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