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 : /*
8 : * NVMe over RDMA transport
9 : */
10 :
11 : #include "spdk/stdinc.h"
12 :
13 : #include "spdk/assert.h"
14 : #include "spdk/dma.h"
15 : #include "spdk/log.h"
16 : #include "spdk/trace.h"
17 : #include "spdk/queue.h"
18 : #include "spdk/nvme.h"
19 : #include "spdk/nvmf_spec.h"
20 : #include "spdk/string.h"
21 : #include "spdk/endian.h"
22 : #include "spdk/likely.h"
23 : #include "spdk/config.h"
24 :
25 : #include "nvme_internal.h"
26 : #include "spdk_internal/rdma_provider.h"
27 : #include "spdk_internal/rdma_utils.h"
28 :
29 : #define NVME_RDMA_TIME_OUT_IN_MS 2000
30 : #define NVME_RDMA_RW_BUFFER_SIZE 131072
31 :
32 : /*
33 : * NVME RDMA qpair Resource Defaults
34 : */
35 : #define NVME_RDMA_DEFAULT_TX_SGE 2
36 : #define NVME_RDMA_DEFAULT_RX_SGE 1
37 :
38 : /* Max number of NVMe-oF SGL descriptors supported by the host */
39 : #define NVME_RDMA_MAX_SGL_DESCRIPTORS 16
40 :
41 : /* number of STAILQ entries for holding pending RDMA CM events. */
42 : #define NVME_RDMA_NUM_CM_EVENTS 256
43 :
44 : /* The default size for a shared rdma completion queue. */
45 : #define DEFAULT_NVME_RDMA_CQ_SIZE 4096
46 :
47 : /*
48 : * In the special case of a stale connection we don't expose a mechanism
49 : * for the user to retry the connection so we need to handle it internally.
50 : */
51 : #define NVME_RDMA_STALE_CONN_RETRY_MAX 5
52 : #define NVME_RDMA_STALE_CONN_RETRY_DELAY_US 10000
53 :
54 : /*
55 : * Maximum value of transport_retry_count used by RDMA controller
56 : */
57 : #define NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT 7
58 :
59 : /*
60 : * Maximum value of transport_ack_timeout used by RDMA controller
61 : */
62 : #define NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT 31
63 :
64 : /*
65 : * Number of microseconds to wait until the lingering qpair becomes quiet.
66 : */
67 : #define NVME_RDMA_DISCONNECTED_QPAIR_TIMEOUT_US 1000000ull
68 :
69 : /*
70 : * The max length of keyed SGL data block (3 bytes)
71 : */
72 : #define NVME_RDMA_MAX_KEYED_SGL_LENGTH ((1u << 24u) - 1)
73 :
74 : #define WC_PER_QPAIR(queue_depth) (queue_depth * 2)
75 :
76 : #define NVME_RDMA_POLL_GROUP_CHECK_QPN(_rqpair, qpn) \
77 : ((_rqpair)->rdma_qp && (_rqpair)->rdma_qp->qp->qp_num == (qpn)) \
78 :
79 : enum nvme_rdma_wr_type {
80 : RDMA_WR_TYPE_RECV,
81 : RDMA_WR_TYPE_SEND,
82 : };
83 :
84 : struct nvme_rdma_wr {
85 : /* Using this instead of the enum allows this struct to only occupy one byte. */
86 : uint8_t type;
87 : };
88 :
89 : struct spdk_nvmf_cmd {
90 : struct spdk_nvme_cmd cmd;
91 : struct spdk_nvme_sgl_descriptor sgl[NVME_RDMA_MAX_SGL_DESCRIPTORS];
92 : };
93 :
94 : struct spdk_nvme_rdma_hooks g_nvme_hooks = {};
95 :
96 : /* STAILQ wrapper for cm events. */
97 : struct nvme_rdma_cm_event_entry {
98 : struct rdma_cm_event *evt;
99 : STAILQ_ENTRY(nvme_rdma_cm_event_entry) link;
100 : };
101 :
102 : /* NVMe RDMA transport extensions for spdk_nvme_ctrlr */
103 : struct nvme_rdma_ctrlr {
104 : struct spdk_nvme_ctrlr ctrlr;
105 :
106 : uint16_t max_sge;
107 :
108 : struct rdma_event_channel *cm_channel;
109 :
110 : STAILQ_HEAD(, nvme_rdma_cm_event_entry) pending_cm_events;
111 :
112 : STAILQ_HEAD(, nvme_rdma_cm_event_entry) free_cm_events;
113 :
114 : struct nvme_rdma_cm_event_entry *cm_events;
115 : };
116 :
117 : struct nvme_rdma_poller_stats {
118 : uint64_t polls;
119 : uint64_t idle_polls;
120 : uint64_t queued_requests;
121 : uint64_t completions;
122 : struct spdk_rdma_provider_qp_stats rdma_stats;
123 : };
124 :
125 : struct nvme_rdma_poll_group;
126 : struct nvme_rdma_rsps;
127 :
128 : struct nvme_rdma_poller {
129 : struct ibv_context *device;
130 : struct ibv_cq *cq;
131 : struct spdk_rdma_provider_srq *srq;
132 : struct nvme_rdma_rsps *rsps;
133 : struct ibv_pd *pd;
134 : struct spdk_rdma_utils_mem_map *mr_map;
135 : uint32_t refcnt;
136 : int required_num_wc;
137 : int current_num_wc;
138 : struct nvme_rdma_poller_stats stats;
139 : struct nvme_rdma_poll_group *group;
140 : STAILQ_ENTRY(nvme_rdma_poller) link;
141 : };
142 :
143 : struct nvme_rdma_qpair;
144 :
145 : struct nvme_rdma_poll_group {
146 : struct spdk_nvme_transport_poll_group group;
147 : STAILQ_HEAD(, nvme_rdma_poller) pollers;
148 : uint32_t num_pollers;
149 : TAILQ_HEAD(, nvme_rdma_qpair) connecting_qpairs;
150 : TAILQ_HEAD(, nvme_rdma_qpair) active_qpairs;
151 : };
152 :
153 : enum nvme_rdma_qpair_state {
154 : NVME_RDMA_QPAIR_STATE_INVALID = 0,
155 : NVME_RDMA_QPAIR_STATE_STALE_CONN,
156 : NVME_RDMA_QPAIR_STATE_INITIALIZING,
157 : NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_SEND,
158 : NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_POLL,
159 : NVME_RDMA_QPAIR_STATE_AUTHENTICATING,
160 : NVME_RDMA_QPAIR_STATE_RUNNING,
161 : NVME_RDMA_QPAIR_STATE_EXITING,
162 : NVME_RDMA_QPAIR_STATE_LINGERING,
163 : NVME_RDMA_QPAIR_STATE_EXITED,
164 : };
165 :
166 : typedef int (*nvme_rdma_cm_event_cb)(struct nvme_rdma_qpair *rqpair, int ret);
167 :
168 : struct nvme_rdma_rsp_opts {
169 : uint16_t num_entries;
170 : struct nvme_rdma_qpair *rqpair;
171 : struct spdk_rdma_provider_srq *srq;
172 : struct spdk_rdma_utils_mem_map *mr_map;
173 : };
174 :
175 : struct nvme_rdma_rsps {
176 : /* Parallel arrays of response buffers + response SGLs of size num_entries */
177 : struct ibv_sge *rsp_sgls;
178 : struct spdk_nvme_rdma_rsp *rsps;
179 :
180 : struct ibv_recv_wr *rsp_recv_wrs;
181 :
182 : /* Count of outstanding recv objects */
183 : uint16_t current_num_recvs;
184 :
185 : uint16_t num_entries;
186 : };
187 :
188 : /* NVMe RDMA qpair extensions for spdk_nvme_qpair */
189 : struct nvme_rdma_qpair {
190 : struct spdk_nvme_qpair qpair;
191 :
192 : struct spdk_rdma_provider_qp *rdma_qp;
193 : struct rdma_cm_id *cm_id;
194 : struct ibv_cq *cq;
195 : struct spdk_rdma_provider_srq *srq;
196 :
197 : struct spdk_nvme_rdma_req *rdma_reqs;
198 :
199 : uint32_t max_send_sge;
200 :
201 : uint16_t num_entries;
202 :
203 : bool delay_cmd_submit;
204 :
205 : uint32_t num_completions;
206 : uint32_t num_outstanding_reqs;
207 :
208 : struct nvme_rdma_rsps *rsps;
209 :
210 : /*
211 : * Array of num_entries NVMe commands registered as RDMA message buffers.
212 : * Indexed by rdma_req->id.
213 : */
214 : struct spdk_nvmf_cmd *cmds;
215 :
216 : struct spdk_rdma_utils_mem_map *mr_map;
217 :
218 : TAILQ_HEAD(, spdk_nvme_rdma_req) free_reqs;
219 : TAILQ_HEAD(, spdk_nvme_rdma_req) outstanding_reqs;
220 :
221 : /* Count of outstanding send objects */
222 : uint16_t current_num_sends;
223 :
224 : TAILQ_ENTRY(nvme_rdma_qpair) link_active;
225 :
226 : /* Placed at the end of the struct since it is not used frequently */
227 : struct rdma_cm_event *evt;
228 : struct nvme_rdma_poller *poller;
229 :
230 : uint64_t evt_timeout_ticks;
231 : nvme_rdma_cm_event_cb evt_cb;
232 : enum rdma_cm_event_type expected_evt_type;
233 :
234 : enum nvme_rdma_qpair_state state;
235 :
236 : bool in_connect_poll;
237 :
238 : uint8_t stale_conn_retry_count;
239 : bool need_destroy;
240 : TAILQ_ENTRY(nvme_rdma_qpair) link_connecting;
241 : };
242 :
243 : enum NVME_RDMA_COMPLETION_FLAGS {
244 : NVME_RDMA_SEND_COMPLETED = 1u << 0,
245 : NVME_RDMA_RECV_COMPLETED = 1u << 1,
246 : };
247 :
248 : struct spdk_nvme_rdma_req {
249 : uint16_t id;
250 : uint16_t completion_flags: 2;
251 : uint16_t in_progress_accel: 1;
252 : uint16_t reserved: 13;
253 : /* if completion of RDMA_RECV received before RDMA_SEND, we will complete nvme request
254 : * during processing of RDMA_SEND. To complete the request we must know the response
255 : * received in RDMA_RECV, so store it in this field */
256 : struct spdk_nvme_rdma_rsp *rdma_rsp;
257 :
258 : struct nvme_rdma_wr rdma_wr;
259 :
260 : struct ibv_send_wr send_wr;
261 :
262 : struct nvme_request *req;
263 :
264 : struct ibv_sge send_sgl[NVME_RDMA_DEFAULT_TX_SGE];
265 :
266 : TAILQ_ENTRY(spdk_nvme_rdma_req) link;
267 :
268 : /* Fields below are not used in regular IO path, keep them last */
269 : spdk_memory_domain_data_cpl_cb transfer_cpl_cb;
270 : void *transfer_cpl_cb_arg;
271 : /* Accel sequence API works with iovec pointer, we need to store result of next_sge callback */
272 : struct iovec iovs[NVME_RDMA_MAX_SGL_DESCRIPTORS];
273 : };
274 :
275 : struct spdk_nvme_rdma_rsp {
276 : struct spdk_nvme_cpl cpl;
277 : struct nvme_rdma_qpair *rqpair;
278 : struct ibv_recv_wr *recv_wr;
279 : struct nvme_rdma_wr rdma_wr;
280 : };
281 :
282 : struct nvme_rdma_memory_translation_ctx {
283 : void *addr;
284 : size_t length;
285 : uint32_t lkey;
286 : uint32_t rkey;
287 : };
288 :
289 : static const char *rdma_cm_event_str[] = {
290 : "RDMA_CM_EVENT_ADDR_RESOLVED",
291 : "RDMA_CM_EVENT_ADDR_ERROR",
292 : "RDMA_CM_EVENT_ROUTE_RESOLVED",
293 : "RDMA_CM_EVENT_ROUTE_ERROR",
294 : "RDMA_CM_EVENT_CONNECT_REQUEST",
295 : "RDMA_CM_EVENT_CONNECT_RESPONSE",
296 : "RDMA_CM_EVENT_CONNECT_ERROR",
297 : "RDMA_CM_EVENT_UNREACHABLE",
298 : "RDMA_CM_EVENT_REJECTED",
299 : "RDMA_CM_EVENT_ESTABLISHED",
300 : "RDMA_CM_EVENT_DISCONNECTED",
301 : "RDMA_CM_EVENT_DEVICE_REMOVAL",
302 : "RDMA_CM_EVENT_MULTICAST_JOIN",
303 : "RDMA_CM_EVENT_MULTICAST_ERROR",
304 : "RDMA_CM_EVENT_ADDR_CHANGE",
305 : "RDMA_CM_EVENT_TIMEWAIT_EXIT"
306 : };
307 :
308 : static struct nvme_rdma_poller *nvme_rdma_poll_group_get_poller(struct nvme_rdma_poll_group *group,
309 : struct ibv_context *device);
310 : static void nvme_rdma_poll_group_put_poller(struct nvme_rdma_poll_group *group,
311 : struct nvme_rdma_poller *poller);
312 :
313 : static int nvme_rdma_ctrlr_delete_io_qpair(struct spdk_nvme_ctrlr *ctrlr,
314 : struct spdk_nvme_qpair *qpair);
315 :
316 : static inline int nvme_rdma_memory_domain_transfer_data(struct spdk_memory_domain *dst_domain,
317 : void *dst_domain_ctx,
318 : struct iovec *dst_iov, uint32_t dst_iovcnt,
319 : struct spdk_memory_domain *src_domain, void *src_domain_ctx,
320 : struct iovec *src_iov, uint32_t src_iovcnt,
321 : struct spdk_memory_domain_translation_result *translation,
322 : spdk_memory_domain_data_cpl_cb cpl_cb, void *cpl_cb_arg);
323 :
324 : static inline int _nvme_rdma_qpair_submit_request(struct nvme_rdma_qpair *rqpair,
325 : struct spdk_nvme_rdma_req *rdma_req);
326 :
327 : static inline struct nvme_rdma_qpair *
328 18 : nvme_rdma_qpair(struct spdk_nvme_qpair *qpair)
329 : {
330 18 : assert(qpair->trtype == SPDK_NVME_TRANSPORT_RDMA);
331 18 : return SPDK_CONTAINEROF(qpair, struct nvme_rdma_qpair, qpair);
332 : }
333 :
334 : static inline struct nvme_rdma_poll_group *
335 8 : nvme_rdma_poll_group(struct spdk_nvme_transport_poll_group *group)
336 : {
337 8 : return (SPDK_CONTAINEROF(group, struct nvme_rdma_poll_group, group));
338 : }
339 :
340 : static inline struct nvme_rdma_ctrlr *
341 8 : nvme_rdma_ctrlr(struct spdk_nvme_ctrlr *ctrlr)
342 : {
343 8 : assert(ctrlr->trid.trtype == SPDK_NVME_TRANSPORT_RDMA);
344 8 : return SPDK_CONTAINEROF(ctrlr, struct nvme_rdma_ctrlr, ctrlr);
345 : }
346 :
347 : static inline struct spdk_nvme_rdma_req *
348 3 : nvme_rdma_req_get(struct nvme_rdma_qpair *rqpair)
349 : {
350 : struct spdk_nvme_rdma_req *rdma_req;
351 :
352 3 : rdma_req = TAILQ_FIRST(&rqpair->free_reqs);
353 3 : if (spdk_likely(rdma_req)) {
354 2 : TAILQ_REMOVE(&rqpair->free_reqs, rdma_req, link);
355 2 : }
356 :
357 3 : return rdma_req;
358 : }
359 :
360 : static inline void
361 1 : nvme_rdma_req_put(struct nvme_rdma_qpair *rqpair, struct spdk_nvme_rdma_req *rdma_req)
362 : {
363 1 : rdma_req->completion_flags = 0;
364 1 : rdma_req->req = NULL;
365 1 : rdma_req->rdma_rsp = NULL;
366 1 : assert(rdma_req->transfer_cpl_cb == NULL);
367 1 : TAILQ_INSERT_HEAD(&rqpair->free_reqs, rdma_req, link);
368 1 : }
369 :
370 : static inline void
371 0 : nvme_rdma_finish_data_transfer(struct spdk_nvme_rdma_req *rdma_req, int rc)
372 : {
373 0 : spdk_memory_domain_data_cpl_cb cb = rdma_req->transfer_cpl_cb;
374 :
375 0 : SPDK_DEBUGLOG(nvme, "req %p, finish data transfer, rc %d\n", rdma_req, rc);
376 0 : rdma_req->transfer_cpl_cb = NULL;
377 0 : assert(cb);
378 0 : cb(rdma_req->transfer_cpl_cb_arg, rc);
379 0 : }
380 :
381 : static void
382 0 : nvme_rdma_req_complete(struct spdk_nvme_rdma_req *rdma_req,
383 : struct spdk_nvme_cpl *rsp,
384 : bool print_on_error)
385 : {
386 0 : struct nvme_request *req = rdma_req->req;
387 : struct nvme_rdma_qpair *rqpair;
388 : struct spdk_nvme_qpair *qpair;
389 : bool error, print_error;
390 :
391 0 : assert(req != NULL);
392 :
393 0 : qpair = req->qpair;
394 0 : rqpair = nvme_rdma_qpair(qpair);
395 :
396 0 : error = spdk_nvme_cpl_is_error(rsp);
397 0 : print_error = error && print_on_error && !qpair->ctrlr->opts.disable_error_logging;
398 :
399 0 : if (print_error) {
400 0 : spdk_nvme_qpair_print_command(qpair, &req->cmd);
401 0 : }
402 :
403 0 : if (print_error || SPDK_DEBUGLOG_FLAG_ENABLED("nvme")) {
404 0 : spdk_nvme_qpair_print_completion(qpair, rsp);
405 0 : }
406 :
407 0 : assert(rqpair->num_outstanding_reqs > 0);
408 0 : rqpair->num_outstanding_reqs--;
409 :
410 0 : TAILQ_REMOVE(&rqpair->outstanding_reqs, rdma_req, link);
411 :
412 0 : nvme_complete_request(req->cb_fn, req->cb_arg, qpair, req, rsp);
413 0 : nvme_rdma_req_put(rqpair, rdma_req);
414 0 : }
415 :
416 : static const char *
417 4 : nvme_rdma_cm_event_str_get(uint32_t event)
418 : {
419 4 : if (event < SPDK_COUNTOF(rdma_cm_event_str)) {
420 4 : return rdma_cm_event_str[event];
421 : } else {
422 0 : return "Undefined";
423 : }
424 4 : }
425 :
426 :
427 : static int
428 12 : nvme_rdma_qpair_process_cm_event(struct nvme_rdma_qpair *rqpair)
429 : {
430 12 : struct rdma_cm_event *event = rqpair->evt;
431 : struct spdk_nvmf_rdma_accept_private_data *accept_data;
432 12 : int rc = 0;
433 :
434 12 : if (event) {
435 12 : switch (event->event) {
436 : case RDMA_CM_EVENT_ADDR_RESOLVED:
437 : case RDMA_CM_EVENT_ADDR_ERROR:
438 : case RDMA_CM_EVENT_ROUTE_RESOLVED:
439 : case RDMA_CM_EVENT_ROUTE_ERROR:
440 1 : break;
441 : case RDMA_CM_EVENT_CONNECT_REQUEST:
442 1 : break;
443 : case RDMA_CM_EVENT_CONNECT_ERROR:
444 1 : break;
445 : case RDMA_CM_EVENT_UNREACHABLE:
446 : case RDMA_CM_EVENT_REJECTED:
447 1 : break;
448 : case RDMA_CM_EVENT_CONNECT_RESPONSE:
449 2 : rc = spdk_rdma_provider_qp_complete_connect(rqpair->rdma_qp);
450 : /* fall through */
451 : case RDMA_CM_EVENT_ESTABLISHED:
452 2 : accept_data = (struct spdk_nvmf_rdma_accept_private_data *)event->param.conn.private_data;
453 2 : if (accept_data == NULL) {
454 1 : rc = -1;
455 1 : } else {
456 1 : SPDK_DEBUGLOG(nvme, "Requested queue depth %d. Target receive queue depth %d.\n",
457 : rqpair->num_entries + 1, accept_data->crqsize);
458 : }
459 2 : break;
460 : case RDMA_CM_EVENT_DISCONNECTED:
461 1 : rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_REMOTE;
462 1 : break;
463 : case RDMA_CM_EVENT_DEVICE_REMOVAL:
464 1 : rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_LOCAL;
465 1 : rqpair->need_destroy = true;
466 1 : break;
467 : case RDMA_CM_EVENT_MULTICAST_JOIN:
468 : case RDMA_CM_EVENT_MULTICAST_ERROR:
469 1 : break;
470 : case RDMA_CM_EVENT_ADDR_CHANGE:
471 1 : rqpair->qpair.transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_LOCAL;
472 1 : break;
473 : case RDMA_CM_EVENT_TIMEWAIT_EXIT:
474 1 : break;
475 : default:
476 1 : SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event);
477 1 : break;
478 : }
479 12 : rqpair->evt = NULL;
480 12 : rdma_ack_cm_event(event);
481 12 : }
482 :
483 12 : return rc;
484 : }
485 :
486 : /*
487 : * This function must be called under the nvme controller's lock
488 : * because it touches global controller variables. The lock is taken
489 : * by the generic transport code before invoking a few of the functions
490 : * in this file: nvme_rdma_ctrlr_connect_qpair, nvme_rdma_ctrlr_delete_io_qpair,
491 : * and conditionally nvme_rdma_qpair_process_completions when it is calling
492 : * completions on the admin qpair. When adding a new call to this function, please
493 : * verify that it is in a situation where it falls under the lock.
494 : */
495 : static int
496 0 : nvme_rdma_poll_events(struct nvme_rdma_ctrlr *rctrlr)
497 : {
498 : struct nvme_rdma_cm_event_entry *entry, *tmp;
499 : struct nvme_rdma_qpair *event_qpair;
500 : struct rdma_cm_event *event;
501 0 : struct rdma_event_channel *channel = rctrlr->cm_channel;
502 :
503 0 : STAILQ_FOREACH_SAFE(entry, &rctrlr->pending_cm_events, link, tmp) {
504 0 : event_qpair = entry->evt->id->context;
505 0 : if (event_qpair->evt == NULL) {
506 0 : event_qpair->evt = entry->evt;
507 0 : STAILQ_REMOVE(&rctrlr->pending_cm_events, entry, nvme_rdma_cm_event_entry, link);
508 0 : STAILQ_INSERT_HEAD(&rctrlr->free_cm_events, entry, link);
509 0 : }
510 0 : }
511 :
512 0 : while (rdma_get_cm_event(channel, &event) == 0) {
513 0 : event_qpair = event->id->context;
514 0 : if (event_qpair->evt == NULL) {
515 0 : event_qpair->evt = event;
516 0 : } else {
517 0 : assert(rctrlr == nvme_rdma_ctrlr(event_qpair->qpair.ctrlr));
518 0 : entry = STAILQ_FIRST(&rctrlr->free_cm_events);
519 0 : if (entry == NULL) {
520 0 : rdma_ack_cm_event(event);
521 0 : return -ENOMEM;
522 : }
523 0 : STAILQ_REMOVE(&rctrlr->free_cm_events, entry, nvme_rdma_cm_event_entry, link);
524 0 : entry->evt = event;
525 0 : STAILQ_INSERT_TAIL(&rctrlr->pending_cm_events, entry, link);
526 : }
527 : }
528 :
529 : /* rdma_get_cm_event() returns -1 on error. If an error occurs, errno
530 : * will be set to indicate the failure reason. So return negated errno here.
531 : */
532 0 : return -errno;
533 0 : }
534 :
535 : static int
536 4 : nvme_rdma_validate_cm_event(enum rdma_cm_event_type expected_evt_type,
537 : struct rdma_cm_event *reaped_evt)
538 : {
539 4 : int rc = -EBADMSG;
540 :
541 4 : if (expected_evt_type == reaped_evt->event) {
542 1 : return 0;
543 : }
544 :
545 3 : switch (expected_evt_type) {
546 : case RDMA_CM_EVENT_ESTABLISHED:
547 : /*
548 : * There is an enum ib_cm_rej_reason in the kernel headers that sets 10 as
549 : * IB_CM_REJ_STALE_CONN. I can't find the corresponding userspace but we get
550 : * the same values here.
551 : */
552 2 : if (reaped_evt->event == RDMA_CM_EVENT_REJECTED && reaped_evt->status == 10) {
553 1 : rc = -ESTALE;
554 2 : } else if (reaped_evt->event == RDMA_CM_EVENT_CONNECT_RESPONSE) {
555 : /*
556 : * If we are using a qpair which is not created using rdma cm API
557 : * then we will receive RDMA_CM_EVENT_CONNECT_RESPONSE instead of
558 : * RDMA_CM_EVENT_ESTABLISHED.
559 : */
560 1 : return 0;
561 : }
562 1 : break;
563 : default:
564 1 : break;
565 : }
566 :
567 2 : SPDK_ERRLOG("Expected %s but received %s (%d) from CM event channel (status = %d)\n",
568 : nvme_rdma_cm_event_str_get(expected_evt_type),
569 : nvme_rdma_cm_event_str_get(reaped_evt->event), reaped_evt->event,
570 : reaped_evt->status);
571 2 : return rc;
572 4 : }
573 :
574 : static int
575 0 : nvme_rdma_process_event_start(struct nvme_rdma_qpair *rqpair,
576 : enum rdma_cm_event_type evt,
577 : nvme_rdma_cm_event_cb evt_cb)
578 : {
579 : int rc;
580 :
581 0 : assert(evt_cb != NULL);
582 :
583 0 : if (rqpair->evt != NULL) {
584 0 : rc = nvme_rdma_qpair_process_cm_event(rqpair);
585 0 : if (rc) {
586 0 : return rc;
587 : }
588 0 : }
589 :
590 0 : rqpair->expected_evt_type = evt;
591 0 : rqpair->evt_cb = evt_cb;
592 0 : rqpair->evt_timeout_ticks = (g_spdk_nvme_transport_opts.rdma_cm_event_timeout_ms * 1000 *
593 0 : spdk_get_ticks_hz()) / SPDK_SEC_TO_USEC + spdk_get_ticks();
594 :
595 0 : return 0;
596 0 : }
597 :
598 : static int
599 0 : nvme_rdma_process_event_poll(struct nvme_rdma_qpair *rqpair)
600 : {
601 : struct nvme_rdma_ctrlr *rctrlr;
602 0 : int rc = 0, rc2;
603 :
604 0 : rctrlr = nvme_rdma_ctrlr(rqpair->qpair.ctrlr);
605 0 : assert(rctrlr != NULL);
606 :
607 0 : if (!rqpair->evt && spdk_get_ticks() < rqpair->evt_timeout_ticks) {
608 0 : rc = nvme_rdma_poll_events(rctrlr);
609 0 : if (rc == -EAGAIN || rc == -EWOULDBLOCK) {
610 0 : return rc;
611 : }
612 0 : }
613 :
614 0 : if (rqpair->evt == NULL) {
615 0 : rc = -EADDRNOTAVAIL;
616 0 : goto exit;
617 : }
618 :
619 0 : rc = nvme_rdma_validate_cm_event(rqpair->expected_evt_type, rqpair->evt);
620 :
621 0 : rc2 = nvme_rdma_qpair_process_cm_event(rqpair);
622 : /* bad message takes precedence over the other error codes from processing the event. */
623 0 : rc = rc == 0 ? rc2 : rc;
624 :
625 : exit:
626 0 : assert(rqpair->evt_cb != NULL);
627 0 : return rqpair->evt_cb(rqpair, rc);
628 0 : }
629 :
630 : static int
631 3 : nvme_rdma_resize_cq(struct nvme_rdma_qpair *rqpair, struct nvme_rdma_poller *poller)
632 : {
633 : int current_num_wc, required_num_wc;
634 : int max_cq_size;
635 :
636 3 : required_num_wc = poller->required_num_wc + WC_PER_QPAIR(rqpair->num_entries);
637 3 : current_num_wc = poller->current_num_wc;
638 3 : if (current_num_wc < required_num_wc) {
639 2 : current_num_wc = spdk_max(current_num_wc * 2, required_num_wc);
640 2 : }
641 :
642 3 : max_cq_size = g_spdk_nvme_transport_opts.rdma_max_cq_size;
643 3 : if (max_cq_size != 0 && current_num_wc > max_cq_size) {
644 0 : current_num_wc = max_cq_size;
645 0 : }
646 :
647 3 : if (poller->current_num_wc != current_num_wc) {
648 2 : SPDK_DEBUGLOG(nvme, "Resize RDMA CQ from %d to %d\n", poller->current_num_wc,
649 : current_num_wc);
650 2 : if (ibv_resize_cq(poller->cq, current_num_wc)) {
651 1 : SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno));
652 1 : return -1;
653 : }
654 :
655 1 : poller->current_num_wc = current_num_wc;
656 1 : }
657 :
658 2 : poller->required_num_wc = required_num_wc;
659 2 : return 0;
660 3 : }
661 :
662 : static int
663 5 : nvme_rdma_qpair_set_poller(struct spdk_nvme_qpair *qpair)
664 : {
665 5 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
666 5 : struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group);
667 : struct nvme_rdma_poller *poller;
668 :
669 5 : assert(rqpair->cq == NULL);
670 :
671 5 : poller = nvme_rdma_poll_group_get_poller(group, rqpair->cm_id->verbs);
672 5 : if (!poller) {
673 2 : SPDK_ERRLOG("Unable to find a cq for qpair %p on poll group %p\n", qpair, qpair->poll_group);
674 2 : return -EINVAL;
675 : }
676 :
677 3 : if (!poller->srq) {
678 3 : if (nvme_rdma_resize_cq(rqpair, poller)) {
679 1 : nvme_rdma_poll_group_put_poller(group, poller);
680 1 : return -EPROTO;
681 : }
682 2 : }
683 :
684 2 : rqpair->cq = poller->cq;
685 2 : rqpair->srq = poller->srq;
686 2 : if (rqpair->srq) {
687 0 : rqpair->rsps = poller->rsps;
688 0 : }
689 2 : rqpair->poller = poller;
690 2 : return 0;
691 5 : }
692 :
693 : static int
694 1 : nvme_rdma_qpair_init(struct nvme_rdma_qpair *rqpair)
695 : {
696 : int rc;
697 1 : struct spdk_rdma_provider_qp_init_attr attr = {};
698 : struct ibv_device_attr dev_attr;
699 : struct nvme_rdma_ctrlr *rctrlr;
700 : uint32_t num_cqe, max_num_cqe;
701 :
702 1 : rc = ibv_query_device(rqpair->cm_id->verbs, &dev_attr);
703 1 : if (rc != 0) {
704 0 : SPDK_ERRLOG("Failed to query RDMA device attributes.\n");
705 0 : return -1;
706 : }
707 :
708 1 : if (rqpair->qpair.poll_group) {
709 0 : assert(!rqpair->cq);
710 0 : rc = nvme_rdma_qpair_set_poller(&rqpair->qpair);
711 0 : if (rc) {
712 0 : SPDK_ERRLOG("Unable to activate the rdmaqpair.\n");
713 0 : return -1;
714 : }
715 0 : assert(rqpair->cq);
716 0 : } else {
717 1 : num_cqe = rqpair->num_entries * 2;
718 1 : max_num_cqe = g_spdk_nvme_transport_opts.rdma_max_cq_size;
719 1 : if (max_num_cqe != 0 && num_cqe > max_num_cqe) {
720 0 : num_cqe = max_num_cqe;
721 0 : }
722 1 : rqpair->cq = ibv_create_cq(rqpair->cm_id->verbs, num_cqe, rqpair, NULL, 0);
723 1 : if (!rqpair->cq) {
724 0 : SPDK_ERRLOG("Unable to create completion queue: errno %d: %s\n", errno, spdk_strerror(errno));
725 0 : return -1;
726 : }
727 : }
728 :
729 1 : rctrlr = nvme_rdma_ctrlr(rqpair->qpair.ctrlr);
730 1 : if (g_nvme_hooks.get_ibv_pd) {
731 0 : attr.pd = g_nvme_hooks.get_ibv_pd(&rctrlr->ctrlr.trid, rqpair->cm_id->verbs);
732 0 : } else {
733 1 : attr.pd = spdk_rdma_utils_get_pd(rqpair->cm_id->verbs);
734 : }
735 :
736 1 : attr.stats = rqpair->poller ? &rqpair->poller->stats.rdma_stats : NULL;
737 1 : attr.send_cq = rqpair->cq;
738 1 : attr.recv_cq = rqpair->cq;
739 1 : attr.cap.max_send_wr = rqpair->num_entries; /* SEND operations */
740 1 : if (rqpair->srq) {
741 0 : attr.srq = rqpair->srq->srq;
742 0 : } else {
743 1 : attr.cap.max_recv_wr = rqpair->num_entries; /* RECV operations */
744 : }
745 1 : attr.cap.max_send_sge = spdk_min(NVME_RDMA_DEFAULT_TX_SGE, dev_attr.max_sge);
746 1 : attr.cap.max_recv_sge = spdk_min(NVME_RDMA_DEFAULT_RX_SGE, dev_attr.max_sge);
747 1 : attr.domain_transfer = spdk_rdma_provider_accel_sequence_supported() ?
748 : nvme_rdma_memory_domain_transfer_data : NULL;
749 :
750 1 : rqpair->rdma_qp = spdk_rdma_provider_qp_create(rqpair->cm_id, &attr);
751 :
752 1 : if (!rqpair->rdma_qp) {
753 0 : return -1;
754 : }
755 :
756 : /* ibv_create_qp will change the values in attr.cap. Make sure we store the proper value. */
757 1 : rqpair->max_send_sge = spdk_min(NVME_RDMA_DEFAULT_TX_SGE, attr.cap.max_send_sge);
758 1 : rqpair->current_num_sends = 0;
759 :
760 1 : rqpair->cm_id->context = rqpair;
761 :
762 1 : return 0;
763 1 : }
764 :
765 : static void
766 0 : nvme_rdma_reset_failed_sends(struct nvme_rdma_qpair *rqpair,
767 : struct ibv_send_wr *bad_send_wr)
768 : {
769 0 : while (bad_send_wr != NULL) {
770 0 : assert(rqpair->current_num_sends > 0);
771 0 : rqpair->current_num_sends--;
772 0 : bad_send_wr = bad_send_wr->next;
773 : }
774 0 : }
775 :
776 : static void
777 0 : nvme_rdma_reset_failed_recvs(struct nvme_rdma_rsps *rsps,
778 : struct ibv_recv_wr *bad_recv_wr, int rc)
779 : {
780 0 : SPDK_ERRLOG("Failed to post WRs on receive queue, errno %d (%s), bad_wr %p\n",
781 : rc, spdk_strerror(rc), bad_recv_wr);
782 0 : while (bad_recv_wr != NULL) {
783 0 : assert(rsps->current_num_recvs > 0);
784 0 : rsps->current_num_recvs--;
785 0 : bad_recv_wr = bad_recv_wr->next;
786 : }
787 0 : }
788 :
789 : static inline int
790 1 : nvme_rdma_qpair_submit_sends(struct nvme_rdma_qpair *rqpair)
791 : {
792 1 : struct ibv_send_wr *bad_send_wr = NULL;
793 : int rc;
794 :
795 1 : rc = spdk_rdma_provider_qp_flush_send_wrs(rqpair->rdma_qp, &bad_send_wr);
796 :
797 1 : if (spdk_unlikely(rc)) {
798 0 : SPDK_ERRLOG("Failed to post WRs on send queue, errno %d (%s), bad_wr %p\n",
799 : rc, spdk_strerror(rc), bad_send_wr);
800 0 : nvme_rdma_reset_failed_sends(rqpair, bad_send_wr);
801 0 : }
802 :
803 1 : return rc;
804 : }
805 :
806 : static inline int
807 0 : nvme_rdma_qpair_submit_recvs(struct nvme_rdma_qpair *rqpair)
808 : {
809 : struct ibv_recv_wr *bad_recv_wr;
810 0 : int rc = 0;
811 :
812 0 : rc = spdk_rdma_provider_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr);
813 0 : if (spdk_unlikely(rc)) {
814 0 : nvme_rdma_reset_failed_recvs(rqpair->rsps, bad_recv_wr, rc);
815 0 : }
816 :
817 0 : return rc;
818 : }
819 :
820 : static inline int
821 0 : nvme_rdma_poller_submit_recvs(struct nvme_rdma_poller *poller)
822 : {
823 : struct ibv_recv_wr *bad_recv_wr;
824 : int rc;
825 :
826 0 : rc = spdk_rdma_provider_srq_flush_recv_wrs(poller->srq, &bad_recv_wr);
827 0 : if (spdk_unlikely(rc)) {
828 0 : nvme_rdma_reset_failed_recvs(poller->rsps, bad_recv_wr, rc);
829 0 : }
830 :
831 0 : return rc;
832 : }
833 :
834 : #define nvme_rdma_trace_ibv_sge(sg_list) \
835 : if (sg_list) { \
836 : SPDK_DEBUGLOG(nvme, "local addr %p length 0x%x lkey 0x%x\n", \
837 : (void *)(sg_list)->addr, (sg_list)->length, (sg_list)->lkey); \
838 : }
839 :
840 : static void
841 3 : nvme_rdma_free_rsps(struct nvme_rdma_rsps *rsps)
842 : {
843 3 : if (!rsps) {
844 1 : return;
845 : }
846 :
847 2 : spdk_free(rsps->rsps);
848 2 : spdk_free(rsps->rsp_sgls);
849 2 : spdk_free(rsps->rsp_recv_wrs);
850 2 : spdk_free(rsps);
851 3 : }
852 :
853 : static struct nvme_rdma_rsps *
854 2 : nvme_rdma_create_rsps(struct nvme_rdma_rsp_opts *opts)
855 : {
856 : struct nvme_rdma_rsps *rsps;
857 : struct spdk_rdma_utils_memory_translation translation;
858 : uint16_t i;
859 : int rc;
860 :
861 2 : rsps = spdk_zmalloc(sizeof(*rsps), 0, NULL, SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA);
862 2 : if (!rsps) {
863 0 : SPDK_ERRLOG("Failed to allocate rsps object\n");
864 0 : return NULL;
865 : }
866 :
867 2 : rsps->rsp_sgls = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsp_sgls), 0, NULL,
868 : SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA);
869 2 : if (!rsps->rsp_sgls) {
870 1 : SPDK_ERRLOG("Failed to allocate rsp_sgls\n");
871 1 : goto fail;
872 : }
873 :
874 1 : rsps->rsp_recv_wrs = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsp_recv_wrs), 0, NULL,
875 : SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA);
876 1 : if (!rsps->rsp_recv_wrs) {
877 0 : SPDK_ERRLOG("Failed to allocate rsp_recv_wrs\n");
878 0 : goto fail;
879 : }
880 :
881 1 : rsps->rsps = spdk_zmalloc(opts->num_entries * sizeof(*rsps->rsps), 0, NULL,
882 : SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA);
883 1 : if (!rsps->rsps) {
884 0 : SPDK_ERRLOG("can not allocate rdma rsps\n");
885 0 : goto fail;
886 : }
887 :
888 2 : for (i = 0; i < opts->num_entries; i++) {
889 1 : struct ibv_sge *rsp_sgl = &rsps->rsp_sgls[i];
890 1 : struct spdk_nvme_rdma_rsp *rsp = &rsps->rsps[i];
891 1 : struct ibv_recv_wr *recv_wr = &rsps->rsp_recv_wrs[i];
892 :
893 1 : rsp->rqpair = opts->rqpair;
894 1 : rsp->rdma_wr.type = RDMA_WR_TYPE_RECV;
895 1 : rsp->recv_wr = recv_wr;
896 1 : rsp_sgl->addr = (uint64_t)rsp;
897 1 : rsp_sgl->length = sizeof(struct spdk_nvme_cpl);
898 1 : rc = spdk_rdma_utils_get_translation(opts->mr_map, rsp, sizeof(*rsp), &translation);
899 1 : if (rc) {
900 0 : goto fail;
901 : }
902 1 : rsp_sgl->lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation);
903 :
904 1 : recv_wr->wr_id = (uint64_t)&rsp->rdma_wr;
905 1 : recv_wr->next = NULL;
906 1 : recv_wr->sg_list = rsp_sgl;
907 1 : recv_wr->num_sge = 1;
908 :
909 1 : nvme_rdma_trace_ibv_sge(recv_wr->sg_list);
910 :
911 1 : if (opts->rqpair) {
912 1 : spdk_rdma_provider_qp_queue_recv_wrs(opts->rqpair->rdma_qp, recv_wr);
913 1 : } else {
914 0 : spdk_rdma_provider_srq_queue_recv_wrs(opts->srq, recv_wr);
915 : }
916 1 : }
917 :
918 1 : rsps->num_entries = opts->num_entries;
919 1 : rsps->current_num_recvs = opts->num_entries;
920 :
921 1 : return rsps;
922 : fail:
923 1 : nvme_rdma_free_rsps(rsps);
924 1 : return NULL;
925 2 : }
926 :
927 : static void
928 3 : nvme_rdma_free_reqs(struct nvme_rdma_qpair *rqpair)
929 : {
930 3 : if (!rqpair->rdma_reqs) {
931 2 : return;
932 : }
933 :
934 1 : spdk_free(rqpair->cmds);
935 1 : rqpair->cmds = NULL;
936 :
937 1 : spdk_free(rqpair->rdma_reqs);
938 1 : rqpair->rdma_reqs = NULL;
939 3 : }
940 :
941 : static int
942 4 : nvme_rdma_create_reqs(struct nvme_rdma_qpair *rqpair)
943 : {
944 : struct spdk_rdma_utils_memory_translation translation;
945 : uint16_t i;
946 : int rc;
947 :
948 4 : assert(!rqpair->rdma_reqs);
949 4 : rqpair->rdma_reqs = spdk_zmalloc(rqpair->num_entries * sizeof(struct spdk_nvme_rdma_req), 0, NULL,
950 : SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA);
951 4 : if (rqpair->rdma_reqs == NULL) {
952 1 : SPDK_ERRLOG("Failed to allocate rdma_reqs\n");
953 1 : goto fail;
954 : }
955 :
956 3 : assert(!rqpair->cmds);
957 3 : rqpair->cmds = spdk_zmalloc(rqpair->num_entries * sizeof(*rqpair->cmds), 0, NULL,
958 : SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA);
959 3 : if (!rqpair->cmds) {
960 0 : SPDK_ERRLOG("Failed to allocate RDMA cmds\n");
961 0 : goto fail;
962 : }
963 :
964 3 : TAILQ_INIT(&rqpair->free_reqs);
965 3 : TAILQ_INIT(&rqpair->outstanding_reqs);
966 10 : for (i = 0; i < rqpair->num_entries; i++) {
967 : struct spdk_nvme_rdma_req *rdma_req;
968 : struct spdk_nvmf_cmd *cmd;
969 :
970 7 : rdma_req = &rqpair->rdma_reqs[i];
971 7 : rdma_req->rdma_wr.type = RDMA_WR_TYPE_SEND;
972 7 : cmd = &rqpair->cmds[i];
973 :
974 7 : rdma_req->id = i;
975 :
976 7 : rc = spdk_rdma_utils_get_translation(rqpair->mr_map, cmd, sizeof(*cmd), &translation);
977 7 : if (rc) {
978 0 : goto fail;
979 : }
980 7 : rdma_req->send_sgl[0].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation);
981 :
982 : /* The first RDMA sgl element will always point
983 : * at this data structure. Depending on whether
984 : * an NVMe-oF SGL is required, the length of
985 : * this element may change. */
986 7 : rdma_req->send_sgl[0].addr = (uint64_t)cmd;
987 7 : rdma_req->send_wr.wr_id = (uint64_t)&rdma_req->rdma_wr;
988 7 : rdma_req->send_wr.next = NULL;
989 7 : rdma_req->send_wr.opcode = IBV_WR_SEND;
990 7 : rdma_req->send_wr.send_flags = IBV_SEND_SIGNALED;
991 7 : rdma_req->send_wr.sg_list = rdma_req->send_sgl;
992 7 : rdma_req->send_wr.imm_data = 0;
993 :
994 7 : TAILQ_INSERT_TAIL(&rqpair->free_reqs, rdma_req, link);
995 7 : }
996 :
997 3 : return 0;
998 : fail:
999 1 : nvme_rdma_free_reqs(rqpair);
1000 1 : return -ENOMEM;
1001 4 : }
1002 :
1003 : static int nvme_rdma_connect(struct nvme_rdma_qpair *rqpair);
1004 :
1005 : static int
1006 0 : nvme_rdma_route_resolved(struct nvme_rdma_qpair *rqpair, int ret)
1007 : {
1008 0 : if (ret) {
1009 0 : SPDK_ERRLOG("RDMA route resolution error\n");
1010 0 : return -1;
1011 : }
1012 :
1013 0 : ret = nvme_rdma_qpair_init(rqpair);
1014 0 : if (ret < 0) {
1015 0 : SPDK_ERRLOG("nvme_rdma_qpair_init() failed\n");
1016 0 : return -1;
1017 : }
1018 :
1019 0 : return nvme_rdma_connect(rqpair);
1020 0 : }
1021 :
1022 : static int
1023 0 : nvme_rdma_addr_resolved(struct nvme_rdma_qpair *rqpair, int ret)
1024 : {
1025 0 : if (ret) {
1026 0 : SPDK_ERRLOG("RDMA address resolution error\n");
1027 0 : return -1;
1028 : }
1029 :
1030 0 : if (rqpair->qpair.ctrlr->opts.transport_ack_timeout != SPDK_NVME_TRANSPORT_ACK_TIMEOUT_DISABLED) {
1031 : #ifdef SPDK_CONFIG_RDMA_SET_ACK_TIMEOUT
1032 : uint8_t timeout = rqpair->qpair.ctrlr->opts.transport_ack_timeout;
1033 : ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID,
1034 : RDMA_OPTION_ID_ACK_TIMEOUT,
1035 : &timeout, sizeof(timeout));
1036 : if (ret) {
1037 : SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_ACK_TIMEOUT %d, ret %d\n", timeout, ret);
1038 : }
1039 : #else
1040 0 : SPDK_DEBUGLOG(nvme, "transport_ack_timeout is not supported\n");
1041 : #endif
1042 0 : }
1043 :
1044 0 : if (rqpair->qpair.ctrlr->opts.transport_tos != SPDK_NVME_TRANSPORT_TOS_DISABLED) {
1045 : #ifdef SPDK_CONFIG_RDMA_SET_TOS
1046 0 : uint8_t tos = rqpair->qpair.ctrlr->opts.transport_tos;
1047 0 : ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID, RDMA_OPTION_ID_TOS, &tos, sizeof(tos));
1048 0 : if (ret) {
1049 0 : SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_TOS %u, ret %d\n", tos, ret);
1050 0 : }
1051 : #else
1052 : SPDK_DEBUGLOG(nvme, "transport_tos is not supported\n");
1053 : #endif
1054 0 : }
1055 :
1056 0 : ret = rdma_resolve_route(rqpair->cm_id, NVME_RDMA_TIME_OUT_IN_MS);
1057 0 : if (ret) {
1058 0 : SPDK_ERRLOG("rdma_resolve_route\n");
1059 0 : return ret;
1060 : }
1061 :
1062 0 : return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ROUTE_RESOLVED,
1063 : nvme_rdma_route_resolved);
1064 0 : }
1065 :
1066 : static int
1067 0 : nvme_rdma_resolve_addr(struct nvme_rdma_qpair *rqpair,
1068 : struct sockaddr *src_addr,
1069 : struct sockaddr *dst_addr)
1070 : {
1071 : int ret;
1072 :
1073 0 : if (src_addr) {
1074 0 : int reuse = 1;
1075 :
1076 0 : ret = rdma_set_option(rqpair->cm_id, RDMA_OPTION_ID, RDMA_OPTION_ID_REUSEADDR,
1077 : &reuse, sizeof(reuse));
1078 0 : if (ret) {
1079 0 : SPDK_NOTICELOG("Can't apply RDMA_OPTION_ID_REUSEADDR %d, ret %d\n",
1080 : reuse, ret);
1081 : /* It is likely that rdma_resolve_addr() returns -EADDRINUSE, but
1082 : * we may missing something. We rely on rdma_resolve_addr().
1083 : */
1084 0 : }
1085 0 : }
1086 :
1087 0 : ret = rdma_resolve_addr(rqpair->cm_id, src_addr, dst_addr,
1088 : NVME_RDMA_TIME_OUT_IN_MS);
1089 0 : if (ret) {
1090 0 : SPDK_ERRLOG("rdma_resolve_addr, %d\n", errno);
1091 0 : return ret;
1092 : }
1093 :
1094 0 : return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ADDR_RESOLVED,
1095 : nvme_rdma_addr_resolved);
1096 0 : }
1097 :
1098 : static int nvme_rdma_stale_conn_retry(struct nvme_rdma_qpair *rqpair);
1099 :
1100 : static int
1101 0 : nvme_rdma_connect_established(struct nvme_rdma_qpair *rqpair, int ret)
1102 : {
1103 0 : struct nvme_rdma_rsp_opts opts = {};
1104 :
1105 0 : if (ret == -ESTALE) {
1106 0 : return nvme_rdma_stale_conn_retry(rqpair);
1107 0 : } else if (ret) {
1108 0 : SPDK_ERRLOG("RDMA connect error %d\n", ret);
1109 0 : return ret;
1110 : }
1111 :
1112 0 : assert(!rqpair->mr_map);
1113 0 : rqpair->mr_map = spdk_rdma_utils_create_mem_map(rqpair->rdma_qp->qp->pd, &g_nvme_hooks,
1114 : IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_READ | IBV_ACCESS_REMOTE_WRITE);
1115 0 : if (!rqpair->mr_map) {
1116 0 : SPDK_ERRLOG("Unable to register RDMA memory translation map\n");
1117 0 : return -1;
1118 : }
1119 :
1120 0 : ret = nvme_rdma_create_reqs(rqpair);
1121 0 : SPDK_DEBUGLOG(nvme, "rc =%d\n", ret);
1122 0 : if (ret) {
1123 0 : SPDK_ERRLOG("Unable to create rqpair RDMA requests\n");
1124 0 : return -1;
1125 : }
1126 0 : SPDK_DEBUGLOG(nvme, "RDMA requests created\n");
1127 :
1128 0 : if (!rqpair->srq) {
1129 0 : opts.num_entries = rqpair->num_entries;
1130 0 : opts.rqpair = rqpair;
1131 0 : opts.srq = NULL;
1132 0 : opts.mr_map = rqpair->mr_map;
1133 :
1134 0 : assert(!rqpair->rsps);
1135 0 : rqpair->rsps = nvme_rdma_create_rsps(&opts);
1136 0 : if (!rqpair->rsps) {
1137 0 : SPDK_ERRLOG("Unable to create rqpair RDMA responses\n");
1138 0 : return -1;
1139 : }
1140 0 : SPDK_DEBUGLOG(nvme, "RDMA responses created\n");
1141 :
1142 0 : ret = nvme_rdma_qpair_submit_recvs(rqpair);
1143 0 : SPDK_DEBUGLOG(nvme, "rc =%d\n", ret);
1144 0 : if (ret) {
1145 0 : SPDK_ERRLOG("Unable to submit rqpair RDMA responses\n");
1146 0 : return -1;
1147 : }
1148 0 : SPDK_DEBUGLOG(nvme, "RDMA responses submitted\n");
1149 0 : }
1150 :
1151 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_SEND;
1152 :
1153 0 : return 0;
1154 0 : }
1155 :
1156 : static int
1157 0 : nvme_rdma_connect(struct nvme_rdma_qpair *rqpair)
1158 : {
1159 0 : struct rdma_conn_param param = {};
1160 0 : struct spdk_nvmf_rdma_request_private_data request_data = {};
1161 : struct ibv_device_attr attr;
1162 : int ret;
1163 : struct spdk_nvme_ctrlr *ctrlr;
1164 :
1165 0 : ret = ibv_query_device(rqpair->cm_id->verbs, &attr);
1166 0 : if (ret != 0) {
1167 0 : SPDK_ERRLOG("Failed to query RDMA device attributes.\n");
1168 0 : return ret;
1169 : }
1170 :
1171 0 : param.responder_resources = attr.max_qp_rd_atom;
1172 :
1173 0 : ctrlr = rqpair->qpair.ctrlr;
1174 0 : if (!ctrlr) {
1175 0 : return -1;
1176 : }
1177 :
1178 0 : request_data.qid = rqpair->qpair.id;
1179 0 : request_data.hrqsize = rqpair->num_entries + 1;
1180 0 : request_data.hsqsize = rqpair->num_entries;
1181 0 : request_data.cntlid = ctrlr->cntlid;
1182 :
1183 0 : param.private_data = &request_data;
1184 0 : param.private_data_len = sizeof(request_data);
1185 0 : param.retry_count = ctrlr->opts.transport_retry_count;
1186 0 : param.rnr_retry_count = 7;
1187 :
1188 : /* Fields below are ignored by rdma cm if qpair has been
1189 : * created using rdma cm API. */
1190 0 : param.srq = 0;
1191 0 : param.qp_num = rqpair->rdma_qp->qp->qp_num;
1192 :
1193 0 : ret = rdma_connect(rqpair->cm_id, ¶m);
1194 0 : if (ret) {
1195 0 : SPDK_ERRLOG("nvme rdma connect error\n");
1196 0 : return ret;
1197 : }
1198 :
1199 0 : ctrlr->numa.id_valid = 1;
1200 0 : ctrlr->numa.id = spdk_rdma_cm_id_get_numa_id(rqpair->cm_id);
1201 :
1202 0 : return nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_ESTABLISHED,
1203 : nvme_rdma_connect_established);
1204 0 : }
1205 :
1206 : static int
1207 0 : nvme_rdma_ctrlr_connect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair)
1208 : {
1209 : struct sockaddr_storage dst_addr;
1210 : struct sockaddr_storage src_addr;
1211 : bool src_addr_specified;
1212 0 : long int port, src_port = 0;
1213 : int rc;
1214 : struct nvme_rdma_ctrlr *rctrlr;
1215 : struct nvme_rdma_qpair *rqpair;
1216 : struct nvme_rdma_poll_group *group;
1217 : int family;
1218 :
1219 0 : rqpair = nvme_rdma_qpair(qpair);
1220 0 : rctrlr = nvme_rdma_ctrlr(ctrlr);
1221 0 : assert(rctrlr != NULL);
1222 :
1223 0 : switch (ctrlr->trid.adrfam) {
1224 : case SPDK_NVMF_ADRFAM_IPV4:
1225 0 : family = AF_INET;
1226 0 : break;
1227 : case SPDK_NVMF_ADRFAM_IPV6:
1228 0 : family = AF_INET6;
1229 0 : break;
1230 : default:
1231 0 : SPDK_ERRLOG("Unhandled ADRFAM %d\n", ctrlr->trid.adrfam);
1232 0 : return -1;
1233 : }
1234 :
1235 0 : SPDK_DEBUGLOG(nvme, "adrfam %d ai_family %d\n", ctrlr->trid.adrfam, family);
1236 :
1237 0 : memset(&dst_addr, 0, sizeof(dst_addr));
1238 :
1239 0 : SPDK_DEBUGLOG(nvme, "trsvcid is %s\n", ctrlr->trid.trsvcid);
1240 0 : rc = nvme_parse_addr(&dst_addr, family, ctrlr->trid.traddr, ctrlr->trid.trsvcid, &port);
1241 0 : if (rc != 0) {
1242 0 : SPDK_ERRLOG("dst_addr nvme_parse_addr() failed\n");
1243 0 : return -1;
1244 : }
1245 :
1246 0 : if (ctrlr->opts.src_addr[0] || ctrlr->opts.src_svcid[0]) {
1247 0 : memset(&src_addr, 0, sizeof(src_addr));
1248 0 : rc = nvme_parse_addr(&src_addr, family,
1249 0 : ctrlr->opts.src_addr[0] ? ctrlr->opts.src_addr : NULL,
1250 0 : ctrlr->opts.src_svcid[0] ? ctrlr->opts.src_svcid : NULL,
1251 : &src_port);
1252 0 : if (rc != 0) {
1253 0 : SPDK_ERRLOG("src_addr nvme_parse_addr() failed\n");
1254 0 : return -1;
1255 : }
1256 0 : src_addr_specified = true;
1257 0 : } else {
1258 0 : src_addr_specified = false;
1259 : }
1260 :
1261 0 : rc = rdma_create_id(rctrlr->cm_channel, &rqpair->cm_id, rqpair, RDMA_PS_TCP);
1262 0 : if (rc < 0) {
1263 0 : SPDK_ERRLOG("rdma_create_id() failed\n");
1264 0 : return -1;
1265 : }
1266 :
1267 0 : rc = nvme_rdma_resolve_addr(rqpair,
1268 0 : src_addr_specified ? (struct sockaddr *)&src_addr : NULL,
1269 : (struct sockaddr *)&dst_addr);
1270 0 : if (rc < 0) {
1271 0 : SPDK_ERRLOG("nvme_rdma_resolve_addr() failed\n");
1272 0 : return -1;
1273 : }
1274 :
1275 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_INITIALIZING;
1276 :
1277 0 : if (qpair->poll_group != NULL) {
1278 0 : group = nvme_rdma_poll_group(qpair->poll_group);
1279 0 : TAILQ_INSERT_TAIL(&group->connecting_qpairs, rqpair, link_connecting);
1280 0 : }
1281 :
1282 0 : return 0;
1283 0 : }
1284 :
1285 : static int
1286 0 : nvme_rdma_stale_conn_reconnect(struct nvme_rdma_qpair *rqpair)
1287 : {
1288 0 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
1289 :
1290 0 : if (spdk_get_ticks() < rqpair->evt_timeout_ticks) {
1291 0 : return -EAGAIN;
1292 : }
1293 :
1294 0 : return nvme_rdma_ctrlr_connect_qpair(qpair->ctrlr, qpair);
1295 0 : }
1296 :
1297 : static int
1298 0 : nvme_rdma_ctrlr_connect_qpair_poll(struct spdk_nvme_ctrlr *ctrlr,
1299 : struct spdk_nvme_qpair *qpair)
1300 : {
1301 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
1302 : int rc;
1303 :
1304 0 : if (rqpair->in_connect_poll) {
1305 0 : return -EAGAIN;
1306 : }
1307 :
1308 0 : rqpair->in_connect_poll = true;
1309 :
1310 0 : switch (rqpair->state) {
1311 : case NVME_RDMA_QPAIR_STATE_INVALID:
1312 0 : rc = -EAGAIN;
1313 0 : break;
1314 :
1315 : case NVME_RDMA_QPAIR_STATE_INITIALIZING:
1316 : case NVME_RDMA_QPAIR_STATE_EXITING:
1317 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
1318 0 : nvme_ctrlr_lock(ctrlr);
1319 0 : }
1320 :
1321 0 : rc = nvme_rdma_process_event_poll(rqpair);
1322 :
1323 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
1324 0 : nvme_ctrlr_unlock(ctrlr);
1325 0 : }
1326 :
1327 0 : if (rc == 0) {
1328 0 : rc = -EAGAIN;
1329 0 : }
1330 0 : rqpair->in_connect_poll = false;
1331 :
1332 0 : return rc;
1333 :
1334 : case NVME_RDMA_QPAIR_STATE_STALE_CONN:
1335 0 : rc = nvme_rdma_stale_conn_reconnect(rqpair);
1336 0 : if (rc == 0) {
1337 0 : rc = -EAGAIN;
1338 0 : }
1339 0 : break;
1340 : case NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_SEND:
1341 0 : rc = nvme_fabric_qpair_connect_async(qpair, rqpair->num_entries + 1);
1342 0 : if (rc == 0) {
1343 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_POLL;
1344 0 : rc = -EAGAIN;
1345 0 : } else {
1346 0 : SPDK_ERRLOG("Failed to send an NVMe-oF Fabric CONNECT command\n");
1347 : }
1348 0 : break;
1349 : case NVME_RDMA_QPAIR_STATE_FABRIC_CONNECT_POLL:
1350 0 : rc = nvme_fabric_qpair_connect_poll(qpair);
1351 0 : if (rc == 0) {
1352 0 : if (nvme_fabric_qpair_auth_required(qpair)) {
1353 0 : rc = nvme_fabric_qpair_authenticate_async(qpair);
1354 0 : if (rc == 0) {
1355 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_AUTHENTICATING;
1356 0 : rc = -EAGAIN;
1357 0 : }
1358 0 : } else {
1359 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_RUNNING;
1360 0 : nvme_qpair_set_state(qpair, NVME_QPAIR_CONNECTED);
1361 : }
1362 0 : } else if (rc != -EAGAIN) {
1363 0 : SPDK_ERRLOG("Failed to poll NVMe-oF Fabric CONNECT command\n");
1364 0 : }
1365 0 : break;
1366 : case NVME_RDMA_QPAIR_STATE_AUTHENTICATING:
1367 0 : rc = nvme_fabric_qpair_authenticate_poll(qpair);
1368 0 : if (rc == 0) {
1369 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_RUNNING;
1370 0 : nvme_qpair_set_state(qpair, NVME_QPAIR_CONNECTED);
1371 0 : }
1372 0 : break;
1373 : case NVME_RDMA_QPAIR_STATE_RUNNING:
1374 0 : rc = 0;
1375 0 : break;
1376 : default:
1377 0 : assert(false);
1378 : rc = -EINVAL;
1379 : break;
1380 : }
1381 :
1382 0 : rqpair->in_connect_poll = false;
1383 :
1384 0 : return rc;
1385 0 : }
1386 :
1387 : static inline int
1388 28 : nvme_rdma_get_memory_translation(struct nvme_request *req, struct nvme_rdma_qpair *rqpair,
1389 : struct nvme_rdma_memory_translation_ctx *_ctx)
1390 : {
1391 : struct spdk_memory_domain_translation_ctx ctx;
1392 28 : struct spdk_memory_domain_translation_result dma_translation = {.iov_count = 0};
1393 : struct spdk_rdma_utils_memory_translation rdma_translation;
1394 : int rc;
1395 :
1396 28 : assert(req);
1397 28 : assert(rqpair);
1398 28 : assert(_ctx);
1399 :
1400 28 : if (req->payload.opts && req->payload.opts->memory_domain) {
1401 2 : ctx.size = sizeof(struct spdk_memory_domain_translation_ctx);
1402 2 : ctx.rdma.ibv_qp = rqpair->rdma_qp->qp;
1403 2 : dma_translation.size = sizeof(struct spdk_memory_domain_translation_result);
1404 :
1405 4 : rc = spdk_memory_domain_translate_data(req->payload.opts->memory_domain,
1406 2 : req->payload.opts->memory_domain_ctx,
1407 2 : rqpair->rdma_qp->domain, &ctx, _ctx->addr,
1408 2 : _ctx->length, &dma_translation);
1409 2 : if (spdk_unlikely(rc) || dma_translation.iov_count != 1) {
1410 1 : SPDK_ERRLOG("DMA memory translation failed, rc %d, iov count %u\n", rc, dma_translation.iov_count);
1411 1 : return rc;
1412 : }
1413 :
1414 1 : _ctx->lkey = dma_translation.rdma.lkey;
1415 1 : _ctx->rkey = dma_translation.rdma.rkey;
1416 1 : _ctx->addr = dma_translation.iov.iov_base;
1417 1 : _ctx->length = dma_translation.iov.iov_len;
1418 1 : } else {
1419 26 : rc = spdk_rdma_utils_get_translation(rqpair->mr_map, _ctx->addr, _ctx->length, &rdma_translation);
1420 26 : if (spdk_unlikely(rc)) {
1421 2 : SPDK_ERRLOG("RDMA memory translation failed, rc %d\n", rc);
1422 2 : return rc;
1423 : }
1424 24 : if (rdma_translation.translation_type == SPDK_RDMA_UTILS_TRANSLATION_MR) {
1425 24 : _ctx->lkey = rdma_translation.mr_or_key.mr->lkey;
1426 24 : _ctx->rkey = rdma_translation.mr_or_key.mr->rkey;
1427 24 : } else {
1428 0 : _ctx->lkey = _ctx->rkey = (uint32_t)rdma_translation.mr_or_key.key;
1429 : }
1430 : }
1431 :
1432 25 : return 0;
1433 28 : }
1434 :
1435 :
1436 : /*
1437 : * Build SGL describing empty payload.
1438 : */
1439 : static int
1440 2 : nvme_rdma_build_null_request(struct spdk_nvme_rdma_req *rdma_req)
1441 : {
1442 2 : struct nvme_request *req = rdma_req->req;
1443 :
1444 2 : req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG;
1445 :
1446 : /* The first element of this SGL is pointing at an
1447 : * spdk_nvmf_cmd object. For this particular command,
1448 : * we only need the first 64 bytes corresponding to
1449 : * the NVMe command. */
1450 2 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd);
1451 :
1452 : /* The RDMA SGL needs one element describing the NVMe command. */
1453 2 : rdma_req->send_wr.num_sge = 1;
1454 :
1455 2 : req->cmd.dptr.sgl1.keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
1456 2 : req->cmd.dptr.sgl1.keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
1457 2 : req->cmd.dptr.sgl1.keyed.length = 0;
1458 2 : req->cmd.dptr.sgl1.keyed.key = 0;
1459 2 : req->cmd.dptr.sgl1.address = 0;
1460 :
1461 2 : return 0;
1462 : }
1463 :
1464 : static inline void
1465 3 : nvme_rdma_configure_contig_inline_request(struct spdk_nvme_rdma_req *rdma_req,
1466 : struct nvme_request *req, struct nvme_rdma_memory_translation_ctx *ctx)
1467 : {
1468 3 : rdma_req->send_sgl[1].lkey = ctx->lkey;
1469 :
1470 : /* The first element of this SGL is pointing at an
1471 : * spdk_nvmf_cmd object. For this particular command,
1472 : * we only need the first 64 bytes corresponding to
1473 : * the NVMe command. */
1474 3 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd);
1475 :
1476 3 : rdma_req->send_sgl[1].addr = (uint64_t)ctx->addr;
1477 3 : rdma_req->send_sgl[1].length = (uint32_t)ctx->length;
1478 :
1479 : /* The RDMA SGL contains two elements. The first describes
1480 : * the NVMe command and the second describes the data
1481 : * payload. */
1482 3 : rdma_req->send_wr.num_sge = 2;
1483 :
1484 3 : req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG;
1485 3 : req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK;
1486 3 : req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
1487 3 : req->cmd.dptr.sgl1.unkeyed.length = (uint32_t)ctx->length;
1488 : /* Inline only supported for icdoff == 0 currently. This function will
1489 : * not get called for controllers with other values. */
1490 3 : req->cmd.dptr.sgl1.address = (uint64_t)0;
1491 3 : }
1492 :
1493 : /*
1494 : * Build inline SGL describing contiguous payload buffer.
1495 : */
1496 : static inline int
1497 3 : nvme_rdma_build_contig_inline_request(struct nvme_rdma_qpair *rqpair,
1498 : struct spdk_nvme_rdma_req *rdma_req)
1499 : {
1500 3 : struct nvme_request *req = rdma_req->req;
1501 9 : struct nvme_rdma_memory_translation_ctx ctx = {
1502 3 : .addr = (uint8_t *)req->payload.contig_or_cb_arg + req->payload_offset,
1503 3 : .length = req->payload_size
1504 : };
1505 : int rc;
1506 :
1507 3 : assert(ctx.length != 0);
1508 3 : assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_CONTIG);
1509 :
1510 3 : rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx);
1511 3 : if (spdk_unlikely(rc)) {
1512 0 : return -1;
1513 : }
1514 :
1515 3 : nvme_rdma_configure_contig_inline_request(rdma_req, req, &ctx);
1516 :
1517 3 : return 0;
1518 3 : }
1519 :
1520 : static inline void
1521 2 : nvme_rdma_configure_contig_request(struct spdk_nvme_rdma_req *rdma_req, struct nvme_request *req,
1522 : struct nvme_rdma_memory_translation_ctx *ctx)
1523 : {
1524 2 : req->cmd.dptr.sgl1.keyed.key = ctx->rkey;
1525 :
1526 : /* The first element of this SGL is pointing at an
1527 : * spdk_nvmf_cmd object. For this particular command,
1528 : * we only need the first 64 bytes corresponding to
1529 : * the NVMe command. */
1530 2 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd);
1531 :
1532 : /* The RDMA SGL needs one element describing the NVMe command. */
1533 2 : rdma_req->send_wr.num_sge = 1;
1534 :
1535 2 : req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG;
1536 2 : req->cmd.dptr.sgl1.keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
1537 2 : req->cmd.dptr.sgl1.keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
1538 2 : req->cmd.dptr.sgl1.keyed.length = (uint32_t)ctx->length;
1539 2 : req->cmd.dptr.sgl1.address = (uint64_t)ctx->addr;
1540 2 : }
1541 :
1542 : /*
1543 : * Build SGL describing contiguous payload buffer.
1544 : */
1545 : static inline int
1546 3 : nvme_rdma_build_contig_request(struct nvme_rdma_qpair *rqpair,
1547 : struct spdk_nvme_rdma_req *rdma_req)
1548 : {
1549 3 : struct nvme_request *req = rdma_req->req;
1550 9 : struct nvme_rdma_memory_translation_ctx ctx = {
1551 3 : .addr = (uint8_t *)req->payload.contig_or_cb_arg + req->payload_offset,
1552 3 : .length = req->payload_size
1553 : };
1554 : int rc;
1555 :
1556 3 : assert(req->payload_size != 0);
1557 3 : assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_CONTIG);
1558 :
1559 3 : if (spdk_unlikely(req->payload_size > NVME_RDMA_MAX_KEYED_SGL_LENGTH)) {
1560 1 : SPDK_ERRLOG("SGL length %u exceeds max keyed SGL block size %u\n",
1561 : req->payload_size, NVME_RDMA_MAX_KEYED_SGL_LENGTH);
1562 1 : return -1;
1563 : }
1564 :
1565 2 : rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx);
1566 2 : if (spdk_unlikely(rc)) {
1567 0 : return -1;
1568 : }
1569 :
1570 2 : nvme_rdma_configure_contig_request(rdma_req, req, &ctx);
1571 :
1572 2 : return 0;
1573 3 : }
1574 :
1575 : /*
1576 : * Build SGL describing scattered payload buffer.
1577 : */
1578 : static inline int
1579 7 : nvme_rdma_build_sgl_request(struct nvme_rdma_qpair *rqpair,
1580 : struct spdk_nvme_rdma_req *rdma_req)
1581 : {
1582 7 : struct nvme_request *req = rdma_req->req;
1583 7 : struct spdk_nvmf_cmd *cmd = &rqpair->cmds[rdma_req->id];
1584 : struct nvme_rdma_memory_translation_ctx ctx;
1585 : uint32_t remaining_size;
1586 : uint32_t sge_length;
1587 : int rc, max_num_sgl, num_sgl_desc;
1588 :
1589 7 : assert(req->payload_size != 0);
1590 7 : assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_SGL);
1591 7 : assert(req->payload.reset_sgl_fn != NULL);
1592 7 : assert(req->payload.next_sge_fn != NULL);
1593 7 : req->payload.reset_sgl_fn(req->payload.contig_or_cb_arg, req->payload_offset);
1594 :
1595 7 : max_num_sgl = req->qpair->ctrlr->max_sges;
1596 :
1597 7 : remaining_size = req->payload_size;
1598 7 : num_sgl_desc = 0;
1599 7 : do {
1600 18 : rc = req->payload.next_sge_fn(req->payload.contig_or_cb_arg, &ctx.addr, &sge_length);
1601 18 : if (spdk_unlikely(rc)) {
1602 1 : return -1;
1603 : }
1604 :
1605 17 : sge_length = spdk_min(remaining_size, sge_length);
1606 :
1607 17 : if (spdk_unlikely(sge_length > NVME_RDMA_MAX_KEYED_SGL_LENGTH)) {
1608 1 : SPDK_ERRLOG("SGL length %u exceeds max keyed SGL block size %u\n",
1609 : sge_length, NVME_RDMA_MAX_KEYED_SGL_LENGTH);
1610 1 : return -1;
1611 : }
1612 16 : ctx.length = sge_length;
1613 16 : rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx);
1614 16 : if (spdk_unlikely(rc)) {
1615 1 : return -1;
1616 : }
1617 :
1618 15 : cmd->sgl[num_sgl_desc].keyed.key = ctx.rkey;
1619 15 : cmd->sgl[num_sgl_desc].keyed.type = SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK;
1620 15 : cmd->sgl[num_sgl_desc].keyed.subtype = SPDK_NVME_SGL_SUBTYPE_ADDRESS;
1621 15 : cmd->sgl[num_sgl_desc].keyed.length = (uint32_t)ctx.length;
1622 15 : cmd->sgl[num_sgl_desc].address = (uint64_t)ctx.addr;
1623 :
1624 15 : remaining_size -= ctx.length;
1625 15 : num_sgl_desc++;
1626 15 : } while (remaining_size > 0 && num_sgl_desc < max_num_sgl);
1627 :
1628 :
1629 : /* Should be impossible if we did our sgl checks properly up the stack, but do a sanity check here. */
1630 4 : if (spdk_unlikely(remaining_size > 0)) {
1631 0 : return -1;
1632 : }
1633 :
1634 4 : req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG;
1635 :
1636 : /* The RDMA SGL needs one element describing some portion
1637 : * of the spdk_nvmf_cmd structure. */
1638 4 : rdma_req->send_wr.num_sge = 1;
1639 :
1640 : /*
1641 : * If only one SGL descriptor is required, it can be embedded directly in the command
1642 : * as a data block descriptor.
1643 : */
1644 4 : if (num_sgl_desc == 1) {
1645 : /* The first element of this SGL is pointing at an
1646 : * spdk_nvmf_cmd object. For this particular command,
1647 : * we only need the first 64 bytes corresponding to
1648 : * the NVMe command. */
1649 2 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd);
1650 :
1651 2 : req->cmd.dptr.sgl1.keyed.type = cmd->sgl[0].keyed.type;
1652 2 : req->cmd.dptr.sgl1.keyed.subtype = cmd->sgl[0].keyed.subtype;
1653 2 : req->cmd.dptr.sgl1.keyed.length = cmd->sgl[0].keyed.length;
1654 2 : req->cmd.dptr.sgl1.keyed.key = cmd->sgl[0].keyed.key;
1655 2 : req->cmd.dptr.sgl1.address = cmd->sgl[0].address;
1656 2 : } else {
1657 : /*
1658 : * Otherwise, The SGL descriptor embedded in the command must point to the list of
1659 : * SGL descriptors used to describe the operation. In that case it is a last segment descriptor.
1660 : */
1661 2 : uint32_t descriptors_size = sizeof(struct spdk_nvme_sgl_descriptor) * num_sgl_desc;
1662 :
1663 2 : if (spdk_unlikely(descriptors_size > rqpair->qpair.ctrlr->ioccsz_bytes)) {
1664 1 : SPDK_ERRLOG("Size of SGL descriptors (%u) exceeds ICD (%u)\n",
1665 : descriptors_size, rqpair->qpair.ctrlr->ioccsz_bytes);
1666 1 : return -1;
1667 : }
1668 1 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd) + descriptors_size;
1669 :
1670 1 : req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_LAST_SEGMENT;
1671 1 : req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
1672 1 : req->cmd.dptr.sgl1.unkeyed.length = descriptors_size;
1673 1 : req->cmd.dptr.sgl1.address = (uint64_t)0;
1674 : }
1675 :
1676 3 : return 0;
1677 7 : }
1678 :
1679 : /*
1680 : * Build inline SGL describing sgl payload buffer.
1681 : */
1682 : static inline int
1683 3 : nvme_rdma_build_sgl_inline_request(struct nvme_rdma_qpair *rqpair,
1684 : struct spdk_nvme_rdma_req *rdma_req)
1685 : {
1686 3 : struct nvme_request *req = rdma_req->req;
1687 : struct nvme_rdma_memory_translation_ctx ctx;
1688 : uint32_t length;
1689 : int rc;
1690 :
1691 3 : assert(req->payload_size != 0);
1692 3 : assert(nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_SGL);
1693 3 : assert(req->payload.reset_sgl_fn != NULL);
1694 3 : assert(req->payload.next_sge_fn != NULL);
1695 3 : req->payload.reset_sgl_fn(req->payload.contig_or_cb_arg, req->payload_offset);
1696 :
1697 3 : rc = req->payload.next_sge_fn(req->payload.contig_or_cb_arg, &ctx.addr, &length);
1698 3 : if (spdk_unlikely(rc)) {
1699 0 : return -1;
1700 : }
1701 :
1702 3 : if (length < req->payload_size) {
1703 0 : SPDK_DEBUGLOG(nvme, "Inline SGL request split so sending separately.\n");
1704 0 : return nvme_rdma_build_sgl_request(rqpair, rdma_req);
1705 : }
1706 :
1707 3 : if (length > req->payload_size) {
1708 0 : length = req->payload_size;
1709 0 : }
1710 :
1711 3 : ctx.length = length;
1712 3 : rc = nvme_rdma_get_memory_translation(req, rqpair, &ctx);
1713 3 : if (spdk_unlikely(rc)) {
1714 0 : return -1;
1715 : }
1716 :
1717 3 : rdma_req->send_sgl[1].addr = (uint64_t)ctx.addr;
1718 3 : rdma_req->send_sgl[1].length = (uint32_t)ctx.length;
1719 3 : rdma_req->send_sgl[1].lkey = ctx.lkey;
1720 :
1721 3 : rdma_req->send_wr.num_sge = 2;
1722 :
1723 : /* The first element of this SGL is pointing at an
1724 : * spdk_nvmf_cmd object. For this particular command,
1725 : * we only need the first 64 bytes corresponding to
1726 : * the NVMe command. */
1727 3 : rdma_req->send_sgl[0].length = sizeof(struct spdk_nvme_cmd);
1728 :
1729 3 : req->cmd.psdt = SPDK_NVME_PSDT_SGL_MPTR_CONTIG;
1730 3 : req->cmd.dptr.sgl1.unkeyed.type = SPDK_NVME_SGL_TYPE_DATA_BLOCK;
1731 3 : req->cmd.dptr.sgl1.unkeyed.subtype = SPDK_NVME_SGL_SUBTYPE_OFFSET;
1732 3 : req->cmd.dptr.sgl1.unkeyed.length = (uint32_t)ctx.length;
1733 : /* Inline only supported for icdoff == 0 currently. This function will
1734 : * not get called for controllers with other values. */
1735 3 : req->cmd.dptr.sgl1.address = (uint64_t)0;
1736 :
1737 3 : return 0;
1738 3 : }
1739 :
1740 : static inline int
1741 0 : nvme_rdma_accel_append_copy(struct spdk_nvme_poll_group *pg, void **seq,
1742 : struct spdk_memory_domain *rdma_domain, struct spdk_nvme_rdma_req *rdma_req,
1743 : struct iovec *iovs, uint32_t iovcnt,
1744 : struct spdk_memory_domain *src_domain, void *src_domain_ctx)
1745 : {
1746 0 : return pg->accel_fn_table.append_copy(pg->ctx, seq, iovs, iovcnt, rdma_domain, rdma_req, iovs,
1747 0 : iovcnt, src_domain, src_domain_ctx, NULL, NULL);
1748 : }
1749 :
1750 : static inline void
1751 0 : nvme_rdma_accel_reverse(struct spdk_nvme_poll_group *pg, void *seq)
1752 : {
1753 0 : pg->accel_fn_table.reverse_sequence(seq);
1754 0 : }
1755 :
1756 : static inline void
1757 0 : nvme_rdma_accel_finish(struct spdk_nvme_poll_group *pg, void *seq,
1758 : spdk_nvme_accel_completion_cb cb_fn, void *cb_arg)
1759 : {
1760 0 : pg->accel_fn_table.finish_sequence(seq, cb_fn, cb_arg);
1761 0 : }
1762 :
1763 : static inline void
1764 0 : nvme_rdma_accel_completion_cb(void *cb_arg, int status)
1765 : {
1766 0 : struct spdk_nvme_rdma_req *rdma_req = cb_arg;
1767 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(rdma_req->req->qpair);
1768 : struct spdk_nvme_cpl cpl;
1769 : enum spdk_nvme_generic_command_status_code sc;
1770 0 : uint16_t dnr = 0;
1771 :
1772 0 : rdma_req->in_progress_accel = 0;
1773 0 : rdma_req->req->accel_sequence = NULL;
1774 0 : SPDK_DEBUGLOG(nvme, "rdma_req %p qpair %p, accel completion rc %d\n", rdma_req, rqpair, status);
1775 :
1776 : /* nvme_rdma driver may fail data transfer on WC_FLUSH error completion which is expected.
1777 : * To prevent false errors from accel, first check if qpair is in the process of disconnect */
1778 0 : if (spdk_unlikely(!spdk_nvme_qpair_is_connected(&rqpair->qpair))) {
1779 0 : SPDK_DEBUGLOG(nvme, "qpair %p, req %p accel cpl in disconnecting, outstanding %u\n",
1780 : rqpair, rdma_req, rqpair->qpair.num_outstanding_reqs);
1781 0 : sc = SPDK_NVME_SC_ABORTED_SQ_DELETION;
1782 0 : goto fail_req;
1783 : }
1784 0 : if (spdk_unlikely(status)) {
1785 0 : SPDK_ERRLOG("qpair %p, req %p, accel sequence status %d\n", rdma_req->req->qpair, rdma_req, status);
1786 0 : sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR;
1787 : /* Something wrong happened, let the upper layer know that retry is no desired */
1788 0 : dnr = 1;
1789 0 : goto fail_req;
1790 : }
1791 :
1792 0 : nvme_rdma_req_complete(rdma_req, &rdma_req->rdma_rsp->cpl, true);
1793 0 : return;
1794 :
1795 : fail_req:
1796 0 : memset(&cpl, 0, sizeof(cpl));
1797 0 : cpl.status.sc = sc;
1798 0 : cpl.status.sct = SPDK_NVME_SCT_GENERIC;
1799 0 : cpl.status.dnr = dnr;
1800 0 : nvme_rdma_req_complete(rdma_req, &cpl, true);
1801 0 : }
1802 :
1803 : static inline int
1804 0 : nvme_rdma_apply_accel_sequence(struct nvme_rdma_qpair *rqpair, struct nvme_request *req,
1805 : struct spdk_nvme_rdma_req *rdma_req)
1806 : {
1807 0 : struct spdk_nvme_poll_group *pg = rqpair->qpair.poll_group->group;
1808 : struct spdk_memory_domain *src_domain;
1809 : void *src_domain_ctx;
1810 0 : void *accel_seq = req->accel_sequence;
1811 0 : uint32_t iovcnt = 0;
1812 : int rc;
1813 :
1814 0 : SPDK_DEBUGLOG(nvme, "req %p, start accel seq %p\n", rdma_req, accel_seq);
1815 0 : if (nvme_payload_type(&req->payload) == NVME_PAYLOAD_TYPE_SGL) {
1816 : void *addr;
1817 : uint32_t sge_length, payload_size;
1818 :
1819 0 : payload_size = req->payload_size;
1820 0 : assert(payload_size);
1821 0 : req->payload.reset_sgl_fn(req->payload.contig_or_cb_arg, req->payload_offset);
1822 0 : do {
1823 0 : rc = req->payload.next_sge_fn(req->payload.contig_or_cb_arg, &addr, &sge_length);
1824 0 : if (spdk_unlikely(rc)) {
1825 0 : return -1;
1826 : }
1827 0 : sge_length = spdk_min(payload_size, sge_length);
1828 0 : rdma_req->iovs[iovcnt].iov_base = addr;
1829 0 : rdma_req->iovs[iovcnt].iov_len = sge_length;
1830 0 : iovcnt++;
1831 0 : payload_size -= sge_length;
1832 0 : } while (payload_size && iovcnt < NVME_RDMA_MAX_SGL_DESCRIPTORS);
1833 :
1834 0 : if (spdk_unlikely(payload_size)) {
1835 0 : SPDK_ERRLOG("not enough iovs to handle req %p, remaining len %u\n", rdma_req, payload_size);
1836 0 : return -E2BIG;
1837 : }
1838 0 : } else {
1839 0 : rdma_req->iovs[iovcnt].iov_base = req->payload.contig_or_cb_arg;
1840 0 : rdma_req->iovs[iovcnt].iov_len = req->payload_size;
1841 0 : iovcnt = 1;
1842 : }
1843 0 : if (req->payload.opts && req->payload.opts->memory_domain) {
1844 0 : if (accel_seq) {
1845 0 : src_domain = rqpair->rdma_qp->domain;
1846 0 : src_domain_ctx = rdma_req;
1847 0 : } else {
1848 0 : src_domain = req->payload.opts->memory_domain;
1849 0 : src_domain_ctx = req->payload.opts->memory_domain_ctx;
1850 : }
1851 0 : } else {
1852 0 : src_domain = NULL;
1853 0 : src_domain_ctx = NULL;
1854 : }
1855 :
1856 0 : rc = nvme_rdma_accel_append_copy(pg, &accel_seq, rqpair->rdma_qp->domain, rdma_req, rdma_req->iovs,
1857 0 : iovcnt, src_domain, src_domain_ctx);
1858 0 : if (spdk_unlikely(rc)) {
1859 0 : return rc;
1860 : }
1861 :
1862 0 : if (spdk_nvme_opc_get_data_transfer(req->cmd.opc) == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
1863 0 : nvme_rdma_accel_reverse(pg, accel_seq);
1864 0 : }
1865 :
1866 0 : rdma_req->in_progress_accel = 1;
1867 0 : TAILQ_INSERT_TAIL(&rqpair->outstanding_reqs, rdma_req, link);
1868 0 : rqpair->num_outstanding_reqs++;
1869 :
1870 0 : SPDK_DEBUGLOG(nvme, "req %p, finish accel seq %p\n", rdma_req, accel_seq);
1871 0 : nvme_rdma_accel_finish(pg, accel_seq, nvme_rdma_accel_completion_cb, rdma_req);
1872 :
1873 0 : return 0;
1874 0 : }
1875 :
1876 : static inline int
1877 0 : nvme_rdma_memory_domain_transfer_data(struct spdk_memory_domain *dst_domain, void *dst_domain_ctx,
1878 : struct iovec *dst_iov, uint32_t dst_iovcnt,
1879 : struct spdk_memory_domain *src_domain, void *src_domain_ctx,
1880 : struct iovec *src_iov, uint32_t src_iovcnt,
1881 : struct spdk_memory_domain_translation_result *translation,
1882 : spdk_memory_domain_data_cpl_cb cpl_cb, void *cpl_cb_arg)
1883 : {
1884 : struct nvme_rdma_memory_translation_ctx ctx;
1885 0 : struct spdk_nvme_rdma_req *rdma_req = dst_domain_ctx;
1886 0 : struct nvme_request *req = rdma_req->req;
1887 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(rdma_req->req->qpair);
1888 0 : struct spdk_nvme_ctrlr *ctrlr = rqpair->qpair.ctrlr;
1889 : bool icd_supported;
1890 :
1891 0 : assert(dst_domain == rqpair->rdma_qp->domain);
1892 0 : assert(src_domain);
1893 0 : assert(spdk_memory_domain_get_dma_device_type(src_domain) == SPDK_DMA_DEVICE_TYPE_RDMA);
1894 : /* We expect "inplace" operation */
1895 0 : assert(dst_iov == src_iov);
1896 0 : assert(dst_iovcnt == src_iovcnt);
1897 :
1898 0 : if (spdk_unlikely(!src_domain ||
1899 : spdk_memory_domain_get_dma_device_type(src_domain) != SPDK_DMA_DEVICE_TYPE_RDMA)) {
1900 0 : SPDK_ERRLOG("Unexpected source memory domain %p, type %d\n", src_domain,
1901 : src_domain ? (int)spdk_memory_domain_get_dma_device_type(src_domain) : -1);
1902 0 : return -ENOTSUP;
1903 : }
1904 0 : if (spdk_unlikely(dst_iovcnt != 1 || !translation || translation->iov_count != 1)) {
1905 0 : SPDK_ERRLOG("Unexpected iovcnt %u or missed translation, rdma_req %p\n", dst_iovcnt, rdma_req);
1906 0 : return -ENOTSUP;
1907 : }
1908 0 : ctx.addr = translation->iov.iov_base;
1909 0 : ctx.length = translation->iov.iov_len;
1910 0 : ctx.lkey = translation->rdma.lkey;
1911 0 : ctx.rkey = translation->rdma.rkey;
1912 :
1913 0 : SPDK_DEBUGLOG(nvme, "req %p, addr %p, len %zu, key %u\n", rdma_req, ctx.addr, ctx.length, ctx.rkey);
1914 0 : icd_supported = spdk_nvme_opc_get_data_transfer(req->cmd.opc) == SPDK_NVME_DATA_HOST_TO_CONTROLLER
1915 0 : && req->payload_size <= ctrlr->ioccsz_bytes && ctrlr->icdoff == 0;
1916 :
1917 : /* We expect that result of accel sequence is a Memory Key which describes a virtually contig address space.
1918 : * That means we prepare a contig request even if original payload was scattered */
1919 0 : if (icd_supported) {
1920 0 : nvme_rdma_configure_contig_inline_request(rdma_req, req, &ctx);
1921 0 : } else {
1922 0 : nvme_rdma_configure_contig_request(rdma_req, req, &ctx);
1923 : }
1924 0 : rdma_req->transfer_cpl_cb = cpl_cb;
1925 0 : rdma_req->transfer_cpl_cb_arg = cpl_cb_arg;
1926 :
1927 0 : memcpy(&rqpair->cmds[rdma_req->id], &req->cmd, sizeof(req->cmd));
1928 :
1929 0 : return _nvme_rdma_qpair_submit_request(rqpair, rdma_req);
1930 0 : }
1931 :
1932 : static inline int
1933 6 : nvme_rdma_req_init(struct nvme_rdma_qpair *rqpair, struct spdk_nvme_rdma_req *rdma_req)
1934 : {
1935 6 : struct nvme_request *req = rdma_req->req;
1936 6 : struct spdk_nvme_ctrlr *ctrlr = rqpair->qpair.ctrlr;
1937 : enum nvme_payload_type payload_type;
1938 : bool icd_supported;
1939 6 : int rc = -1;
1940 :
1941 6 : payload_type = nvme_payload_type(&req->payload);
1942 : /*
1943 : * Check if icdoff is non zero, to avoid interop conflicts with
1944 : * targets with non-zero icdoff. Both SPDK and the Linux kernel
1945 : * targets use icdoff = 0. For targets with non-zero icdoff, we
1946 : * will currently just not use inline data for now.
1947 : */
1948 12 : icd_supported = spdk_nvme_opc_get_data_transfer(req->cmd.opc) == SPDK_NVME_DATA_HOST_TO_CONTROLLER
1949 6 : && req->payload_size <= ctrlr->ioccsz_bytes && ctrlr->icdoff == 0;
1950 :
1951 6 : if (spdk_unlikely(req->payload_size == 0)) {
1952 2 : rc = nvme_rdma_build_null_request(rdma_req);
1953 6 : } else if (payload_type == NVME_PAYLOAD_TYPE_CONTIG) {
1954 2 : if (icd_supported) {
1955 1 : rc = nvme_rdma_build_contig_inline_request(rqpair, rdma_req);
1956 1 : } else {
1957 1 : rc = nvme_rdma_build_contig_request(rqpair, rdma_req);
1958 : }
1959 4 : } else if (payload_type == NVME_PAYLOAD_TYPE_SGL) {
1960 2 : if (icd_supported) {
1961 1 : rc = nvme_rdma_build_sgl_inline_request(rqpair, rdma_req);
1962 1 : } else {
1963 1 : rc = nvme_rdma_build_sgl_request(rqpair, rdma_req);
1964 : }
1965 2 : }
1966 :
1967 6 : if (spdk_unlikely(rc)) {
1968 0 : return rc;
1969 : }
1970 :
1971 6 : memcpy(&rqpair->cmds[rdma_req->id], &req->cmd, sizeof(req->cmd));
1972 6 : return 0;
1973 6 : }
1974 :
1975 : static struct spdk_nvme_qpair *
1976 5 : nvme_rdma_ctrlr_create_qpair(struct spdk_nvme_ctrlr *ctrlr,
1977 : uint16_t qid, uint32_t qsize,
1978 : enum spdk_nvme_qprio qprio,
1979 : uint32_t num_requests,
1980 : bool delay_cmd_submit,
1981 : bool async)
1982 : {
1983 : struct nvme_rdma_qpair *rqpair;
1984 : struct spdk_nvme_qpair *qpair;
1985 : int rc;
1986 :
1987 5 : if (qsize < SPDK_NVME_QUEUE_MIN_ENTRIES) {
1988 2 : SPDK_ERRLOG("Failed to create qpair with size %u. Minimum queue size is %d.\n",
1989 : qsize, SPDK_NVME_QUEUE_MIN_ENTRIES);
1990 2 : return NULL;
1991 : }
1992 :
1993 3 : rqpair = spdk_zmalloc(sizeof(struct nvme_rdma_qpair), 0, NULL, SPDK_ENV_NUMA_ID_ANY,
1994 : SPDK_MALLOC_DMA);
1995 3 : if (!rqpair) {
1996 0 : SPDK_ERRLOG("failed to get create rqpair\n");
1997 0 : return NULL;
1998 : }
1999 :
2000 : /* Set num_entries one less than queue size. According to NVMe
2001 : * and NVMe-oF specs we can not submit queue size requests,
2002 : * one slot shall always remain empty.
2003 : */
2004 3 : rqpair->num_entries = qsize - 1;
2005 3 : rqpair->delay_cmd_submit = delay_cmd_submit;
2006 3 : rqpair->state = NVME_RDMA_QPAIR_STATE_INVALID;
2007 3 : qpair = &rqpair->qpair;
2008 3 : rc = nvme_qpair_init(qpair, qid, ctrlr, qprio, num_requests, async);
2009 3 : if (rc != 0) {
2010 0 : spdk_free(rqpair);
2011 0 : return NULL;
2012 : }
2013 :
2014 3 : return qpair;
2015 5 : }
2016 :
2017 : static void
2018 1 : nvme_rdma_qpair_destroy(struct nvme_rdma_qpair *rqpair)
2019 : {
2020 1 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
2021 : struct nvme_rdma_ctrlr *rctrlr;
2022 : struct nvme_rdma_cm_event_entry *entry, *tmp;
2023 :
2024 1 : spdk_rdma_utils_free_mem_map(&rqpair->mr_map);
2025 :
2026 1 : if (rqpair->evt) {
2027 0 : rdma_ack_cm_event(rqpair->evt);
2028 0 : rqpair->evt = NULL;
2029 0 : }
2030 :
2031 : /*
2032 : * This works because we have the controller lock both in
2033 : * this function and in the function where we add new events.
2034 : */
2035 1 : if (qpair->ctrlr != NULL) {
2036 1 : rctrlr = nvme_rdma_ctrlr(qpair->ctrlr);
2037 1 : STAILQ_FOREACH_SAFE(entry, &rctrlr->pending_cm_events, link, tmp) {
2038 0 : if (entry->evt->id->context == rqpair) {
2039 0 : STAILQ_REMOVE(&rctrlr->pending_cm_events, entry, nvme_rdma_cm_event_entry, link);
2040 0 : rdma_ack_cm_event(entry->evt);
2041 0 : STAILQ_INSERT_HEAD(&rctrlr->free_cm_events, entry, link);
2042 0 : }
2043 0 : }
2044 1 : }
2045 :
2046 1 : if (rqpair->cm_id) {
2047 0 : if (rqpair->rdma_qp) {
2048 0 : spdk_rdma_utils_put_pd(rqpair->rdma_qp->qp->pd);
2049 0 : spdk_rdma_provider_qp_destroy(rqpair->rdma_qp);
2050 0 : rqpair->rdma_qp = NULL;
2051 0 : }
2052 0 : }
2053 :
2054 1 : if (rqpair->poller) {
2055 : struct nvme_rdma_poll_group *group;
2056 :
2057 0 : assert(qpair->poll_group);
2058 0 : group = nvme_rdma_poll_group(qpair->poll_group);
2059 :
2060 0 : nvme_rdma_poll_group_put_poller(group, rqpair->poller);
2061 :
2062 0 : rqpair->poller = NULL;
2063 0 : rqpair->cq = NULL;
2064 0 : if (rqpair->srq) {
2065 0 : rqpair->srq = NULL;
2066 0 : rqpair->rsps = NULL;
2067 0 : }
2068 1 : } else if (rqpair->cq) {
2069 0 : ibv_destroy_cq(rqpair->cq);
2070 0 : rqpair->cq = NULL;
2071 0 : }
2072 :
2073 1 : nvme_rdma_free_reqs(rqpair);
2074 1 : nvme_rdma_free_rsps(rqpair->rsps);
2075 1 : rqpair->rsps = NULL;
2076 :
2077 : /* destroy cm_id last so cma device will not be freed before we destroy the cq. */
2078 1 : if (rqpair->cm_id) {
2079 0 : rdma_destroy_id(rqpair->cm_id);
2080 0 : rqpair->cm_id = NULL;
2081 0 : }
2082 1 : }
2083 :
2084 : static void nvme_rdma_qpair_abort_reqs(struct spdk_nvme_qpair *qpair, uint32_t dnr);
2085 :
2086 : static void
2087 0 : nvme_rdma_qpair_flush_send_wrs(struct nvme_rdma_qpair *rqpair)
2088 : {
2089 0 : struct ibv_send_wr *bad_wr = NULL;
2090 : int rc;
2091 :
2092 0 : rc = spdk_rdma_provider_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr);
2093 0 : if (rc) {
2094 0 : nvme_rdma_reset_failed_sends(rqpair, bad_wr);
2095 0 : }
2096 0 : }
2097 :
2098 : static int
2099 1 : nvme_rdma_qpair_disconnected(struct nvme_rdma_qpair *rqpair, int ret)
2100 : {
2101 1 : if (ret) {
2102 0 : SPDK_DEBUGLOG(nvme, "Target did not respond to qpair disconnect.\n");
2103 0 : goto quiet;
2104 : }
2105 :
2106 1 : if (rqpair->poller == NULL) {
2107 : /* If poller is not used, cq is not shared.
2108 : * So complete disconnecting qpair immediately.
2109 : */
2110 1 : goto quiet;
2111 : }
2112 :
2113 0 : if (rqpair->rsps == NULL) {
2114 0 : goto quiet;
2115 : }
2116 :
2117 0 : nvme_rdma_qpair_flush_send_wrs(rqpair);
2118 :
2119 0 : if (rqpair->need_destroy ||
2120 0 : (rqpair->current_num_sends != 0 ||
2121 0 : (!rqpair->srq && rqpair->rsps->current_num_recvs != 0)) ||
2122 0 : ((rqpair->qpair.ctrlr->flags & SPDK_NVME_CTRLR_ACCEL_SEQUENCE_SUPPORTED) &&
2123 0 : (!TAILQ_EMPTY(&rqpair->outstanding_reqs)))) {
2124 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_LINGERING;
2125 0 : rqpair->evt_timeout_ticks = (NVME_RDMA_DISCONNECTED_QPAIR_TIMEOUT_US * spdk_get_ticks_hz()) /
2126 0 : SPDK_SEC_TO_USEC + spdk_get_ticks();
2127 :
2128 0 : return -EAGAIN;
2129 : }
2130 :
2131 : quiet:
2132 1 : rqpair->state = NVME_RDMA_QPAIR_STATE_EXITED;
2133 :
2134 1 : nvme_rdma_qpair_abort_reqs(&rqpair->qpair, rqpair->qpair.abort_dnr);
2135 1 : nvme_rdma_qpair_destroy(rqpair);
2136 1 : nvme_transport_ctrlr_disconnect_qpair_done(&rqpair->qpair);
2137 :
2138 1 : return 0;
2139 1 : }
2140 :
2141 : static int
2142 0 : nvme_rdma_qpair_wait_until_quiet(struct nvme_rdma_qpair *rqpair)
2143 : {
2144 0 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
2145 0 : struct spdk_nvme_ctrlr *ctrlr = qpair->ctrlr;
2146 :
2147 0 : if (spdk_get_ticks() < rqpair->evt_timeout_ticks &&
2148 0 : (rqpair->current_num_sends != 0 ||
2149 0 : (!rqpair->srq && rqpair->rsps->current_num_recvs != 0))) {
2150 0 : return -EAGAIN;
2151 : }
2152 :
2153 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_EXITED;
2154 0 : nvme_rdma_qpair_abort_reqs(qpair, qpair->abort_dnr);
2155 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
2156 0 : nvme_robust_mutex_lock(&ctrlr->ctrlr_lock);
2157 0 : }
2158 0 : nvme_rdma_qpair_destroy(rqpair);
2159 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
2160 0 : nvme_robust_mutex_unlock(&ctrlr->ctrlr_lock);
2161 0 : }
2162 0 : nvme_transport_ctrlr_disconnect_qpair_done(&rqpair->qpair);
2163 :
2164 0 : return 0;
2165 0 : }
2166 :
2167 : static void
2168 0 : _nvme_rdma_ctrlr_disconnect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair,
2169 : nvme_rdma_cm_event_cb disconnected_qpair_cb)
2170 : {
2171 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2172 : int rc;
2173 :
2174 0 : assert(disconnected_qpair_cb != NULL);
2175 :
2176 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_EXITING;
2177 :
2178 0 : if (rqpair->cm_id) {
2179 0 : if (rqpair->rdma_qp) {
2180 0 : rc = spdk_rdma_provider_qp_disconnect(rqpair->rdma_qp);
2181 0 : if ((qpair->ctrlr != NULL) && (rc == 0)) {
2182 0 : rc = nvme_rdma_process_event_start(rqpair, RDMA_CM_EVENT_DISCONNECTED,
2183 0 : disconnected_qpair_cb);
2184 0 : if (rc == 0) {
2185 0 : return;
2186 : }
2187 0 : }
2188 0 : }
2189 0 : }
2190 :
2191 0 : disconnected_qpair_cb(rqpair, 0);
2192 0 : }
2193 :
2194 : static int
2195 0 : nvme_rdma_ctrlr_disconnect_qpair_poll(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair)
2196 : {
2197 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2198 : int rc;
2199 :
2200 0 : switch (rqpair->state) {
2201 : case NVME_RDMA_QPAIR_STATE_EXITING:
2202 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
2203 0 : nvme_ctrlr_lock(ctrlr);
2204 0 : }
2205 :
2206 0 : rc = nvme_rdma_process_event_poll(rqpair);
2207 :
2208 0 : if (!nvme_qpair_is_admin_queue(qpair)) {
2209 0 : nvme_ctrlr_unlock(ctrlr);
2210 0 : }
2211 0 : break;
2212 :
2213 : case NVME_RDMA_QPAIR_STATE_LINGERING:
2214 0 : rc = nvme_rdma_qpair_wait_until_quiet(rqpair);
2215 0 : break;
2216 : case NVME_RDMA_QPAIR_STATE_EXITED:
2217 0 : rc = 0;
2218 0 : break;
2219 :
2220 : default:
2221 0 : assert(false);
2222 : rc = -EAGAIN;
2223 : break;
2224 : }
2225 :
2226 0 : return rc;
2227 : }
2228 :
2229 : static void
2230 0 : nvme_rdma_ctrlr_disconnect_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair)
2231 : {
2232 : int rc;
2233 :
2234 0 : _nvme_rdma_ctrlr_disconnect_qpair(ctrlr, qpair, nvme_rdma_qpair_disconnected);
2235 :
2236 : /* If the async mode is disabled, poll the qpair until it is actually disconnected.
2237 : * It is ensured that poll_group_process_completions() calls disconnected_qpair_cb
2238 : * for any disconnected qpair. Hence, we do not have to check if the qpair is in
2239 : * a poll group or not.
2240 : * At the same time, if the qpair is being destroyed, i.e. this function is called by
2241 : * spdk_nvme_ctrlr_free_io_qpair then we need to wait until qpair is disconnected, otherwise
2242 : * we may leak some resources.
2243 : */
2244 0 : if (qpair->async && !qpair->destroy_in_progress) {
2245 0 : return;
2246 : }
2247 :
2248 0 : while (1) {
2249 0 : rc = nvme_rdma_ctrlr_disconnect_qpair_poll(ctrlr, qpair);
2250 0 : if (rc != -EAGAIN) {
2251 0 : break;
2252 : }
2253 : }
2254 0 : }
2255 :
2256 : static int
2257 0 : nvme_rdma_stale_conn_disconnected(struct nvme_rdma_qpair *rqpair, int ret)
2258 : {
2259 0 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
2260 :
2261 0 : if (ret) {
2262 0 : SPDK_DEBUGLOG(nvme, "Target did not respond to qpair disconnect.\n");
2263 0 : }
2264 :
2265 0 : nvme_rdma_qpair_destroy(rqpair);
2266 :
2267 0 : qpair->last_transport_failure_reason = qpair->transport_failure_reason;
2268 0 : qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_NONE;
2269 :
2270 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_STALE_CONN;
2271 0 : rqpair->evt_timeout_ticks = (NVME_RDMA_STALE_CONN_RETRY_DELAY_US * spdk_get_ticks_hz()) /
2272 0 : SPDK_SEC_TO_USEC + spdk_get_ticks();
2273 :
2274 0 : return 0;
2275 : }
2276 :
2277 : static int
2278 0 : nvme_rdma_stale_conn_retry(struct nvme_rdma_qpair *rqpair)
2279 : {
2280 0 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
2281 :
2282 0 : if (rqpair->stale_conn_retry_count >= NVME_RDMA_STALE_CONN_RETRY_MAX) {
2283 0 : SPDK_ERRLOG("Retry failed %d times, give up stale connection to qpair (cntlid:%u, qid:%u).\n",
2284 : NVME_RDMA_STALE_CONN_RETRY_MAX, qpair->ctrlr->cntlid, qpair->id);
2285 0 : return -ESTALE;
2286 : }
2287 :
2288 0 : rqpair->stale_conn_retry_count++;
2289 :
2290 0 : SPDK_NOTICELOG("%d times, retry stale connection to qpair (cntlid:%u, qid:%u).\n",
2291 : rqpair->stale_conn_retry_count, qpair->ctrlr->cntlid, qpair->id);
2292 :
2293 0 : _nvme_rdma_ctrlr_disconnect_qpair(qpair->ctrlr, qpair, nvme_rdma_stale_conn_disconnected);
2294 :
2295 0 : return 0;
2296 0 : }
2297 :
2298 : static int
2299 1 : nvme_rdma_ctrlr_delete_io_qpair(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_qpair *qpair)
2300 : {
2301 : struct nvme_rdma_qpair *rqpair;
2302 :
2303 1 : assert(qpair != NULL);
2304 1 : rqpair = nvme_rdma_qpair(qpair);
2305 :
2306 1 : if (rqpair->state != NVME_RDMA_QPAIR_STATE_EXITED) {
2307 : int rc __attribute__((unused));
2308 :
2309 : /* qpair was removed from the poll group while the disconnect is not finished.
2310 : * Destroy rdma resources forcefully. */
2311 1 : rc = nvme_rdma_qpair_disconnected(rqpair, 0);
2312 1 : assert(rc == 0);
2313 1 : }
2314 :
2315 1 : nvme_rdma_qpair_abort_reqs(qpair, qpair->abort_dnr);
2316 1 : nvme_qpair_deinit(qpair);
2317 :
2318 1 : spdk_free(rqpair);
2319 :
2320 1 : return 0;
2321 : }
2322 :
2323 : static struct spdk_nvme_qpair *
2324 0 : nvme_rdma_ctrlr_create_io_qpair(struct spdk_nvme_ctrlr *ctrlr, uint16_t qid,
2325 : const struct spdk_nvme_io_qpair_opts *opts)
2326 : {
2327 0 : return nvme_rdma_ctrlr_create_qpair(ctrlr, qid, opts->io_queue_size, opts->qprio,
2328 0 : opts->io_queue_requests,
2329 0 : opts->delay_cmd_submit,
2330 0 : opts->async_mode);
2331 : }
2332 :
2333 : static int
2334 0 : nvme_rdma_ctrlr_enable(struct spdk_nvme_ctrlr *ctrlr)
2335 : {
2336 : /* do nothing here */
2337 0 : return 0;
2338 : }
2339 :
2340 : static int nvme_rdma_ctrlr_destruct(struct spdk_nvme_ctrlr *ctrlr);
2341 :
2342 : /* We have to use the typedef in the function declaration to appease astyle. */
2343 : typedef struct spdk_nvme_ctrlr spdk_nvme_ctrlr_t;
2344 :
2345 : static spdk_nvme_ctrlr_t *
2346 1 : nvme_rdma_ctrlr_construct(const struct spdk_nvme_transport_id *trid,
2347 : const struct spdk_nvme_ctrlr_opts *opts,
2348 : void *devhandle)
2349 : {
2350 : struct nvme_rdma_ctrlr *rctrlr;
2351 : struct ibv_context **contexts;
2352 : struct ibv_device_attr dev_attr;
2353 : int i, flag, rc;
2354 :
2355 1 : rctrlr = spdk_zmalloc(sizeof(struct nvme_rdma_ctrlr), 0, NULL, SPDK_ENV_NUMA_ID_ANY,
2356 : SPDK_MALLOC_DMA);
2357 1 : if (rctrlr == NULL) {
2358 0 : SPDK_ERRLOG("could not allocate ctrlr\n");
2359 0 : return NULL;
2360 : }
2361 :
2362 1 : rctrlr->ctrlr.opts = *opts;
2363 1 : rctrlr->ctrlr.trid = *trid;
2364 :
2365 1 : if (opts->transport_retry_count > NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT) {
2366 1 : SPDK_NOTICELOG("transport_retry_count exceeds max value %d, use max value\n",
2367 : NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT);
2368 1 : rctrlr->ctrlr.opts.transport_retry_count = NVME_RDMA_CTRLR_MAX_TRANSPORT_RETRY_COUNT;
2369 1 : }
2370 :
2371 1 : if (opts->transport_ack_timeout > NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT) {
2372 1 : SPDK_NOTICELOG("transport_ack_timeout exceeds max value %d, use max value\n",
2373 : NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT);
2374 1 : rctrlr->ctrlr.opts.transport_ack_timeout = NVME_RDMA_CTRLR_MAX_TRANSPORT_ACK_TIMEOUT;
2375 1 : }
2376 :
2377 1 : contexts = rdma_get_devices(NULL);
2378 1 : if (contexts == NULL) {
2379 0 : SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno);
2380 0 : spdk_free(rctrlr);
2381 0 : return NULL;
2382 : }
2383 :
2384 1 : i = 0;
2385 1 : rctrlr->max_sge = NVME_RDMA_MAX_SGL_DESCRIPTORS;
2386 :
2387 3 : while (contexts[i] != NULL) {
2388 2 : rc = ibv_query_device(contexts[i], &dev_attr);
2389 2 : if (rc < 0) {
2390 0 : SPDK_ERRLOG("Failed to query RDMA device attributes.\n");
2391 0 : rdma_free_devices(contexts);
2392 0 : spdk_free(rctrlr);
2393 0 : return NULL;
2394 : }
2395 2 : rctrlr->max_sge = spdk_min(rctrlr->max_sge, (uint16_t)dev_attr.max_sge);
2396 2 : i++;
2397 : }
2398 :
2399 1 : rdma_free_devices(contexts);
2400 :
2401 1 : rc = nvme_ctrlr_construct(&rctrlr->ctrlr);
2402 1 : if (rc != 0) {
2403 0 : spdk_free(rctrlr);
2404 0 : return NULL;
2405 : }
2406 :
2407 1 : STAILQ_INIT(&rctrlr->pending_cm_events);
2408 1 : STAILQ_INIT(&rctrlr->free_cm_events);
2409 1 : rctrlr->cm_events = spdk_zmalloc(NVME_RDMA_NUM_CM_EVENTS * sizeof(*rctrlr->cm_events), 0, NULL,
2410 : SPDK_ENV_NUMA_ID_ANY, SPDK_MALLOC_DMA);
2411 1 : if (rctrlr->cm_events == NULL) {
2412 0 : SPDK_ERRLOG("unable to allocate buffers to hold CM events.\n");
2413 0 : goto destruct_ctrlr;
2414 : }
2415 :
2416 257 : for (i = 0; i < NVME_RDMA_NUM_CM_EVENTS; i++) {
2417 256 : STAILQ_INSERT_TAIL(&rctrlr->free_cm_events, &rctrlr->cm_events[i], link);
2418 256 : }
2419 :
2420 1 : rctrlr->cm_channel = rdma_create_event_channel();
2421 1 : if (rctrlr->cm_channel == NULL) {
2422 0 : SPDK_ERRLOG("rdma_create_event_channel() failed\n");
2423 0 : goto destruct_ctrlr;
2424 : }
2425 :
2426 1 : flag = fcntl(rctrlr->cm_channel->fd, F_GETFL);
2427 1 : if (fcntl(rctrlr->cm_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) {
2428 0 : SPDK_ERRLOG("Cannot set event channel to non blocking\n");
2429 0 : goto destruct_ctrlr;
2430 : }
2431 :
2432 2 : rctrlr->ctrlr.adminq = nvme_rdma_ctrlr_create_qpair(&rctrlr->ctrlr, 0,
2433 1 : rctrlr->ctrlr.opts.admin_queue_size, 0,
2434 1 : rctrlr->ctrlr.opts.admin_queue_size, false, true);
2435 1 : if (!rctrlr->ctrlr.adminq) {
2436 0 : SPDK_ERRLOG("failed to create admin qpair\n");
2437 0 : goto destruct_ctrlr;
2438 : }
2439 1 : if (spdk_rdma_provider_accel_sequence_supported()) {
2440 0 : rctrlr->ctrlr.flags |= SPDK_NVME_CTRLR_ACCEL_SEQUENCE_SUPPORTED;
2441 0 : }
2442 :
2443 1 : if (nvme_ctrlr_add_process(&rctrlr->ctrlr, 0) != 0) {
2444 0 : SPDK_ERRLOG("nvme_ctrlr_add_process() failed\n");
2445 0 : goto destruct_ctrlr;
2446 : }
2447 :
2448 1 : SPDK_DEBUGLOG(nvme, "successfully initialized the nvmf ctrlr\n");
2449 1 : return &rctrlr->ctrlr;
2450 :
2451 : destruct_ctrlr:
2452 0 : nvme_ctrlr_destruct(&rctrlr->ctrlr);
2453 0 : return NULL;
2454 1 : }
2455 :
2456 : static int
2457 1 : nvme_rdma_ctrlr_destruct(struct spdk_nvme_ctrlr *ctrlr)
2458 : {
2459 1 : struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(ctrlr);
2460 : struct nvme_rdma_cm_event_entry *entry;
2461 :
2462 1 : if (ctrlr->adminq) {
2463 1 : nvme_rdma_ctrlr_delete_io_qpair(ctrlr, ctrlr->adminq);
2464 1 : }
2465 :
2466 1 : STAILQ_FOREACH(entry, &rctrlr->pending_cm_events, link) {
2467 0 : rdma_ack_cm_event(entry->evt);
2468 0 : }
2469 :
2470 1 : STAILQ_INIT(&rctrlr->free_cm_events);
2471 1 : STAILQ_INIT(&rctrlr->pending_cm_events);
2472 1 : spdk_free(rctrlr->cm_events);
2473 :
2474 1 : if (rctrlr->cm_channel) {
2475 1 : rdma_destroy_event_channel(rctrlr->cm_channel);
2476 1 : rctrlr->cm_channel = NULL;
2477 1 : }
2478 :
2479 1 : nvme_ctrlr_destruct_finish(ctrlr);
2480 :
2481 1 : spdk_free(rctrlr);
2482 :
2483 1 : return 0;
2484 : }
2485 :
2486 : static inline int
2487 1 : _nvme_rdma_qpair_submit_request(struct nvme_rdma_qpair *rqpair,
2488 : struct spdk_nvme_rdma_req *rdma_req)
2489 : {
2490 1 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
2491 : struct ibv_send_wr *wr;
2492 : struct nvme_rdma_poll_group *group;
2493 :
2494 1 : if (!rqpair->link_active.tqe_prev && qpair->poll_group) {
2495 0 : group = nvme_rdma_poll_group(qpair->poll_group);
2496 0 : TAILQ_INSERT_TAIL(&group->active_qpairs, rqpair, link_active);
2497 0 : }
2498 1 : assert(rqpair->current_num_sends < rqpair->num_entries);
2499 1 : rqpair->current_num_sends++;
2500 :
2501 1 : wr = &rdma_req->send_wr;
2502 1 : wr->next = NULL;
2503 1 : nvme_rdma_trace_ibv_sge(wr->sg_list);
2504 :
2505 1 : spdk_rdma_provider_qp_queue_send_wrs(rqpair->rdma_qp, wr);
2506 :
2507 1 : if (!rqpair->delay_cmd_submit) {
2508 1 : return nvme_rdma_qpair_submit_sends(rqpair);
2509 : }
2510 :
2511 0 : return 0;
2512 1 : }
2513 :
2514 : static int
2515 2 : nvme_rdma_qpair_submit_request(struct spdk_nvme_qpair *qpair,
2516 : struct nvme_request *req)
2517 : {
2518 : struct nvme_rdma_qpair *rqpair;
2519 : struct spdk_nvme_rdma_req *rdma_req;
2520 : int rc;
2521 :
2522 2 : rqpair = nvme_rdma_qpair(qpair);
2523 2 : assert(rqpair != NULL);
2524 2 : assert(req != NULL);
2525 :
2526 2 : rdma_req = nvme_rdma_req_get(rqpair);
2527 2 : if (spdk_unlikely(!rdma_req)) {
2528 1 : if (rqpair->poller) {
2529 1 : rqpair->poller->stats.queued_requests++;
2530 1 : }
2531 : /* Inform the upper layer to try again later. */
2532 1 : return -EAGAIN;
2533 : }
2534 :
2535 1 : assert(rdma_req->req == NULL);
2536 1 : rdma_req->req = req;
2537 1 : req->cmd.cid = rdma_req->id;
2538 1 : if (req->accel_sequence) {
2539 0 : assert(spdk_rdma_provider_accel_sequence_supported());
2540 0 : assert(rqpair->qpair.poll_group->group);
2541 0 : assert(rqpair->qpair.poll_group->group->accel_fn_table.append_copy);
2542 0 : assert(rqpair->qpair.poll_group->group->accel_fn_table.reverse_sequence);
2543 0 : assert(rqpair->qpair.poll_group->group->accel_fn_table.finish_sequence);
2544 :
2545 0 : rc = nvme_rdma_apply_accel_sequence(rqpair, req, rdma_req);
2546 0 : if (spdk_unlikely(rc)) {
2547 0 : SPDK_ERRLOG("failed to apply accel seq, rqpair %p, req %p, rc %d\n", rqpair, rdma_req, rc);
2548 0 : nvme_rdma_req_put(rqpair, rdma_req);
2549 0 : return rc;
2550 : }
2551 : /* Capsule will be sent in data_transfer callback */
2552 0 : return 0;
2553 : }
2554 :
2555 1 : rc = nvme_rdma_req_init(rqpair, rdma_req);
2556 1 : if (spdk_unlikely(rc)) {
2557 0 : SPDK_ERRLOG("nvme_rdma_req_init() failed\n");
2558 0 : nvme_rdma_req_put(rqpair, rdma_req);
2559 0 : return -1;
2560 : }
2561 :
2562 1 : TAILQ_INSERT_TAIL(&rqpair->outstanding_reqs, rdma_req, link);
2563 1 : rqpair->num_outstanding_reqs++;
2564 :
2565 1 : return _nvme_rdma_qpair_submit_request(rqpair, rdma_req);
2566 2 : }
2567 :
2568 : static int
2569 0 : nvme_rdma_qpair_reset(struct spdk_nvme_qpair *qpair)
2570 : {
2571 : /* Currently, doing nothing here */
2572 0 : return 0;
2573 : }
2574 :
2575 : static void
2576 2 : nvme_rdma_qpair_abort_reqs(struct spdk_nvme_qpair *qpair, uint32_t dnr)
2577 : {
2578 : struct spdk_nvme_rdma_req *rdma_req, *tmp;
2579 : struct spdk_nvme_cpl cpl;
2580 2 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2581 :
2582 2 : cpl.sqid = qpair->id;
2583 2 : cpl.status.sc = SPDK_NVME_SC_ABORTED_SQ_DELETION;
2584 2 : cpl.status.sct = SPDK_NVME_SCT_GENERIC;
2585 2 : cpl.status.dnr = dnr;
2586 :
2587 : /*
2588 : * We cannot abort requests at the RDMA layer without
2589 : * unregistering them. If we do, we can still get error
2590 : * free completions on the shared completion queue.
2591 : */
2592 2 : if (nvme_qpair_get_state(qpair) > NVME_QPAIR_DISCONNECTING &&
2593 0 : nvme_qpair_get_state(qpair) != NVME_QPAIR_DESTROYING) {
2594 0 : nvme_ctrlr_disconnect_qpair(qpair);
2595 0 : }
2596 :
2597 2 : TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) {
2598 0 : if (rdma_req->in_progress_accel) {
2599 : /* We should wait for accel completion */
2600 0 : continue;
2601 : }
2602 0 : nvme_rdma_req_complete(rdma_req, &cpl, true);
2603 0 : }
2604 2 : }
2605 :
2606 : static void
2607 0 : nvme_rdma_qpair_check_timeout(struct spdk_nvme_qpair *qpair)
2608 : {
2609 : uint64_t t02;
2610 : struct spdk_nvme_rdma_req *rdma_req, *tmp;
2611 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2612 0 : struct spdk_nvme_ctrlr *ctrlr = qpair->ctrlr;
2613 : struct spdk_nvme_ctrlr_process *active_proc;
2614 :
2615 : /* Don't check timeouts during controller initialization. */
2616 0 : if (ctrlr->state != NVME_CTRLR_STATE_READY) {
2617 0 : return;
2618 : }
2619 :
2620 0 : if (nvme_qpair_is_admin_queue(qpair)) {
2621 0 : active_proc = nvme_ctrlr_get_current_process(ctrlr);
2622 0 : } else {
2623 0 : active_proc = qpair->active_proc;
2624 : }
2625 :
2626 : /* Only check timeouts if the current process has a timeout callback. */
2627 0 : if (active_proc == NULL || active_proc->timeout_cb_fn == NULL) {
2628 0 : return;
2629 : }
2630 :
2631 0 : t02 = spdk_get_ticks();
2632 0 : TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) {
2633 0 : assert(rdma_req->req != NULL);
2634 :
2635 0 : if (nvme_request_check_timeout(rdma_req->req, rdma_req->id, active_proc, t02)) {
2636 : /*
2637 : * The requests are in order, so as soon as one has not timed out,
2638 : * stop iterating.
2639 : */
2640 0 : break;
2641 : }
2642 0 : }
2643 0 : }
2644 :
2645 : static inline void
2646 0 : nvme_rdma_request_ready(struct nvme_rdma_qpair *rqpair, struct spdk_nvme_rdma_req *rdma_req)
2647 : {
2648 0 : struct spdk_nvme_rdma_rsp *rdma_rsp = rdma_req->rdma_rsp;
2649 0 : struct ibv_recv_wr *recv_wr = rdma_rsp->recv_wr;
2650 :
2651 0 : if (rdma_req->transfer_cpl_cb) {
2652 0 : int rc = 0;
2653 :
2654 0 : if (spdk_unlikely(spdk_nvme_cpl_is_error(&rdma_rsp->cpl))) {
2655 0 : SPDK_WARNLOG("req %p, error cpl sct %d, sc %d\n", rdma_req, rdma_rsp->cpl.status.sct,
2656 : rdma_rsp->cpl.status.sc);
2657 0 : rc = -EIO;
2658 0 : }
2659 0 : nvme_rdma_finish_data_transfer(rdma_req, rc);
2660 0 : } else {
2661 0 : nvme_rdma_req_complete(rdma_req, &rdma_rsp->cpl, true);
2662 : }
2663 :
2664 0 : if (spdk_unlikely(rqpair->state >= NVME_RDMA_QPAIR_STATE_EXITING && !rqpair->srq)) {
2665 : /* Skip posting back recv wr if we are in a disconnection process. We may never get
2666 : * a WC and we may end up stuck in LINGERING state until the timeout. */
2667 0 : return;
2668 : }
2669 :
2670 0 : assert(rqpair->rsps->current_num_recvs < rqpair->rsps->num_entries);
2671 0 : rqpair->rsps->current_num_recvs++;
2672 :
2673 0 : recv_wr->next = NULL;
2674 0 : nvme_rdma_trace_ibv_sge(recv_wr->sg_list);
2675 :
2676 0 : if (!rqpair->srq) {
2677 0 : spdk_rdma_provider_qp_queue_recv_wrs(rqpair->rdma_qp, recv_wr);
2678 0 : } else {
2679 0 : spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, recv_wr);
2680 : }
2681 0 : }
2682 :
2683 : #define MAX_COMPLETIONS_PER_POLL 128
2684 :
2685 : static void
2686 0 : nvme_rdma_fail_qpair(struct spdk_nvme_qpair *qpair, int failure_reason)
2687 : {
2688 0 : if (failure_reason == IBV_WC_RETRY_EXC_ERR) {
2689 0 : qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_REMOTE;
2690 0 : } else if (qpair->transport_failure_reason == SPDK_NVME_QPAIR_FAILURE_NONE) {
2691 0 : qpair->transport_failure_reason = SPDK_NVME_QPAIR_FAILURE_UNKNOWN;
2692 0 : }
2693 :
2694 0 : nvme_ctrlr_disconnect_qpair(qpair);
2695 0 : }
2696 :
2697 : static struct nvme_rdma_qpair *
2698 4 : get_rdma_qpair_from_wc(struct nvme_rdma_poll_group *group, struct ibv_wc *wc)
2699 : {
2700 : struct spdk_nvme_qpair *qpair;
2701 : struct nvme_rdma_qpair *rqpair;
2702 :
2703 5 : STAILQ_FOREACH(qpair, &group->group.connected_qpairs, poll_group_stailq) {
2704 2 : rqpair = nvme_rdma_qpair(qpair);
2705 2 : if (NVME_RDMA_POLL_GROUP_CHECK_QPN(rqpair, wc->qp_num)) {
2706 1 : return rqpair;
2707 : }
2708 1 : }
2709 :
2710 4 : STAILQ_FOREACH(qpair, &group->group.disconnected_qpairs, poll_group_stailq) {
2711 2 : rqpair = nvme_rdma_qpair(qpair);
2712 2 : if (NVME_RDMA_POLL_GROUP_CHECK_QPN(rqpair, wc->qp_num)) {
2713 1 : return rqpair;
2714 : }
2715 1 : }
2716 :
2717 2 : return NULL;
2718 4 : }
2719 :
2720 : static inline void
2721 0 : nvme_rdma_log_wc_status(struct nvme_rdma_qpair *rqpair, struct ibv_wc *wc)
2722 : {
2723 0 : struct nvme_rdma_wr *rdma_wr = (struct nvme_rdma_wr *)wc->wr_id;
2724 :
2725 0 : if (wc->status == IBV_WC_WR_FLUSH_ERR) {
2726 : /* If qpair is in ERR state, we will receive completions for all posted and not completed
2727 : * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */
2728 0 : SPDK_DEBUGLOG(nvme, "WC error, qid %u, qp state %d, request 0x%lu type %d, status: (%d): %s\n",
2729 : rqpair->qpair.id, rqpair->qpair.state, wc->wr_id, rdma_wr->type, wc->status,
2730 : ibv_wc_status_str(wc->status));
2731 0 : } else {
2732 0 : SPDK_ERRLOG("WC error, qid %u, qp state %d, request 0x%lu type %d, status: (%d): %s\n",
2733 : rqpair->qpair.id, rqpair->qpair.state, wc->wr_id, rdma_wr->type, wc->status,
2734 : ibv_wc_status_str(wc->status));
2735 : }
2736 0 : }
2737 :
2738 : static inline int
2739 0 : nvme_rdma_process_recv_completion(struct nvme_rdma_poller *poller, struct ibv_wc *wc,
2740 : struct nvme_rdma_wr *rdma_wr)
2741 : {
2742 : struct nvme_rdma_qpair *rqpair;
2743 : struct spdk_nvme_rdma_req *rdma_req;
2744 : struct spdk_nvme_rdma_rsp *rdma_rsp;
2745 :
2746 0 : rdma_rsp = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvme_rdma_rsp, rdma_wr);
2747 :
2748 0 : if (poller && poller->srq) {
2749 0 : rqpair = get_rdma_qpair_from_wc(poller->group, wc);
2750 0 : if (spdk_unlikely(!rqpair)) {
2751 : /* Since we do not handle the LAST_WQE_REACHED event, we do not know when
2752 : * a Receive Queue in a QP, that is associated with an SRQ, is flushed.
2753 : * We may get a WC for a already destroyed QP.
2754 : *
2755 : * However, for the SRQ, this is not any error. Hence, just re-post the
2756 : * receive request to the SRQ to reuse for other QPs, and return 0.
2757 : */
2758 0 : spdk_rdma_provider_srq_queue_recv_wrs(poller->srq, rdma_rsp->recv_wr);
2759 0 : return 0;
2760 : }
2761 0 : } else {
2762 0 : rqpair = rdma_rsp->rqpair;
2763 0 : if (spdk_unlikely(!rqpair)) {
2764 : /* TODO: Fix forceful QP destroy when it is not async mode.
2765 : * CQ itself did not cause any error. Hence, return 0 for now.
2766 : */
2767 0 : SPDK_WARNLOG("QP might be already destroyed.\n");
2768 0 : return 0;
2769 : }
2770 : }
2771 :
2772 :
2773 0 : assert(rqpair->rsps->current_num_recvs > 0);
2774 0 : rqpair->rsps->current_num_recvs--;
2775 :
2776 0 : if (spdk_unlikely(wc->status)) {
2777 0 : nvme_rdma_log_wc_status(rqpair, wc);
2778 0 : goto err_wc;
2779 : }
2780 :
2781 0 : SPDK_DEBUGLOG(nvme, "CQ recv completion\n");
2782 :
2783 0 : if (spdk_unlikely(wc->byte_len < sizeof(struct spdk_nvme_cpl))) {
2784 0 : SPDK_ERRLOG("recv length %u less than expected response size\n", wc->byte_len);
2785 0 : goto err_wc;
2786 : }
2787 0 : rdma_req = &rqpair->rdma_reqs[rdma_rsp->cpl.cid];
2788 0 : rdma_req->completion_flags |= NVME_RDMA_RECV_COMPLETED;
2789 0 : rdma_req->rdma_rsp = rdma_rsp;
2790 :
2791 0 : if ((rdma_req->completion_flags & NVME_RDMA_SEND_COMPLETED) == 0) {
2792 0 : return 0;
2793 : }
2794 :
2795 0 : rqpair->num_completions++;
2796 :
2797 0 : nvme_rdma_request_ready(rqpair, rdma_req);
2798 :
2799 0 : if (!rqpair->delay_cmd_submit) {
2800 0 : if (spdk_unlikely(nvme_rdma_qpair_submit_recvs(rqpair))) {
2801 0 : SPDK_ERRLOG("Unable to re-post rx descriptor\n");
2802 0 : nvme_rdma_fail_qpair(&rqpair->qpair, 0);
2803 0 : return -ENXIO;
2804 : }
2805 0 : }
2806 :
2807 0 : return 1;
2808 :
2809 : err_wc:
2810 0 : nvme_rdma_fail_qpair(&rqpair->qpair, 0);
2811 0 : if (poller && poller->srq) {
2812 0 : spdk_rdma_provider_srq_queue_recv_wrs(poller->srq, rdma_rsp->recv_wr);
2813 0 : }
2814 0 : rdma_req = &rqpair->rdma_reqs[rdma_rsp->cpl.cid];
2815 0 : if (rdma_req->transfer_cpl_cb) {
2816 0 : nvme_rdma_finish_data_transfer(rdma_req, -ENXIO);
2817 0 : }
2818 0 : return -ENXIO;
2819 0 : }
2820 :
2821 : static inline int
2822 0 : nvme_rdma_process_send_completion(struct nvme_rdma_poller *poller,
2823 : struct nvme_rdma_qpair *rdma_qpair,
2824 : struct ibv_wc *wc, struct nvme_rdma_wr *rdma_wr)
2825 : {
2826 : struct nvme_rdma_qpair *rqpair;
2827 : struct spdk_nvme_rdma_req *rdma_req;
2828 :
2829 0 : rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvme_rdma_req, rdma_wr);
2830 0 : rqpair = rdma_req->req ? nvme_rdma_qpair(rdma_req->req->qpair) : NULL;
2831 0 : if (spdk_unlikely(!rqpair)) {
2832 0 : rqpair = rdma_qpair != NULL ? rdma_qpair : get_rdma_qpair_from_wc(poller->group, wc);
2833 0 : }
2834 :
2835 : /* If we are flushing I/O */
2836 0 : if (spdk_unlikely(wc->status)) {
2837 0 : if (!rqpair) {
2838 : /* When poll_group is used, several qpairs share the same CQ and it is possible to
2839 : * receive a completion with error (e.g. IBV_WC_WR_FLUSH_ERR) for already disconnected qpair
2840 : * That happens due to qpair is destroyed while there are submitted but not completed send/receive
2841 : * Work Requests */
2842 0 : assert(poller);
2843 0 : return 0;
2844 : }
2845 0 : assert(rqpair->current_num_sends > 0);
2846 0 : rqpair->current_num_sends--;
2847 0 : nvme_rdma_log_wc_status(rqpair, wc);
2848 0 : nvme_rdma_fail_qpair(&rqpair->qpair, 0);
2849 0 : if (rdma_req->rdma_rsp && poller && poller->srq) {
2850 0 : spdk_rdma_provider_srq_queue_recv_wrs(poller->srq, rdma_req->rdma_rsp->recv_wr);
2851 0 : }
2852 0 : if (rdma_req->transfer_cpl_cb) {
2853 0 : nvme_rdma_finish_data_transfer(rdma_req, -ENXIO);
2854 0 : }
2855 0 : return -ENXIO;
2856 : }
2857 :
2858 : /* We do not support Soft Roce anymore. Other than Soft Roce's bug, we should not
2859 : * receive a completion without error status after qpair is disconnected/destroyed.
2860 : */
2861 0 : if (spdk_unlikely(rdma_req->req == NULL)) {
2862 : /*
2863 : * Some infiniband drivers do not guarantee the previous assumption after we
2864 : * received a RDMA_CM_EVENT_DEVICE_REMOVAL event.
2865 : */
2866 0 : SPDK_ERRLOG("Received malformed completion: request 0x%"PRIx64" type %d\n", wc->wr_id,
2867 : rdma_wr->type);
2868 0 : if (!rqpair || !rqpair->need_destroy) {
2869 0 : assert(0);
2870 : }
2871 0 : return -ENXIO;
2872 : }
2873 :
2874 0 : rdma_req->completion_flags |= NVME_RDMA_SEND_COMPLETED;
2875 0 : assert(rqpair->current_num_sends > 0);
2876 0 : rqpair->current_num_sends--;
2877 :
2878 0 : if ((rdma_req->completion_flags & NVME_RDMA_RECV_COMPLETED) == 0) {
2879 0 : return 0;
2880 : }
2881 :
2882 0 : rqpair->num_completions++;
2883 :
2884 0 : nvme_rdma_request_ready(rqpair, rdma_req);
2885 :
2886 0 : if (!rqpair->delay_cmd_submit) {
2887 0 : if (spdk_unlikely(nvme_rdma_qpair_submit_recvs(rqpair))) {
2888 0 : SPDK_ERRLOG("Unable to re-post rx descriptor\n");
2889 0 : nvme_rdma_fail_qpair(&rqpair->qpair, 0);
2890 0 : return -ENXIO;
2891 : }
2892 0 : }
2893 :
2894 0 : return 1;
2895 0 : }
2896 :
2897 : static inline int
2898 0 : nvme_rdma_cq_process_completions(struct ibv_cq *cq, uint32_t batch_size,
2899 : struct nvme_rdma_poller *poller,
2900 : struct nvme_rdma_qpair *rdma_qpair,
2901 : uint64_t *rdma_completions)
2902 : {
2903 : struct ibv_wc wc[MAX_COMPLETIONS_PER_POLL];
2904 : struct nvme_rdma_wr *rdma_wr;
2905 0 : uint32_t reaped = 0;
2906 0 : int completion_rc = 0;
2907 : int rc, _rc, i;
2908 :
2909 0 : rc = ibv_poll_cq(cq, batch_size, wc);
2910 0 : if (spdk_unlikely(rc < 0)) {
2911 0 : SPDK_ERRLOG("Error polling CQ! (%d): %s\n",
2912 : errno, spdk_strerror(errno));
2913 0 : return -ECANCELED;
2914 0 : } else if (rc == 0) {
2915 0 : return 0;
2916 : }
2917 :
2918 0 : for (i = 0; i < rc; i++) {
2919 0 : rdma_wr = (struct nvme_rdma_wr *)wc[i].wr_id;
2920 0 : switch (rdma_wr->type) {
2921 : case RDMA_WR_TYPE_RECV:
2922 0 : _rc = nvme_rdma_process_recv_completion(poller, &wc[i], rdma_wr);
2923 0 : break;
2924 :
2925 : case RDMA_WR_TYPE_SEND:
2926 0 : _rc = nvme_rdma_process_send_completion(poller, rdma_qpair, &wc[i], rdma_wr);
2927 0 : break;
2928 :
2929 : default:
2930 0 : SPDK_ERRLOG("Received an unexpected opcode on the CQ: %d\n", rdma_wr->type);
2931 0 : return -ECANCELED;
2932 : }
2933 0 : if (spdk_likely(_rc >= 0)) {
2934 0 : reaped += _rc;
2935 0 : } else {
2936 0 : completion_rc = _rc;
2937 : }
2938 0 : }
2939 :
2940 0 : *rdma_completions += rc;
2941 :
2942 0 : if (spdk_unlikely(completion_rc)) {
2943 0 : return completion_rc;
2944 : }
2945 :
2946 0 : return reaped;
2947 0 : }
2948 :
2949 : static void
2950 0 : dummy_disconnected_qpair_cb(struct spdk_nvme_qpair *qpair, void *poll_group_ctx)
2951 : {
2952 :
2953 0 : }
2954 :
2955 : static int
2956 0 : nvme_rdma_qpair_process_completions(struct spdk_nvme_qpair *qpair,
2957 : uint32_t max_completions)
2958 : {
2959 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
2960 0 : struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(qpair->ctrlr);
2961 0 : int rc = 0, batch_size;
2962 : struct ibv_cq *cq;
2963 0 : uint64_t rdma_completions = 0;
2964 :
2965 : /*
2966 : * This is used during the connection phase. It's possible that we are still reaping error completions
2967 : * from other qpairs so we need to call the poll group function. Also, it's more correct since the cq
2968 : * is shared.
2969 : */
2970 0 : if (qpair->poll_group != NULL) {
2971 0 : return spdk_nvme_poll_group_process_completions(qpair->poll_group->group, max_completions,
2972 : dummy_disconnected_qpair_cb);
2973 : }
2974 :
2975 0 : if (max_completions == 0) {
2976 0 : max_completions = rqpair->num_entries;
2977 0 : } else {
2978 0 : max_completions = spdk_min(max_completions, rqpair->num_entries);
2979 : }
2980 :
2981 0 : switch (nvme_qpair_get_state(qpair)) {
2982 : case NVME_QPAIR_CONNECTING:
2983 0 : rc = nvme_rdma_ctrlr_connect_qpair_poll(qpair->ctrlr, qpair);
2984 0 : if (rc == 0) {
2985 : /* Once the connection is completed, we can submit queued requests */
2986 0 : nvme_qpair_resubmit_requests(qpair, rqpair->num_entries);
2987 0 : } else if (rc != -EAGAIN) {
2988 0 : SPDK_ERRLOG("Failed to connect rqpair=%p\n", rqpair);
2989 0 : goto failed;
2990 0 : } else if (rqpair->state <= NVME_RDMA_QPAIR_STATE_INITIALIZING) {
2991 0 : return 0;
2992 : }
2993 0 : break;
2994 :
2995 : case NVME_QPAIR_DISCONNECTING:
2996 0 : nvme_rdma_ctrlr_disconnect_qpair_poll(qpair->ctrlr, qpair);
2997 0 : return -ENXIO;
2998 :
2999 : default:
3000 0 : if (nvme_qpair_is_admin_queue(qpair)) {
3001 0 : nvme_rdma_poll_events(rctrlr);
3002 0 : }
3003 0 : nvme_rdma_qpair_process_cm_event(rqpair);
3004 0 : break;
3005 : }
3006 :
3007 0 : if (spdk_unlikely(qpair->transport_failure_reason != SPDK_NVME_QPAIR_FAILURE_NONE)) {
3008 0 : goto failed;
3009 : }
3010 :
3011 0 : cq = rqpair->cq;
3012 :
3013 0 : rqpair->num_completions = 0;
3014 0 : do {
3015 0 : batch_size = spdk_min((max_completions - rqpair->num_completions), MAX_COMPLETIONS_PER_POLL);
3016 0 : rc = nvme_rdma_cq_process_completions(cq, batch_size, NULL, rqpair, &rdma_completions);
3017 :
3018 0 : if (rc == 0) {
3019 0 : break;
3020 : /* Handle the case where we fail to poll the cq. */
3021 0 : } else if (rc == -ECANCELED) {
3022 0 : goto failed;
3023 0 : } else if (rc == -ENXIO) {
3024 0 : return rc;
3025 : }
3026 0 : } while (rqpair->num_completions < max_completions);
3027 :
3028 0 : if (spdk_unlikely(nvme_rdma_qpair_submit_sends(rqpair) ||
3029 : nvme_rdma_qpair_submit_recvs(rqpair))) {
3030 0 : goto failed;
3031 : }
3032 :
3033 0 : if (spdk_unlikely(qpair->ctrlr->timeout_enabled)) {
3034 0 : nvme_rdma_qpair_check_timeout(qpair);
3035 0 : }
3036 :
3037 0 : return rqpair->num_completions;
3038 :
3039 : failed:
3040 0 : nvme_rdma_fail_qpair(qpair, 0);
3041 0 : return -ENXIO;
3042 0 : }
3043 :
3044 : static uint32_t
3045 0 : nvme_rdma_ctrlr_get_max_xfer_size(struct spdk_nvme_ctrlr *ctrlr)
3046 : {
3047 : /* max_mr_size by ibv_query_device indicates the largest value that we can
3048 : * set for a registered memory region. It is independent from the actual
3049 : * I/O size and is very likely to be larger than 2 MiB which is the
3050 : * granularity we currently register memory regions. Hence return
3051 : * UINT32_MAX here and let the generic layer use the controller data to
3052 : * moderate this value.
3053 : */
3054 0 : return UINT32_MAX;
3055 : }
3056 :
3057 : static uint16_t
3058 5 : nvme_rdma_ctrlr_get_max_sges(struct spdk_nvme_ctrlr *ctrlr)
3059 : {
3060 5 : struct nvme_rdma_ctrlr *rctrlr = nvme_rdma_ctrlr(ctrlr);
3061 5 : uint32_t max_sge = rctrlr->max_sge;
3062 10 : uint32_t max_in_capsule_sge = (ctrlr->cdata.nvmf_specific.ioccsz * 16 -
3063 5 : sizeof(struct spdk_nvme_cmd)) /
3064 : sizeof(struct spdk_nvme_sgl_descriptor);
3065 :
3066 : /* Max SGE is limited by capsule size */
3067 5 : max_sge = spdk_min(max_sge, max_in_capsule_sge);
3068 : /* Max SGE may be limited by MSDBD */
3069 5 : if (ctrlr->cdata.nvmf_specific.msdbd != 0) {
3070 5 : max_sge = spdk_min(max_sge, ctrlr->cdata.nvmf_specific.msdbd);
3071 5 : }
3072 :
3073 : /* Max SGE can't be less than 1 */
3074 5 : max_sge = spdk_max(1, max_sge);
3075 5 : return max_sge;
3076 : }
3077 :
3078 : static int
3079 0 : nvme_rdma_qpair_iterate_requests(struct spdk_nvme_qpair *qpair,
3080 : int (*iter_fn)(struct nvme_request *req, void *arg),
3081 : void *arg)
3082 : {
3083 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
3084 : struct spdk_nvme_rdma_req *rdma_req, *tmp;
3085 : int rc;
3086 :
3087 0 : assert(iter_fn != NULL);
3088 :
3089 0 : TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) {
3090 0 : assert(rdma_req->req != NULL);
3091 :
3092 0 : rc = iter_fn(rdma_req->req, arg);
3093 0 : if (rc != 0) {
3094 0 : return rc;
3095 : }
3096 0 : }
3097 :
3098 0 : return 0;
3099 0 : }
3100 :
3101 : static int
3102 0 : nvme_rdma_qpair_authenticate(struct spdk_nvme_qpair *qpair)
3103 : {
3104 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
3105 : int rc;
3106 :
3107 : /* If the qpair is still connecting, it'll be forced to authenticate later on */
3108 0 : if (rqpair->state < NVME_RDMA_QPAIR_STATE_RUNNING) {
3109 0 : return 0;
3110 0 : } else if (rqpair->state != NVME_RDMA_QPAIR_STATE_RUNNING) {
3111 0 : return -ENOTCONN;
3112 : }
3113 :
3114 0 : rc = nvme_fabric_qpair_authenticate_async(qpair);
3115 0 : if (rc == 0) {
3116 0 : nvme_qpair_set_state(qpair, NVME_QPAIR_CONNECTING);
3117 0 : rqpair->state = NVME_RDMA_QPAIR_STATE_AUTHENTICATING;
3118 0 : }
3119 :
3120 0 : return rc;
3121 0 : }
3122 :
3123 : static void
3124 0 : nvme_rdma_admin_qpair_abort_aers(struct spdk_nvme_qpair *qpair)
3125 : {
3126 : struct spdk_nvme_rdma_req *rdma_req, *tmp;
3127 : struct spdk_nvme_cpl cpl;
3128 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
3129 :
3130 0 : cpl.status.sc = SPDK_NVME_SC_ABORTED_SQ_DELETION;
3131 0 : cpl.status.sct = SPDK_NVME_SCT_GENERIC;
3132 :
3133 0 : TAILQ_FOREACH_SAFE(rdma_req, &rqpair->outstanding_reqs, link, tmp) {
3134 0 : assert(rdma_req->req != NULL);
3135 :
3136 0 : if (rdma_req->req->cmd.opc != SPDK_NVME_OPC_ASYNC_EVENT_REQUEST) {
3137 0 : continue;
3138 : }
3139 :
3140 0 : nvme_rdma_req_complete(rdma_req, &cpl, false);
3141 0 : }
3142 0 : }
3143 :
3144 : static void
3145 9 : nvme_rdma_poller_destroy(struct nvme_rdma_poller *poller)
3146 : {
3147 9 : if (poller->cq) {
3148 7 : ibv_destroy_cq(poller->cq);
3149 7 : }
3150 9 : if (poller->rsps) {
3151 0 : nvme_rdma_free_rsps(poller->rsps);
3152 0 : }
3153 9 : if (poller->srq) {
3154 0 : spdk_rdma_provider_srq_destroy(poller->srq);
3155 0 : }
3156 9 : if (poller->mr_map) {
3157 0 : spdk_rdma_utils_free_mem_map(&poller->mr_map);
3158 0 : }
3159 9 : if (poller->pd) {
3160 0 : spdk_rdma_utils_put_pd(poller->pd);
3161 0 : }
3162 9 : free(poller);
3163 9 : }
3164 :
3165 : static struct nvme_rdma_poller *
3166 9 : nvme_rdma_poller_create(struct nvme_rdma_poll_group *group, struct ibv_context *ctx)
3167 : {
3168 : struct nvme_rdma_poller *poller;
3169 : struct ibv_device_attr dev_attr;
3170 9 : struct spdk_rdma_provider_srq_init_attr srq_init_attr = {};
3171 : struct nvme_rdma_rsp_opts opts;
3172 : int num_cqe, max_num_cqe;
3173 : int rc;
3174 :
3175 9 : poller = calloc(1, sizeof(*poller));
3176 9 : if (poller == NULL) {
3177 0 : SPDK_ERRLOG("Unable to allocate poller.\n");
3178 0 : return NULL;
3179 : }
3180 :
3181 9 : poller->group = group;
3182 9 : poller->device = ctx;
3183 :
3184 9 : if (g_spdk_nvme_transport_opts.rdma_srq_size != 0) {
3185 0 : rc = ibv_query_device(ctx, &dev_attr);
3186 0 : if (rc) {
3187 0 : SPDK_ERRLOG("Unable to query RDMA device.\n");
3188 0 : goto fail;
3189 : }
3190 :
3191 0 : poller->pd = spdk_rdma_utils_get_pd(ctx);
3192 0 : if (poller->pd == NULL) {
3193 0 : SPDK_ERRLOG("Unable to get PD.\n");
3194 0 : goto fail;
3195 : }
3196 :
3197 0 : poller->mr_map = spdk_rdma_utils_create_mem_map(poller->pd, &g_nvme_hooks,
3198 : IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_READ | IBV_ACCESS_REMOTE_WRITE);
3199 0 : if (poller->mr_map == NULL) {
3200 0 : SPDK_ERRLOG("Unable to create memory map.\n");
3201 0 : goto fail;
3202 : }
3203 :
3204 0 : srq_init_attr.stats = &poller->stats.rdma_stats.recv;
3205 0 : srq_init_attr.pd = poller->pd;
3206 0 : srq_init_attr.srq_init_attr.attr.max_wr = spdk_min((uint32_t)dev_attr.max_srq_wr,
3207 : g_spdk_nvme_transport_opts.rdma_srq_size);
3208 0 : srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(dev_attr.max_sge,
3209 : NVME_RDMA_DEFAULT_RX_SGE);
3210 :
3211 0 : poller->srq = spdk_rdma_provider_srq_create(&srq_init_attr);
3212 0 : if (poller->srq == NULL) {
3213 0 : SPDK_ERRLOG("Unable to create SRQ.\n");
3214 0 : goto fail;
3215 : }
3216 :
3217 0 : opts.num_entries = g_spdk_nvme_transport_opts.rdma_srq_size;
3218 0 : opts.rqpair = NULL;
3219 0 : opts.srq = poller->srq;
3220 0 : opts.mr_map = poller->mr_map;
3221 :
3222 0 : poller->rsps = nvme_rdma_create_rsps(&opts);
3223 0 : if (poller->rsps == NULL) {
3224 0 : SPDK_ERRLOG("Unable to create poller RDMA responses.\n");
3225 0 : goto fail;
3226 : }
3227 :
3228 0 : rc = nvme_rdma_poller_submit_recvs(poller);
3229 0 : if (rc) {
3230 0 : SPDK_ERRLOG("Unable to submit poller RDMA responses.\n");
3231 0 : goto fail;
3232 : }
3233 :
3234 : /*
3235 : * When using an srq, fix the size of the completion queue at startup.
3236 : * The initiator sends only send and recv WRs. Hence, the multiplier is 2.
3237 : * (The target sends also data WRs. Hence, the multiplier is 3.)
3238 : */
3239 0 : num_cqe = g_spdk_nvme_transport_opts.rdma_srq_size * 2;
3240 0 : } else {
3241 9 : num_cqe = DEFAULT_NVME_RDMA_CQ_SIZE;
3242 : }
3243 :
3244 9 : max_num_cqe = g_spdk_nvme_transport_opts.rdma_max_cq_size;
3245 9 : if (max_num_cqe != 0 && num_cqe > max_num_cqe) {
3246 0 : num_cqe = max_num_cqe;
3247 0 : }
3248 :
3249 9 : poller->cq = ibv_create_cq(poller->device, num_cqe, group, NULL, 0);
3250 :
3251 9 : if (poller->cq == NULL) {
3252 2 : SPDK_ERRLOG("Unable to create CQ, errno %d.\n", errno);
3253 2 : goto fail;
3254 : }
3255 :
3256 7 : STAILQ_INSERT_HEAD(&group->pollers, poller, link);
3257 7 : group->num_pollers++;
3258 7 : poller->current_num_wc = num_cqe;
3259 7 : poller->required_num_wc = 0;
3260 7 : return poller;
3261 :
3262 : fail:
3263 2 : nvme_rdma_poller_destroy(poller);
3264 2 : return NULL;
3265 9 : }
3266 :
3267 : static void
3268 3 : nvme_rdma_poll_group_free_pollers(struct nvme_rdma_poll_group *group)
3269 : {
3270 : struct nvme_rdma_poller *poller, *tmp_poller;
3271 :
3272 5 : STAILQ_FOREACH_SAFE(poller, &group->pollers, link, tmp_poller) {
3273 2 : assert(poller->refcnt == 0);
3274 2 : if (poller->refcnt) {
3275 0 : SPDK_WARNLOG("Destroying poller with non-zero ref count: poller %p, refcnt %d\n",
3276 : poller, poller->refcnt);
3277 0 : }
3278 :
3279 2 : STAILQ_REMOVE(&group->pollers, poller, nvme_rdma_poller, link);
3280 2 : nvme_rdma_poller_destroy(poller);
3281 2 : }
3282 3 : }
3283 :
3284 : static struct nvme_rdma_poller *
3285 8 : nvme_rdma_poll_group_get_poller(struct nvme_rdma_poll_group *group, struct ibv_context *device)
3286 : {
3287 8 : struct nvme_rdma_poller *poller = NULL;
3288 :
3289 10 : STAILQ_FOREACH(poller, &group->pollers, link) {
3290 3 : if (poller->device == device) {
3291 1 : break;
3292 : }
3293 2 : }
3294 :
3295 8 : if (!poller) {
3296 7 : poller = nvme_rdma_poller_create(group, device);
3297 7 : if (!poller) {
3298 2 : SPDK_ERRLOG("Failed to create a poller for device %p\n", device);
3299 2 : return NULL;
3300 : }
3301 5 : }
3302 :
3303 6 : poller->refcnt++;
3304 6 : return poller;
3305 8 : }
3306 :
3307 : static void
3308 6 : nvme_rdma_poll_group_put_poller(struct nvme_rdma_poll_group *group, struct nvme_rdma_poller *poller)
3309 : {
3310 6 : assert(poller->refcnt > 0);
3311 6 : if (--poller->refcnt == 0) {
3312 5 : STAILQ_REMOVE(&group->pollers, poller, nvme_rdma_poller, link);
3313 5 : group->num_pollers--;
3314 5 : nvme_rdma_poller_destroy(poller);
3315 5 : }
3316 6 : }
3317 :
3318 : static struct spdk_nvme_transport_poll_group *
3319 1 : nvme_rdma_poll_group_create(void)
3320 : {
3321 : struct nvme_rdma_poll_group *group;
3322 :
3323 1 : group = calloc(1, sizeof(*group));
3324 1 : if (group == NULL) {
3325 0 : SPDK_ERRLOG("Unable to allocate poll group.\n");
3326 0 : return NULL;
3327 : }
3328 :
3329 1 : STAILQ_INIT(&group->pollers);
3330 1 : TAILQ_INIT(&group->connecting_qpairs);
3331 1 : TAILQ_INIT(&group->active_qpairs);
3332 1 : return &group->group;
3333 1 : }
3334 :
3335 : static int
3336 0 : nvme_rdma_poll_group_connect_qpair(struct spdk_nvme_qpair *qpair)
3337 : {
3338 0 : return 0;
3339 : }
3340 :
3341 : static int
3342 0 : nvme_rdma_poll_group_disconnect_qpair(struct spdk_nvme_qpair *qpair)
3343 : {
3344 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
3345 0 : struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group);
3346 :
3347 0 : if (rqpair->link_connecting.tqe_prev) {
3348 0 : TAILQ_REMOVE(&group->connecting_qpairs, rqpair, link_connecting);
3349 : /* We use prev pointer to check if qpair is in connecting list or not .
3350 : * TAILQ_REMOVE doesn't do it. So, we do it manually.
3351 : */
3352 0 : rqpair->link_connecting.tqe_prev = NULL;
3353 0 : }
3354 :
3355 0 : return 0;
3356 : }
3357 :
3358 : static int
3359 0 : nvme_rdma_poll_group_add(struct spdk_nvme_transport_poll_group *tgroup,
3360 : struct spdk_nvme_qpair *qpair)
3361 : {
3362 0 : return 0;
3363 : }
3364 :
3365 : static int
3366 0 : nvme_rdma_poll_group_remove(struct spdk_nvme_transport_poll_group *tgroup,
3367 : struct spdk_nvme_qpair *qpair)
3368 : {
3369 0 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(qpair);
3370 0 : struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(qpair->poll_group);
3371 :
3372 0 : if (rqpair->poller) {
3373 : /* A qpair may skip transport disconnect part if it was already disconnecting. But on RDMA level a qpair
3374 : * may still have a poller reference. In that case we should continue transport disconnect here
3375 : * because a poller depends on the poll group reference which is going to be removed */
3376 0 : SPDK_INFOLOG(nvme, "qpair %p, id %u, nvme state %d, rdma state %d, force disconnect\n",
3377 : qpair, qpair->id, qpair->state, rqpair->state);
3378 0 : nvme_rdma_ctrlr_disconnect_qpair(qpair->ctrlr, qpair);
3379 0 : }
3380 :
3381 0 : if (rqpair->link_active.tqe_prev) {
3382 0 : TAILQ_REMOVE(&group->active_qpairs, rqpair, link_active);
3383 0 : rqpair->link_active.tqe_prev = NULL;
3384 0 : }
3385 :
3386 0 : return 0;
3387 : }
3388 :
3389 : static inline void
3390 0 : nvme_rdma_qpair_process_submits(struct nvme_rdma_poll_group *group,
3391 : struct nvme_rdma_qpair *rqpair)
3392 : {
3393 0 : struct spdk_nvme_qpair *qpair = &rqpair->qpair;
3394 :
3395 0 : assert(rqpair->link_active.tqe_prev != NULL);
3396 :
3397 0 : if (spdk_unlikely(rqpair->state <= NVME_RDMA_QPAIR_STATE_INITIALIZING ||
3398 : rqpair->state >= NVME_RDMA_QPAIR_STATE_EXITING)) {
3399 0 : return;
3400 : }
3401 :
3402 0 : if (spdk_unlikely(qpair->ctrlr->timeout_enabled)) {
3403 0 : nvme_rdma_qpair_check_timeout(qpair);
3404 0 : }
3405 :
3406 0 : nvme_rdma_qpair_submit_sends(rqpair);
3407 0 : if (!rqpair->srq) {
3408 0 : nvme_rdma_qpair_submit_recvs(rqpair);
3409 0 : }
3410 0 : if (rqpair->num_completions > 0) {
3411 0 : nvme_qpair_resubmit_requests(qpair, rqpair->num_completions);
3412 0 : rqpair->num_completions = 0;
3413 0 : }
3414 :
3415 0 : if (rqpair->num_outstanding_reqs == 0 && STAILQ_EMPTY(&qpair->queued_req)) {
3416 0 : TAILQ_REMOVE(&group->active_qpairs, rqpair, link_active);
3417 : /* We use prev pointer to check if qpair is in active list or not.
3418 : * TAILQ_REMOVE doesn't do it. So, we do it manually.
3419 : */
3420 0 : rqpair->link_active.tqe_prev = NULL;
3421 0 : }
3422 0 : }
3423 :
3424 : static int64_t
3425 0 : nvme_rdma_poll_group_process_completions(struct spdk_nvme_transport_poll_group *tgroup,
3426 : uint32_t completions_per_qpair, spdk_nvme_disconnected_qpair_cb disconnected_qpair_cb)
3427 : {
3428 : struct spdk_nvme_qpair *qpair, *tmp_qpair;
3429 : struct nvme_rdma_qpair *rqpair, *tmp_rqpair;
3430 : struct nvme_rdma_poll_group *group;
3431 : struct nvme_rdma_poller *poller;
3432 0 : int batch_size, rc, rc2 = 0;
3433 0 : int64_t total_completions = 0;
3434 0 : uint64_t completions_allowed = 0;
3435 0 : uint64_t completions_per_poller = 0;
3436 0 : uint64_t poller_completions = 0;
3437 : uint64_t rdma_completions;
3438 :
3439 0 : if (completions_per_qpair == 0) {
3440 0 : completions_per_qpair = MAX_COMPLETIONS_PER_POLL;
3441 0 : }
3442 :
3443 0 : group = nvme_rdma_poll_group(tgroup);
3444 :
3445 0 : STAILQ_FOREACH_SAFE(qpair, &tgroup->disconnected_qpairs, poll_group_stailq, tmp_qpair) {
3446 0 : rc = nvme_rdma_ctrlr_disconnect_qpair_poll(qpair->ctrlr, qpair);
3447 0 : if (rc == 0) {
3448 0 : disconnected_qpair_cb(qpair, tgroup->group->ctx);
3449 0 : }
3450 0 : }
3451 :
3452 0 : TAILQ_FOREACH_SAFE(rqpair, &group->connecting_qpairs, link_connecting, tmp_rqpair) {
3453 0 : qpair = &rqpair->qpair;
3454 :
3455 0 : rc = nvme_rdma_ctrlr_connect_qpair_poll(qpair->ctrlr, qpair);
3456 0 : if (rc == 0 || rc != -EAGAIN) {
3457 0 : TAILQ_REMOVE(&group->connecting_qpairs, rqpair, link_connecting);
3458 : /* We use prev pointer to check if qpair is in connecting list or not.
3459 : * TAILQ_REMOVE does not do it. So, we do it manually.
3460 : */
3461 0 : rqpair->link_connecting.tqe_prev = NULL;
3462 :
3463 0 : if (rc == 0) {
3464 : /* Once the connection is completed, we can submit queued requests */
3465 0 : nvme_qpair_resubmit_requests(qpair, rqpair->num_entries);
3466 0 : } else if (rc != -EAGAIN) {
3467 0 : SPDK_ERRLOG("Failed to connect rqpair=%p\n", rqpair);
3468 0 : nvme_rdma_fail_qpair(qpair, 0);
3469 0 : }
3470 0 : }
3471 0 : }
3472 :
3473 0 : STAILQ_FOREACH_SAFE(qpair, &tgroup->connected_qpairs, poll_group_stailq, tmp_qpair) {
3474 0 : rqpair = nvme_rdma_qpair(qpair);
3475 :
3476 0 : if (spdk_likely(nvme_qpair_get_state(qpair) != NVME_QPAIR_CONNECTING)) {
3477 0 : nvme_rdma_qpair_process_cm_event(rqpair);
3478 0 : }
3479 :
3480 0 : if (spdk_unlikely(qpair->transport_failure_reason != SPDK_NVME_QPAIR_FAILURE_NONE)) {
3481 0 : rc2 = -ENXIO;
3482 0 : nvme_rdma_fail_qpair(qpair, 0);
3483 0 : }
3484 0 : }
3485 :
3486 0 : completions_allowed = completions_per_qpair * tgroup->num_connected_qpairs;
3487 0 : if (spdk_likely(group->num_pollers)) {
3488 0 : completions_per_poller = spdk_max(completions_allowed / group->num_pollers, 1);
3489 0 : }
3490 :
3491 0 : STAILQ_FOREACH(poller, &group->pollers, link) {
3492 0 : poller_completions = 0;
3493 0 : rdma_completions = 0;
3494 0 : do {
3495 0 : poller->stats.polls++;
3496 0 : batch_size = spdk_min((completions_per_poller - poller_completions), MAX_COMPLETIONS_PER_POLL);
3497 0 : rc = nvme_rdma_cq_process_completions(poller->cq, batch_size, poller, NULL, &rdma_completions);
3498 0 : if (rc <= 0) {
3499 0 : if (rc == -ECANCELED) {
3500 0 : return -EIO;
3501 0 : } else if (rc == 0) {
3502 0 : poller->stats.idle_polls++;
3503 0 : }
3504 0 : break;
3505 : }
3506 :
3507 0 : poller_completions += rc;
3508 0 : } while (poller_completions < completions_per_poller);
3509 0 : total_completions += poller_completions;
3510 0 : poller->stats.completions += rdma_completions;
3511 0 : if (poller->srq) {
3512 0 : nvme_rdma_poller_submit_recvs(poller);
3513 0 : }
3514 0 : }
3515 :
3516 0 : TAILQ_FOREACH_SAFE(rqpair, &group->active_qpairs, link_active, tmp_rqpair) {
3517 0 : nvme_rdma_qpair_process_submits(group, rqpair);
3518 0 : }
3519 :
3520 0 : return rc2 != 0 ? rc2 : total_completions;
3521 0 : }
3522 :
3523 : /*
3524 : * Handle disconnected qpairs when interrupt support gets added.
3525 : */
3526 : static void
3527 0 : nvme_rdma_poll_group_check_disconnected_qpairs(struct spdk_nvme_transport_poll_group *tgroup,
3528 : spdk_nvme_disconnected_qpair_cb disconnected_qpair_cb)
3529 : {
3530 0 : }
3531 :
3532 : static int
3533 1 : nvme_rdma_poll_group_destroy(struct spdk_nvme_transport_poll_group *tgroup)
3534 : {
3535 1 : struct nvme_rdma_poll_group *group = nvme_rdma_poll_group(tgroup);
3536 :
3537 1 : if (!STAILQ_EMPTY(&tgroup->connected_qpairs) || !STAILQ_EMPTY(&tgroup->disconnected_qpairs)) {
3538 0 : return -EBUSY;
3539 : }
3540 :
3541 1 : nvme_rdma_poll_group_free_pollers(group);
3542 1 : free(group);
3543 :
3544 1 : return 0;
3545 1 : }
3546 :
3547 : static int
3548 3 : nvme_rdma_poll_group_get_stats(struct spdk_nvme_transport_poll_group *tgroup,
3549 : struct spdk_nvme_transport_poll_group_stat **_stats)
3550 : {
3551 : struct nvme_rdma_poll_group *group;
3552 : struct spdk_nvme_transport_poll_group_stat *stats;
3553 : struct spdk_nvme_rdma_device_stat *device_stat;
3554 : struct nvme_rdma_poller *poller;
3555 3 : uint32_t i = 0;
3556 :
3557 3 : if (tgroup == NULL || _stats == NULL) {
3558 2 : SPDK_ERRLOG("Invalid stats or group pointer\n");
3559 2 : return -EINVAL;
3560 : }
3561 :
3562 1 : group = nvme_rdma_poll_group(tgroup);
3563 1 : stats = calloc(1, sizeof(*stats));
3564 1 : if (!stats) {
3565 0 : SPDK_ERRLOG("Can't allocate memory for RDMA stats\n");
3566 0 : return -ENOMEM;
3567 : }
3568 1 : stats->trtype = SPDK_NVME_TRANSPORT_RDMA;
3569 1 : stats->rdma.num_devices = group->num_pollers;
3570 :
3571 1 : if (stats->rdma.num_devices == 0) {
3572 0 : *_stats = stats;
3573 0 : return 0;
3574 : }
3575 :
3576 1 : stats->rdma.device_stats = calloc(stats->rdma.num_devices, sizeof(*stats->rdma.device_stats));
3577 1 : if (!stats->rdma.device_stats) {
3578 0 : SPDK_ERRLOG("Can't allocate memory for RDMA device stats\n");
3579 0 : free(stats);
3580 0 : return -ENOMEM;
3581 : }
3582 :
3583 3 : STAILQ_FOREACH(poller, &group->pollers, link) {
3584 2 : device_stat = &stats->rdma.device_stats[i];
3585 2 : device_stat->name = poller->device->device->name;
3586 2 : device_stat->polls = poller->stats.polls;
3587 2 : device_stat->idle_polls = poller->stats.idle_polls;
3588 2 : device_stat->completions = poller->stats.completions;
3589 2 : device_stat->queued_requests = poller->stats.queued_requests;
3590 2 : device_stat->total_send_wrs = poller->stats.rdma_stats.send.num_submitted_wrs;
3591 2 : device_stat->send_doorbell_updates = poller->stats.rdma_stats.send.doorbell_updates;
3592 2 : device_stat->total_recv_wrs = poller->stats.rdma_stats.recv.num_submitted_wrs;
3593 2 : device_stat->recv_doorbell_updates = poller->stats.rdma_stats.recv.doorbell_updates;
3594 2 : i++;
3595 2 : }
3596 :
3597 1 : *_stats = stats;
3598 :
3599 1 : return 0;
3600 3 : }
3601 :
3602 : static void
3603 1 : nvme_rdma_poll_group_free_stats(struct spdk_nvme_transport_poll_group *tgroup,
3604 : struct spdk_nvme_transport_poll_group_stat *stats)
3605 : {
3606 1 : if (stats) {
3607 1 : free(stats->rdma.device_stats);
3608 1 : }
3609 1 : free(stats);
3610 1 : }
3611 :
3612 : static int
3613 4 : nvme_rdma_ctrlr_get_memory_domains(const struct spdk_nvme_ctrlr *ctrlr,
3614 : struct spdk_memory_domain **domains, int array_size)
3615 : {
3616 4 : struct nvme_rdma_qpair *rqpair = nvme_rdma_qpair(ctrlr->adminq);
3617 :
3618 4 : if (domains && array_size > 0) {
3619 1 : domains[0] = rqpair->rdma_qp->domain;
3620 1 : }
3621 :
3622 4 : return 1;
3623 : }
3624 :
3625 : void
3626 0 : spdk_nvme_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks)
3627 : {
3628 0 : g_nvme_hooks = *hooks;
3629 0 : }
3630 :
3631 : const struct spdk_nvme_transport_ops rdma_ops = {
3632 : .name = "RDMA",
3633 : .type = SPDK_NVME_TRANSPORT_RDMA,
3634 : .ctrlr_construct = nvme_rdma_ctrlr_construct,
3635 : .ctrlr_scan = nvme_fabric_ctrlr_scan,
3636 : .ctrlr_destruct = nvme_rdma_ctrlr_destruct,
3637 : .ctrlr_enable = nvme_rdma_ctrlr_enable,
3638 :
3639 : .ctrlr_set_reg_4 = nvme_fabric_ctrlr_set_reg_4,
3640 : .ctrlr_set_reg_8 = nvme_fabric_ctrlr_set_reg_8,
3641 : .ctrlr_get_reg_4 = nvme_fabric_ctrlr_get_reg_4,
3642 : .ctrlr_get_reg_8 = nvme_fabric_ctrlr_get_reg_8,
3643 : .ctrlr_set_reg_4_async = nvme_fabric_ctrlr_set_reg_4_async,
3644 : .ctrlr_set_reg_8_async = nvme_fabric_ctrlr_set_reg_8_async,
3645 : .ctrlr_get_reg_4_async = nvme_fabric_ctrlr_get_reg_4_async,
3646 : .ctrlr_get_reg_8_async = nvme_fabric_ctrlr_get_reg_8_async,
3647 :
3648 : .ctrlr_get_max_xfer_size = nvme_rdma_ctrlr_get_max_xfer_size,
3649 : .ctrlr_get_max_sges = nvme_rdma_ctrlr_get_max_sges,
3650 :
3651 : .ctrlr_create_io_qpair = nvme_rdma_ctrlr_create_io_qpair,
3652 : .ctrlr_delete_io_qpair = nvme_rdma_ctrlr_delete_io_qpair,
3653 : .ctrlr_connect_qpair = nvme_rdma_ctrlr_connect_qpair,
3654 : .ctrlr_disconnect_qpair = nvme_rdma_ctrlr_disconnect_qpair,
3655 :
3656 : .ctrlr_get_memory_domains = nvme_rdma_ctrlr_get_memory_domains,
3657 :
3658 : .qpair_abort_reqs = nvme_rdma_qpair_abort_reqs,
3659 : .qpair_reset = nvme_rdma_qpair_reset,
3660 : .qpair_submit_request = nvme_rdma_qpair_submit_request,
3661 : .qpair_process_completions = nvme_rdma_qpair_process_completions,
3662 : .qpair_iterate_requests = nvme_rdma_qpair_iterate_requests,
3663 : .qpair_authenticate = nvme_rdma_qpair_authenticate,
3664 : .admin_qpair_abort_aers = nvme_rdma_admin_qpair_abort_aers,
3665 :
3666 : .poll_group_create = nvme_rdma_poll_group_create,
3667 : .poll_group_connect_qpair = nvme_rdma_poll_group_connect_qpair,
3668 : .poll_group_disconnect_qpair = nvme_rdma_poll_group_disconnect_qpair,
3669 : .poll_group_add = nvme_rdma_poll_group_add,
3670 : .poll_group_remove = nvme_rdma_poll_group_remove,
3671 : .poll_group_process_completions = nvme_rdma_poll_group_process_completions,
3672 : .poll_group_check_disconnected_qpairs = nvme_rdma_poll_group_check_disconnected_qpairs,
3673 : .poll_group_destroy = nvme_rdma_poll_group_destroy,
3674 : .poll_group_get_stats = nvme_rdma_poll_group_get_stats,
3675 : .poll_group_free_stats = nvme_rdma_poll_group_free_stats,
3676 : };
3677 :
3678 1 : SPDK_NVME_TRANSPORT_REGISTER(rdma, &rdma_ops);
|
