i *
737 128 : opts->in_capsule_data_size));
738 : }
739 :
740 128 : rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV;
741 :
742 128 : rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i];
743 128 : rdma_recv->sgl[0].length = sizeof(resources->cmds[i]);
744 128 : rc = spdk_rdma_utils_get_translation(opts->map, &resources->cmds[i], sizeof(resources->cmds[i]),
745 : &translation);
746 128 : if (rc) {
747 0 : goto cleanup;
748 : }
749 128 : rdma_recv->sgl[0].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation);
750 128 : rdma_recv->wr.num_sge = 1;
751 :
752 128 : if (rdma_recv->buf) {
753 128 : rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf;
754 128 : rdma_recv->sgl[1].length = opts->in_capsule_data_size;
755 128 : rc = spdk_rdma_utils_get_translation(opts->map, rdma_recv->buf, opts->in_capsule_data_size,
756 : &translation);
757 128 : if (rc) {
758 0 : goto cleanup;
759 : }
760 128 : rdma_recv->sgl[1].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation);
761 128 : rdma_recv->wr.num_sge++;
762 : }
763 :
764 128 : rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr;
765 128 : rdma_recv->wr.sg_list = rdma_recv->sgl;
766 128 : if (srq) {
767 0 : spdk_rdma_provider_srq_queue_recv_wrs(srq, &rdma_recv->wr);
768 : } else {
769 128 : spdk_rdma_provider_qp_queue_recv_wrs(qp, &rdma_recv->wr);
770 : }
771 : }
772 :
773 129 : for (i = 0; i < opts->max_queue_depth; i++) {
774 128 : rdma_req = &resources->reqs[i];
775 :
776 128 : if (opts->qpair != NULL) {
777 128 : rdma_req->req.qpair = &opts->qpair->qpair;
778 : } else {
779 0 : rdma_req->req.qpair = NULL;
780 : }
781 128 : rdma_req->req.cmd = NULL;
782 128 : rdma_req->req.iovcnt = 0;
783 128 : rdma_req->req.stripped_data = NULL;
784 :
785 : /* Set up memory to send responses */
786 128 : rdma_req->req.rsp = &resources->cpls[i];
787 :
788 128 : rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i];
789 128 : rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]);
790 128 : rc = spdk_rdma_utils_get_translation(opts->map, &resources->cpls[i], sizeof(resources->cpls[i]),
791 : &translation);
792 128 : if (rc) {
793 0 : goto cleanup;
794 : }
795 128 : rdma_req->rsp.sgl[0].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation);
796 :
797 128 : rdma_req->rsp_wr.type = RDMA_WR_TYPE_SEND;
798 128 : rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp_wr;
799 128 : rdma_req->rsp.wr.next = NULL;
800 128 : rdma_req->rsp.wr.opcode = IBV_WR_SEND;
801 128 : rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED;
802 128 : rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl;
803 128 : rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl);
804 :
805 : /* Set up memory for data buffers */
806 128 : rdma_req->data_wr.type = RDMA_WR_TYPE_DATA;
807 128 : rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data_wr;
808 128 : rdma_req->data.wr.next = NULL;
809 128 : rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED;
810 128 : rdma_req->data.wr.sg_list = rdma_req->data.sgl;
811 128 : rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl);
812 :
813 : /* Initialize request state to FREE */
814 128 : rdma_req->state = RDMA_REQUEST_STATE_FREE;
815 128 : STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link);
816 : }
817 :
818 1 : if (srq) {
819 0 : rc = spdk_rdma_provider_srq_flush_recv_wrs(srq, &bad_wr);
820 : } else {
821 1 : rc = spdk_rdma_provider_qp_flush_recv_wrs(qp, &bad_wr);
822 : }
823 :
824 1 : if (rc) {
825 0 : goto cleanup;
826 : }
827 :
828 1 : return resources;
829 :
830 0 : cleanup:
831 0 : nvmf_rdma_resources_destroy(resources);
832 0 : return NULL;
833 : }
834 :
835 : static void
836 0 : nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair)
837 : {
838 : struct spdk_nvmf_rdma_ibv_event_ctx *ctx, *tctx;
839 0 : STAILQ_FOREACH_SAFE(ctx, &rqpair->ibv_events, link, tctx) {
840 0 : ctx->rqpair = NULL;
841 : /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_ibv_event */
842 0 : STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link);
843 : }
844 0 : }
845 :
846 : static void nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller);
847 :
848 : static void
849 0 : nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair)
850 : {
851 : struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp;
852 0 : struct ibv_recv_wr *bad_recv_wr = NULL;
853 : int rc;
854 :
855 0 : spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair);
856 :
857 0 : if (rqpair->qd != 0) {
858 0 : struct spdk_nvmf_qpair *qpair = &rqpair->qpair;
859 0 : struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport,
860 : struct spdk_nvmf_rdma_transport, transport);
861 : struct spdk_nvmf_rdma_request *req;
862 0 : uint32_t i, max_req_count = 0;
863 :
864 0 : SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd);
865 :
866 0 : if (rqpair->srq == NULL) {
867 0 : nvmf_rdma_dump_qpair_contents(rqpair);
868 0 : max_req_count = rqpair->max_queue_depth;
869 0 : } else if (rqpair->poller && rqpair->resources) {
870 0 : max_req_count = rqpair->poller->max_srq_depth;
871 : }
872 :
873 0 : SPDK_DEBUGLOG(rdma, "Release incomplete requests\n");
874 0 : for (i = 0; i < max_req_count; i++) {
875 0 : req = &rqpair->resources->reqs[i];
876 0 : if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) {
877 : /* nvmf_rdma_request_process checks qpair ibv and internal state
878 : * and completes a request */
879 0 : nvmf_rdma_request_process(rtransport, req);
880 : }
881 : }
882 0 : assert(rqpair->qd == 0);
883 : }
884 :
885 0 : if (rqpair->poller) {
886 0 : RB_REMOVE(qpairs_tree, &rqpair->poller->qpairs, rqpair);
887 :
888 0 : if (rqpair->srq != NULL && rqpair->resources != NULL) {
889 : /* Drop all received but unprocessed commands for this queue and return them to SRQ */
890 0 : STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) {
891 0 : if (rqpair == rdma_recv->qpair) {
892 0 : STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link);
893 0 : spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, &rdma_recv->wr);
894 0 : rc = spdk_rdma_provider_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr);
895 0 : if (rc) {
896 0 : SPDK_ERRLOG("Unable to re-post rx descriptor\n");
897 : }
898 : }
899 : }
900 : }
901 : }
902 :
903 0 : if (rqpair->cm_id) {
904 0 : if (rqpair->rdma_qp != NULL) {
905 0 : spdk_rdma_provider_qp_destroy(rqpair->rdma_qp);
906 0 : rqpair->rdma_qp = NULL;
907 : }
908 :
909 0 : if (rqpair->poller != NULL && rqpair->srq == NULL) {
910 0 : rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth);
911 : }
912 : }
913 :
914 0 : if (rqpair->srq == NULL && rqpair->resources != NULL) {
915 0 : nvmf_rdma_resources_destroy(rqpair->resources);
916 : }
917 :
918 0 : nvmf_rdma_qpair_clean_ibv_events(rqpair);
919 :
920 0 : if (rqpair->destruct_channel) {
921 0 : spdk_put_io_channel(rqpair->destruct_channel);
922 0 : rqpair->destruct_channel = NULL;
923 : }
924 :
925 0 : if (rqpair->poller && rqpair->poller->need_destroy && RB_EMPTY(&rqpair->poller->qpairs)) {
926 0 : nvmf_rdma_poller_destroy(rqpair->poller);
927 : }
928 :
929 : /* destroy cm_id last so cma device will not be freed before we destroy the cq. */
930 0 : if (rqpair->cm_id) {
931 0 : rdma_destroy_id(rqpair->cm_id);
932 : }
933 :
934 0 : free(rqpair);
935 0 : }
936 :
937 : static int
938 5 : nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device)
939 : {
940 : struct spdk_nvmf_rdma_poller *rpoller;
941 : int rc, num_cqe, required_num_wr;
942 :
943 : /* Enlarge CQ size dynamically */
944 5 : rpoller = rqpair->poller;
945 5 : required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth);
946 5 : num_cqe = rpoller->num_cqe;
947 5 : if (num_cqe < required_num_wr) {
948 4 : num_cqe = spdk_max(num_cqe * 2, required_num_wr);
949 4 : num_cqe = spdk_min(num_cqe, device->attr.max_cqe);
950 : }
951 :
952 5 : if (rpoller->num_cqe != num_cqe) {
953 4 : if (device->context->device->transport_type == IBV_TRANSPORT_IWARP) {
954 1 : SPDK_ERRLOG("iWARP doesn't support CQ resize. Current capacity %u, required %u\n"
955 : "Using CQ of insufficient size may lead to CQ overrun\n", rpoller->num_cqe, num_cqe);
956 1 : return -1;
957 : }
958 3 : if (required_num_wr > device->attr.max_cqe) {
959 1 : SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n",
960 : required_num_wr, device->attr.max_cqe);
961 1 : return -1;
962 : }
963 :
964 2 : SPDK_DEBUGLOG(rdma, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe);
965 2 : rc = ibv_resize_cq(rpoller->cq, num_cqe);
966 2 : if (rc) {
967 1 : SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno));
968 1 : return -1;
969 : }
970 :
971 1 : rpoller->num_cqe = num_cqe;
972 : }
973 :
974 2 : rpoller->required_num_wr = required_num_wr;
975 2 : return 0;
976 : }
977 :
978 : static int
979 0 : nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair)
980 : {
981 : struct spdk_nvmf_rdma_qpair *rqpair;
982 : struct spdk_nvmf_rdma_transport *rtransport;
983 : struct spdk_nvmf_transport *transport;
984 0 : struct spdk_nvmf_rdma_resource_opts opts;
985 : struct spdk_nvmf_rdma_device *device;
986 0 : struct spdk_rdma_provider_qp_init_attr qp_init_attr = {};
987 :
988 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
989 0 : device = rqpair->device;
990 :
991 0 : qp_init_attr.qp_context = rqpair;
992 0 : qp_init_attr.pd = device->pd;
993 0 : qp_init_attr.send_cq = rqpair->poller->cq;
994 0 : qp_init_attr.recv_cq = rqpair->poller->cq;
995 :
996 0 : if (rqpair->srq) {
997 0 : qp_init_attr.srq = rqpair->srq->srq;
998 : } else {
999 0 : qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth;
1000 : }
1001 :
1002 : /* SEND, READ, and WRITE operations */
1003 0 : qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2;
1004 0 : qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE);
1005 0 : qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE);
1006 0 : qp_init_attr.stats = &rqpair->poller->stat.qp_stats;
1007 :
1008 0 : if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) {
1009 0 : SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n");
1010 0 : goto error;
1011 : }
1012 :
1013 0 : rqpair->rdma_qp = spdk_rdma_provider_qp_create(rqpair->cm_id, &qp_init_attr);
1014 0 : if (!rqpair->rdma_qp) {
1015 0 : goto error;
1016 : }
1017 :
1018 0 : rqpair->qp_num = rqpair->rdma_qp->qp->qp_num;
1019 :
1020 0 : rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2),
1021 : qp_init_attr.cap.max_send_wr);
1022 0 : rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge);
1023 0 : rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge);
1024 0 : spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair);
1025 0 : SPDK_DEBUGLOG(rdma, "New RDMA Connection: %p\n", qpair);
1026 :
1027 0 : if (rqpair->poller->srq == NULL) {
1028 0 : rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
1029 0 : transport = &rtransport->transport;
1030 :
1031 0 : opts.qp = rqpair->rdma_qp;
1032 0 : opts.map = device->map;
1033 0 : opts.qpair = rqpair;
1034 0 : opts.shared = false;
1035 0 : opts.max_queue_depth = rqpair->max_queue_depth;
1036 0 : opts.in_capsule_data_size = transport->opts.in_capsule_data_size;
1037 :
1038 0 : rqpair->resources = nvmf_rdma_resources_create(&opts);
1039 :
1040 0 : if (!rqpair->resources) {
1041 0 : SPDK_ERRLOG("Unable to allocate resources for receive queue.\n");
1042 0 : rdma_destroy_qp(rqpair->cm_id);
1043 0 : goto error;
1044 : }
1045 : } else {
1046 0 : rqpair->resources = rqpair->poller->resources;
1047 : }
1048 :
1049 0 : rqpair->current_recv_depth = 0;
1050 0 : STAILQ_INIT(&rqpair->pending_rdma_read_queue);
1051 0 : STAILQ_INIT(&rqpair->pending_rdma_write_queue);
1052 0 : STAILQ_INIT(&rqpair->pending_rdma_send_queue);
1053 0 : rqpair->qpair.queue_depth = 0;
1054 :
1055 0 : return 0;
1056 :
1057 0 : error:
1058 0 : rdma_destroy_id(rqpair->cm_id);
1059 0 : rqpair->cm_id = NULL;
1060 0 : return -1;
1061 : }
1062 :
1063 : /* Append the given recv wr structure to the resource structs outstanding recvs list. */
1064 : /* This function accepts either a single wr or the first wr in a linked list. */
1065 : static void
1066 6 : nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first)
1067 : {
1068 6 : struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
1069 : struct spdk_nvmf_rdma_transport, transport);
1070 :
1071 6 : if (rqpair->srq != NULL) {
1072 0 : spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, first);
1073 : } else {
1074 6 : if (spdk_rdma_provider_qp_queue_recv_wrs(rqpair->rdma_qp, first)) {
1075 6 : STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link);
1076 : }
1077 : }
1078 :
1079 6 : if (rtransport->rdma_opts.no_wr_batching) {
1080 0 : _poller_submit_recvs(rtransport, rqpair->poller);
1081 : }
1082 6 : }
1083 :
1084 : static int
1085 4 : request_transfer_in(struct spdk_nvmf_request *req)
1086 : {
1087 : struct spdk_nvmf_rdma_request *rdma_req;
1088 : struct spdk_nvmf_qpair *qpair;
1089 : struct spdk_nvmf_rdma_qpair *rqpair;
1090 : struct spdk_nvmf_rdma_transport *rtransport;
1091 :
1092 4 : qpair = req->qpair;
1093 4 : rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
1094 4 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
1095 4 : rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
1096 : struct spdk_nvmf_rdma_transport, transport);
1097 :
1098 4 : assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
1099 4 : assert(rdma_req != NULL);
1100 :
1101 4 : if (spdk_rdma_provider_qp_queue_send_wrs(rqpair->rdma_qp, rdma_req->transfer_wr)) {
1102 4 : STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link);
1103 : }
1104 4 : if (rtransport->rdma_opts.no_wr_batching) {
1105 0 : _poller_submit_sends(rtransport, rqpair->poller);
1106 : }
1107 :
1108 4 : assert(rqpair->current_read_depth + rdma_req->num_outstanding_data_wr <= rqpair->max_read_depth);
1109 4 : rqpair->current_read_depth += rdma_req->num_outstanding_data_wr;
1110 4 : assert(rqpair->current_send_depth + rdma_req->num_outstanding_data_wr <= rqpair->max_send_depth);
1111 4 : rqpair->current_send_depth += rdma_req->num_outstanding_data_wr;
1112 4 : return 0;
1113 : }
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 : struct spdk_nvmf_rdma_request *rdma_req;
1131 : struct ibv_send_wr *wr;
1132 : 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 : }
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 : }
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 : struct spdk_nvmf_rdma_request *rdma_req;
1160 : struct spdk_nvmf_qpair *qpair;
1161 : struct spdk_nvmf_rdma_qpair *rqpair;
1162 : struct spdk_nvme_cpl *rsp;
1163 6 : struct ibv_send_wr *first = NULL;
1164 : 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 : } 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 5 : } 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 : }
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 : }
1210 6 : if (rtransport->rdma_opts.no_wr_batching) {
1211 0 : _poller_submit_sends(rtransport, rqpair->poller);
1212 : }
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 : }
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 : 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 : }
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 : } else {
1256 0 : SPDK_DEBUGLOG(rdma, "Sent back the accept\n");
1257 : }
1258 :
1259 0 : return rc;
1260 : }
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 : struct spdk_nvmf_rdma_transport *rtransport;
1277 0 : struct spdk_nvmf_rdma_qpair *rqpair = NULL;
1278 : 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 : uint16_t max_queue_depth;
1282 : 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 : } 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 : }
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 0 : STAILQ_INIT(&rqpair->ibv_events);
1382 : /* use qid from the private data to determine the qpair type
1383 : qid will be set to the appropriate value when the controller is created */
1384 0 : rqpair->qpair.qid = private_data->qid;
1385 :
1386 0 : event->id->context = &rqpair->qpair;
1387 :
1388 0 : spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair);
1389 :
1390 0 : return 0;
1391 : }
1392 :
1393 : static inline void
1394 28 : nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next,
1395 : enum spdk_nvme_data_transfer xfer)
1396 : {
1397 28 : if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
1398 24 : wr->opcode = IBV_WR_RDMA_WRITE;
1399 24 : wr->send_flags = 0;
1400 24 : wr->next = next;
1401 4 : } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) {
1402 4 : wr->opcode = IBV_WR_RDMA_READ;
1403 4 : wr->send_flags = IBV_SEND_SIGNALED;
1404 4 : wr->next = NULL;
1405 : } else {
1406 0 : assert(0);
1407 : }
1408 28 : }
1409 :
1410 : static int
1411 6 : nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport,
1412 : struct spdk_nvmf_rdma_request *rdma_req,
1413 : uint32_t num_sgl_descriptors)
1414 : {
1415 6 : struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES];
1416 : struct spdk_nvmf_rdma_request_data *current_data_wr;
1417 : uint32_t i;
1418 :
1419 6 : if (spdk_unlikely(num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES)) {
1420 0 : SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n",
1421 : num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES);
1422 0 : return -EINVAL;
1423 : }
1424 :
1425 6 : if (spdk_unlikely(spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests,
1426 : num_sgl_descriptors))) {
1427 0 : return -ENOMEM;
1428 : }
1429 :
1430 6 : current_data_wr = &rdma_req->data;
1431 :
1432 12 : for (i = 0; i < num_sgl_descriptors; i++) {
1433 6 : nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer);
1434 6 : current_data_wr->wr.next = &work_requests[i]->wr;
1435 6 : current_data_wr = work_requests[i];
1436 6 : current_data_wr->wr.sg_list = current_data_wr->sgl;
1437 6 : current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id;
1438 : }
1439 :
1440 6 : nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer);
1441 :
1442 6 : return 0;
1443 : }
1444 :
1445 : static inline void
1446 16 : nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req)
1447 : {
1448 16 : struct ibv_send_wr *wr = &rdma_req->data.wr;
1449 16 : struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1;
1450 :
1451 16 : wr->wr.rdma.rkey = sgl->keyed.key;
1452 16 : wr->wr.rdma.remote_addr = sgl->address;
1453 16 : nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer);
1454 16 : }
1455 :
1456 : static inline void
1457 1 : nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs)
1458 : {
1459 1 : struct ibv_send_wr *wr = &rdma_req->data.wr;
1460 1 : struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1;
1461 : uint32_t i;
1462 : int j;
1463 1 : uint64_t remote_addr_offset = 0;
1464 :
1465 3 : for (i = 0; i < num_wrs; ++i) {
1466 2 : wr->wr.rdma.rkey = sgl->keyed.key;
1467 2 : wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset;
1468 19 : for (j = 0; j < wr->num_sge; ++j) {
1469 17 : remote_addr_offset += wr->sg_list[j].length;
1470 : }
1471 2 : wr = wr->next;
1472 : }
1473 1 : }
1474 :
1475 : static int
1476 15 : nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_device *device,
1477 : struct spdk_nvmf_rdma_request *rdma_req,
1478 : struct ibv_send_wr *wr,
1479 : uint32_t total_length)
1480 : {
1481 15 : struct spdk_rdma_utils_memory_translation mem_translation;
1482 : struct ibv_sge *sg_ele;
1483 : struct iovec *iov;
1484 : uint32_t lkey, remaining;
1485 : int rc;
1486 :
1487 15 : wr->num_sge = 0;
1488 :
1489 74 : while (total_length && wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES) {
1490 59 : iov = &rdma_req->req.iov[rdma_req->iovpos];
1491 59 : rc = spdk_rdma_utils_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation);
1492 59 : if (spdk_unlikely(rc)) {
1493 0 : return rc;
1494 : }
1495 :
1496 59 : lkey = spdk_rdma_utils_memory_translation_get_lkey(&mem_translation);
1497 59 : sg_ele = &wr->sg_list[wr->num_sge];
1498 59 : remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length);
1499 :
1500 59 : sg_ele->lkey = lkey;
1501 59 : sg_ele->addr = (uintptr_t)iov->iov_base + rdma_req->offset;
1502 59 : sg_ele->length = remaining;
1503 59 : SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, sg_ele->addr,
1504 : sg_ele->length);
1505 59 : rdma_req->offset += sg_ele->length;
1506 59 : total_length -= sg_ele->length;
1507 59 : wr->num_sge++;
1508 :
1509 59 : if (rdma_req->offset == iov->iov_len) {
1510 57 : rdma_req->offset = 0;
1511 57 : rdma_req->iovpos++;
1512 : }
1513 : }
1514 :
1515 15 : if (spdk_unlikely(total_length)) {
1516 0 : SPDK_ERRLOG("Not enough SG entries to hold data buffer\n");
1517 0 : return -EINVAL;
1518 : }
1519 :
1520 15 : return 0;
1521 : }
1522 :
1523 : static int
1524 10 : nvmf_rdma_fill_wr_sgl_with_dif(struct spdk_nvmf_rdma_device *device,
1525 : struct spdk_nvmf_rdma_request *rdma_req,
1526 : struct ibv_send_wr *wr,
1527 : uint32_t total_length,
1528 : uint32_t num_extra_wrs)
1529 : {
1530 10 : struct spdk_rdma_utils_memory_translation mem_translation;
1531 10 : struct spdk_dif_ctx *dif_ctx = &rdma_req->req.dif.dif_ctx;
1532 : struct ibv_sge *sg_ele;
1533 : struct iovec *iov;
1534 : struct iovec *rdma_iov;
1535 : uint32_t lkey, remaining;
1536 : uint32_t remaining_data_block, data_block_size, md_size;
1537 : uint32_t sge_len;
1538 : int rc;
1539 :
1540 10 : data_block_size = dif_ctx->block_size - dif_ctx->md_size;
1541 :
1542 10 : if (spdk_likely(!rdma_req->req.stripped_data)) {
1543 5 : rdma_iov = rdma_req->req.iov;
1544 5 : remaining_data_block = data_block_size;
1545 5 : md_size = dif_ctx->md_size;
1546 : } else {
1547 5 : rdma_iov = rdma_req->req.stripped_data->iov;
1548 5 : total_length = total_length / dif_ctx->block_size * data_block_size;
1549 5 : remaining_data_block = total_length;
1550 5 : md_size = 0;
1551 : }
1552 :
1553 10 : wr->num_sge = 0;
1554 :
1555 25 : while (total_length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) {
1556 15 : iov = rdma_iov + rdma_req->iovpos;
1557 15 : rc = spdk_rdma_utils_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation);
1558 15 : if (spdk_unlikely(rc)) {
1559 0 : return rc;
1560 : }
1561 :
1562 15 : lkey = spdk_rdma_utils_memory_translation_get_lkey(&mem_translation);
1563 15 : sg_ele = &wr->sg_list[wr->num_sge];
1564 15 : remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length);
1565 :
1566 53 : while (remaining) {
1567 38 : if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) {
1568 1 : if (num_extra_wrs > 0 && wr->next) {
1569 1 : wr = wr->next;
1570 1 : wr->num_sge = 0;
1571 1 : sg_ele = &wr->sg_list[wr->num_sge];
1572 1 : num_extra_wrs--;
1573 : } else {
1574 : break;
1575 : }
1576 : }
1577 38 : sg_ele->lkey = lkey;
1578 38 : sg_ele->addr = (uintptr_t)((char *)iov->iov_base + rdma_req->offset);
1579 38 : sge_len = spdk_min(remaining, remaining_data_block);
1580 38 : sg_ele->length = sge_len;
1581 38 : SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele,
1582 : sg_ele->addr, sg_ele->length);
1583 38 : remaining -= sge_len;
1584 38 : remaining_data_block -= sge_len;
1585 38 : rdma_req->offset += sge_len;
1586 38 : total_length -= sge_len;
1587 :
1588 38 : sg_ele++;
1589 38 : wr->num_sge++;
1590 :
1591 38 : if (remaining_data_block == 0) {
1592 : /* skip metadata */
1593 34 : rdma_req->offset += md_size;
1594 34 : total_length -= md_size;
1595 : /* Metadata that do not fit this IO buffer will be included in the next IO buffer */
1596 34 : remaining -= spdk_min(remaining, md_size);
1597 34 : remaining_data_block = data_block_size;
1598 : }
1599 :
1600 38 : if (remaining == 0) {
1601 : /* By subtracting the size of the last IOV from the offset, we ensure that we skip
1602 : the remaining metadata bits at the beginning of the next buffer */
1603 15 : rdma_req->offset -= spdk_min(iov->iov_len, rdma_req->offset);
1604 15 : rdma_req->iovpos++;
1605 : }
1606 : }
1607 : }
1608 :
1609 10 : if (spdk_unlikely(total_length)) {
1610 0 : SPDK_ERRLOG("Not enough SG entries to hold data buffer\n");
1611 0 : return -EINVAL;
1612 : }
1613 :
1614 10 : return 0;
1615 : }
1616 :
1617 : static inline uint32_t
1618 8 : nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size)
1619 : {
1620 : /* estimate the number of SG entries and WRs needed to process the request */
1621 8 : uint32_t num_sge = 0;
1622 : uint32_t i;
1623 8 : uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size);
1624 :
1625 23 : for (i = 0; i < num_buffers && length > 0; i++) {
1626 15 : uint32_t buffer_len = spdk_min(length, io_unit_size);
1627 15 : uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size);
1628 :
1629 15 : if (num_sge_in_block * block_size > buffer_len) {
1630 11 : ++num_sge_in_block;
1631 : }
1632 15 : num_sge += num_sge_in_block;
1633 15 : length -= buffer_len;
1634 : }
1635 8 : return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES);
1636 : }
1637 :
1638 : static int
1639 16 : nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport,
1640 : struct spdk_nvmf_rdma_device *device,
1641 : struct spdk_nvmf_rdma_request *rdma_req)
1642 : {
1643 : struct spdk_nvmf_rdma_qpair *rqpair;
1644 : struct spdk_nvmf_rdma_poll_group *rgroup;
1645 16 : struct spdk_nvmf_request *req = &rdma_req->req;
1646 16 : struct ibv_send_wr *wr = &rdma_req->data.wr;
1647 : int rc;
1648 16 : uint32_t num_wrs = 1;
1649 : uint32_t length;
1650 :
1651 16 : rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair);
1652 16 : rgroup = rqpair->poller->group;
1653 :
1654 : /* rdma wr specifics */
1655 16 : nvmf_rdma_setup_request(rdma_req);
1656 :
1657 16 : length = req->length;
1658 16 : if (spdk_unlikely(req->dif_enabled)) {
1659 8 : req->dif.orig_length = length;
1660 8 : length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx);
1661 8 : req->dif.elba_length = length;
1662 : }
1663 :
1664 16 : rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport,
1665 : length);
1666 16 : if (spdk_unlikely(rc != 0)) {
1667 1 : return rc;
1668 : }
1669 :
1670 15 : assert(req->iovcnt <= rqpair->max_send_sge);
1671 :
1672 : /* When dif_insert_or_strip is true and the I/O data length is greater than one block,
1673 : * the stripped_buffers are got for DIF stripping. */
1674 15 : if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST)
1675 : && (req->dif.elba_length > req->dif.dif_ctx.block_size))) {
1676 7 : rc = nvmf_request_get_stripped_buffers(req, &rgroup->group,
1677 : &rtransport->transport, req->dif.orig_length);
1678 7 : if (rc != 0) {
1679 4 : SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc);
1680 : }
1681 : }
1682 :
1683 15 : rdma_req->iovpos = 0;
1684 :
1685 15 : if (spdk_unlikely(req->dif_enabled)) {
1686 8 : num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size,
1687 : req->dif.dif_ctx.block_size);
1688 8 : if (num_wrs > 1) {
1689 1 : rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1);
1690 1 : if (spdk_unlikely(rc != 0)) {
1691 0 : goto err_exit;
1692 : }
1693 : }
1694 :
1695 8 : rc = nvmf_rdma_fill_wr_sgl_with_dif(device, rdma_req, wr, length, num_wrs - 1);
1696 8 : if (spdk_unlikely(rc != 0)) {
1697 0 : goto err_exit;
1698 : }
1699 :
1700 8 : if (num_wrs > 1) {
1701 1 : nvmf_rdma_update_remote_addr(rdma_req, num_wrs);
1702 : }
1703 : } else {
1704 7 : rc = nvmf_rdma_fill_wr_sgl(device, rdma_req, wr, length);
1705 7 : if (spdk_unlikely(rc != 0)) {
1706 0 : goto err_exit;
1707 : }
1708 : }
1709 :
1710 : /* set the number of outstanding data WRs for this request. */
1711 15 : rdma_req->num_outstanding_data_wr = num_wrs;
1712 :
1713 15 : return rc;
1714 :
1715 0 : err_exit:
1716 0 : spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport);
1717 0 : nvmf_rdma_request_free_data(rdma_req, rtransport);
1718 0 : req->iovcnt = 0;
1719 0 : return rc;
1720 : }
1721 :
1722 : static int
1723 5 : nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport,
1724 : struct spdk_nvmf_rdma_device *device,
1725 : struct spdk_nvmf_rdma_request *rdma_req)
1726 : {
1727 : struct spdk_nvmf_rdma_qpair *rqpair;
1728 : struct spdk_nvmf_rdma_poll_group *rgroup;
1729 : struct ibv_send_wr *current_wr;
1730 5 : struct spdk_nvmf_request *req = &rdma_req->req;
1731 : struct spdk_nvme_sgl_descriptor *inline_segment, *desc;
1732 : uint32_t num_sgl_descriptors;
1733 5 : uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES], total_length = 0;
1734 : uint32_t i;
1735 : int rc;
1736 :
1737 5 : rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
1738 5 : rgroup = rqpair->poller->group;
1739 :
1740 5 : inline_segment = &req->cmd->nvme_cmd.dptr.sgl1;
1741 5 : assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT);
1742 5 : assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET);
1743 :
1744 5 : num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor);
1745 5 : assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES);
1746 :
1747 5 : desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address;
1748 15 : for (i = 0; i < num_sgl_descriptors; i++) {
1749 10 : if (spdk_likely(!req->dif_enabled)) {
1750 8 : lengths[i] = desc->keyed.length;
1751 : } else {
1752 2 : req->dif.orig_length += desc->keyed.length;
1753 2 : lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx);
1754 2 : req->dif.elba_length += lengths[i];
1755 : }
1756 10 : total_length += lengths[i];
1757 10 : desc++;
1758 : }
1759 :
1760 5 : if (spdk_unlikely(total_length > rtransport->transport.opts.max_io_size)) {
1761 0 : SPDK_ERRLOG("Multi SGL length 0x%x exceeds max io size 0x%x\n",
1762 : total_length, rtransport->transport.opts.max_io_size);
1763 0 : req->rsp->nvme_cpl.status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
1764 0 : return -EINVAL;
1765 : }
1766 :
1767 5 : rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1);
1768 5 : if (spdk_unlikely(rc != 0)) {
1769 0 : return -ENOMEM;
1770 : }
1771 :
1772 5 : rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, total_length);
1773 5 : if (spdk_unlikely(rc != 0)) {
1774 0 : nvmf_rdma_request_free_data(rdma_req, rtransport);
1775 0 : return rc;
1776 : }
1777 :
1778 : /* When dif_insert_or_strip is true and the I/O data length is greater than one block,
1779 : * the stripped_buffers are got for DIF stripping. */
1780 5 : if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST)
1781 : && (req->dif.elba_length > req->dif.dif_ctx.block_size))) {
1782 1 : rc = nvmf_request_get_stripped_buffers(req, &rgroup->group,
1783 : &rtransport->transport, req->dif.orig_length);
1784 1 : if (spdk_unlikely(rc != 0)) {
1785 0 : SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc);
1786 : }
1787 : }
1788 :
1789 : /* The first WR must always be the embedded data WR. This is how we unwind them later. */
1790 5 : current_wr = &rdma_req->data.wr;
1791 5 : assert(current_wr != NULL);
1792 :
1793 5 : req->length = 0;
1794 5 : rdma_req->iovpos = 0;
1795 5 : desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address;
1796 15 : for (i = 0; i < num_sgl_descriptors; i++) {
1797 : /* The descriptors must be keyed data block descriptors with an address, not an offset. */
1798 10 : if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK ||
1799 : desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) {
1800 0 : rc = -EINVAL;
1801 0 : goto err_exit;
1802 : }
1803 :
1804 10 : if (spdk_likely(!req->dif_enabled)) {
1805 8 : rc = nvmf_rdma_fill_wr_sgl(device, rdma_req, current_wr, lengths[i]);
1806 : } else {
1807 2 : rc = nvmf_rdma_fill_wr_sgl_with_dif(device, rdma_req, current_wr,
1808 : lengths[i], 0);
1809 : }
1810 10 : if (spdk_unlikely(rc != 0)) {
1811 0 : rc = -ENOMEM;
1812 0 : goto err_exit;
1813 : }
1814 :
1815 10 : req->length += desc->keyed.length;
1816 10 : current_wr->wr.rdma.rkey = desc->keyed.key;
1817 10 : current_wr->wr.rdma.remote_addr = desc->address;
1818 10 : current_wr = current_wr->next;
1819 10 : desc++;
1820 : }
1821 :
1822 : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
1823 : /* Go back to the last descriptor in the list. */
1824 5 : desc--;
1825 5 : if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) {
1826 0 : if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) {
1827 0 : rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV;
1828 0 : rdma_req->rsp.wr.imm_data = desc->keyed.key;
1829 : }
1830 : }
1831 : #endif
1832 :
1833 5 : rdma_req->num_outstanding_data_wr = num_sgl_descriptors;
1834 :
1835 5 : return 0;
1836 :
1837 0 : err_exit:
1838 0 : spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport);
1839 0 : nvmf_rdma_request_free_data(rdma_req, rtransport);
1840 0 : return rc;
1841 : }
1842 :
1843 : static int
1844 25 : nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport,
1845 : struct spdk_nvmf_rdma_device *device,
1846 : struct spdk_nvmf_rdma_request *rdma_req)
1847 : {
1848 25 : struct spdk_nvmf_request *req = &rdma_req->req;
1849 : struct spdk_nvme_cpl *rsp;
1850 : struct spdk_nvme_sgl_descriptor *sgl;
1851 : int rc;
1852 : uint32_t length;
1853 :
1854 25 : rsp = &req->rsp->nvme_cpl;
1855 25 : sgl = &req->cmd->nvme_cmd.dptr.sgl1;
1856 :
1857 25 : if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK &&
1858 17 : (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS ||
1859 0 : sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) {
1860 :
1861 17 : length = sgl->keyed.length;
1862 17 : if (spdk_unlikely(length > rtransport->transport.opts.max_io_size)) {
1863 1 : SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n",
1864 : length, rtransport->transport.opts.max_io_size);
1865 1 : rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
1866 1 : return -1;
1867 : }
1868 : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
1869 16 : if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) {
1870 0 : if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) {
1871 0 : rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV;
1872 0 : rdma_req->rsp.wr.imm_data = sgl->keyed.key;
1873 : }
1874 : }
1875 : #endif
1876 :
1877 : /* fill request length and populate iovs */
1878 16 : req->length = length;
1879 :
1880 16 : rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req);
1881 16 : if (spdk_unlikely(rc < 0)) {
1882 1 : if (rc == -EINVAL) {
1883 0 : SPDK_ERRLOG("SGL length exceeds the max I/O size\n");
1884 0 : rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
1885 0 : return -1;
1886 : }
1887 : /* No available buffers. Queue this request up. */
1888 1 : SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req);
1889 1 : return 0;
1890 : }
1891 :
1892 15 : SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req,
1893 : req->iovcnt);
1894 :
1895 15 : return 0;
1896 8 : } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK &&
1897 3 : sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) {
1898 3 : uint64_t offset = sgl->address;
1899 3 : uint32_t max_len = rtransport->transport.opts.in_capsule_data_size;
1900 :
1901 3 : SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n",
1902 : offset, sgl->unkeyed.length);
1903 :
1904 3 : if (spdk_unlikely(offset > max_len)) {
1905 0 : SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n",
1906 : offset, max_len);
1907 0 : rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET;
1908 0 : return -1;
1909 : }
1910 3 : max_len -= (uint32_t)offset;
1911 :
1912 3 : if (spdk_unlikely(sgl->unkeyed.length > max_len)) {
1913 2 : SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n",
1914 : sgl->unkeyed.length, max_len);
1915 2 : rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
1916 2 : return -1;
1917 : }
1918 :
1919 1 : rdma_req->num_outstanding_data_wr = 0;
1920 1 : req->data_from_pool = false;
1921 1 : req->length = sgl->unkeyed.length;
1922 :
1923 1 : req->iov[0].iov_base = rdma_req->recv->buf + offset;
1924 1 : req->iov[0].iov_len = req->length;
1925 1 : req->iovcnt = 1;
1926 :
1927 1 : return 0;
1928 5 : } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT &&
1929 5 : sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) {
1930 :
1931 5 : rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req);
1932 5 : if (spdk_unlikely(rc == -ENOMEM)) {
1933 0 : SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req);
1934 0 : return 0;
1935 5 : } else if (spdk_unlikely(rc == -EINVAL)) {
1936 0 : SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n");
1937 0 : rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
1938 0 : return -1;
1939 : }
1940 :
1941 5 : SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req,
1942 : req->iovcnt);
1943 :
1944 5 : return 0;
1945 : }
1946 :
1947 0 : SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n",
1948 : sgl->generic.type, sgl->generic.subtype);
1949 0 : rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID;
1950 0 : return -1;
1951 : }
1952 :
1953 : static void
1954 6 : _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req,
1955 : struct spdk_nvmf_rdma_transport *rtransport)
1956 : {
1957 : struct spdk_nvmf_rdma_qpair *rqpair;
1958 : struct spdk_nvmf_rdma_poll_group *rgroup;
1959 :
1960 6 : rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
1961 6 : if (rdma_req->req.data_from_pool) {
1962 5 : rgroup = rqpair->poller->group;
1963 :
1964 5 : spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport);
1965 : }
1966 6 : if (rdma_req->req.stripped_data) {
1967 0 : nvmf_request_free_stripped_buffers(&rdma_req->req,
1968 0 : &rqpair->poller->group->group,
1969 : &rtransport->transport);
1970 : }
1971 6 : nvmf_rdma_request_free_data(rdma_req, rtransport);
1972 6 : rdma_req->req.length = 0;
1973 6 : rdma_req->req.iovcnt = 0;
1974 6 : rdma_req->offset = 0;
1975 6 : rdma_req->req.dif_enabled = false;
1976 6 : rdma_req->fused_failed = false;
1977 6 : rdma_req->transfer_wr = NULL;
1978 6 : if (rdma_req->fused_pair) {
1979 : /* This req was part of a valid fused pair, but failed before it got to
1980 : * READ_TO_EXECUTE state. This means we need to fail the other request
1981 : * in the pair, because it is no longer part of a valid pair. If the pair
1982 : * already reached READY_TO_EXECUTE state, we need to kick it.
1983 : */
1984 0 : rdma_req->fused_pair->fused_failed = true;
1985 0 : if (rdma_req->fused_pair->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) {
1986 0 : nvmf_rdma_request_process(rtransport, rdma_req->fused_pair);
1987 : }
1988 0 : rdma_req->fused_pair = NULL;
1989 : }
1990 6 : memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif));
1991 6 : rqpair->qd--;
1992 :
1993 6 : STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link);
1994 6 : rqpair->qpair.queue_depth--;
1995 6 : rdma_req->state = RDMA_REQUEST_STATE_FREE;
1996 6 : }
1997 :
1998 : static void
1999 6 : nvmf_rdma_check_fused_ordering(struct spdk_nvmf_rdma_transport *rtransport,
2000 : struct spdk_nvmf_rdma_qpair *rqpair,
2001 : struct spdk_nvmf_rdma_request *rdma_req)
2002 : {
2003 : enum spdk_nvme_cmd_fuse last, next;
2004 :
2005 6 : last = rqpair->fused_first ? rqpair->fused_first->req.cmd->nvme_cmd.fuse : SPDK_NVME_CMD_FUSE_NONE;
2006 6 : next = rdma_req->req.cmd->nvme_cmd.fuse;
2007 :
2008 6 : assert(last != SPDK_NVME_CMD_FUSE_SECOND);
2009 :
2010 6 : if (spdk_likely(last == SPDK_NVME_CMD_FUSE_NONE && next == SPDK_NVME_CMD_FUSE_NONE)) {
2011 6 : return;
2012 : }
2013 :
2014 0 : if (last == SPDK_NVME_CMD_FUSE_FIRST) {
2015 0 : if (next == SPDK_NVME_CMD_FUSE_SECOND) {
2016 : /* This is a valid pair of fused commands. Point them at each other
2017 : * so they can be submitted consecutively once ready to be executed.
2018 : */
2019 0 : rqpair->fused_first->fused_pair = rdma_req;
2020 0 : rdma_req->fused_pair = rqpair->fused_first;
2021 0 : rqpair->fused_first = NULL;
2022 0 : return;
2023 : } else {
2024 : /* Mark the last req as failed since it wasn't followed by a SECOND. */
2025 0 : rqpair->fused_first->fused_failed = true;
2026 :
2027 : /* If the last req is in READY_TO_EXECUTE state, then call
2028 : * nvmf_rdma_request_process(), otherwise nothing else will kick it.
2029 : */
2030 0 : if (rqpair->fused_first->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) {
2031 0 : nvmf_rdma_request_process(rtransport, rqpair->fused_first);
2032 : }
2033 :
2034 0 : rqpair->fused_first = NULL;
2035 : }
2036 : }
2037 :
2038 0 : if (next == SPDK_NVME_CMD_FUSE_FIRST) {
2039 : /* Set rqpair->fused_first here so that we know to check that the next request
2040 : * is a SECOND (and to fail this one if it isn't).
2041 : */
2042 0 : rqpair->fused_first = rdma_req;
2043 0 : } else if (next == SPDK_NVME_CMD_FUSE_SECOND) {
2044 : /* Mark this req failed since it ia SECOND and the last one was not a FIRST. */
2045 0 : rdma_req->fused_failed = true;
2046 : }
2047 : }
2048 :
2049 : bool
2050 23 : nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport,
2051 : struct spdk_nvmf_rdma_request *rdma_req)
2052 : {
2053 : struct spdk_nvmf_rdma_qpair *rqpair;
2054 : struct spdk_nvmf_rdma_device *device;
2055 : struct spdk_nvmf_rdma_poll_group *rgroup;
2056 23 : struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl;
2057 : int rc;
2058 : struct spdk_nvmf_rdma_recv *rdma_recv;
2059 : enum spdk_nvmf_rdma_request_state prev_state;
2060 23 : bool progress = false;
2061 23 : int data_posted;
2062 : uint32_t num_blocks, num_rdma_reads_available, qdepth;
2063 :
2064 23 : rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
2065 23 : device = rqpair->device;
2066 23 : rgroup = rqpair->poller->group;
2067 :
2068 23 : assert(rdma_req->state != RDMA_REQUEST_STATE_FREE);
2069 :
2070 : /* If the queue pair is in an error state, force the request to the completed state
2071 : * to release resources. */
2072 23 : if (spdk_unlikely(rqpair->ibv_in_error_state || !spdk_nvmf_qpair_is_active(&rqpair->qpair))) {
2073 0 : switch (rdma_req->state) {
2074 0 : case RDMA_REQUEST_STATE_NEED_BUFFER:
2075 0 : STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link);
2076 0 : break;
2077 0 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING:
2078 0 : STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
2079 0 : break;
2080 0 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING:
2081 0 : STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
2082 0 : break;
2083 0 : case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING:
2084 0 : STAILQ_REMOVE(&rqpair->pending_rdma_send_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
2085 0 : break;
2086 0 : default:
2087 0 : break;
2088 : }
2089 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
2090 : }
2091 :
2092 : /* The loop here is to allow for several back-to-back state changes. */
2093 : do {
2094 66 : prev_state = rdma_req->state;
2095 :
2096 66 : SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state);
2097 :
2098 66 : switch (rdma_req->state) {
2099 6 : case RDMA_REQUEST_STATE_FREE:
2100 : /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW
2101 : * to escape this state. */
2102 6 : break;
2103 6 : case RDMA_REQUEST_STATE_NEW:
2104 6 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0,
2105 : (uintptr_t)rdma_req, (uintptr_t)rqpair, rqpair->qpair.queue_depth);
2106 6 : rdma_recv = rdma_req->recv;
2107 :
2108 : /* The first element of the SGL is the NVMe command */
2109 6 : rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr;
2110 6 : memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp));
2111 6 : rdma_req->transfer_wr = &rdma_req->data.wr;
2112 :
2113 6 : if (spdk_unlikely(rqpair->ibv_in_error_state || !spdk_nvmf_qpair_is_active(&rqpair->qpair))) {
2114 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
2115 0 : break;
2116 : }
2117 :
2118 6 : if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) {
2119 0 : rdma_req->req.dif_enabled = true;
2120 : }
2121 :
2122 6 : nvmf_rdma_check_fused_ordering(rtransport, rqpair, rdma_req);
2123 :
2124 : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
2125 6 : rdma_req->rsp.wr.opcode = IBV_WR_SEND;
2126 6 : rdma_req->rsp.wr.imm_data = 0;
2127 : #endif
2128 :
2129 : /* The next state transition depends on the data transfer needs of this request. */
2130 6 : rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req);
2131 :
2132 6 : if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) {
2133 1 : rsp->status.sct = SPDK_NVME_SCT_GENERIC;
2134 1 : rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE;
2135 1 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
2136 1 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2137 1 : SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req);
2138 1 : break;
2139 : }
2140 :
2141 : /* If no data to transfer, ready to execute. */
2142 5 : if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) {
2143 0 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
2144 0 : break;
2145 : }
2146 :
2147 5 : rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER;
2148 5 : STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link);
2149 5 : break;
2150 5 : case RDMA_REQUEST_STATE_NEED_BUFFER:
2151 5 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0,
2152 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2153 :
2154 5 : assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE);
2155 :
2156 5 : if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) {
2157 : /* This request needs to wait in line to obtain a buffer */
2158 0 : break;
2159 : }
2160 :
2161 : /* Try to get a data buffer */
2162 5 : rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req);
2163 5 : if (spdk_unlikely(rc < 0)) {
2164 0 : STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link);
2165 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
2166 0 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2167 0 : break;
2168 : }
2169 :
2170 5 : if (rdma_req->req.iovcnt == 0) {
2171 : /* No buffers available. */
2172 0 : rgroup->stat.pending_data_buffer++;
2173 0 : break;
2174 : }
2175 :
2176 5 : STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link);
2177 :
2178 : /* If data is transferring from host to controller and the data didn't
2179 : * arrive using in capsule data, we need to do a transfer from the host.
2180 : */
2181 5 : if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER &&
2182 : rdma_req->req.data_from_pool) {
2183 4 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link);
2184 4 : rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING;
2185 4 : break;
2186 : }
2187 :
2188 1 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
2189 1 : break;
2190 4 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING:
2191 4 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0,
2192 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2193 :
2194 4 : if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) {
2195 : /* This request needs to wait in line to perform RDMA */
2196 0 : break;
2197 : }
2198 4 : assert(rqpair->max_send_depth >= rqpair->current_send_depth);
2199 4 : qdepth = rqpair->max_send_depth - rqpair->current_send_depth;
2200 4 : assert(rqpair->max_read_depth >= rqpair->current_read_depth);
2201 4 : num_rdma_reads_available = rqpair->max_read_depth - rqpair->current_read_depth;
2202 4 : if (rdma_req->num_outstanding_data_wr > qdepth ||
2203 4 : rdma_req->num_outstanding_data_wr > num_rdma_reads_available) {
2204 0 : if (num_rdma_reads_available && qdepth) {
2205 : /* Send as much as we can */
2206 0 : request_prepare_transfer_in_part(&rdma_req->req, spdk_min(num_rdma_reads_available, qdepth));
2207 : } else {
2208 : /* We can only have so many WRs outstanding. we have to wait until some finish. */
2209 0 : rqpair->poller->stat.pending_rdma_read++;
2210 0 : break;
2211 : }
2212 : }
2213 :
2214 : /* We have already verified that this request is the head of the queue. */
2215 4 : if (rdma_req->num_remaining_data_wr == 0) {
2216 4 : STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link);
2217 : }
2218 :
2219 4 : rc = request_transfer_in(&rdma_req->req);
2220 4 : if (spdk_likely(rc == 0)) {
2221 4 : rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER;
2222 : } else {
2223 0 : rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR;
2224 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
2225 0 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2226 : }
2227 4 : break;
2228 4 : case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
2229 4 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0,
2230 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2231 : /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE
2232 : * to escape this state. */
2233 4 : break;
2234 5 : case RDMA_REQUEST_STATE_READY_TO_EXECUTE:
2235 5 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0,
2236 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2237 :
2238 5 : if (spdk_unlikely(rdma_req->req.dif_enabled)) {
2239 0 : if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) {
2240 : /* generate DIF for write operation */
2241 0 : num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size);
2242 0 : assert(num_blocks > 0);
2243 :
2244 0 : rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt,
2245 0 : num_blocks, &rdma_req->req.dif.dif_ctx);
2246 0 : if (rc != 0) {
2247 0 : SPDK_ERRLOG("DIF generation failed\n");
2248 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
2249 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
2250 0 : break;
2251 : }
2252 : }
2253 :
2254 0 : assert(rdma_req->req.dif.elba_length >= rdma_req->req.length);
2255 : /* set extended length before IO operation */
2256 0 : rdma_req->req.length = rdma_req->req.dif.elba_length;
2257 : }
2258 :
2259 5 : if (rdma_req->req.cmd->nvme_cmd.fuse != SPDK_NVME_CMD_FUSE_NONE) {
2260 0 : if (rdma_req->fused_failed) {
2261 : /* This request failed FUSED semantics. Fail it immediately, without
2262 : * even sending it to the target layer.
2263 : */
2264 0 : rsp->status.sct = SPDK_NVME_SCT_GENERIC;
2265 0 : rsp->status.sc = SPDK_NVME_SC_ABORTED_MISSING_FUSED;
2266 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
2267 0 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2268 0 : break;
2269 : }
2270 :
2271 0 : if (rdma_req->fused_pair == NULL ||
2272 0 : rdma_req->fused_pair->state != RDMA_REQUEST_STATE_READY_TO_EXECUTE) {
2273 : /* This request is ready to execute, but either we don't know yet if it's
2274 : * valid - i.e. this is a FIRST but we haven't received the next
2275 : * request yet or the other request of this fused pair isn't ready to
2276 : * execute. So break here and this request will get processed later either
2277 : * when the other request is ready or we find that this request isn't valid.
2278 : */
2279 : break;
2280 : }
2281 : }
2282 :
2283 : /* If we get to this point, and this request is a fused command, we know that
2284 : * it is part of valid sequence (FIRST followed by a SECOND) and that both
2285 : * requests are READY_TO_EXECUTE. So call spdk_nvmf_request_exec() both on this
2286 : * request, and the other request of the fused pair, in the correct order.
2287 : * Also clear the ->fused_pair pointers on both requests, since after this point
2288 : * we no longer need to maintain the relationship between these two requests.
2289 : */
2290 5 : if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_SECOND) {
2291 0 : assert(rdma_req->fused_pair != NULL);
2292 0 : assert(rdma_req->fused_pair->fused_pair != NULL);
2293 0 : rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING;
2294 0 : spdk_nvmf_request_exec(&rdma_req->fused_pair->req);
2295 0 : rdma_req->fused_pair->fused_pair = NULL;
2296 0 : rdma_req->fused_pair = NULL;
2297 : }
2298 5 : rdma_req->state = RDMA_REQUEST_STATE_EXECUTING;
2299 5 : spdk_nvmf_request_exec(&rdma_req->req);
2300 5 : if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_FIRST) {
2301 0 : assert(rdma_req->fused_pair != NULL);
2302 0 : assert(rdma_req->fused_pair->fused_pair != NULL);
2303 0 : rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING;
2304 0 : spdk_nvmf_request_exec(&rdma_req->fused_pair->req);
2305 0 : rdma_req->fused_pair->fused_pair = NULL;
2306 0 : rdma_req->fused_pair = NULL;
2307 : }
2308 5 : break;
2309 5 : case RDMA_REQUEST_STATE_EXECUTING:
2310 5 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0,
2311 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2312 : /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED
2313 : * to escape this state. */
2314 5 : break;
2315 5 : case RDMA_REQUEST_STATE_EXECUTED:
2316 5 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0,
2317 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2318 5 : if (rsp->status.sc == SPDK_NVME_SC_SUCCESS &&
2319 5 : rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
2320 1 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link);
2321 1 : rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING;
2322 : } else {
2323 4 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
2324 4 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2325 : }
2326 5 : if (spdk_unlikely(rdma_req->req.dif_enabled)) {
2327 : /* restore the original length */
2328 0 : rdma_req->req.length = rdma_req->req.dif.orig_length;
2329 :
2330 0 : if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
2331 0 : struct spdk_dif_error error_blk;
2332 :
2333 0 : num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size);
2334 0 : if (!rdma_req->req.stripped_data) {
2335 0 : rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks,
2336 0 : &rdma_req->req.dif.dif_ctx, &error_blk);
2337 : } else {
2338 0 : rc = spdk_dif_verify_copy(rdma_req->req.stripped_data->iov,
2339 0 : rdma_req->req.stripped_data->iovcnt,
2340 0 : rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks,
2341 0 : &rdma_req->req.dif.dif_ctx, &error_blk);
2342 : }
2343 0 : if (rc) {
2344 0 : struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl;
2345 :
2346 0 : SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type,
2347 : error_blk.err_offset);
2348 0 : rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR;
2349 0 : rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type);
2350 0 : STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
2351 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
2352 0 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2353 : }
2354 : }
2355 : }
2356 5 : break;
2357 1 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING:
2358 1 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0,
2359 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2360 :
2361 1 : if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) {
2362 : /* This request needs to wait in line to perform RDMA */
2363 0 : break;
2364 : }
2365 1 : if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) >
2366 1 : rqpair->max_send_depth) {
2367 : /* We can only have so many WRs outstanding. we have to wait until some finish.
2368 : * +1 since each request has an additional wr in the resp. */
2369 0 : rqpair->poller->stat.pending_rdma_write++;
2370 0 : break;
2371 : }
2372 :
2373 : /* We have already verified that this request is the head of the queue. */
2374 1 : STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link);
2375 :
2376 : /* The data transfer will be kicked off from
2377 : * RDMA_REQUEST_STATE_READY_TO_COMPLETE state.
2378 : * We verified that data + response fit into send queue, so we can go to the next state directly
2379 : */
2380 1 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE;
2381 1 : break;
2382 7 : case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING:
2383 7 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING, 0, 0,
2384 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2385 :
2386 7 : if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_send_queue)) {
2387 : /* This request needs to wait in line to send the completion */
2388 0 : break;
2389 : }
2390 :
2391 7 : assert(rqpair->current_send_depth <= rqpair->max_send_depth);
2392 7 : if (rqpair->current_send_depth == rqpair->max_send_depth) {
2393 : /* We can only have so many WRs outstanding. we have to wait until some finish */
2394 2 : rqpair->poller->stat.pending_rdma_send++;
2395 2 : break;
2396 : }
2397 :
2398 : /* We have already verified that this request is the head of the queue. */
2399 5 : STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_send_queue, state_link);
2400 :
2401 : /* The response sending will be kicked off from
2402 : * RDMA_REQUEST_STATE_READY_TO_COMPLETE state.
2403 : */
2404 5 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE;
2405 5 : break;
2406 6 : case RDMA_REQUEST_STATE_READY_TO_COMPLETE:
2407 6 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0,
2408 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2409 6 : rc = request_transfer_out(&rdma_req->req, &data_posted);
2410 6 : assert(rc == 0); /* No good way to handle this currently */
2411 6 : if (spdk_unlikely(rc)) {
2412 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
2413 : } else {
2414 6 : rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST :
2415 : RDMA_REQUEST_STATE_COMPLETING;
2416 : }
2417 6 : break;
2418 1 : case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST:
2419 1 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0,
2420 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2421 : /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED
2422 : * to escape this state. */
2423 1 : break;
2424 5 : case RDMA_REQUEST_STATE_COMPLETING:
2425 5 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0,
2426 : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2427 : /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED
2428 : * to escape this state. */
2429 5 : break;
2430 6 : case RDMA_REQUEST_STATE_COMPLETED:
2431 6 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0,
2432 : (uintptr_t)rdma_req, (uintptr_t)rqpair, rqpair->qpair.queue_depth);
2433 :
2434 6 : rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc;
2435 6 : _nvmf_rdma_request_free(rdma_req, rtransport);
2436 6 : break;
2437 0 : case RDMA_REQUEST_NUM_STATES:
2438 : default:
2439 0 : assert(0);
2440 : break;
2441 : }
2442 :
2443 66 : if (rdma_req->state != prev_state) {
2444 43 : progress = true;
2445 : }
2446 66 : } while (rdma_req->state != prev_state);
2447 :
2448 23 : return progress;
2449 : }
2450 :
2451 : /* Public API callbacks begin here */
2452 :
2453 : #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128
2454 : #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128
2455 : #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096
2456 : #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128
2457 : #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096
2458 : #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072
2459 : #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES)
2460 : #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095
2461 : #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE UINT32_MAX
2462 : #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false
2463 : #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false
2464 : #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100
2465 : #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1
2466 : #define SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING false
2467 : #define SPDK_NVMF_RDMA_DEFAULT_DATA_WR_POOL_SIZE 4095
2468 :
2469 : static void
2470 1 : nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts)
2471 : {
2472 1 : opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH;
2473 1 : opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR;
2474 1 : opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE;
2475 1 : opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE;
2476 1 : opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE;
2477 1 : opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH;
2478 1 : opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS;
2479 1 : opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE;
2480 1 : opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP;
2481 1 : opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC;
2482 1 : opts->transport_specific = NULL;
2483 1 : opts->data_wr_pool_size = SPDK_NVMF_RDMA_DEFAULT_DATA_WR_POOL_SIZE;
2484 1 : }
2485 :
2486 : static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport,
2487 : spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg);
2488 :
2489 : static inline bool
2490 0 : nvmf_rdma_is_rxe_device(struct spdk_nvmf_rdma_device *device)
2491 : {
2492 0 : return device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_OLD ||
2493 0 : device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_NEW;
2494 : }
2495 :
2496 : static int nvmf_rdma_accept(void *ctx);
2497 : static bool nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport);
2498 : static void destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport,
2499 : struct spdk_nvmf_rdma_device *device);
2500 :
2501 : static int
2502 0 : create_ib_device(struct spdk_nvmf_rdma_transport *rtransport, struct ibv_context *context,
2503 : struct spdk_nvmf_rdma_device **new_device)
2504 : {
2505 : struct spdk_nvmf_rdma_device *device;
2506 0 : int flag = 0;
2507 0 : int rc = 0;
2508 :
2509 0 : device = calloc(1, sizeof(*device));
2510 0 : if (!device) {
2511 0 : SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n");
2512 0 : return -ENOMEM;
2513 : }
2514 0 : device->context = context;
2515 0 : rc = ibv_query_device(device->context, &device->attr);
2516 0 : if (rc < 0) {
2517 0 : SPDK_ERRLOG("Failed to query RDMA device attributes.\n");
2518 0 : free(device);
2519 0 : return rc;
2520 : }
2521 :
2522 : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
2523 0 : if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) {
2524 0 : SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,");
2525 0 : SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id);
2526 : }
2527 :
2528 : /**
2529 : * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE.
2530 : * The Soft-RoCE RXE driver does not currently support send with invalidate,
2531 : * but incorrectly reports that it does. There are changes making their way
2532 : * through the kernel now that will enable this feature. When they are merged,
2533 : * we can conditionally enable this feature.
2534 : *
2535 : * TODO: enable this for versions of the kernel rxe driver that support it.
2536 : */
2537 0 : if (nvmf_rdma_is_rxe_device(device)) {
2538 0 : device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS);
2539 : }
2540 : #endif
2541 :
2542 : /* set up device context async ev fd as NON_BLOCKING */
2543 0 : flag = fcntl(device->context->async_fd, F_GETFL);
2544 0 : rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK);
2545 0 : if (rc < 0) {
2546 0 : SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n");
2547 0 : free(device);
2548 0 : return rc;
2549 : }
2550 :
2551 0 : TAILQ_INSERT_TAIL(&rtransport->devices, device, link);
2552 0 : SPDK_DEBUGLOG(rdma, "New device %p is added to RDMA trasport\n", device);
2553 :
2554 0 : if (g_nvmf_hooks.get_ibv_pd) {
2555 0 : device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context);
2556 : } else {
2557 0 : device->pd = ibv_alloc_pd(device->context);
2558 : }
2559 :
2560 0 : if (!device->pd) {
2561 0 : SPDK_ERRLOG("Unable to allocate protection domain.\n");
2562 0 : destroy_ib_device(rtransport, device);
2563 0 : return -ENOMEM;
2564 : }
2565 :
2566 0 : assert(device->map == NULL);
2567 :
2568 0 : device->map = spdk_rdma_utils_create_mem_map(device->pd, &g_nvmf_hooks, IBV_ACCESS_LOCAL_WRITE);
2569 0 : if (!device->map) {
2570 0 : SPDK_ERRLOG("Unable to allocate memory map for listen address\n");
2571 0 : destroy_ib_device(rtransport, device);
2572 0 : return -ENOMEM;
2573 : }
2574 :
2575 0 : assert(device->map != NULL);
2576 0 : assert(device->pd != NULL);
2577 :
2578 0 : if (new_device) {
2579 0 : *new_device = device;
2580 : }
2581 0 : SPDK_NOTICELOG("Create IB device %s(%p/%p) succeed.\n", ibv_get_device_name(context->device),
2582 : device, context);
2583 :
2584 0 : return 0;
2585 : }
2586 :
2587 : static void
2588 0 : free_poll_fds(struct spdk_nvmf_rdma_transport *rtransport)
2589 : {
2590 0 : if (rtransport->poll_fds) {
2591 0 : free(rtransport->poll_fds);
2592 0 : rtransport->poll_fds = NULL;
2593 : }
2594 0 : rtransport->npoll_fds = 0;
2595 0 : }
2596 :
2597 : static int
2598 0 : generate_poll_fds(struct spdk_nvmf_rdma_transport *rtransport)
2599 : {
2600 : /* Set up poll descriptor array to monitor events from RDMA and IB
2601 : * in a single poll syscall
2602 : */
2603 0 : int device_count = 0;
2604 0 : int i = 0;
2605 : struct spdk_nvmf_rdma_device *device, *tmp;
2606 :
2607 0 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) {
2608 0 : device_count++;
2609 : }
2610 :
2611 0 : rtransport->npoll_fds = device_count + 1;
2612 :
2613 0 : rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd));
2614 0 : if (rtransport->poll_fds == NULL) {
2615 0 : SPDK_ERRLOG("poll_fds allocation failed\n");
2616 0 : return -ENOMEM;
2617 : }
2618 :
2619 0 : rtransport->poll_fds[i].fd = rtransport->event_channel->fd;
2620 0 : rtransport->poll_fds[i++].events = POLLIN;
2621 :
2622 0 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) {
2623 0 : rtransport->poll_fds[i].fd = device->context->async_fd;
2624 0 : rtransport->poll_fds[i++].events = POLLIN;
2625 : }
2626 :
2627 0 : return 0;
2628 : }
2629 :
2630 : static struct spdk_nvmf_transport *
2631 0 : nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts)
2632 : {
2633 : int rc;
2634 : struct spdk_nvmf_rdma_transport *rtransport;
2635 0 : struct spdk_nvmf_rdma_device *device;
2636 : struct ibv_context **contexts;
2637 : size_t data_wr_pool_size;
2638 : uint32_t i;
2639 : int flag;
2640 : uint32_t sge_count;
2641 : uint32_t min_shared_buffers;
2642 : uint32_t min_in_capsule_data_size;
2643 0 : int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
2644 :
2645 0 : rtransport = calloc(1, sizeof(*rtransport));
2646 0 : if (!rtransport) {
2647 0 : return NULL;
2648 : }
2649 :
2650 0 : TAILQ_INIT(&rtransport->devices);
2651 0 : TAILQ_INIT(&rtransport->ports);
2652 0 : TAILQ_INIT(&rtransport->poll_groups);
2653 0 : TAILQ_INIT(&rtransport->retry_ports);
2654 :
2655 0 : rtransport->transport.ops = &spdk_nvmf_transport_rdma;
2656 0 : rtransport->rdma_opts.num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE;
2657 0 : rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH;
2658 0 : rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ;
2659 0 : rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG;
2660 0 : rtransport->rdma_opts.no_wr_batching = SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING;
2661 0 : if (opts->transport_specific != NULL &&
2662 0 : spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder,
2663 : SPDK_COUNTOF(rdma_transport_opts_decoder),
2664 0 : &rtransport->rdma_opts)) {
2665 0 : SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n");
2666 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2667 0 : return NULL;
2668 : }
2669 :
2670 0 : SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n"
2671 : " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n"
2672 : " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n"
2673 : " in_capsule_data_size=%d, max_aq_depth=%d,\n"
2674 : " num_shared_buffers=%d, num_cqe=%d, max_srq_depth=%d, no_srq=%d,"
2675 : " acceptor_backlog=%d, no_wr_batching=%d abort_timeout_sec=%d\n",
2676 : opts->max_queue_depth,
2677 : opts->max_io_size,
2678 : opts->max_qpairs_per_ctrlr - 1,
2679 : opts->io_unit_size,
2680 : opts->in_capsule_data_size,
2681 : opts->max_aq_depth,
2682 : opts->num_shared_buffers,
2683 : rtransport->rdma_opts.num_cqe,
2684 : rtransport->rdma_opts.max_srq_depth,
2685 : rtransport->rdma_opts.no_srq,
2686 : rtransport->rdma_opts.acceptor_backlog,
2687 : rtransport->rdma_opts.no_wr_batching,
2688 : opts->abort_timeout_sec);
2689 :
2690 : /* I/O unit size cannot be larger than max I/O size */
2691 0 : if (opts->io_unit_size > opts->max_io_size) {
2692 0 : opts->io_unit_size = opts->max_io_size;
2693 : }
2694 :
2695 0 : if (rtransport->rdma_opts.acceptor_backlog <= 0) {
2696 0 : SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n",
2697 : SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG);
2698 0 : rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG;
2699 : }
2700 :
2701 0 : if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) {
2702 0 : SPDK_ERRLOG("The number of shared data buffers (%d) is less than"
2703 : "the minimum number required to guarantee that forward progress can be made (%d)\n",
2704 : opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2));
2705 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2706 0 : return NULL;
2707 : }
2708 :
2709 : /* If buf_cache_size == UINT32_MAX, we will dynamically pick a cache size later that we know will fit. */
2710 0 : if (opts->buf_cache_size < UINT32_MAX) {
2711 0 : min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size;
2712 0 : if (min_shared_buffers > opts->num_shared_buffers) {
2713 0 : SPDK_ERRLOG("There are not enough buffers to satisfy"
2714 : "per-poll group caches for each thread. (%" PRIu32 ")"
2715 : "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers);
2716 0 : SPDK_ERRLOG("Please specify a larger number of shared buffers\n");
2717 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2718 0 : return NULL;
2719 : }
2720 : }
2721 :
2722 0 : sge_count = opts->max_io_size / opts->io_unit_size;
2723 0 : if (sge_count > NVMF_DEFAULT_TX_SGE) {
2724 0 : SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size);
2725 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2726 0 : return NULL;
2727 : }
2728 :
2729 0 : min_in_capsule_data_size = sizeof(struct spdk_nvme_sgl_descriptor) * SPDK_NVMF_MAX_SGL_ENTRIES;
2730 0 : if (opts->in_capsule_data_size < min_in_capsule_data_size) {
2731 0 : SPDK_WARNLOG("In capsule data size is set to %u, this is minimum size required to support msdbd=16\n",
2732 : min_in_capsule_data_size);
2733 0 : opts->in_capsule_data_size = min_in_capsule_data_size;
2734 : }
2735 :
2736 0 : rtransport->event_channel = rdma_create_event_channel();
2737 0 : if (rtransport->event_channel == NULL) {
2738 0 : SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno));
2739 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2740 0 : return NULL;
2741 : }
2742 :
2743 0 : flag = fcntl(rtransport->event_channel->fd, F_GETFL);
2744 0 : if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) {
2745 0 : SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n",
2746 : rtransport->event_channel->fd, spdk_strerror(errno));
2747 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2748 0 : return NULL;
2749 : }
2750 :
2751 0 : data_wr_pool_size = opts->data_wr_pool_size;
2752 0 : if (data_wr_pool_size < SPDK_NVMF_MAX_SGL_ENTRIES * 2 * spdk_env_get_core_count()) {
2753 0 : data_wr_pool_size = SPDK_NVMF_MAX_SGL_ENTRIES * 2 * spdk_env_get_core_count();
2754 0 : SPDK_NOTICELOG("data_wr_pool_size is changed to %zu to guarantee enough cache for handling "
2755 : "at least one IO in each core\n", data_wr_pool_size);
2756 : }
2757 0 : rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", data_wr_pool_size,
2758 : sizeof(struct spdk_nvmf_rdma_request_data), SPDK_MEMPOOL_DEFAULT_CACHE_SIZE,
2759 : SPDK_ENV_SOCKET_ID_ANY);
2760 0 : if (!rtransport->data_wr_pool) {
2761 0 : if (spdk_mempool_lookup("spdk_nvmf_rdma_wr_data") != NULL) {
2762 0 : SPDK_ERRLOG("Unable to allocate work request pool for poll group: already exists\n");
2763 0 : SPDK_ERRLOG("Probably running in multiprocess environment, which is "
2764 : "unsupported by the nvmf library\n");
2765 : } else {
2766 0 : SPDK_ERRLOG("Unable to allocate work request pool for poll group\n");
2767 : }
2768 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2769 0 : return NULL;
2770 : }
2771 :
2772 0 : contexts = rdma_get_devices(NULL);
2773 0 : if (contexts == NULL) {
2774 0 : SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno);
2775 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2776 0 : return NULL;
2777 : }
2778 :
2779 0 : i = 0;
2780 0 : rc = 0;
2781 0 : while (contexts[i] != NULL) {
2782 0 : rc = create_ib_device(rtransport, contexts[i], &device);
2783 0 : if (rc < 0) {
2784 0 : break;
2785 : }
2786 0 : i++;
2787 0 : max_device_sge = spdk_min(max_device_sge, device->attr.max_sge);
2788 0 : device->is_ready = true;
2789 : }
2790 0 : rdma_free_devices(contexts);
2791 :
2792 0 : if (opts->io_unit_size * max_device_sge < opts->max_io_size) {
2793 : /* divide and round up. */
2794 0 : opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge;
2795 :
2796 : /* round up to the nearest 4k. */
2797 0 : opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK;
2798 :
2799 0 : opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE);
2800 0 : SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n",
2801 : opts->io_unit_size);
2802 : }
2803 :
2804 0 : if (rc < 0) {
2805 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2806 0 : return NULL;
2807 : }
2808 :
2809 0 : rc = generate_poll_fds(rtransport);
2810 0 : if (rc < 0) {
2811 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2812 0 : return NULL;
2813 : }
2814 :
2815 0 : rtransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_rdma_accept, &rtransport->transport,
2816 : opts->acceptor_poll_rate);
2817 0 : if (!rtransport->accept_poller) {
2818 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2819 0 : return NULL;
2820 : }
2821 :
2822 0 : return &rtransport->transport;
2823 : }
2824 :
2825 : static void
2826 0 : destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport,
2827 : struct spdk_nvmf_rdma_device *device)
2828 : {
2829 0 : TAILQ_REMOVE(&rtransport->devices, device, link);
2830 0 : spdk_rdma_utils_free_mem_map(&device->map);
2831 0 : if (device->pd) {
2832 0 : if (!g_nvmf_hooks.get_ibv_pd) {
2833 0 : ibv_dealloc_pd(device->pd);
2834 : }
2835 : }
2836 0 : SPDK_DEBUGLOG(rdma, "IB device [%p] is destroyed.\n", device);
2837 0 : free(device);
2838 0 : }
2839 :
2840 : static void
2841 0 : nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w)
2842 : {
2843 : struct spdk_nvmf_rdma_transport *rtransport;
2844 0 : assert(w != NULL);
2845 :
2846 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
2847 0 : spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth);
2848 0 : spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq);
2849 0 : if (rtransport->rdma_opts.no_srq == true) {
2850 0 : spdk_json_write_named_int32(w, "num_cqe", rtransport->rdma_opts.num_cqe);
2851 : }
2852 0 : spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog);
2853 0 : spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching);
2854 0 : }
2855 :
2856 : static int
2857 0 : nvmf_rdma_destroy(struct spdk_nvmf_transport *transport,
2858 : spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg)
2859 : {
2860 : struct spdk_nvmf_rdma_transport *rtransport;
2861 : struct spdk_nvmf_rdma_port *port, *port_tmp;
2862 : struct spdk_nvmf_rdma_device *device, *device_tmp;
2863 :
2864 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
2865 :
2866 0 : TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, port_tmp) {
2867 0 : TAILQ_REMOVE(&rtransport->retry_ports, port, link);
2868 0 : free(port);
2869 : }
2870 :
2871 0 : TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) {
2872 0 : TAILQ_REMOVE(&rtransport->ports, port, link);
2873 0 : rdma_destroy_id(port->id);
2874 0 : free(port);
2875 : }
2876 :
2877 0 : free_poll_fds(rtransport);
2878 :
2879 0 : if (rtransport->event_channel != NULL) {
2880 0 : rdma_destroy_event_channel(rtransport->event_channel);
2881 : }
2882 :
2883 0 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) {
2884 0 : destroy_ib_device(rtransport, device);
2885 : }
2886 :
2887 0 : if (rtransport->data_wr_pool != NULL) {
2888 0 : if (spdk_mempool_count(rtransport->data_wr_pool) != transport->opts.data_wr_pool_size) {
2889 0 : SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n",
2890 : spdk_mempool_count(rtransport->data_wr_pool),
2891 : transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES);
2892 : }
2893 : }
2894 :
2895 0 : spdk_mempool_free(rtransport->data_wr_pool);
2896 :
2897 0 : spdk_poller_unregister(&rtransport->accept_poller);
2898 0 : free(rtransport);
2899 :
2900 0 : if (cb_fn) {
2901 0 : cb_fn(cb_arg);
2902 : }
2903 0 : return 0;
2904 : }
2905 :
2906 : static int nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id,
2907 : struct spdk_nvme_transport_id *trid,
2908 : bool peer);
2909 :
2910 : static bool nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport);
2911 :
2912 : static int
2913 0 : nvmf_rdma_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid,
2914 : struct spdk_nvmf_listen_opts *listen_opts)
2915 : {
2916 : struct spdk_nvmf_rdma_transport *rtransport;
2917 : struct spdk_nvmf_rdma_device *device;
2918 : struct spdk_nvmf_rdma_port *port, *tmp_port;
2919 0 : struct addrinfo *res;
2920 0 : struct addrinfo hints;
2921 : int family;
2922 : int rc;
2923 : long int port_val;
2924 0 : bool is_retry = false;
2925 :
2926 0 : if (!strlen(trid->trsvcid)) {
2927 0 : SPDK_ERRLOG("Service id is required\n");
2928 0 : return -EINVAL;
2929 : }
2930 :
2931 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
2932 0 : assert(rtransport->event_channel != NULL);
2933 :
2934 0 : port = calloc(1, sizeof(*port));
2935 0 : if (!port) {
2936 0 : SPDK_ERRLOG("Port allocation failed\n");
2937 0 : return -ENOMEM;
2938 : }
2939 :
2940 0 : port->trid = trid;
2941 :
2942 0 : switch (trid->adrfam) {
2943 0 : case SPDK_NVMF_ADRFAM_IPV4:
2944 0 : family = AF_INET;
2945 0 : break;
2946 0 : case SPDK_NVMF_ADRFAM_IPV6:
2947 0 : family = AF_INET6;
2948 0 : break;
2949 0 : default:
2950 0 : SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam);
2951 0 : free(port);
2952 0 : return -EINVAL;
2953 : }
2954 :
2955 0 : memset(&hints, 0, sizeof(hints));
2956 0 : hints.ai_family = family;
2957 0 : hints.ai_flags = AI_NUMERICSERV;
2958 0 : hints.ai_socktype = SOCK_STREAM;
2959 0 : hints.ai_protocol = 0;
2960 :
2961 : /* Range check the trsvcid. Fail in 3 cases:
2962 : * < 0: means that spdk_strtol hit an error
2963 : * 0: this results in ephemeral port which we don't want
2964 : * > 65535: port too high
2965 : */
2966 0 : port_val = spdk_strtol(trid->trsvcid, 10);
2967 0 : if (port_val <= 0 || port_val > 65535) {
2968 0 : SPDK_ERRLOG("invalid trsvcid %s\n", trid->trsvcid);
2969 0 : free(port);
2970 0 : return -EINVAL;
2971 : }
2972 :
2973 0 : rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res);
2974 0 : if (rc) {
2975 0 : SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc);
2976 0 : free(port);
2977 0 : return -(abs(rc));
2978 : }
2979 :
2980 0 : rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP);
2981 0 : if (rc < 0) {
2982 0 : SPDK_ERRLOG("rdma_create_id() failed\n");
2983 0 : freeaddrinfo(res);
2984 0 : free(port);
2985 0 : return rc;
2986 : }
2987 :
2988 0 : rc = rdma_bind_addr(port->id, res->ai_addr);
2989 0 : freeaddrinfo(res);
2990 :
2991 0 : if (rc < 0) {
2992 0 : TAILQ_FOREACH(tmp_port, &rtransport->retry_ports, link) {
2993 0 : if (spdk_nvme_transport_id_compare(tmp_port->trid, trid) == 0) {
2994 0 : is_retry = true;
2995 0 : break;
2996 : }
2997 : }
2998 0 : if (!is_retry) {
2999 0 : SPDK_ERRLOG("rdma_bind_addr() failed\n");
3000 : }
3001 0 : rdma_destroy_id(port->id);
3002 0 : free(port);
3003 0 : return rc;
3004 : }
3005 :
3006 0 : if (!port->id->verbs) {
3007 0 : SPDK_ERRLOG("ibv_context is null\n");
3008 0 : rdma_destroy_id(port->id);
3009 0 : free(port);
3010 0 : return -1;
3011 : }
3012 :
3013 0 : rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog);
3014 0 : if (rc < 0) {
3015 0 : SPDK_ERRLOG("rdma_listen() failed\n");
3016 0 : rdma_destroy_id(port->id);
3017 0 : free(port);
3018 0 : return rc;
3019 : }
3020 :
3021 0 : TAILQ_FOREACH(device, &rtransport->devices, link) {
3022 0 : if (device->context == port->id->verbs && device->is_ready) {
3023 0 : port->device = device;
3024 0 : break;
3025 : }
3026 : }
3027 0 : if (!port->device) {
3028 0 : SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n",
3029 : port->id->verbs);
3030 0 : rdma_destroy_id(port->id);
3031 0 : free(port);
3032 0 : nvmf_rdma_rescan_devices(rtransport);
3033 0 : return -EINVAL;
3034 : }
3035 :
3036 0 : SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n",
3037 : trid->traddr, trid->trsvcid);
3038 :
3039 0 : TAILQ_INSERT_TAIL(&rtransport->ports, port, link);
3040 0 : return 0;
3041 : }
3042 :
3043 : static void
3044 0 : nvmf_rdma_stop_listen_ex(struct spdk_nvmf_transport *transport,
3045 : const struct spdk_nvme_transport_id *trid, bool need_retry)
3046 : {
3047 : struct spdk_nvmf_rdma_transport *rtransport;
3048 : struct spdk_nvmf_rdma_port *port, *tmp;
3049 :
3050 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3051 :
3052 0 : if (!need_retry) {
3053 0 : TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp) {
3054 0 : if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) {
3055 0 : TAILQ_REMOVE(&rtransport->retry_ports, port, link);
3056 0 : free(port);
3057 : }
3058 : }
3059 : }
3060 :
3061 0 : TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) {
3062 0 : if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) {
3063 0 : SPDK_DEBUGLOG(rdma, "Port %s:%s removed. need retry: %d\n",
3064 : port->trid->traddr, port->trid->trsvcid, need_retry);
3065 0 : TAILQ_REMOVE(&rtransport->ports, port, link);
3066 0 : rdma_destroy_id(port->id);
3067 0 : port->id = NULL;
3068 0 : port->device = NULL;
3069 0 : if (need_retry) {
3070 0 : TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link);
3071 : } else {
3072 0 : free(port);
3073 : }
3074 0 : break;
3075 : }
3076 : }
3077 0 : }
3078 :
3079 : static void
3080 0 : nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport,
3081 : const struct spdk_nvme_transport_id *trid)
3082 : {
3083 0 : nvmf_rdma_stop_listen_ex(transport, trid, false);
3084 0 : }
3085 :
3086 : static void _nvmf_rdma_register_poller_in_group(void *c);
3087 : static void _nvmf_rdma_remove_poller_in_group(void *c);
3088 :
3089 : static bool
3090 0 : nvmf_rdma_all_pollers_management_done(void *c)
3091 : {
3092 0 : struct poller_manage_ctx *ctx = c;
3093 : int counter;
3094 :
3095 0 : counter = __atomic_sub_fetch(ctx->inflight_op_counter, 1, __ATOMIC_SEQ_CST);
3096 0 : SPDK_DEBUGLOG(rdma, "nvmf_rdma_all_pollers_management_done called. counter: %d, poller: %p\n",
3097 : counter, ctx->rpoller);
3098 :
3099 0 : if (counter == 0) {
3100 0 : free((void *)ctx->inflight_op_counter);
3101 : }
3102 0 : free(ctx);
3103 :
3104 0 : return counter == 0;
3105 : }
3106 :
3107 : static int
3108 0 : nvmf_rdma_manage_poller(struct spdk_nvmf_rdma_transport *rtransport,
3109 : struct spdk_nvmf_rdma_device *device, bool *has_inflight, bool is_add)
3110 : {
3111 : struct spdk_nvmf_rdma_poll_group *rgroup;
3112 : struct spdk_nvmf_rdma_poller *rpoller;
3113 : struct spdk_nvmf_poll_group *poll_group;
3114 : struct poller_manage_ctx *ctx;
3115 : bool found;
3116 : int *inflight_counter;
3117 : spdk_msg_fn do_fn;
3118 :
3119 0 : *has_inflight = false;
3120 0 : do_fn = is_add ? _nvmf_rdma_register_poller_in_group : _nvmf_rdma_remove_poller_in_group;
3121 0 : inflight_counter = calloc(1, sizeof(int));
3122 0 : if (!inflight_counter) {
3123 0 : SPDK_ERRLOG("Failed to allocate inflight counter when removing pollers\n");
3124 0 : return -ENOMEM;
3125 : }
3126 :
3127 0 : TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) {
3128 0 : (*inflight_counter)++;
3129 : }
3130 :
3131 0 : TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) {
3132 0 : found = false;
3133 0 : TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
3134 0 : if (rpoller->device == device) {
3135 0 : found = true;
3136 0 : break;
3137 : }
3138 : }
3139 0 : if (found == is_add) {
3140 0 : __atomic_fetch_sub(inflight_counter, 1, __ATOMIC_SEQ_CST);
3141 0 : continue;
3142 : }
3143 :
3144 0 : ctx = calloc(1, sizeof(struct poller_manage_ctx));
3145 0 : if (!ctx) {
3146 0 : SPDK_ERRLOG("Failed to allocate poller_manage_ctx when removing pollers\n");
3147 0 : if (!*has_inflight) {
3148 0 : free(inflight_counter);
3149 : }
3150 0 : return -ENOMEM;
3151 : }
3152 :
3153 0 : ctx->rtransport = rtransport;
3154 0 : ctx->rgroup = rgroup;
3155 0 : ctx->rpoller = rpoller;
3156 0 : ctx->device = device;
3157 0 : ctx->thread = spdk_get_thread();
3158 0 : ctx->inflight_op_counter = inflight_counter;
3159 0 : *has_inflight = true;
3160 :
3161 0 : poll_group = rgroup->group.group;
3162 0 : if (poll_group->thread != spdk_get_thread()) {
3163 0 : spdk_thread_send_msg(poll_group->thread, do_fn, ctx);
3164 : } else {
3165 0 : do_fn(ctx);
3166 : }
3167 : }
3168 :
3169 0 : if (!*has_inflight) {
3170 0 : free(inflight_counter);
3171 : }
3172 :
3173 0 : return 0;
3174 : }
3175 :
3176 : static void nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport,
3177 : struct spdk_nvmf_rdma_device *device);
3178 :
3179 : static struct spdk_nvmf_rdma_device *
3180 0 : nvmf_rdma_find_ib_device(struct spdk_nvmf_rdma_transport *rtransport,
3181 : struct ibv_context *context)
3182 : {
3183 : struct spdk_nvmf_rdma_device *device, *tmp_device;
3184 :
3185 0 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp_device) {
3186 0 : if (device->need_destroy) {
3187 0 : continue;
3188 : }
3189 :
3190 0 : if (strcmp(device->context->device->dev_name, context->device->dev_name) == 0) {
3191 0 : return device;
3192 : }
3193 : }
3194 :
3195 0 : return NULL;
3196 : }
3197 :
3198 : static bool
3199 0 : nvmf_rdma_check_devices_context(struct spdk_nvmf_rdma_transport *rtransport,
3200 : struct ibv_context *context)
3201 : {
3202 0 : struct spdk_nvmf_rdma_device *old_device, *new_device;
3203 0 : int rc = 0;
3204 0 : bool has_inflight;
3205 :
3206 0 : old_device = nvmf_rdma_find_ib_device(rtransport, context);
3207 :
3208 0 : if (old_device) {
3209 0 : if (old_device->context != context && !old_device->need_destroy && old_device->is_ready) {
3210 : /* context may not have time to be cleaned when rescan. exactly one context
3211 : * is valid for a device so this context must be invalid and just remove it. */
3212 0 : SPDK_WARNLOG("Device %p has a invalid context %p\n", old_device, old_device->context);
3213 0 : old_device->need_destroy = true;
3214 0 : nvmf_rdma_handle_device_removal(rtransport, old_device);
3215 : }
3216 0 : return false;
3217 : }
3218 :
3219 0 : rc = create_ib_device(rtransport, context, &new_device);
3220 : /* TODO: update transport opts. */
3221 0 : if (rc < 0) {
3222 0 : SPDK_ERRLOG("Failed to create ib device for context: %s(%p)\n",
3223 : ibv_get_device_name(context->device), context);
3224 0 : return false;
3225 : }
3226 :
3227 0 : rc = nvmf_rdma_manage_poller(rtransport, new_device, &has_inflight, true);
3228 0 : if (rc < 0) {
3229 0 : SPDK_ERRLOG("Failed to add poller for device context: %s(%p)\n",
3230 : ibv_get_device_name(context->device), context);
3231 0 : return false;
3232 : }
3233 :
3234 0 : if (has_inflight) {
3235 0 : new_device->is_ready = true;
3236 : }
3237 :
3238 0 : return true;
3239 : }
3240 :
3241 : static bool
3242 0 : nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport)
3243 : {
3244 : struct spdk_nvmf_rdma_device *device;
3245 0 : struct ibv_device **ibv_device_list = NULL;
3246 0 : struct ibv_context **contexts = NULL;
3247 0 : int i = 0;
3248 0 : int num_dev = 0;
3249 0 : bool new_create = false, has_new_device = false;
3250 0 : struct ibv_context *tmp_verbs = NULL;
3251 :
3252 : /* do not rescan when any device is destroying, or context may be freed when
3253 : * regenerating the poll fds.
3254 : */
3255 0 : TAILQ_FOREACH(device, &rtransport->devices, link) {
3256 0 : if (device->need_destroy) {
3257 0 : return false;
3258 : }
3259 : }
3260 :
3261 0 : ibv_device_list = ibv_get_device_list(&num_dev);
3262 :
3263 : /* There is a bug in librdmacm. If verbs init failed in rdma_get_devices, it'll be
3264 : * marked as dead verbs and never be init again. So we need to make sure the
3265 : * verbs is available before we call rdma_get_devices. */
3266 0 : if (num_dev >= 0) {
3267 0 : for (i = 0; i < num_dev; i++) {
3268 0 : tmp_verbs = ibv_open_device(ibv_device_list[i]);
3269 0 : if (!tmp_verbs) {
3270 0 : SPDK_WARNLOG("Failed to init ibv device %p, err %d. Skip rescan.\n", ibv_device_list[i], errno);
3271 0 : break;
3272 : }
3273 0 : if (nvmf_rdma_find_ib_device(rtransport, tmp_verbs) == NULL) {
3274 0 : SPDK_DEBUGLOG(rdma, "Find new verbs init ibv device %p(%s).\n", ibv_device_list[i],
3275 : tmp_verbs->device->dev_name);
3276 0 : has_new_device = true;
3277 : }
3278 0 : ibv_close_device(tmp_verbs);
3279 : }
3280 0 : ibv_free_device_list(ibv_device_list);
3281 0 : if (!tmp_verbs || !has_new_device) {
3282 0 : return false;
3283 : }
3284 : }
3285 :
3286 0 : contexts = rdma_get_devices(NULL);
3287 :
3288 0 : for (i = 0; contexts && contexts[i] != NULL; i++) {
3289 0 : new_create |= nvmf_rdma_check_devices_context(rtransport, contexts[i]);
3290 : }
3291 :
3292 0 : if (new_create) {
3293 0 : free_poll_fds(rtransport);
3294 0 : generate_poll_fds(rtransport);
3295 : }
3296 :
3297 0 : if (contexts) {
3298 0 : rdma_free_devices(contexts);
3299 : }
3300 :
3301 0 : return new_create;
3302 : }
3303 :
3304 : static bool
3305 0 : nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport)
3306 : {
3307 : struct spdk_nvmf_rdma_port *port, *tmp_port;
3308 0 : int rc = 0;
3309 0 : bool new_create = false;
3310 :
3311 0 : if (TAILQ_EMPTY(&rtransport->retry_ports)) {
3312 0 : return false;
3313 : }
3314 :
3315 0 : new_create = nvmf_rdma_rescan_devices(rtransport);
3316 :
3317 0 : TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp_port) {
3318 0 : rc = nvmf_rdma_listen(&rtransport->transport, port->trid, NULL);
3319 :
3320 0 : TAILQ_REMOVE(&rtransport->retry_ports, port, link);
3321 0 : if (rc) {
3322 0 : if (new_create) {
3323 0 : SPDK_ERRLOG("Found new IB device but port %s:%s is still failed(%d) to listen.\n",
3324 : port->trid->traddr, port->trid->trsvcid, rc);
3325 : }
3326 0 : TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link);
3327 0 : break;
3328 : } else {
3329 0 : SPDK_NOTICELOG("Port %s:%s come back\n", port->trid->traddr, port->trid->trsvcid);
3330 0 : free(port);
3331 : }
3332 : }
3333 :
3334 0 : return true;
3335 : }
3336 :
3337 : static void
3338 0 : nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport,
3339 : struct spdk_nvmf_rdma_qpair *rqpair, bool drain)
3340 : {
3341 : struct spdk_nvmf_request *req, *tmp;
3342 : struct spdk_nvmf_rdma_request *rdma_req, *req_tmp;
3343 : struct spdk_nvmf_rdma_resources *resources;
3344 :
3345 : /* First process requests which are waiting for response to be sent */
3346 0 : STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_send_queue, state_link, req_tmp) {
3347 0 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
3348 0 : break;
3349 : }
3350 : }
3351 :
3352 : /* We process I/O in the data transfer pending queue at the highest priority. */
3353 0 : STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) {
3354 0 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
3355 0 : break;
3356 : }
3357 : }
3358 :
3359 : /* Then RDMA writes since reads have stronger restrictions than writes */
3360 0 : STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) {
3361 0 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
3362 0 : break;
3363 : }
3364 : }
3365 :
3366 : /* Then we handle request waiting on memory buffers. */
3367 0 : STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) {
3368 0 : rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
3369 0 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
3370 0 : break;
3371 : }
3372 : }
3373 :
3374 0 : resources = rqpair->resources;
3375 0 : while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) {
3376 0 : rdma_req = STAILQ_FIRST(&resources->free_queue);
3377 0 : STAILQ_REMOVE_HEAD(&resources->free_queue, state_link);
3378 0 : rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue);
3379 0 : STAILQ_REMOVE_HEAD(&resources->incoming_queue, link);
3380 :
3381 0 : if (rqpair->srq != NULL) {
3382 0 : rdma_req->req.qpair = &rdma_req->recv->qpair->qpair;
3383 0 : rdma_req->recv->qpair->qd++;
3384 : } else {
3385 0 : rqpair->qd++;
3386 : }
3387 :
3388 0 : rdma_req->receive_tsc = rdma_req->recv->receive_tsc;
3389 0 : rdma_req->state = RDMA_REQUEST_STATE_NEW;
3390 0 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false) {
3391 0 : break;
3392 : }
3393 : }
3394 0 : if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) {
3395 0 : rqpair->poller->stat.pending_free_request++;
3396 : }
3397 0 : }
3398 :
3399 : static void
3400 0 : nvmf_rdma_poller_process_pending_buf_queue(struct spdk_nvmf_rdma_transport *rtransport,
3401 : struct spdk_nvmf_rdma_poller *rpoller)
3402 : {
3403 : struct spdk_nvmf_request *req, *tmp;
3404 : struct spdk_nvmf_rdma_request *rdma_req;
3405 :
3406 0 : STAILQ_FOREACH_SAFE(req, &rpoller->group->group.pending_buf_queue, buf_link, tmp) {
3407 0 : rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
3408 0 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false) {
3409 0 : break;
3410 : }
3411 : }
3412 0 : }
3413 :
3414 : static inline bool
3415 0 : nvmf_rdma_can_ignore_last_wqe_reached(struct spdk_nvmf_rdma_device *device)
3416 : {
3417 : /* iWARP transport and SoftRoCE driver don't support LAST_WQE_REACHED ibv async event */
3418 0 : return nvmf_rdma_is_rxe_device(device) ||
3419 0 : device->context->device->transport_type == IBV_TRANSPORT_IWARP;
3420 : }
3421 :
3422 : static void
3423 0 : nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair)
3424 : {
3425 0 : struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
3426 : struct spdk_nvmf_rdma_transport, transport);
3427 :
3428 0 : nvmf_rdma_qpair_process_pending(rtransport, rqpair, true);
3429 :
3430 : /* nvmf_rdma_close_qpair is not called */
3431 0 : if (!rqpair->to_close) {
3432 0 : return;
3433 : }
3434 :
3435 : /* device is already destroyed and we should force destroy this qpair. */
3436 0 : if (rqpair->poller && rqpair->poller->need_destroy) {
3437 0 : nvmf_rdma_qpair_destroy(rqpair);
3438 0 : return;
3439 : }
3440 :
3441 : /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */
3442 0 : if (rqpair->current_send_depth != 0) {
3443 0 : return;
3444 : }
3445 :
3446 0 : if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) {
3447 0 : return;
3448 : }
3449 :
3450 0 : if (rqpair->srq != NULL && rqpair->last_wqe_reached == false &&
3451 0 : !nvmf_rdma_can_ignore_last_wqe_reached(rqpair->device)) {
3452 0 : return;
3453 : }
3454 :
3455 0 : assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR);
3456 :
3457 0 : nvmf_rdma_qpair_destroy(rqpair);
3458 : }
3459 :
3460 : static int
3461 0 : nvmf_rdma_disconnect(struct rdma_cm_event *evt, bool *event_acked)
3462 : {
3463 : struct spdk_nvmf_qpair *qpair;
3464 : struct spdk_nvmf_rdma_qpair *rqpair;
3465 :
3466 0 : if (evt->id == NULL) {
3467 0 : SPDK_ERRLOG("disconnect request: missing cm_id\n");
3468 0 : return -1;
3469 : }
3470 :
3471 0 : qpair = evt->id->context;
3472 0 : if (qpair == NULL) {
3473 0 : SPDK_ERRLOG("disconnect request: no active connection\n");
3474 0 : return -1;
3475 : }
3476 :
3477 0 : rdma_ack_cm_event(evt);
3478 0 : *event_acked = true;
3479 :
3480 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
3481 :
3482 0 : spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair);
3483 :
3484 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
3485 :
3486 0 : return 0;
3487 : }
3488 :
3489 : #ifdef DEBUG
3490 : static const char *CM_EVENT_STR[] = {
3491 : "RDMA_CM_EVENT_ADDR_RESOLVED",
3492 : "RDMA_CM_EVENT_ADDR_ERROR",
3493 : "RDMA_CM_EVENT_ROUTE_RESOLVED",
3494 : "RDMA_CM_EVENT_ROUTE_ERROR",
3495 : "RDMA_CM_EVENT_CONNECT_REQUEST",
3496 : "RDMA_CM_EVENT_CONNECT_RESPONSE",
3497 : "RDMA_CM_EVENT_CONNECT_ERROR",
3498 : "RDMA_CM_EVENT_UNREACHABLE",
3499 : "RDMA_CM_EVENT_REJECTED",
3500 : "RDMA_CM_EVENT_ESTABLISHED",
3501 : "RDMA_CM_EVENT_DISCONNECTED",
3502 : "RDMA_CM_EVENT_DEVICE_REMOVAL",
3503 : "RDMA_CM_EVENT_MULTICAST_JOIN",
3504 : "RDMA_CM_EVENT_MULTICAST_ERROR",
3505 : "RDMA_CM_EVENT_ADDR_CHANGE",
3506 : "RDMA_CM_EVENT_TIMEWAIT_EXIT"
3507 : };
3508 : #endif /* DEBUG */
3509 :
3510 : static void
3511 0 : nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport,
3512 : struct spdk_nvmf_rdma_port *port)
3513 : {
3514 : struct spdk_nvmf_rdma_poll_group *rgroup;
3515 : struct spdk_nvmf_rdma_poller *rpoller;
3516 : struct spdk_nvmf_rdma_qpair *rqpair;
3517 :
3518 0 : TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) {
3519 0 : TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
3520 0 : RB_FOREACH(rqpair, qpairs_tree, &rpoller->qpairs) {
3521 0 : if (rqpair->listen_id == port->id) {
3522 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
3523 : }
3524 : }
3525 : }
3526 : }
3527 0 : }
3528 :
3529 : static bool
3530 0 : nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport,
3531 : struct rdma_cm_event *event)
3532 : {
3533 : const struct spdk_nvme_transport_id *trid;
3534 : struct spdk_nvmf_rdma_port *port;
3535 : struct spdk_nvmf_rdma_transport *rtransport;
3536 0 : bool event_acked = false;
3537 :
3538 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3539 0 : TAILQ_FOREACH(port, &rtransport->ports, link) {
3540 0 : if (port->id == event->id) {
3541 0 : SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid);
3542 0 : rdma_ack_cm_event(event);
3543 0 : event_acked = true;
3544 0 : trid = port->trid;
3545 0 : break;
3546 : }
3547 : }
3548 :
3549 0 : if (event_acked) {
3550 0 : nvmf_rdma_disconnect_qpairs_on_port(rtransport, port);
3551 :
3552 0 : nvmf_rdma_stop_listen(transport, trid);
3553 0 : nvmf_rdma_listen(transport, trid, NULL);
3554 : }
3555 :
3556 0 : return event_acked;
3557 : }
3558 :
3559 : static void
3560 0 : nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport,
3561 : struct spdk_nvmf_rdma_device *device)
3562 : {
3563 : struct spdk_nvmf_rdma_port *port, *port_tmp;
3564 : int rc;
3565 0 : bool has_inflight;
3566 :
3567 0 : rc = nvmf_rdma_manage_poller(rtransport, device, &has_inflight, false);
3568 0 : if (rc) {
3569 0 : SPDK_ERRLOG("Failed to handle device removal, rc %d\n", rc);
3570 0 : return;
3571 : }
3572 :
3573 0 : if (!has_inflight) {
3574 : /* no pollers, destroy the device */
3575 0 : device->ready_to_destroy = true;
3576 0 : spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_remove_destroyed_device, rtransport);
3577 : }
3578 :
3579 0 : TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) {
3580 0 : if (port->device == device) {
3581 0 : SPDK_NOTICELOG("Port %s:%s on device %s is being removed.\n",
3582 : port->trid->traddr,
3583 : port->trid->trsvcid,
3584 : ibv_get_device_name(port->device->context->device));
3585 :
3586 : /* keep NVMF listener and only destroy structures of the
3587 : * RDMA transport. when the device comes back we can retry listening
3588 : * and the application's workflow will not be interrupted.
3589 : */
3590 0 : nvmf_rdma_stop_listen_ex(&rtransport->transport, port->trid, true);
3591 : }
3592 : }
3593 : }
3594 :
3595 : static void
3596 0 : nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport,
3597 : struct rdma_cm_event *event)
3598 : {
3599 : struct spdk_nvmf_rdma_port *port, *tmp_port;
3600 : struct spdk_nvmf_rdma_transport *rtransport;
3601 :
3602 0 : port = event->id->context;
3603 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3604 :
3605 0 : rdma_ack_cm_event(event);
3606 :
3607 : /* if device removal happens during ctrl qpair disconnecting, it's possible that we receive
3608 : * an DEVICE_REMOVAL event on qpair but the id->qp is just NULL. So we should make sure that
3609 : * we are handling a port event here.
3610 : */
3611 0 : TAILQ_FOREACH(tmp_port, &rtransport->ports, link) {
3612 0 : if (port == tmp_port && port->device && !port->device->need_destroy) {
3613 0 : port->device->need_destroy = true;
3614 0 : nvmf_rdma_handle_device_removal(rtransport, port->device);
3615 : }
3616 : }
3617 0 : }
3618 :
3619 : static void
3620 0 : nvmf_process_cm_events(struct spdk_nvmf_transport *transport, uint32_t max_events)
3621 : {
3622 : struct spdk_nvmf_rdma_transport *rtransport;
3623 0 : struct rdma_cm_event *event;
3624 : uint32_t i;
3625 : int rc;
3626 0 : bool event_acked;
3627 :
3628 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3629 :
3630 0 : if (rtransport->event_channel == NULL) {
3631 0 : return;
3632 : }
3633 :
3634 0 : for (i = 0; i < max_events; i++) {
3635 0 : event_acked = false;
3636 0 : rc = rdma_get_cm_event(rtransport->event_channel, &event);
3637 0 : if (rc) {
3638 0 : if (errno != EAGAIN && errno != EWOULDBLOCK) {
3639 0 : SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno));
3640 : }
3641 0 : break;
3642 : }
3643 :
3644 0 : SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]);
3645 :
3646 0 : spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event);
3647 :
3648 0 : switch (event->event) {
3649 0 : case RDMA_CM_EVENT_ADDR_RESOLVED:
3650 : case RDMA_CM_EVENT_ADDR_ERROR:
3651 : case RDMA_CM_EVENT_ROUTE_RESOLVED:
3652 : case RDMA_CM_EVENT_ROUTE_ERROR:
3653 : /* No action required. The target never attempts to resolve routes. */
3654 0 : break;
3655 0 : case RDMA_CM_EVENT_CONNECT_REQUEST:
3656 0 : rc = nvmf_rdma_connect(transport, event);
3657 0 : if (rc < 0) {
3658 0 : SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc);
3659 0 : break;
3660 : }
3661 0 : break;
3662 0 : case RDMA_CM_EVENT_CONNECT_RESPONSE:
3663 : /* The target never initiates a new connection. So this will not occur. */
3664 0 : break;
3665 0 : case RDMA_CM_EVENT_CONNECT_ERROR:
3666 : /* Can this happen? The docs say it can, but not sure what causes it. */
3667 0 : break;
3668 0 : case RDMA_CM_EVENT_UNREACHABLE:
3669 : case RDMA_CM_EVENT_REJECTED:
3670 : /* These only occur on the client side. */
3671 0 : break;
3672 0 : case RDMA_CM_EVENT_ESTABLISHED:
3673 : /* TODO: Should we be waiting for this event anywhere? */
3674 0 : break;
3675 0 : case RDMA_CM_EVENT_DISCONNECTED:
3676 0 : rc = nvmf_rdma_disconnect(event, &event_acked);
3677 0 : if (rc < 0) {
3678 0 : SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc);
3679 0 : break;
3680 : }
3681 0 : break;
3682 0 : case RDMA_CM_EVENT_DEVICE_REMOVAL:
3683 : /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL
3684 : * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s.
3685 : * Once these events are sent to SPDK, we should release all IB resources and
3686 : * don't make attempts to call any ibv_query/modify/create functions. We can only call
3687 : * ibv_destroy* functions to release user space memory allocated by IB. All kernel
3688 : * resources are already cleaned. */
3689 0 : if (event->id->qp) {
3690 : /* If rdma_cm event has a valid `qp` pointer then the event refers to the
3691 : * corresponding qpair. Otherwise the event refers to a listening device. */
3692 0 : rc = nvmf_rdma_disconnect(event, &event_acked);
3693 0 : if (rc < 0) {
3694 0 : SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc);
3695 0 : break;
3696 : }
3697 : } else {
3698 0 : nvmf_rdma_handle_cm_event_port_removal(transport, event);
3699 0 : event_acked = true;
3700 : }
3701 0 : break;
3702 0 : case RDMA_CM_EVENT_MULTICAST_JOIN:
3703 : case RDMA_CM_EVENT_MULTICAST_ERROR:
3704 : /* Multicast is not used */
3705 0 : break;
3706 0 : case RDMA_CM_EVENT_ADDR_CHANGE:
3707 0 : event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event);
3708 0 : break;
3709 0 : case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3710 : /* For now, do nothing. The target never re-uses queue pairs. */
3711 0 : break;
3712 0 : default:
3713 0 : SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event);
3714 0 : break;
3715 : }
3716 0 : if (!event_acked) {
3717 0 : rdma_ack_cm_event(event);
3718 : }
3719 : }
3720 : }
3721 :
3722 : static void
3723 0 : nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair)
3724 : {
3725 0 : rqpair->last_wqe_reached = true;
3726 0 : nvmf_rdma_destroy_drained_qpair(rqpair);
3727 0 : }
3728 :
3729 : static void
3730 0 : nvmf_rdma_qpair_process_ibv_event(void *ctx)
3731 : {
3732 0 : struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx;
3733 :
3734 0 : if (event_ctx->rqpair) {
3735 0 : STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link);
3736 0 : if (event_ctx->cb_fn) {
3737 0 : event_ctx->cb_fn(event_ctx->rqpair);
3738 : }
3739 : }
3740 0 : free(event_ctx);
3741 0 : }
3742 :
3743 : static int
3744 0 : nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair,
3745 : spdk_nvmf_rdma_qpair_ibv_event fn)
3746 : {
3747 : struct spdk_nvmf_rdma_ibv_event_ctx *ctx;
3748 0 : struct spdk_thread *thr = NULL;
3749 : int rc;
3750 :
3751 0 : if (rqpair->qpair.group) {
3752 0 : thr = rqpair->qpair.group->thread;
3753 0 : } else if (rqpair->destruct_channel) {
3754 0 : thr = spdk_io_channel_get_thread(rqpair->destruct_channel);
3755 : }
3756 :
3757 0 : if (!thr) {
3758 0 : SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair);
3759 0 : return -EINVAL;
3760 : }
3761 :
3762 0 : ctx = calloc(1, sizeof(*ctx));
3763 0 : if (!ctx) {
3764 0 : return -ENOMEM;
3765 : }
3766 :
3767 0 : ctx->rqpair = rqpair;
3768 0 : ctx->cb_fn = fn;
3769 0 : STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link);
3770 :
3771 0 : rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_ibv_event, ctx);
3772 0 : if (rc) {
3773 0 : STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link);
3774 0 : free(ctx);
3775 : }
3776 :
3777 0 : return rc;
3778 : }
3779 :
3780 : static int
3781 0 : nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device)
3782 : {
3783 : int rc;
3784 0 : struct spdk_nvmf_rdma_qpair *rqpair = NULL;
3785 0 : struct ibv_async_event event;
3786 :
3787 0 : rc = ibv_get_async_event(device->context, &event);
3788 :
3789 0 : if (rc) {
3790 : /* In non-blocking mode -1 means there are no events available */
3791 0 : return rc;
3792 : }
3793 :
3794 0 : switch (event.event_type) {
3795 0 : case IBV_EVENT_QP_FATAL:
3796 : case IBV_EVENT_QP_LAST_WQE_REACHED:
3797 : case IBV_EVENT_QP_REQ_ERR:
3798 : case IBV_EVENT_QP_ACCESS_ERR:
3799 : case IBV_EVENT_COMM_EST:
3800 : case IBV_EVENT_PATH_MIG:
3801 : case IBV_EVENT_PATH_MIG_ERR:
3802 0 : rqpair = event.element.qp->qp_context;
3803 0 : if (!rqpair) {
3804 : /* Any QP event for NVMe-RDMA initiator may be returned. */
3805 0 : SPDK_NOTICELOG("Async QP event for unknown QP: %s\n",
3806 : ibv_event_type_str(event.event_type));
3807 0 : break;
3808 : }
3809 :
3810 0 : switch (event.event_type) {
3811 0 : case IBV_EVENT_QP_FATAL:
3812 0 : SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair);
3813 0 : spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0,
3814 : (uintptr_t)rqpair, event.event_type);
3815 0 : rqpair->ibv_in_error_state = true;
3816 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
3817 0 : break;
3818 0 : case IBV_EVENT_QP_LAST_WQE_REACHED:
3819 : /* This event only occurs for shared receive queues. */
3820 0 : SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair);
3821 0 : rc = nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached);
3822 0 : if (rc) {
3823 0 : SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc);
3824 0 : rqpair->last_wqe_reached = true;
3825 : }
3826 0 : break;
3827 0 : case IBV_EVENT_QP_REQ_ERR:
3828 : case IBV_EVENT_QP_ACCESS_ERR:
3829 : case IBV_EVENT_COMM_EST:
3830 : case IBV_EVENT_PATH_MIG:
3831 : case IBV_EVENT_PATH_MIG_ERR:
3832 0 : SPDK_NOTICELOG("Async QP event: %s\n",
3833 : ibv_event_type_str(event.event_type));
3834 0 : spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0,
3835 : (uintptr_t)rqpair, event.event_type);
3836 0 : rqpair->ibv_in_error_state = true;
3837 0 : break;
3838 0 : default:
3839 0 : break;
3840 : }
3841 0 : break;
3842 0 : case IBV_EVENT_DEVICE_FATAL:
3843 0 : SPDK_ERRLOG("Device Fatal event[%s] received on %s. device: %p\n",
3844 : ibv_event_type_str(event.event_type), ibv_get_device_name(device->context->device), device);
3845 0 : device->need_destroy = true;
3846 0 : break;
3847 0 : case IBV_EVENT_CQ_ERR:
3848 : case IBV_EVENT_PORT_ACTIVE:
3849 : case IBV_EVENT_PORT_ERR:
3850 : case IBV_EVENT_LID_CHANGE:
3851 : case IBV_EVENT_PKEY_CHANGE:
3852 : case IBV_EVENT_SM_CHANGE:
3853 : case IBV_EVENT_SRQ_ERR:
3854 : case IBV_EVENT_SRQ_LIMIT_REACHED:
3855 : case IBV_EVENT_CLIENT_REREGISTER:
3856 : case IBV_EVENT_GID_CHANGE:
3857 : case IBV_EVENT_SQ_DRAINED:
3858 : default:
3859 0 : SPDK_NOTICELOG("Async event: %s\n",
3860 : ibv_event_type_str(event.event_type));
3861 0 : spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type);
3862 0 : break;
3863 : }
3864 0 : ibv_ack_async_event(&event);
3865 :
3866 0 : return 0;
3867 : }
3868 :
3869 : static void
3870 0 : nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events)
3871 : {
3872 0 : int rc = 0;
3873 0 : uint32_t i = 0;
3874 :
3875 0 : for (i = 0; i < max_events; i++) {
3876 0 : rc = nvmf_process_ib_event(device);
3877 0 : if (rc) {
3878 0 : break;
3879 : }
3880 : }
3881 :
3882 0 : SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i);
3883 0 : }
3884 :
3885 : static int
3886 0 : nvmf_rdma_accept(void *ctx)
3887 : {
3888 0 : int nfds, i = 0;
3889 0 : struct spdk_nvmf_transport *transport = ctx;
3890 : struct spdk_nvmf_rdma_transport *rtransport;
3891 : struct spdk_nvmf_rdma_device *device, *tmp;
3892 : uint32_t count;
3893 : short revents;
3894 : bool do_retry;
3895 :
3896 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3897 0 : do_retry = nvmf_rdma_retry_listen_port(rtransport);
3898 :
3899 0 : count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0);
3900 :
3901 0 : if (nfds <= 0) {
3902 0 : return do_retry ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE;
3903 : }
3904 :
3905 : /* The first poll descriptor is RDMA CM event */
3906 0 : if (rtransport->poll_fds[i++].revents & POLLIN) {
3907 0 : nvmf_process_cm_events(transport, NVMF_RDMA_MAX_EVENTS_PER_POLL);
3908 0 : nfds--;
3909 : }
3910 :
3911 0 : if (nfds == 0) {
3912 0 : return SPDK_POLLER_BUSY;
3913 : }
3914 :
3915 : /* Second and subsequent poll descriptors are IB async events */
3916 0 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) {
3917 0 : revents = rtransport->poll_fds[i++].revents;
3918 0 : if (revents & POLLIN) {
3919 0 : if (spdk_likely(!device->need_destroy)) {
3920 0 : nvmf_process_ib_events(device, NVMF_RDMA_MAX_EVENTS_PER_POLL);
3921 0 : if (spdk_unlikely(device->need_destroy)) {
3922 0 : nvmf_rdma_handle_device_removal(rtransport, device);
3923 : }
3924 : }
3925 0 : nfds--;
3926 0 : } else if (revents & POLLNVAL || revents & POLLHUP) {
3927 0 : SPDK_ERRLOG("Receive unknown revent %x on device %p\n", (int)revents, device);
3928 0 : nfds--;
3929 : }
3930 : }
3931 : /* check all flagged fd's have been served */
3932 0 : assert(nfds == 0);
3933 :
3934 0 : return count > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE;
3935 : }
3936 :
3937 : static void
3938 0 : nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem,
3939 : struct spdk_nvmf_ctrlr_data *cdata)
3940 : {
3941 0 : cdata->nvmf_specific.msdbd = NVMF_DEFAULT_MSDBD;
3942 :
3943 : /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled
3944 : since in-capsule data only works with NVME drives that support SGL memory layout */
3945 0 : if (transport->opts.dif_insert_or_strip) {
3946 0 : cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16;
3947 : }
3948 :
3949 0 : if (cdata->nvmf_specific.ioccsz > ((sizeof(struct spdk_nvme_cmd) + 0x1000) / 16)) {
3950 0 : SPDK_WARNLOG("RDMA is configured to support up to 16 SGL entries while in capsule"
3951 : " data is greater than 4KiB.\n");
3952 0 : SPDK_WARNLOG("When used in conjunction with the NVMe-oF initiator from the Linux "
3953 : "kernel between versions 5.4 and 5.12 data corruption may occur for "
3954 : "writes that are not a multiple of 4KiB in size.\n");
3955 : }
3956 0 : }
3957 :
3958 : static void
3959 0 : nvmf_rdma_discover(struct spdk_nvmf_transport *transport,
3960 : struct spdk_nvme_transport_id *trid,
3961 : struct spdk_nvmf_discovery_log_page_entry *entry)
3962 : {
3963 0 : entry->trtype = SPDK_NVMF_TRTYPE_RDMA;
3964 0 : entry->adrfam = trid->adrfam;
3965 0 : entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED;
3966 :
3967 0 : spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' ');
3968 0 : spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' ');
3969 :
3970 0 : entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED;
3971 0 : entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE;
3972 0 : entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM;
3973 0 : }
3974 :
3975 : static int
3976 0 : nvmf_rdma_poller_create(struct spdk_nvmf_rdma_transport *rtransport,
3977 : struct spdk_nvmf_rdma_poll_group *rgroup, struct spdk_nvmf_rdma_device *device,
3978 : struct spdk_nvmf_rdma_poller **out_poller)
3979 : {
3980 : struct spdk_nvmf_rdma_poller *poller;
3981 0 : struct spdk_rdma_provider_srq_init_attr srq_init_attr;
3982 0 : struct spdk_nvmf_rdma_resource_opts opts;
3983 : int num_cqe;
3984 :
3985 0 : poller = calloc(1, sizeof(*poller));
3986 0 : if (!poller) {
3987 0 : SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n");
3988 0 : return -1;
3989 : }
3990 :
3991 0 : poller->device = device;
3992 0 : poller->group = rgroup;
3993 0 : *out_poller = poller;
3994 :
3995 0 : RB_INIT(&poller->qpairs);
3996 0 : STAILQ_INIT(&poller->qpairs_pending_send);
3997 0 : STAILQ_INIT(&poller->qpairs_pending_recv);
3998 :
3999 0 : TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link);
4000 0 : SPDK_DEBUGLOG(rdma, "Create poller %p on device %p in poll group %p.\n", poller, device, rgroup);
4001 0 : if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) {
4002 0 : if ((int)rtransport->rdma_opts.max_srq_depth > device->attr.max_srq_wr) {
4003 0 : SPDK_WARNLOG("Requested SRQ depth %u, max supported by dev %s is %d\n",
4004 : rtransport->rdma_opts.max_srq_depth, device->context->device->name, device->attr.max_srq_wr);
4005 : }
4006 0 : poller->max_srq_depth = spdk_min((int)rtransport->rdma_opts.max_srq_depth, device->attr.max_srq_wr);
4007 :
4008 0 : device->num_srq++;
4009 0 : memset(&srq_init_attr, 0, sizeof(srq_init_attr));
4010 0 : srq_init_attr.pd = device->pd;
4011 0 : srq_init_attr.stats = &poller->stat.qp_stats.recv;
4012 0 : srq_init_attr.srq_init_attr.attr.max_wr = poller->max_srq_depth;
4013 0 : srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE);
4014 0 : poller->srq = spdk_rdma_provider_srq_create(&srq_init_attr);
4015 0 : if (!poller->srq) {
4016 0 : SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno);
4017 0 : return -1;
4018 : }
4019 :
4020 0 : opts.qp = poller->srq;
4021 0 : opts.map = device->map;
4022 0 : opts.qpair = NULL;
4023 0 : opts.shared = true;
4024 0 : opts.max_queue_depth = poller->max_srq_depth;
4025 0 : opts.in_capsule_data_size = rtransport->transport.opts.in_capsule_data_size;
4026 :
4027 0 : poller->resources = nvmf_rdma_resources_create(&opts);
4028 0 : if (!poller->resources) {
4029 0 : SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n");
4030 0 : return -1;
4031 : }
4032 : }
4033 :
4034 : /*
4035 : * When using an srq, we can limit the completion queue at startup.
4036 : * The following formula represents the calculation:
4037 : * num_cqe = num_recv + num_data_wr + num_send_wr.
4038 : * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth
4039 : */
4040 0 : if (poller->srq) {
4041 0 : num_cqe = poller->max_srq_depth * 3;
4042 : } else {
4043 0 : num_cqe = rtransport->rdma_opts.num_cqe;
4044 : }
4045 :
4046 0 : poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0);
4047 0 : if (!poller->cq) {
4048 0 : SPDK_ERRLOG("Unable to create completion queue\n");
4049 0 : return -1;
4050 : }
4051 0 : poller->num_cqe = num_cqe;
4052 0 : return 0;
4053 : }
4054 :
4055 : static void
4056 0 : _nvmf_rdma_register_poller_in_group(void *c)
4057 : {
4058 0 : struct spdk_nvmf_rdma_poller *poller;
4059 0 : struct poller_manage_ctx *ctx = c;
4060 : struct spdk_nvmf_rdma_device *device;
4061 : int rc;
4062 :
4063 0 : rc = nvmf_rdma_poller_create(ctx->rtransport, ctx->rgroup, ctx->device, &poller);
4064 0 : if (rc < 0 && poller) {
4065 0 : nvmf_rdma_poller_destroy(poller);
4066 : }
4067 :
4068 0 : device = ctx->device;
4069 0 : if (nvmf_rdma_all_pollers_management_done(ctx)) {
4070 0 : device->is_ready = true;
4071 : }
4072 0 : }
4073 :
4074 : static void nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group);
4075 :
4076 : static struct spdk_nvmf_transport_poll_group *
4077 5 : nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport,
4078 : struct spdk_nvmf_poll_group *group)
4079 : {
4080 : struct spdk_nvmf_rdma_transport *rtransport;
4081 : struct spdk_nvmf_rdma_poll_group *rgroup;
4082 5 : struct spdk_nvmf_rdma_poller *poller;
4083 : struct spdk_nvmf_rdma_device *device;
4084 : int rc;
4085 :
4086 5 : if (spdk_interrupt_mode_is_enabled()) {
4087 0 : SPDK_ERRLOG("RDMA transport does not support interrupt mode\n");
4088 0 : return NULL;
4089 : }
4090 :
4091 5 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
4092 :
4093 5 : rgroup = calloc(1, sizeof(*rgroup));
4094 5 : if (!rgroup) {
4095 0 : return NULL;
4096 : }
4097 :
4098 5 : TAILQ_INIT(&rgroup->pollers);
4099 :
4100 5 : TAILQ_FOREACH(device, &rtransport->devices, link) {
4101 0 : rc = nvmf_rdma_poller_create(rtransport, rgroup, device, &poller);
4102 0 : if (rc < 0) {
4103 0 : nvmf_rdma_poll_group_destroy(&rgroup->group);
4104 0 : return NULL;
4105 : }
4106 : }
4107 :
4108 5 : TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link);
4109 5 : if (rtransport->conn_sched.next_admin_pg == NULL) {
4110 1 : rtransport->conn_sched.next_admin_pg = rgroup;
4111 1 : rtransport->conn_sched.next_io_pg = rgroup;
4112 : }
4113 :
4114 5 : return &rgroup->group;
4115 : }
4116 :
4117 : static uint32_t
4118 12 : nvmf_poll_group_get_io_qpair_count(struct spdk_nvmf_poll_group *pg)
4119 : {
4120 : uint32_t count;
4121 :
4122 : /* Just assume that unassociated qpairs will eventually be io
4123 : * qpairs. This is close enough for the use cases for this
4124 : * function.
4125 : */
4126 12 : pthread_mutex_lock(&pg->mutex);
4127 12 : count = pg->stat.current_io_qpairs + pg->current_unassociated_qpairs;
4128 12 : pthread_mutex_unlock(&pg->mutex);
4129 :
4130 12 : return count;
4131 : }
4132 :
4133 : static struct spdk_nvmf_transport_poll_group *
4134 14 : nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair)
4135 : {
4136 : struct spdk_nvmf_rdma_transport *rtransport;
4137 : struct spdk_nvmf_rdma_poll_group **pg;
4138 : struct spdk_nvmf_transport_poll_group *result;
4139 : uint32_t count;
4140 :
4141 14 : rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
4142 :
4143 14 : if (TAILQ_EMPTY(&rtransport->poll_groups)) {
4144 2 : return NULL;
4145 : }
4146 :
4147 12 : if (qpair->qid == 0) {
4148 6 : pg = &rtransport->conn_sched.next_admin_pg;
4149 : } else {
4150 : struct spdk_nvmf_rdma_poll_group *pg_min, *pg_start, *pg_current;
4151 : uint32_t min_value;
4152 :
4153 6 : pg = &rtransport->conn_sched.next_io_pg;
4154 6 : pg_min = *pg;
4155 6 : pg_start = *pg;
4156 6 : pg_current = *pg;
4157 6 : min_value = nvmf_poll_group_get_io_qpair_count(pg_current->group.group);
4158 :
4159 : while (1) {
4160 6 : count = nvmf_poll_group_get_io_qpair_count(pg_current->group.group);
4161 :
4162 6 : if (count < min_value) {
4163 0 : min_value = count;
4164 0 : pg_min = pg_current;
4165 : }
4166 :
4167 6 : pg_current = TAILQ_NEXT(pg_current, link);
4168 6 : if (pg_current == NULL) {
4169 2 : pg_current = TAILQ_FIRST(&rtransport->poll_groups);
4170 : }
4171 :
4172 6 : if (pg_current == pg_start || min_value == 0) {
4173 : break;
4174 : }
4175 : }
4176 6 : *pg = pg_min;
4177 : }
4178 :
4179 12 : assert(*pg != NULL);
4180 :
4181 12 : result = &(*pg)->group;
4182 :
4183 12 : *pg = TAILQ_NEXT(*pg, link);
4184 12 : if (*pg == NULL) {
4185 4 : *pg = TAILQ_FIRST(&rtransport->poll_groups);
4186 : }
4187 :
4188 12 : return result;
4189 : }
4190 :
4191 : static void
4192 0 : nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller)
4193 : {
4194 : struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair;
4195 : int rc;
4196 :
4197 0 : TAILQ_REMOVE(&poller->group->pollers, poller, link);
4198 0 : RB_FOREACH_SAFE(qpair, qpairs_tree, &poller->qpairs, tmp_qpair) {
4199 0 : nvmf_rdma_qpair_destroy(qpair);
4200 : }
4201 :
4202 0 : if (poller->srq) {
4203 0 : if (poller->resources) {
4204 0 : nvmf_rdma_resources_destroy(poller->resources);
4205 : }
4206 0 : spdk_rdma_provider_srq_destroy(poller->srq);
4207 0 : SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq);
4208 : }
4209 :
4210 0 : if (poller->cq) {
4211 0 : rc = ibv_destroy_cq(poller->cq);
4212 0 : if (rc != 0) {
4213 0 : SPDK_ERRLOG("Destroy cq return %d, error: %s\n", rc, strerror(errno));
4214 : }
4215 : }
4216 :
4217 0 : if (poller->destroy_cb) {
4218 0 : poller->destroy_cb(poller->destroy_cb_ctx);
4219 0 : poller->destroy_cb = NULL;
4220 : }
4221 :
4222 0 : free(poller);
4223 0 : }
4224 :
4225 : static void
4226 5 : nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group)
4227 : {
4228 : struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup;
4229 : struct spdk_nvmf_rdma_poller *poller, *tmp;
4230 : struct spdk_nvmf_rdma_transport *rtransport;
4231 :
4232 5 : rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
4233 5 : if (!rgroup) {
4234 0 : return;
4235 : }
4236 :
4237 5 : TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) {
4238 0 : nvmf_rdma_poller_destroy(poller);
4239 : }
4240 :
4241 5 : if (rgroup->group.transport == NULL) {
4242 : /* Transport can be NULL when nvmf_rdma_poll_group_create()
4243 : * calls this function directly in a failure path. */
4244 0 : free(rgroup);
4245 0 : return;
4246 : }
4247 :
4248 5 : rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport);
4249 :
4250 5 : next_rgroup = TAILQ_NEXT(rgroup, link);
4251 5 : TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link);
4252 5 : if (next_rgroup == NULL) {
4253 1 : next_rgroup = TAILQ_FIRST(&rtransport->poll_groups);
4254 : }
4255 5 : if (rtransport->conn_sched.next_admin_pg == rgroup) {
4256 5 : rtransport->conn_sched.next_admin_pg = next_rgroup;
4257 : }
4258 5 : if (rtransport->conn_sched.next_io_pg == rgroup) {
4259 5 : rtransport->conn_sched.next_io_pg = next_rgroup;
4260 : }
4261 :
4262 5 : free(rgroup);
4263 : }
4264 :
4265 : static void
4266 0 : nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair)
4267 : {
4268 0 : if (rqpair->cm_id != NULL) {
4269 0 : nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES);
4270 : }
4271 0 : }
4272 :
4273 : static int
4274 0 : nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group,
4275 : struct spdk_nvmf_qpair *qpair)
4276 : {
4277 : struct spdk_nvmf_rdma_poll_group *rgroup;
4278 : struct spdk_nvmf_rdma_qpair *rqpair;
4279 : struct spdk_nvmf_rdma_device *device;
4280 : struct spdk_nvmf_rdma_poller *poller;
4281 : int rc;
4282 :
4283 0 : rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
4284 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4285 :
4286 0 : device = rqpair->device;
4287 :
4288 0 : TAILQ_FOREACH(poller, &rgroup->pollers, link) {
4289 0 : if (poller->device == device) {
4290 0 : break;
4291 : }
4292 : }
4293 :
4294 0 : if (!poller) {
4295 0 : SPDK_ERRLOG("No poller found for device.\n");
4296 0 : return -1;
4297 : }
4298 :
4299 0 : if (poller->need_destroy) {
4300 0 : SPDK_ERRLOG("Poller is destroying.\n");
4301 0 : return -1;
4302 : }
4303 :
4304 0 : rqpair->poller = poller;
4305 0 : rqpair->srq = rqpair->poller->srq;
4306 :
4307 0 : rc = nvmf_rdma_qpair_initialize(qpair);
4308 0 : if (rc < 0) {
4309 0 : SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair);
4310 0 : rqpair->poller = NULL;
4311 0 : rqpair->srq = NULL;
4312 0 : return -1;
4313 : }
4314 :
4315 0 : RB_INSERT(qpairs_tree, &poller->qpairs, rqpair);
4316 :
4317 0 : rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair);
4318 0 : if (rc) {
4319 : /* Try to reject, but we probably can't */
4320 0 : nvmf_rdma_qpair_reject_connection(rqpair);
4321 0 : return -1;
4322 : }
4323 :
4324 0 : return 0;
4325 : }
4326 :
4327 : static int
4328 0 : nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group,
4329 : struct spdk_nvmf_qpair *qpair)
4330 : {
4331 : struct spdk_nvmf_rdma_qpair *rqpair;
4332 :
4333 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4334 0 : assert(group->transport->tgt != NULL);
4335 :
4336 0 : rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt);
4337 :
4338 0 : if (!rqpair->destruct_channel) {
4339 0 : SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair);
4340 0 : return 0;
4341 : }
4342 :
4343 : /* Sanity check that we get io_channel on the correct thread */
4344 0 : if (qpair->group) {
4345 0 : assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel));
4346 : }
4347 :
4348 0 : return 0;
4349 : }
4350 :
4351 : static int
4352 0 : nvmf_rdma_request_free(struct spdk_nvmf_request *req)
4353 : {
4354 0 : struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
4355 0 : struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport,
4356 : struct spdk_nvmf_rdma_transport, transport);
4357 0 : struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair,
4358 : struct spdk_nvmf_rdma_qpair, qpair);
4359 :
4360 : /*
4361 : * AER requests are freed when a qpair is destroyed. The recv corresponding to that request
4362 : * needs to be returned to the shared receive queue or the poll group will eventually be
4363 : * starved of RECV structures.
4364 : */
4365 0 : if (rqpair->srq && rdma_req->recv) {
4366 : int rc;
4367 0 : struct ibv_recv_wr *bad_recv_wr;
4368 :
4369 0 : spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, &rdma_req->recv->wr);
4370 0 : rc = spdk_rdma_provider_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr);
4371 0 : if (rc) {
4372 0 : SPDK_ERRLOG("Unable to re-post rx descriptor\n");
4373 : }
4374 : }
4375 :
4376 0 : _nvmf_rdma_request_free(rdma_req, rtransport);
4377 0 : return 0;
4378 : }
4379 :
4380 : static int
4381 0 : nvmf_rdma_request_complete(struct spdk_nvmf_request *req)
4382 : {
4383 0 : struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport,
4384 : struct spdk_nvmf_rdma_transport, transport);
4385 0 : struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req,
4386 : struct spdk_nvmf_rdma_request, req);
4387 0 : struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair,
4388 : struct spdk_nvmf_rdma_qpair, qpair);
4389 :
4390 0 : if (spdk_unlikely(rqpair->ibv_in_error_state)) {
4391 : /* The connection is dead. Move the request directly to the completed state. */
4392 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
4393 : } else {
4394 : /* The connection is alive, so process the request as normal */
4395 0 : rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
4396 : }
4397 :
4398 0 : nvmf_rdma_request_process(rtransport, rdma_req);
4399 :
4400 0 : return 0;
4401 : }
4402 :
4403 : static void
4404 0 : nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair,
4405 : spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg)
4406 : {
4407 0 : struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4408 :
4409 0 : rqpair->to_close = true;
4410 :
4411 : /* This happens only when the qpair is disconnected before
4412 : * it is added to the poll group. Since there is no poll group,
4413 : * the RDMA qp has not been initialized yet and the RDMA CM
4414 : * event has not yet been acknowledged, so we need to reject it.
4415 : */
4416 0 : if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) {
4417 0 : nvmf_rdma_qpair_reject_connection(rqpair);
4418 0 : nvmf_rdma_qpair_destroy(rqpair);
4419 0 : return;
4420 : }
4421 :
4422 0 : if (rqpair->rdma_qp) {
4423 0 : spdk_rdma_provider_qp_disconnect(rqpair->rdma_qp);
4424 : }
4425 :
4426 0 : nvmf_rdma_destroy_drained_qpair(rqpair);
4427 :
4428 0 : if (cb_fn) {
4429 0 : cb_fn(cb_arg);
4430 : }
4431 : }
4432 :
4433 : static struct spdk_nvmf_rdma_qpair *
4434 0 : get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc)
4435 : {
4436 0 : struct spdk_nvmf_rdma_qpair find;
4437 :
4438 0 : find.qp_num = wc->qp_num;
4439 :
4440 0 : return RB_FIND(qpairs_tree, &rpoller->qpairs, &find);
4441 : }
4442 :
4443 : #ifdef DEBUG
4444 : static int
4445 0 : nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req)
4446 : {
4447 0 : return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST ||
4448 0 : rdma_req->state == RDMA_REQUEST_STATE_COMPLETING;
4449 : }
4450 : #endif
4451 :
4452 : static void
4453 0 : _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr,
4454 : int rc)
4455 : {
4456 : struct spdk_nvmf_rdma_recv *rdma_recv;
4457 : struct spdk_nvmf_rdma_wr *bad_rdma_wr;
4458 :
4459 0 : SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc);
4460 0 : while (bad_recv_wr != NULL) {
4461 0 : bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id;
4462 0 : rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr);
4463 :
4464 0 : rdma_recv->qpair->current_recv_depth++;
4465 0 : bad_recv_wr = bad_recv_wr->next;
4466 0 : SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc);
4467 0 : spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair);
4468 : }
4469 0 : }
4470 :
4471 : static void
4472 0 : _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc)
4473 : {
4474 0 : SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc);
4475 0 : while (bad_recv_wr != NULL) {
4476 0 : bad_recv_wr = bad_recv_wr->next;
4477 0 : rqpair->current_recv_depth++;
4478 : }
4479 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
4480 0 : }
4481 :
4482 : static void
4483 0 : _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport,
4484 : struct spdk_nvmf_rdma_poller *rpoller)
4485 : {
4486 : struct spdk_nvmf_rdma_qpair *rqpair;
4487 0 : struct ibv_recv_wr *bad_recv_wr;
4488 : int rc;
4489 :
4490 0 : if (rpoller->srq) {
4491 0 : rc = spdk_rdma_provider_srq_flush_recv_wrs(rpoller->srq, &bad_recv_wr);
4492 0 : if (spdk_unlikely(rc)) {
4493 0 : _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc);
4494 : }
4495 : } else {
4496 0 : while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) {
4497 0 : rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv);
4498 0 : rc = spdk_rdma_provider_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr);
4499 0 : if (spdk_unlikely(rc)) {
4500 0 : _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc);
4501 : }
4502 0 : STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link);
4503 : }
4504 : }
4505 0 : }
4506 :
4507 : static void
4508 0 : _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport,
4509 : struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc)
4510 : {
4511 : struct spdk_nvmf_rdma_wr *bad_rdma_wr;
4512 0 : struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL;
4513 :
4514 0 : SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc);
4515 0 : for (; bad_wr != NULL; bad_wr = bad_wr->next) {
4516 0 : bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id;
4517 0 : assert(rqpair->current_send_depth > 0);
4518 0 : rqpair->current_send_depth--;
4519 0 : switch (bad_rdma_wr->type) {
4520 0 : case RDMA_WR_TYPE_DATA:
4521 0 : cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data_wr);
4522 0 : if (bad_wr->opcode == IBV_WR_RDMA_READ) {
4523 0 : assert(rqpair->current_read_depth > 0);
4524 0 : rqpair->current_read_depth--;
4525 : }
4526 0 : break;
4527 0 : case RDMA_WR_TYPE_SEND:
4528 0 : cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr);
4529 0 : break;
4530 0 : default:
4531 0 : SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair);
4532 0 : prev_rdma_req = cur_rdma_req;
4533 0 : continue;
4534 : }
4535 :
4536 0 : if (prev_rdma_req == cur_rdma_req) {
4537 : /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */
4538 : /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */
4539 0 : continue;
4540 : }
4541 :
4542 0 : switch (cur_rdma_req->state) {
4543 0 : case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
4544 0 : cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR;
4545 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, cur_rdma_req, state_link);
4546 0 : cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
4547 0 : break;
4548 0 : case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST:
4549 : case RDMA_REQUEST_STATE_COMPLETING:
4550 0 : cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
4551 0 : break;
4552 0 : default:
4553 0 : SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n",
4554 : cur_rdma_req->state, rqpair);
4555 0 : continue;
4556 : }
4557 :
4558 0 : nvmf_rdma_request_process(rtransport, cur_rdma_req);
4559 0 : prev_rdma_req = cur_rdma_req;
4560 : }
4561 :
4562 0 : if (spdk_nvmf_qpair_is_active(&rqpair->qpair)) {
4563 : /* Disconnect the connection. */
4564 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
4565 : }
4566 :
4567 0 : }
4568 :
4569 : static void
4570 0 : _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport,
4571 : struct spdk_nvmf_rdma_poller *rpoller)
4572 : {
4573 : struct spdk_nvmf_rdma_qpair *rqpair;
4574 0 : struct ibv_send_wr *bad_wr = NULL;
4575 : int rc;
4576 :
4577 0 : while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) {
4578 0 : rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send);
4579 0 : rc = spdk_rdma_provider_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr);
4580 :
4581 : /* bad wr always points to the first wr that failed. */
4582 0 : if (spdk_unlikely(rc)) {
4583 0 : _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc);
4584 : }
4585 0 : STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link);
4586 : }
4587 0 : }
4588 :
4589 : static const char *
4590 0 : nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type)
4591 : {
4592 0 : switch (wr_type) {
4593 0 : case RDMA_WR_TYPE_RECV:
4594 0 : return "RECV";
4595 0 : case RDMA_WR_TYPE_SEND:
4596 0 : return "SEND";
4597 0 : case RDMA_WR_TYPE_DATA:
4598 0 : return "DATA";
4599 0 : default:
4600 0 : SPDK_ERRLOG("Unknown WR type %d\n", wr_type);
4601 0 : SPDK_UNREACHABLE();
4602 : }
4603 : }
4604 :
4605 : static inline void
4606 0 : nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc)
4607 : {
4608 0 : enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type;
4609 :
4610 0 : if (wc->status == IBV_WC_WR_FLUSH_ERR) {
4611 : /* If qpair is in ERR state, we will receive completions for all posted and not completed
4612 : * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */
4613 0 : SPDK_DEBUGLOG(rdma,
4614 : "Error on CQ %p, (qp state %d, in_error %d) request 0x%lu, type %s, status: (%d): %s\n",
4615 : rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_in_error_state, wc->wr_id,
4616 : nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status));
4617 : } else {
4618 0 : SPDK_ERRLOG("Error on CQ %p, (qp state %d, in_error %d) request 0x%lu, type %s, status: (%d): %s\n",
4619 : rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_in_error_state, wc->wr_id,
4620 : nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status));
4621 : }
4622 0 : }
4623 :
4624 : static int
4625 0 : nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport,
4626 : struct spdk_nvmf_rdma_poller *rpoller)
4627 : {
4628 0 : struct ibv_wc wc[32];
4629 : struct spdk_nvmf_rdma_wr *rdma_wr;
4630 : struct spdk_nvmf_rdma_request *rdma_req;
4631 : struct spdk_nvmf_rdma_recv *rdma_recv;
4632 : struct spdk_nvmf_rdma_qpair *rqpair, *tmp_rqpair;
4633 : int reaped, i;
4634 0 : int count = 0;
4635 : int rc;
4636 0 : bool error = false;
4637 0 : uint64_t poll_tsc = spdk_get_ticks();
4638 :
4639 0 : if (spdk_unlikely(rpoller->need_destroy)) {
4640 : /* If qpair is closed before poller destroy, nvmf_rdma_destroy_drained_qpair may not
4641 : * be called because we cannot poll anything from cq. So we call that here to force
4642 : * destroy the qpair after to_close turning true.
4643 : */
4644 0 : RB_FOREACH_SAFE(rqpair, qpairs_tree, &rpoller->qpairs, tmp_rqpair) {
4645 0 : nvmf_rdma_destroy_drained_qpair(rqpair);
4646 : }
4647 0 : return 0;
4648 : }
4649 :
4650 : /* Poll for completing operations. */
4651 0 : reaped = ibv_poll_cq(rpoller->cq, 32, wc);
4652 0 : if (spdk_unlikely(reaped < 0)) {
4653 0 : SPDK_ERRLOG("Error polling CQ! (%d): %s\n",
4654 : errno, spdk_strerror(errno));
4655 0 : return -1;
4656 0 : } else if (reaped == 0) {
4657 0 : rpoller->stat.idle_polls++;
4658 : }
4659 :
4660 0 : rpoller->stat.polls++;
4661 0 : rpoller->stat.completions += reaped;
4662 :
4663 0 : for (i = 0; i < reaped; i++) {
4664 :
4665 0 : rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id;
4666 :
4667 0 : switch (rdma_wr->type) {
4668 0 : case RDMA_WR_TYPE_SEND:
4669 0 : rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr);
4670 0 : rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
4671 :
4672 0 : if (spdk_likely(!wc[i].status)) {
4673 0 : count++;
4674 0 : assert(wc[i].opcode == IBV_WC_SEND);
4675 0 : assert(nvmf_rdma_req_is_completing(rdma_req));
4676 : }
4677 :
4678 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
4679 : /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */
4680 0 : assert(rqpair->current_send_depth >= (uint32_t)rdma_req->num_outstanding_data_wr + 1);
4681 0 : rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1;
4682 0 : rdma_req->num_outstanding_data_wr = 0;
4683 :
4684 0 : nvmf_rdma_request_process(rtransport, rdma_req);
4685 0 : break;
4686 0 : case RDMA_WR_TYPE_RECV:
4687 : /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */
4688 0 : rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr);
4689 0 : if (rpoller->srq != NULL) {
4690 0 : rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]);
4691 : /* It is possible that there are still some completions for destroyed QP
4692 : * associated with SRQ. We just ignore these late completions and re-post
4693 : * receive WRs back to SRQ.
4694 : */
4695 0 : if (spdk_unlikely(NULL == rdma_recv->qpair)) {
4696 0 : struct ibv_recv_wr *bad_wr;
4697 :
4698 0 : rdma_recv->wr.next = NULL;
4699 0 : spdk_rdma_provider_srq_queue_recv_wrs(rpoller->srq, &rdma_recv->wr);
4700 0 : rc = spdk_rdma_provider_srq_flush_recv_wrs(rpoller->srq, &bad_wr);
4701 0 : if (rc) {
4702 0 : SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc);
4703 : }
4704 0 : continue;
4705 : }
4706 : }
4707 0 : rqpair = rdma_recv->qpair;
4708 :
4709 0 : assert(rqpair != NULL);
4710 0 : if (spdk_likely(!wc[i].status)) {
4711 0 : assert(wc[i].opcode == IBV_WC_RECV);
4712 0 : if (rqpair->current_recv_depth >= rqpair->max_queue_depth) {
4713 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
4714 0 : break;
4715 : }
4716 : }
4717 :
4718 0 : rdma_recv->wr.next = NULL;
4719 0 : rqpair->current_recv_depth++;
4720 0 : rdma_recv->receive_tsc = poll_tsc;
4721 0 : rpoller->stat.requests++;
4722 0 : STAILQ_INSERT_HEAD(&rqpair->resources->incoming_queue, rdma_recv, link);
4723 0 : rqpair->qpair.queue_depth++;
4724 0 : break;
4725 0 : case RDMA_WR_TYPE_DATA:
4726 0 : rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data_wr);
4727 0 : rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
4728 :
4729 0 : assert(rdma_req->num_outstanding_data_wr > 0);
4730 :
4731 0 : rqpair->current_send_depth--;
4732 0 : rdma_req->num_outstanding_data_wr--;
4733 0 : if (spdk_likely(!wc[i].status)) {
4734 0 : assert(wc[i].opcode == IBV_WC_RDMA_READ);
4735 0 : rqpair->current_read_depth--;
4736 : /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */
4737 0 : if (rdma_req->num_outstanding_data_wr == 0) {
4738 0 : if (rdma_req->num_remaining_data_wr) {
4739 : /* Only part of RDMA_READ operations was submitted, process the rest */
4740 0 : nvmf_rdma_request_reset_transfer_in(rdma_req, rtransport);
4741 0 : rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING;
4742 0 : nvmf_rdma_request_process(rtransport, rdma_req);
4743 0 : break;
4744 : }
4745 0 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
4746 0 : nvmf_rdma_request_process(rtransport, rdma_req);
4747 : }
4748 : } else {
4749 : /* If the data transfer fails still force the queue into the error state,
4750 : * if we were performing an RDMA_READ, we need to force the request into a
4751 : * completed state since it wasn't linked to a send. However, in the RDMA_WRITE
4752 : * case, we should wait for the SEND to complete. */
4753 0 : if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) {
4754 0 : rqpair->current_read_depth--;
4755 0 : if (rdma_req->num_outstanding_data_wr == 0) {
4756 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
4757 : }
4758 : }
4759 : }
4760 0 : break;
4761 0 : default:
4762 0 : SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode);
4763 0 : continue;
4764 : }
4765 :
4766 : /* Handle error conditions */
4767 0 : if (spdk_unlikely(wc[i].status)) {
4768 0 : rqpair->ibv_in_error_state = true;
4769 0 : nvmf_rdma_log_wc_status(rqpair, &wc[i]);
4770 :
4771 0 : error = true;
4772 :
4773 0 : if (spdk_nvmf_qpair_is_active(&rqpair->qpair)) {
4774 : /* Disconnect the connection. */
4775 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
4776 : } else {
4777 0 : nvmf_rdma_destroy_drained_qpair(rqpair);
4778 : }
4779 0 : continue;
4780 : }
4781 :
4782 0 : nvmf_rdma_qpair_process_pending(rtransport, rqpair, false);
4783 :
4784 0 : if (spdk_unlikely(!spdk_nvmf_qpair_is_active(&rqpair->qpair))) {
4785 0 : nvmf_rdma_destroy_drained_qpair(rqpair);
4786 : }
4787 : }
4788 :
4789 0 : if (spdk_unlikely(error == true)) {
4790 0 : return -1;
4791 : }
4792 :
4793 0 : if (reaped == 0) {
4794 : /* In some cases we may not receive any CQE but we still may have pending IO requests waiting for
4795 : * a resource (e.g. a WR from the data_wr_pool).
4796 : * We need to start processing of such requests if no CQE reaped */
4797 0 : nvmf_rdma_poller_process_pending_buf_queue(rtransport, rpoller);
4798 : }
4799 :
4800 : /* submit outstanding work requests. */
4801 0 : _poller_submit_recvs(rtransport, rpoller);
4802 0 : _poller_submit_sends(rtransport, rpoller);
4803 :
4804 0 : return count;
4805 : }
4806 :
4807 : static void
4808 0 : _nvmf_rdma_remove_destroyed_device(void *c)
4809 : {
4810 0 : struct spdk_nvmf_rdma_transport *rtransport = c;
4811 : struct spdk_nvmf_rdma_device *device, *device_tmp;
4812 : int rc;
4813 :
4814 0 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) {
4815 0 : if (device->ready_to_destroy) {
4816 0 : destroy_ib_device(rtransport, device);
4817 : }
4818 : }
4819 :
4820 0 : free_poll_fds(rtransport);
4821 0 : rc = generate_poll_fds(rtransport);
4822 : /* cannot handle fd allocation error here */
4823 0 : if (rc != 0) {
4824 0 : SPDK_ERRLOG("Failed to generate poll fds after remove ib device.\n");
4825 : }
4826 0 : }
4827 :
4828 : static void
4829 0 : _nvmf_rdma_remove_poller_in_group_cb(void *c)
4830 : {
4831 0 : struct poller_manage_ctx *ctx = c;
4832 0 : struct spdk_nvmf_rdma_transport *rtransport = ctx->rtransport;
4833 0 : struct spdk_nvmf_rdma_device *device = ctx->device;
4834 0 : struct spdk_thread *thread = ctx->thread;
4835 :
4836 0 : if (nvmf_rdma_all_pollers_management_done(c)) {
4837 : /* destroy device when last poller is destroyed */
4838 0 : device->ready_to_destroy = true;
4839 0 : spdk_thread_send_msg(thread, _nvmf_rdma_remove_destroyed_device, rtransport);
4840 : }
4841 0 : }
4842 :
4843 : static void
4844 0 : _nvmf_rdma_remove_poller_in_group(void *c)
4845 : {
4846 0 : struct poller_manage_ctx *ctx = c;
4847 :
4848 0 : ctx->rpoller->need_destroy = true;
4849 0 : ctx->rpoller->destroy_cb_ctx = ctx;
4850 0 : ctx->rpoller->destroy_cb = _nvmf_rdma_remove_poller_in_group_cb;
4851 :
4852 : /* qp will be disconnected after receiving a RDMA_CM_EVENT_DEVICE_REMOVAL event. */
4853 0 : if (RB_EMPTY(&ctx->rpoller->qpairs)) {
4854 0 : nvmf_rdma_poller_destroy(ctx->rpoller);
4855 : }
4856 0 : }
4857 :
4858 : static int
4859 0 : nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group)
4860 : {
4861 : struct spdk_nvmf_rdma_transport *rtransport;
4862 : struct spdk_nvmf_rdma_poll_group *rgroup;
4863 : struct spdk_nvmf_rdma_poller *rpoller, *tmp;
4864 0 : int count = 0, rc, rc2 = 0;
4865 :
4866 0 : rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport);
4867 0 : rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
4868 :
4869 0 : TAILQ_FOREACH_SAFE(rpoller, &rgroup->pollers, link, tmp) {
4870 0 : rc = nvmf_rdma_poller_poll(rtransport, rpoller);
4871 0 : if (spdk_unlikely(rc < 0)) {
4872 0 : if (rc2 == 0) {
4873 0 : rc2 = rc;
4874 : }
4875 0 : continue;
4876 : }
4877 0 : count += rc;
4878 : }
4879 :
4880 0 : return rc2 ? rc2 : count;
4881 : }
4882 :
4883 : static int
4884 0 : nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id,
4885 : struct spdk_nvme_transport_id *trid,
4886 : bool peer)
4887 : {
4888 : struct sockaddr *saddr;
4889 : uint16_t port;
4890 :
4891 0 : spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA);
4892 :
4893 0 : if (peer) {
4894 0 : saddr = rdma_get_peer_addr(id);
4895 : } else {
4896 0 : saddr = rdma_get_local_addr(id);
4897 : }
4898 0 : switch (saddr->sa_family) {
4899 0 : case AF_INET: {
4900 0 : struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr;
4901 :
4902 0 : trid->adrfam = SPDK_NVMF_ADRFAM_IPV4;
4903 0 : inet_ntop(AF_INET, &saddr_in->sin_addr,
4904 0 : trid->traddr, sizeof(trid->traddr));
4905 0 : if (peer) {
4906 0 : port = ntohs(rdma_get_dst_port(id));
4907 : } else {
4908 0 : port = ntohs(rdma_get_src_port(id));
4909 : }
4910 0 : snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port);
4911 0 : break;
4912 : }
4913 0 : case AF_INET6: {
4914 0 : struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr;
4915 0 : trid->adrfam = SPDK_NVMF_ADRFAM_IPV6;
4916 0 : inet_ntop(AF_INET6, &saddr_in->sin6_addr,
4917 0 : trid->traddr, sizeof(trid->traddr));
4918 0 : if (peer) {
4919 0 : port = ntohs(rdma_get_dst_port(id));
4920 : } else {
4921 0 : port = ntohs(rdma_get_src_port(id));
4922 : }
4923 0 : snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port);
4924 0 : break;
4925 : }
4926 0 : default:
4927 0 : return -1;
4928 :
4929 : }
4930 :
4931 0 : return 0;
4932 : }
4933 :
4934 : static int
4935 0 : nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair,
4936 : struct spdk_nvme_transport_id *trid)
4937 : {
4938 : struct spdk_nvmf_rdma_qpair *rqpair;
4939 :
4940 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4941 :
4942 0 : return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true);
4943 : }
4944 :
4945 : static int
4946 0 : nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair,
4947 : struct spdk_nvme_transport_id *trid)
4948 : {
4949 : struct spdk_nvmf_rdma_qpair *rqpair;
4950 :
4951 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4952 :
4953 0 : return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false);
4954 : }
4955 :
4956 : static int
4957 0 : nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair,
4958 : struct spdk_nvme_transport_id *trid)
4959 : {
4960 : struct spdk_nvmf_rdma_qpair *rqpair;
4961 :
4962 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4963 :
4964 0 : return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false);
4965 : }
4966 :
4967 : void
4968 0 : spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks)
4969 : {
4970 0 : g_nvmf_hooks = *hooks;
4971 0 : }
4972 :
4973 : static void
4974 0 : nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req,
4975 : struct spdk_nvmf_rdma_request *rdma_req_to_abort,
4976 : struct spdk_nvmf_rdma_qpair *rqpair)
4977 : {
4978 0 : rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC;
4979 0 : rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST;
4980 :
4981 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req_to_abort, state_link);
4982 0 : rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
4983 :
4984 0 : req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */
4985 0 : }
4986 :
4987 : static int
4988 0 : _nvmf_rdma_qpair_abort_request(void *ctx)
4989 : {
4990 0 : struct spdk_nvmf_request *req = ctx;
4991 0 : struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF(
4992 : req->req_to_abort, struct spdk_nvmf_rdma_request, req);
4993 0 : struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair,
4994 : struct spdk_nvmf_rdma_qpair, qpair);
4995 : int rc;
4996 :
4997 0 : spdk_poller_unregister(&req->poller);
4998 :
4999 0 : switch (rdma_req_to_abort->state) {
5000 0 : case RDMA_REQUEST_STATE_EXECUTING:
5001 0 : rc = nvmf_ctrlr_abort_request(req);
5002 0 : if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) {
5003 0 : return SPDK_POLLER_BUSY;
5004 : }
5005 0 : break;
5006 :
5007 0 : case RDMA_REQUEST_STATE_NEED_BUFFER:
5008 0 : STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue,
5009 : &rdma_req_to_abort->req, spdk_nvmf_request, buf_link);
5010 :
5011 0 : nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
5012 0 : break;
5013 :
5014 0 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING:
5015 0 : STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort,
5016 : spdk_nvmf_rdma_request, state_link);
5017 :
5018 0 : nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
5019 0 : break;
5020 :
5021 0 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING:
5022 0 : STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort,
5023 : spdk_nvmf_rdma_request, state_link);
5024 :
5025 0 : nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
5026 0 : break;
5027 :
5028 0 : case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING:
5029 : /* Remove req from the list here to re-use common function */
5030 0 : STAILQ_REMOVE(&rqpair->pending_rdma_send_queue, rdma_req_to_abort,
5031 : spdk_nvmf_rdma_request, state_link);
5032 :
5033 0 : nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
5034 0 : break;
5035 :
5036 0 : case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
5037 0 : if (spdk_get_ticks() < req->timeout_tsc) {
5038 0 : req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0);
5039 0 : return SPDK_POLLER_BUSY;
5040 : }
5041 0 : break;
5042 :
5043 0 : default:
5044 0 : break;
5045 : }
5046 :
5047 0 : spdk_nvmf_request_complete(req);
5048 0 : return SPDK_POLLER_BUSY;
5049 : }
5050 :
5051 : static void
5052 0 : nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair,
5053 : struct spdk_nvmf_request *req)
5054 : {
5055 : struct spdk_nvmf_rdma_qpair *rqpair;
5056 : struct spdk_nvmf_rdma_transport *rtransport;
5057 : struct spdk_nvmf_transport *transport;
5058 : uint16_t cid;
5059 : uint32_t i, max_req_count;
5060 0 : struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL, *rdma_req;
5061 :
5062 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
5063 0 : rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
5064 0 : transport = &rtransport->transport;
5065 :
5066 0 : cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid;
5067 0 : max_req_count = rqpair->srq == NULL ? rqpair->max_queue_depth : rqpair->poller->max_srq_depth;
5068 :
5069 0 : for (i = 0; i < max_req_count; i++) {
5070 0 : rdma_req = &rqpair->resources->reqs[i];
5071 : /* When SRQ == NULL, rqpair has its own requests and req.qpair pointer always points to the qpair
5072 : * When SRQ != NULL all rqpairs share common requests and qpair pointer is assigned when we start to
5073 : * process a request. So in both cases all requests which are not in FREE state have valid qpair ptr */
5074 0 : if (rdma_req->state != RDMA_REQUEST_STATE_FREE && rdma_req->req.cmd->nvme_cmd.cid == cid &&
5075 0 : rdma_req->req.qpair == qpair) {
5076 0 : rdma_req_to_abort = rdma_req;
5077 0 : break;
5078 : }
5079 : }
5080 :
5081 0 : if (rdma_req_to_abort == NULL) {
5082 0 : spdk_nvmf_request_complete(req);
5083 0 : return;
5084 : }
5085 :
5086 0 : req->req_to_abort = &rdma_req_to_abort->req;
5087 0 : req->timeout_tsc = spdk_get_ticks() +
5088 0 : transport->opts.abort_timeout_sec * spdk_get_ticks_hz();
5089 0 : req->poller = NULL;
5090 :
5091 0 : _nvmf_rdma_qpair_abort_request(req);
5092 : }
5093 :
5094 : static void
5095 0 : nvmf_rdma_poll_group_dump_stat(struct spdk_nvmf_transport_poll_group *group,
5096 : struct spdk_json_write_ctx *w)
5097 : {
5098 : struct spdk_nvmf_rdma_poll_group *rgroup;
5099 : struct spdk_nvmf_rdma_poller *rpoller;
5100 :
5101 0 : assert(w != NULL);
5102 :
5103 0 : rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
5104 :
5105 0 : spdk_json_write_named_uint64(w, "pending_data_buffer", rgroup->stat.pending_data_buffer);
5106 :
5107 0 : spdk_json_write_named_array_begin(w, "devices");
5108 :
5109 0 : TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
5110 0 : spdk_json_write_object_begin(w);
5111 0 : spdk_json_write_named_string(w, "name",
5112 0 : ibv_get_device_name(rpoller->device->context->device));
5113 0 : spdk_json_write_named_uint64(w, "polls",
5114 : rpoller->stat.polls);
5115 0 : spdk_json_write_named_uint64(w, "idle_polls",
5116 : rpoller->stat.idle_polls);
5117 0 : spdk_json_write_named_uint64(w, "completions",
5118 : rpoller->stat.completions);
5119 0 : spdk_json_write_named_uint64(w, "requests",
5120 : rpoller->stat.requests);
5121 0 : spdk_json_write_named_uint64(w, "request_latency",
5122 : rpoller->stat.request_latency);
5123 0 : spdk_json_write_named_uint64(w, "pending_free_request",
5124 : rpoller->stat.pending_free_request);
5125 0 : spdk_json_write_named_uint64(w, "pending_rdma_read",
5126 : rpoller->stat.pending_rdma_read);
5127 0 : spdk_json_write_named_uint64(w, "pending_rdma_write",
5128 : rpoller->stat.pending_rdma_write);
5129 0 : spdk_json_write_named_uint64(w, "pending_rdma_send",
5130 : rpoller->stat.pending_rdma_send);
5131 0 : spdk_json_write_named_uint64(w, "total_send_wrs",
5132 : rpoller->stat.qp_stats.send.num_submitted_wrs);
5133 0 : spdk_json_write_named_uint64(w, "send_doorbell_updates",
5134 : rpoller->stat.qp_stats.send.doorbell_updates);
5135 0 : spdk_json_write_named_uint64(w, "total_recv_wrs",
5136 : rpoller->stat.qp_stats.recv.num_submitted_wrs);
5137 0 : spdk_json_write_named_uint64(w, "recv_doorbell_updates",
5138 : rpoller->stat.qp_stats.recv.doorbell_updates);
5139 0 : spdk_json_write_object_end(w);
5140 : }
5141 :
5142 0 : spdk_json_write_array_end(w);
5143 0 : }
5144 :
5145 : const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = {
5146 : .name = "RDMA",
5147 : .type = SPDK_NVME_TRANSPORT_RDMA,
5148 : .opts_init = nvmf_rdma_opts_init,
5149 : .create = nvmf_rdma_create,
5150 : .dump_opts = nvmf_rdma_dump_opts,
5151 : .destroy = nvmf_rdma_destroy,
5152 :
5153 : .listen = nvmf_rdma_listen,
5154 : .stop_listen = nvmf_rdma_stop_listen,
5155 : .cdata_init = nvmf_rdma_cdata_init,
5156 :
5157 : .listener_discover = nvmf_rdma_discover,
5158 :
5159 : .poll_group_create = nvmf_rdma_poll_group_create,
5160 : .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group,
5161 : .poll_group_destroy = nvmf_rdma_poll_group_destroy,
5162 : .poll_group_add = nvmf_rdma_poll_group_add,
5163 : .poll_group_remove = nvmf_rdma_poll_group_remove,
5164 : .poll_group_poll = nvmf_rdma_poll_group_poll,
5165 :
5166 : .req_free = nvmf_rdma_request_free,
5167 : .req_complete = nvmf_rdma_request_complete,
5168 :
5169 : .qpair_fini = nvmf_rdma_close_qpair,
5170 : .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid,
5171 : .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid,
5172 : .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid,
5173 : .qpair_abort_request = nvmf_rdma_qpair_abort_request,
5174 :
5175 : .poll_group_dump_stat = nvmf_rdma_poll_group_dump_stat,
5176 : };
5177 :
5178 2 : SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma);
5179 2 : SPDK_LOG_REGISTER_COMPONENT(rdma)
|