Branch data Line data Source code
1 : : /* SPDX-License-Identifier: BSD-3-Clause
2 : : * Copyright (C) 2016 Intel Corporation. All rights reserved.
3 : : * Copyright (c) 2019-2021 Mellanox Technologies LTD. All rights reserved.
4 : : * Copyright (c) 2021-2024 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
5 : : */
6 : :
7 : : #include "spdk/stdinc.h"
8 : :
9 : : #include "spdk/config.h"
10 : : #include "spdk/thread.h"
11 : : #include "spdk/likely.h"
12 : : #include "spdk/nvmf_transport.h"
13 : : #include "spdk/string.h"
14 : : #include "spdk/trace.h"
15 : : #include "spdk/tree.h"
16 : : #include "spdk/util.h"
17 : :
18 : : #include "spdk_internal/assert.h"
19 : : #include "spdk/log.h"
20 : : #include "spdk_internal/rdma_provider.h"
21 : : #include "spdk_internal/rdma_utils.h"
22 : :
23 : : #include "nvmf_internal.h"
24 : : #include "transport.h"
25 : :
26 : : #include "spdk_internal/trace_defs.h"
27 : :
28 : : struct spdk_nvme_rdma_hooks g_nvmf_hooks = {};
29 : : const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma;
30 : :
31 : : /*
32 : : RDMA Connection Resource Defaults
33 : : */
34 : : #define NVMF_DEFAULT_MSDBD 16
35 : : #define NVMF_DEFAULT_TX_SGE SPDK_NVMF_MAX_SGL_ENTRIES
36 : : #define NVMF_DEFAULT_RSP_SGE 1
37 : : #define NVMF_DEFAULT_RX_SGE 2
38 : :
39 : : #define NVMF_RDMA_MAX_EVENTS_PER_POLL 32
40 : :
41 : : SPDK_STATIC_ASSERT(NVMF_DEFAULT_MSDBD <= SPDK_NVMF_MAX_SGL_ENTRIES,
42 : : "MSDBD must not exceed SPDK_NVMF_MAX_SGL_ENTRIES");
43 : :
44 : : /* The RDMA completion queue size */
45 : : #define DEFAULT_NVMF_RDMA_CQ_SIZE 4096
46 : : #define MAX_WR_PER_QP(queue_depth) (queue_depth * 3 + 2)
47 : :
48 : : enum spdk_nvmf_rdma_request_state {
49 : : /* The request is not currently in use */
50 : : RDMA_REQUEST_STATE_FREE = 0,
51 : :
52 : : /* Initial state when request first received */
53 : : RDMA_REQUEST_STATE_NEW,
54 : :
55 : : /* The request is queued until a data buffer is available. */
56 : : RDMA_REQUEST_STATE_NEED_BUFFER,
57 : :
58 : : /* The request is waiting on RDMA queue depth availability
59 : : * to transfer data from the host to the controller.
60 : : */
61 : : RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING,
62 : :
63 : : /* The request is currently transferring data from the host to the controller. */
64 : : RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER,
65 : :
66 : : /* The request is ready to execute at the block device */
67 : : RDMA_REQUEST_STATE_READY_TO_EXECUTE,
68 : :
69 : : /* The request is currently executing at the block device */
70 : : RDMA_REQUEST_STATE_EXECUTING,
71 : :
72 : : /* The request finished executing at the block device */
73 : : RDMA_REQUEST_STATE_EXECUTED,
74 : :
75 : : /* The request is waiting on RDMA queue depth availability
76 : : * to transfer data from the controller to the host.
77 : : */
78 : : RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING,
79 : :
80 : : /* The request is waiting on RDMA queue depth availability
81 : : * to send response to the host.
82 : : */
83 : : RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING,
84 : :
85 : : /* The request is ready to send a completion */
86 : : RDMA_REQUEST_STATE_READY_TO_COMPLETE,
87 : :
88 : : /* The request is currently transferring data from the controller to the host. */
89 : : RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST,
90 : :
91 : : /* The request currently has an outstanding completion without an
92 : : * associated data transfer.
93 : : */
94 : : RDMA_REQUEST_STATE_COMPLETING,
95 : :
96 : : /* The request completed and can be marked free. */
97 : : RDMA_REQUEST_STATE_COMPLETED,
98 : :
99 : : /* Terminator */
100 : : RDMA_REQUEST_NUM_STATES,
101 : : };
102 : :
103 : : static void
104 : 914 : nvmf_trace(void)
105 : : {
106 : 914 : spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r');
107 : :
108 : 914 : struct spdk_trace_tpoint_opts opts[] = {
109 : : {
110 : : "RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW,
111 : : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 1,
112 : : {
113 : : { "qpair", SPDK_TRACE_ARG_TYPE_PTR, 8 },
114 : : { "qd", SPDK_TRACE_ARG_TYPE_INT, 4 }
115 : : }
116 : : },
117 : : {
118 : : "RDMA_REQ_COMPLETED", TRACE_RDMA_REQUEST_STATE_COMPLETED,
119 : : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
120 : : {
121 : : { "qpair", SPDK_TRACE_ARG_TYPE_PTR, 8 },
122 : : { "qd", SPDK_TRACE_ARG_TYPE_INT, 4 }
123 : : }
124 : : },
125 : : };
126 : :
127 : 914 : spdk_trace_register_description_ext(opts, SPDK_COUNTOF(opts));
128 : 914 : spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER,
129 : : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
130 : : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
131 : 914 : spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H",
132 : : TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING,
133 : : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
134 : : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
135 : 914 : spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C",
136 : : TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING,
137 : : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
138 : : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
139 : 914 : spdk_trace_register_description("RDMA_REQ_TX_H2C",
140 : : TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER,
141 : : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
142 : : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
143 : 914 : spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE",
144 : : TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE,
145 : : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
146 : : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
147 : 914 : spdk_trace_register_description("RDMA_REQ_EXECUTING",
148 : : TRACE_RDMA_REQUEST_STATE_EXECUTING,
149 : : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
150 : : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
151 : 914 : spdk_trace_register_description("RDMA_REQ_EXECUTED",
152 : : TRACE_RDMA_REQUEST_STATE_EXECUTED,
153 : : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
154 : : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
155 : 914 : spdk_trace_register_description("RDMA_REQ_RDY2COMPL_PEND",
156 : : TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING,
157 : : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
158 : : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
159 : 914 : spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL",
160 : : TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE,
161 : : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
162 : : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
163 : 914 : spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H",
164 : : TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST,
165 : : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
166 : : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
167 : 914 : spdk_trace_register_description("RDMA_REQ_COMPLETING",
168 : : TRACE_RDMA_REQUEST_STATE_COMPLETING,
169 : : OWNER_TYPE_NONE, OBJECT_NVMF_RDMA_IO, 0,
170 : : SPDK_TRACE_ARG_TYPE_PTR, "qpair");
171 : :
172 : 914 : spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE,
173 : : OWNER_TYPE_NONE, OBJECT_NONE, 0,
174 : : SPDK_TRACE_ARG_TYPE_INT, "");
175 : 914 : spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT,
176 : : OWNER_TYPE_NONE, OBJECT_NONE, 0,
177 : : SPDK_TRACE_ARG_TYPE_INT, "type");
178 : 914 : spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT,
179 : : OWNER_TYPE_NONE, OBJECT_NONE, 0,
180 : : SPDK_TRACE_ARG_TYPE_INT, "type");
181 : 914 : spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT,
182 : : OWNER_TYPE_NONE, OBJECT_NONE, 0,
183 : : SPDK_TRACE_ARG_TYPE_INT, "");
184 : 914 : spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY,
185 : : OWNER_TYPE_NONE, OBJECT_NONE, 0,
186 : : SPDK_TRACE_ARG_TYPE_INT, "");
187 : :
188 : 914 : spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_START, OBJECT_NVMF_RDMA_IO, 1);
189 : 914 : spdk_trace_tpoint_register_relation(TRACE_BDEV_IO_DONE, OBJECT_NVMF_RDMA_IO, 0);
190 : 914 : }
191 : 987 : SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA)
192 : :
193 : : enum spdk_nvmf_rdma_wr_type {
194 : : RDMA_WR_TYPE_RECV,
195 : : RDMA_WR_TYPE_SEND,
196 : : RDMA_WR_TYPE_DATA,
197 : : };
198 : :
199 : : struct spdk_nvmf_rdma_wr {
200 : : /* Uses enum spdk_nvmf_rdma_wr_type */
201 : : uint8_t type;
202 : : };
203 : :
204 : : /* This structure holds commands as they are received off the wire.
205 : : * It must be dynamically paired with a full request object
206 : : * (spdk_nvmf_rdma_request) to service a request. It is separate
207 : : * from the request because RDMA does not appear to order
208 : : * completions, so occasionally we'll get a new incoming
209 : : * command when there aren't any free request objects.
210 : : */
211 : : struct spdk_nvmf_rdma_recv {
212 : : struct ibv_recv_wr wr;
213 : : struct ibv_sge sgl[NVMF_DEFAULT_RX_SGE];
214 : :
215 : : struct spdk_nvmf_rdma_qpair *qpair;
216 : :
217 : : /* In-capsule data buffer */
218 : : uint8_t *buf;
219 : :
220 : : struct spdk_nvmf_rdma_wr rdma_wr;
221 : : uint64_t receive_tsc;
222 : :
223 : : STAILQ_ENTRY(spdk_nvmf_rdma_recv) link;
224 : : };
225 : :
226 : : struct spdk_nvmf_rdma_request_data {
227 : : struct ibv_send_wr wr;
228 : : struct ibv_sge sgl[SPDK_NVMF_MAX_SGL_ENTRIES];
229 : : };
230 : :
231 : : struct spdk_nvmf_rdma_request {
232 : : struct spdk_nvmf_request req;
233 : :
234 : : bool fused_failed;
235 : :
236 : : struct spdk_nvmf_rdma_wr data_wr;
237 : : struct spdk_nvmf_rdma_wr rsp_wr;
238 : :
239 : : /* Uses enum spdk_nvmf_rdma_request_state */
240 : : uint8_t state;
241 : :
242 : : /* Data offset in req.iov */
243 : : uint32_t offset;
244 : :
245 : : struct spdk_nvmf_rdma_recv *recv;
246 : :
247 : : struct {
248 : : struct ibv_send_wr wr;
249 : : struct ibv_sge sgl[NVMF_DEFAULT_RSP_SGE];
250 : : } rsp;
251 : :
252 : : uint16_t iovpos;
253 : : uint16_t num_outstanding_data_wr;
254 : : /* Used to split Write IO with multi SGL payload */
255 : : uint16_t num_remaining_data_wr;
256 : : uint64_t receive_tsc;
257 : : struct spdk_nvmf_rdma_request *fused_pair;
258 : : STAILQ_ENTRY(spdk_nvmf_rdma_request) state_link;
259 : : struct ibv_send_wr *remaining_tranfer_in_wrs;
260 : : struct ibv_send_wr *transfer_wr;
261 : : struct spdk_nvmf_rdma_request_data data;
262 : : };
263 : :
264 : : struct spdk_nvmf_rdma_resource_opts {
265 : : struct spdk_nvmf_rdma_qpair *qpair;
266 : : /* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */
267 : : void *qp;
268 : : struct spdk_rdma_utils_mem_map *map;
269 : : uint32_t max_queue_depth;
270 : : uint32_t in_capsule_data_size;
271 : : bool shared;
272 : : };
273 : :
274 : : struct spdk_nvmf_rdma_resources {
275 : : /* Array of size "max_queue_depth" containing RDMA requests. */
276 : : struct spdk_nvmf_rdma_request *reqs;
277 : :
278 : : /* Array of size "max_queue_depth" containing RDMA recvs. */
279 : : struct spdk_nvmf_rdma_recv *recvs;
280 : :
281 : : /* Array of size "max_queue_depth" containing 64 byte capsules
282 : : * used for receive.
283 : : */
284 : : union nvmf_h2c_msg *cmds;
285 : :
286 : : /* Array of size "max_queue_depth" containing 16 byte completions
287 : : * to be sent back to the user.
288 : : */
289 : : union nvmf_c2h_msg *cpls;
290 : :
291 : : /* Array of size "max_queue_depth * InCapsuleDataSize" containing
292 : : * buffers to be used for in capsule data.
293 : : */
294 : : void *bufs;
295 : :
296 : : /* Receives that are waiting for a request object */
297 : : STAILQ_HEAD(, spdk_nvmf_rdma_recv) incoming_queue;
298 : :
299 : : /* Queue to track free requests */
300 : : STAILQ_HEAD(, spdk_nvmf_rdma_request) free_queue;
301 : : };
302 : :
303 : : typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair);
304 : :
305 : : typedef void (*spdk_poller_destroy_cb)(void *ctx);
306 : :
307 : : struct spdk_nvmf_rdma_ibv_event_ctx {
308 : : struct spdk_nvmf_rdma_qpair *rqpair;
309 : : };
310 : :
311 : : struct spdk_nvmf_rdma_qpair {
312 : : struct spdk_nvmf_qpair qpair;
313 : :
314 : : struct spdk_nvmf_rdma_device *device;
315 : : struct spdk_nvmf_rdma_poller *poller;
316 : :
317 : : struct spdk_rdma_provider_qp *rdma_qp;
318 : : struct rdma_cm_id *cm_id;
319 : : struct spdk_rdma_provider_srq *srq;
320 : : struct rdma_cm_id *listen_id;
321 : :
322 : : /* Cache the QP number to improve QP search by RB tree. */
323 : : uint32_t qp_num;
324 : :
325 : : /* The maximum number of I/O outstanding on this connection at one time */
326 : : uint16_t max_queue_depth;
327 : :
328 : : /* The maximum number of active RDMA READ and ATOMIC operations at one time */
329 : : uint16_t max_read_depth;
330 : :
331 : : /* The maximum number of RDMA SEND operations at one time */
332 : : uint32_t max_send_depth;
333 : :
334 : : /* The current number of outstanding WRs from this qpair's
335 : : * recv queue. Should not exceed device->attr.max_queue_depth.
336 : : */
337 : : uint16_t current_recv_depth;
338 : :
339 : : /* The current number of active RDMA READ operations */
340 : : uint16_t current_read_depth;
341 : :
342 : : /* The current number of posted WRs from this qpair's
343 : : * send queue. Should not exceed max_send_depth.
344 : : */
345 : : uint32_t current_send_depth;
346 : :
347 : : /* The maximum number of SGEs per WR on the send queue */
348 : : uint32_t max_send_sge;
349 : :
350 : : /* The maximum number of SGEs per WR on the recv queue */
351 : : uint32_t max_recv_sge;
352 : :
353 : : struct spdk_nvmf_rdma_resources *resources;
354 : :
355 : : STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_read_queue;
356 : :
357 : : STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_write_queue;
358 : :
359 : : STAILQ_HEAD(, spdk_nvmf_rdma_request) pending_rdma_send_queue;
360 : :
361 : : /* Number of requests not in the free state */
362 : : uint32_t qd;
363 : :
364 : : bool ibv_in_error_state;
365 : :
366 : : RB_ENTRY(spdk_nvmf_rdma_qpair) node;
367 : :
368 : : STAILQ_ENTRY(spdk_nvmf_rdma_qpair) recv_link;
369 : :
370 : : STAILQ_ENTRY(spdk_nvmf_rdma_qpair) send_link;
371 : :
372 : : /* Points to the a request that has fuse bits set to
373 : : * SPDK_NVME_CMD_FUSE_FIRST, when the qpair is waiting
374 : : * for the request that has SPDK_NVME_CMD_FUSE_SECOND.
375 : : */
376 : : struct spdk_nvmf_rdma_request *fused_first;
377 : :
378 : : /*
379 : : * io_channel which is used to destroy qpair when it is removed from poll group
380 : : */
381 : : struct spdk_io_channel *destruct_channel;
382 : :
383 : : /* ctx for async processing of last_wqe_reached event */
384 : : struct spdk_nvmf_rdma_ibv_event_ctx *last_wqe_reached_ctx;
385 : :
386 : : /* Lets us know that we have received the last_wqe event. */
387 : : bool last_wqe_reached;
388 : :
389 : : /* Indicate that nvmf_rdma_close_qpair is called */
390 : : bool to_close;
391 : : };
392 : :
393 : : struct spdk_nvmf_rdma_poller_stat {
394 : : uint64_t completions;
395 : : uint64_t polls;
396 : : uint64_t idle_polls;
397 : : uint64_t requests;
398 : : uint64_t request_latency;
399 : : uint64_t pending_free_request;
400 : : uint64_t pending_rdma_read;
401 : : uint64_t pending_rdma_write;
402 : : uint64_t pending_rdma_send;
403 : : struct spdk_rdma_provider_qp_stats qp_stats;
404 : : };
405 : :
406 : : struct spdk_nvmf_rdma_poller {
407 : : struct spdk_nvmf_rdma_device *device;
408 : : struct spdk_nvmf_rdma_poll_group *group;
409 : :
410 : : int num_cqe;
411 : : int required_num_wr;
412 : : struct ibv_cq *cq;
413 : :
414 : : /* The maximum number of I/O outstanding on the shared receive queue at one time */
415 : : uint16_t max_srq_depth;
416 : : bool need_destroy;
417 : :
418 : : /* Shared receive queue */
419 : : struct spdk_rdma_provider_srq *srq;
420 : :
421 : : struct spdk_nvmf_rdma_resources *resources;
422 : : struct spdk_nvmf_rdma_poller_stat stat;
423 : :
424 : : spdk_poller_destroy_cb destroy_cb;
425 : : void *destroy_cb_ctx;
426 : :
427 : : RB_HEAD(qpairs_tree, spdk_nvmf_rdma_qpair) qpairs;
428 : :
429 : : STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_recv;
430 : :
431 : : STAILQ_HEAD(, spdk_nvmf_rdma_qpair) qpairs_pending_send;
432 : :
433 : : TAILQ_ENTRY(spdk_nvmf_rdma_poller) link;
434 : : };
435 : :
436 : : struct spdk_nvmf_rdma_poll_group_stat {
437 : : uint64_t pending_data_buffer;
438 : : };
439 : :
440 : : struct spdk_nvmf_rdma_poll_group {
441 : : struct spdk_nvmf_transport_poll_group group;
442 : : struct spdk_nvmf_rdma_poll_group_stat stat;
443 : : TAILQ_HEAD(, spdk_nvmf_rdma_poller) pollers;
444 : : TAILQ_ENTRY(spdk_nvmf_rdma_poll_group) link;
445 : : };
446 : :
447 : : struct spdk_nvmf_rdma_conn_sched {
448 : : struct spdk_nvmf_rdma_poll_group *next_admin_pg;
449 : : struct spdk_nvmf_rdma_poll_group *next_io_pg;
450 : : };
451 : :
452 : : /* Assuming rdma_cm uses just one protection domain per ibv_context. */
453 : : struct spdk_nvmf_rdma_device {
454 : : struct ibv_device_attr attr;
455 : : struct ibv_context *context;
456 : :
457 : : struct spdk_rdma_utils_mem_map *map;
458 : : struct ibv_pd *pd;
459 : :
460 : : int num_srq;
461 : : bool need_destroy;
462 : : bool ready_to_destroy;
463 : : bool is_ready;
464 : :
465 : : TAILQ_ENTRY(spdk_nvmf_rdma_device) link;
466 : : };
467 : :
468 : : struct spdk_nvmf_rdma_port {
469 : : const struct spdk_nvme_transport_id *trid;
470 : : struct rdma_cm_id *id;
471 : : struct spdk_nvmf_rdma_device *device;
472 : : TAILQ_ENTRY(spdk_nvmf_rdma_port) link;
473 : : };
474 : :
475 : : struct rdma_transport_opts {
476 : : int num_cqe;
477 : : uint32_t max_srq_depth;
478 : : bool no_srq;
479 : : bool no_wr_batching;
480 : : int acceptor_backlog;
481 : : };
482 : :
483 : : struct spdk_nvmf_rdma_transport {
484 : : struct spdk_nvmf_transport transport;
485 : : struct rdma_transport_opts rdma_opts;
486 : :
487 : : struct spdk_nvmf_rdma_conn_sched conn_sched;
488 : :
489 : : struct rdma_event_channel *event_channel;
490 : :
491 : : struct spdk_mempool *data_wr_pool;
492 : :
493 : : struct spdk_poller *accept_poller;
494 : :
495 : : /* fields used to poll RDMA/IB events */
496 : : nfds_t npoll_fds;
497 : : struct pollfd *poll_fds;
498 : :
499 : : TAILQ_HEAD(, spdk_nvmf_rdma_device) devices;
500 : : TAILQ_HEAD(, spdk_nvmf_rdma_port) ports;
501 : : TAILQ_HEAD(, spdk_nvmf_rdma_poll_group) poll_groups;
502 : :
503 : : /* ports that are removed unexpectedly and need retry listen */
504 : : TAILQ_HEAD(, spdk_nvmf_rdma_port) retry_ports;
505 : : };
506 : :
507 : : struct poller_manage_ctx {
508 : : struct spdk_nvmf_rdma_transport *rtransport;
509 : : struct spdk_nvmf_rdma_poll_group *rgroup;
510 : : struct spdk_nvmf_rdma_poller *rpoller;
511 : : struct spdk_nvmf_rdma_device *device;
512 : :
513 : : struct spdk_thread *thread;
514 : : volatile int *inflight_op_counter;
515 : : };
516 : :
517 : : static const struct spdk_json_object_decoder rdma_transport_opts_decoder[] = {
518 : : {
519 : : "num_cqe", offsetof(struct rdma_transport_opts, num_cqe),
520 : : spdk_json_decode_int32, true
521 : : },
522 : : {
523 : : "max_srq_depth", offsetof(struct rdma_transport_opts, max_srq_depth),
524 : : spdk_json_decode_uint32, true
525 : : },
526 : : {
527 : : "no_srq", offsetof(struct rdma_transport_opts, no_srq),
528 : : spdk_json_decode_bool, true
529 : : },
530 : : {
531 : : "no_wr_batching", offsetof(struct rdma_transport_opts, no_wr_batching),
532 : : spdk_json_decode_bool, true
533 : : },
534 : : {
535 : : "acceptor_backlog", offsetof(struct rdma_transport_opts, acceptor_backlog),
536 : : spdk_json_decode_int32, true
537 : : },
538 : : };
539 : :
540 : : static int
541 : 16646663 : nvmf_rdma_qpair_compare(struct spdk_nvmf_rdma_qpair *rqpair1, struct spdk_nvmf_rdma_qpair *rqpair2)
542 : : {
543 [ + + # # : 16646663 : return rqpair1->qp_num < rqpair2->qp_num ? -1 : rqpair1->qp_num > rqpair2->qp_num;
# # # # #
# # # # #
# # # # ]
544 : : }
545 : :
546 [ + + + + : 16658728 : RB_GENERATE_STATIC(qpairs_tree, spdk_nvmf_rdma_qpair, node, nvmf_rdma_qpair_compare);
+ + + + +
+ + + + +
+ + + + -
+ - + - +
+ + + + +
+ + + + -
+ + + + -
+ + + + +
- + + + +
- - + - +
+ + + + +
- + + # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # ]
547 : :
548 : : static bool nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport,
549 : : struct spdk_nvmf_rdma_request *rdma_req);
550 : :
551 : : static void _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport,
552 : : struct spdk_nvmf_rdma_poller *rpoller);
553 : :
554 : : static void _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport,
555 : : struct spdk_nvmf_rdma_poller *rpoller);
556 : :
557 : : static void _nvmf_rdma_remove_destroyed_device(void *c);
558 : :
559 : : static inline enum spdk_nvme_media_error_status_code
560 : 0 : nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) {
561 : : enum spdk_nvme_media_error_status_code result;
562 [ # # # # ]: 0 : switch (err_type)
563 : : {
564 : 0 : case SPDK_DIF_REFTAG_ERROR:
565 : 0 : result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR;
566 : 0 : break;
567 : 0 : case SPDK_DIF_APPTAG_ERROR:
568 : 0 : result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR;
569 : 0 : break;
570 : 0 : case SPDK_DIF_GUARD_ERROR:
571 : 0 : result = SPDK_NVME_SC_GUARD_CHECK_ERROR;
572 : 0 : break;
573 : 0 : default:
574 [ # # ]: 0 : SPDK_UNREACHABLE();
575 : : }
576 : :
577 : 0 : return result;
578 : : }
579 : :
580 : : /*
581 : : * Return data_wrs to pool starting from \b data_wr
582 : : * Request's own response and data WR are excluded
583 : : */
584 : : static void
585 : 7990372 : _nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req,
586 : : struct ibv_send_wr *data_wr,
587 : : struct spdk_mempool *pool)
588 : : {
589 : 21 : struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES];
590 : : struct spdk_nvmf_rdma_request_data *nvmf_data;
591 : : struct ibv_send_wr *next_send_wr;
592 [ # # ]: 7990372 : uint64_t req_wrid = (uint64_t)&rdma_req->data_wr;
593 : 7990372 : uint32_t num_wrs = 0;
594 : :
595 [ + + + + : 17186910 : while (data_wr && data_wr->wr_id == req_wrid) {
# # # # ]
596 : 9196538 : nvmf_data = SPDK_CONTAINEROF(data_wr, struct spdk_nvmf_rdma_request_data, wr);
597 [ - + # # : 9196538 : memset(nvmf_data->sgl, 0, sizeof(data_wr->sg_list[0]) * data_wr->num_sge);
# # # # ]
598 [ # # # # ]: 9196538 : data_wr->num_sge = 0;
599 [ # # # # ]: 9196538 : next_send_wr = data_wr->next;
600 [ + + # # : 9196538 : if (data_wr != &rdma_req->data.wr) {
# # ]
601 [ # # # # ]: 1272421 : data_wr->next = NULL;
602 [ - + # # ]: 1272421 : assert(num_wrs < SPDK_NVMF_MAX_SGL_ENTRIES);
603 [ # # # # : 1272421 : work_requests[num_wrs] = nvmf_data;
# # ]
604 : 1272421 : num_wrs++;
605 : 0 : }
606 [ + + + + : 9196538 : data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL : next_send_wr;
# # # # ]
607 : : }
608 : :
609 [ + + ]: 7990372 : if (num_wrs) {
610 : 200310 : spdk_mempool_put_bulk(pool, (void **) work_requests, num_wrs);
611 : 0 : }
612 : 7990372 : }
613 : :
614 : : static void
615 : 7924534 : nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req,
616 : : struct spdk_nvmf_rdma_transport *rtransport)
617 : : {
618 [ # # # # ]: 7924534 : rdma_req->num_outstanding_data_wr = 0;
619 : :
620 [ # # # # : 7924534 : _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool);
# # # # ]
621 : :
622 [ - + # # : 7924534 : if (rdma_req->remaining_tranfer_in_wrs) {
# # ]
623 [ # # # # ]: 0 : _nvmf_rdma_request_free_data(rdma_req, rdma_req->remaining_tranfer_in_wrs,
624 [ # # # # ]: 0 : rtransport->data_wr_pool);
625 [ # # # # ]: 0 : rdma_req->remaining_tranfer_in_wrs = NULL;
626 : 0 : }
627 : :
628 [ # # # # : 7924534 : rdma_req->data.wr.next = NULL;
# # # # ]
629 [ # # # # : 7924534 : rdma_req->rsp.wr.next = NULL;
# # # # ]
630 : 7924534 : }
631 : :
632 : : static void
633 : 0 : nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req)
634 : : {
635 [ # # # # : 0 : SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool);
# # # # ]
636 [ # # # # : 0 : if (req->req.cmd) {
# # # # ]
637 [ # # # # : 0 : SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode);
# # # # #
# # # ]
638 : 0 : }
639 [ # # # # : 0 : if (req->recv) {
# # ]
640 [ # # # # : 0 : SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id);
# # # # #
# ]
641 : 0 : }
642 : 0 : }
643 : :
644 : : static void
645 : 0 : nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair)
646 : : {
647 : : int i;
648 : :
649 [ # # # # : 0 : SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid);
# # ]
650 [ # # # # : 0 : for (i = 0; i < rqpair->max_queue_depth; i++) {
# # # # ]
651 [ # # # # : 0 : if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) {
# # # # #
# # # # #
# # ]
652 [ # # # # : 0 : nvmf_rdma_dump_request(&rqpair->resources->reqs[i]);
# # # # #
# ]
653 : 0 : }
654 : 0 : }
655 : 0 : }
656 : :
657 : : static void
658 : 285 : nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources)
659 : : {
660 [ # # # # ]: 285 : spdk_free(resources->cmds);
661 [ # # # # ]: 285 : spdk_free(resources->cpls);
662 [ # # # # ]: 285 : spdk_free(resources->bufs);
663 [ # # # # ]: 285 : spdk_free(resources->reqs);
664 [ # # # # ]: 285 : spdk_free(resources->recvs);
665 : 285 : free(resources);
666 : 285 : }
667 : :
668 : :
669 : : static struct spdk_nvmf_rdma_resources *
670 : 285 : nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts)
671 : : {
672 : : struct spdk_nvmf_rdma_resources *resources;
673 : : struct spdk_nvmf_rdma_request *rdma_req;
674 : : struct spdk_nvmf_rdma_recv *rdma_recv;
675 : 285 : struct spdk_rdma_provider_qp *qp = NULL;
676 : 285 : struct spdk_rdma_provider_srq *srq = NULL;
677 : 285 : struct ibv_recv_wr *bad_wr = NULL;
678 : 3 : struct spdk_rdma_utils_memory_translation translation;
679 : : uint32_t i;
680 : 285 : int rc = 0;
681 : :
682 : 285 : resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources));
683 [ - + ]: 285 : if (!resources) {
684 : 0 : SPDK_ERRLOG("Unable to allocate resources for receive queue.\n");
685 : 0 : return NULL;
686 : : }
687 : :
688 [ # # # # : 285 : resources->reqs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->reqs),
# # # # ]
689 : : 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
690 [ # # # # : 285 : resources->recvs = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->recvs),
# # # # ]
691 : : 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
692 [ # # # # : 285 : resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds),
# # # # ]
693 : : 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
694 [ # # # # : 285 : resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls),
# # # # ]
695 : : 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
696 : :
697 [ + - # # : 285 : if (opts->in_capsule_data_size > 0) {
# # ]
698 [ # # # # : 285 : resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size,
# # # # #
# # # ]
699 : : 0x1000, NULL, SPDK_ENV_LCORE_ID_ANY,
700 : : SPDK_MALLOC_DMA);
701 : 0 : }
702 : :
703 [ + - + - : 285 : if (!resources->reqs || !resources->recvs || !resources->cmds ||
+ - # # #
# # # # #
# # # # #
# ]
704 [ + - + - : 285 : !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) {
- + # # #
# # # # #
# # ]
705 : 0 : SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n");
706 : 0 : goto cleanup;
707 : : }
708 : :
709 [ - + - + : 285 : SPDK_DEBUGLOG(rdma, "Command Array: %p Length: %lx\n",
# # # # #
# # # #
# ]
710 : : resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds));
711 [ - + - + : 285 : SPDK_DEBUGLOG(rdma, "Completion Array: %p Length: %lx\n",
# # # # #
# # # #
# ]
712 : : resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls));
713 [ + - # # : 285 : if (resources->bufs) {
# # ]
714 [ - + - + : 285 : SPDK_DEBUGLOG(rdma, "In Capsule Data Array: %p Length: %x\n",
# # # # #
# # # # #
# # # # ]
715 : : resources->bufs, opts->max_queue_depth *
716 : : opts->in_capsule_data_size);
717 : 0 : }
718 : :
719 : : /* Initialize queues */
720 [ # # # # : 285 : STAILQ_INIT(&resources->incoming_queue);
# # # # #
# # # # #
# # ]
721 [ # # # # : 285 : STAILQ_INIT(&resources->free_queue);
# # # # #
# # # # #
# # ]
722 : :
723 [ + + + - : 285 : if (opts->shared) {
# # # # ]
724 [ # # # # ]: 285 : srq = (struct spdk_rdma_provider_srq *)opts->qp;
725 : 0 : } else {
726 [ # # # # ]: 0 : qp = (struct spdk_rdma_provider_qp *)opts->qp;
727 : : }
728 : :
729 [ + + # # : 1131165 : for (i = 0; i < opts->max_queue_depth; i++) {
# # ]
730 [ # # # # : 1130880 : rdma_recv = &resources->recvs[i];
# # ]
731 [ # # # # : 1130880 : rdma_recv->qpair = opts->qpair;
# # # # ]
732 : :
733 : : /* Set up memory to receive commands */
734 [ + - # # : 1130880 : if (resources->bufs) {
# # ]
735 [ # # # # : 2261632 : rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i *
# # # # ]
736 [ # # # # ]: 1130880 : opts->in_capsule_data_size));
737 : 0 : }
738 : :
739 [ # # # # : 1130880 : rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV;
# # ]
740 : :
741 [ # # # # : 1130880 : rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i];
# # # # #
# # # # #
# # ]
742 [ # # # # : 1130880 : rdma_recv->sgl[0].length = sizeof(resources->cmds[i]);
# # # # #
# ]
743 [ # # # # : 1130880 : rc = spdk_rdma_utils_get_translation(opts->map, &resources->cmds[i], sizeof(resources->cmds[i]),
# # # # #
# ]
744 : : &translation);
745 [ - + ]: 1130880 : if (rc) {
746 : 0 : goto cleanup;
747 : : }
748 [ # # # # : 1130880 : rdma_recv->sgl[0].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation);
# # # # #
# ]
749 [ # # # # : 1130880 : rdma_recv->wr.num_sge = 1;
# # ]
750 : :
751 [ + - # # : 1130880 : if (rdma_recv->buf) {
# # ]
752 [ # # # # : 1130880 : rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf;
# # # # #
# # # #
# ]
753 [ # # # # : 1130880 : rdma_recv->sgl[1].length = opts->in_capsule_data_size;
# # # # #
# # # #
# ]
754 [ # # # # : 1130880 : rc = spdk_rdma_utils_get_translation(opts->map, rdma_recv->buf, opts->in_capsule_data_size,
# # # # #
# # # ]
755 : : &translation);
756 [ - + ]: 1130880 : if (rc) {
757 : 0 : goto cleanup;
758 : : }
759 [ # # # # : 1130880 : rdma_recv->sgl[1].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation);
# # # # #
# ]
760 [ # # # # : 1130880 : rdma_recv->wr.num_sge++;
# # ]
761 : 0 : }
762 : :
763 [ # # # # : 1130880 : rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr;
# # # # ]
764 [ # # # # : 1130880 : rdma_recv->wr.sg_list = rdma_recv->sgl;
# # # # ]
765 [ + + ]: 1130880 : if (srq) {
766 [ # # ]: 1130496 : spdk_rdma_provider_srq_queue_recv_wrs(srq, &rdma_recv->wr);
767 : 0 : } else {
768 [ # # ]: 384 : spdk_rdma_provider_qp_queue_recv_wrs(qp, &rdma_recv->wr);
769 : : }
770 : 0 : }
771 : :
772 [ + + # # : 1131165 : for (i = 0; i < opts->max_queue_depth; i++) {
# # ]
773 [ # # # # : 1130880 : rdma_req = &resources->reqs[i];
# # ]
774 : :
775 [ + + # # : 1130880 : if (opts->qpair != NULL) {
# # ]
776 [ # # # # : 384 : rdma_req->req.qpair = &opts->qpair->qpair;
# # # # #
# # # ]
777 : 0 : } else {
778 [ # # # # : 1130496 : rdma_req->req.qpair = NULL;
# # ]
779 : : }
780 [ # # # # : 1130880 : rdma_req->req.cmd = NULL;
# # ]
781 [ # # # # : 1130880 : rdma_req->req.iovcnt = 0;
# # ]
782 [ # # # # : 1130880 : rdma_req->req.stripped_data = NULL;
# # ]
783 : :
784 : : /* Set up memory to send responses */
785 [ # # # # : 1130880 : rdma_req->req.rsp = &resources->cpls[i];
# # # # #
# # # ]
786 : :
787 [ # # # # : 1130880 : rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i];
# # # # #
# # # # #
# # ]
788 [ # # # # : 1130880 : rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]);
# # # # #
# ]
789 [ # # # # : 1130880 : rc = spdk_rdma_utils_get_translation(opts->map, &resources->cpls[i], sizeof(resources->cpls[i]),
# # # # #
# ]
790 : : &translation);
791 [ - + ]: 1130880 : if (rc) {
792 : 0 : goto cleanup;
793 : : }
794 [ # # # # : 1130880 : rdma_req->rsp.sgl[0].lkey = spdk_rdma_utils_memory_translation_get_lkey(&translation);
# # # # #
# ]
795 : :
796 [ # # # # : 1130880 : rdma_req->rsp_wr.type = RDMA_WR_TYPE_SEND;
# # ]
797 [ # # # # : 1130880 : rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp_wr;
# # # # #
# ]
798 [ # # # # : 1130880 : rdma_req->rsp.wr.next = NULL;
# # # # ]
799 [ # # # # : 1130880 : rdma_req->rsp.wr.opcode = IBV_WR_SEND;
# # # # ]
800 [ # # # # : 1130880 : rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED;
# # # # ]
801 [ # # # # : 1130880 : rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl;
# # # # #
# # # ]
802 [ # # # # : 1130880 : rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl);
# # # # ]
803 : :
804 : : /* Set up memory for data buffers */
805 [ # # # # : 1130880 : rdma_req->data_wr.type = RDMA_WR_TYPE_DATA;
# # ]
806 [ # # # # : 1130880 : rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data_wr;
# # # # #
# ]
807 [ # # # # : 1130880 : rdma_req->data.wr.next = NULL;
# # # # ]
808 [ # # # # : 1130880 : rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED;
# # # # ]
809 [ # # # # : 1130880 : rdma_req->data.wr.sg_list = rdma_req->data.sgl;
# # # # #
# # # ]
810 [ # # # # : 1130880 : rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl);
# # # # ]
811 : :
812 : : /* Initialize request state to FREE */
813 [ # # # # ]: 1130880 : rdma_req->state = RDMA_REQUEST_STATE_FREE;
814 [ # # # # : 1130880 : STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link);
# # # # #
# # # # #
# # # # #
# # # #
# ]
815 : 0 : }
816 : :
817 [ + + ]: 285 : if (srq) {
818 : 282 : rc = spdk_rdma_provider_srq_flush_recv_wrs(srq, &bad_wr);
819 : 0 : } else {
820 : 3 : rc = spdk_rdma_provider_qp_flush_recv_wrs(qp, &bad_wr);
821 : : }
822 : :
823 [ - + ]: 285 : if (rc) {
824 : 0 : goto cleanup;
825 : : }
826 : :
827 : 285 : return resources;
828 : :
829 : 0 : cleanup:
830 : 0 : nvmf_rdma_resources_destroy(resources);
831 : 0 : return NULL;
832 : 0 : }
833 : :
834 : : static void
835 : 2575 : nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair)
836 : : {
837 : : struct spdk_nvmf_rdma_ibv_event_ctx *ctx;
838 : :
839 [ # # # # ]: 2575 : ctx = rqpair->last_wqe_reached_ctx;
840 [ - + ]: 2575 : if (ctx) {
841 [ # # # # ]: 0 : ctx->rqpair = NULL;
842 : : /* Memory allocated for ctx is freed in nvmf_rdma_qpair_process_last_wqe_event */
843 [ # # # # ]: 0 : rqpair->last_wqe_reached_ctx = NULL;
844 : 0 : }
845 : 2575 : }
846 : :
847 : : static void nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller);
848 : :
849 : : static void
850 : 2575 : nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair)
851 : : {
852 : : struct spdk_nvmf_rdma_recv *rdma_recv, *recv_tmp;
853 : 2575 : struct ibv_recv_wr *bad_recv_wr = NULL;
854 : : int rc;
855 : :
856 [ + + + + : 2575 : spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair);
# # # # #
# # # # #
# # ]
857 : :
858 [ - + # # : 2575 : if (rqpair->qd != 0) {
# # ]
859 [ # # ]: 0 : struct spdk_nvmf_qpair *qpair = &rqpair->qpair;
860 [ # # # # ]: 0 : struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(qpair->transport,
861 : : struct spdk_nvmf_rdma_transport, transport);
862 : : struct spdk_nvmf_rdma_request *req;
863 : 0 : uint32_t i, max_req_count = 0;
864 : :
865 [ # # # # ]: 0 : SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd);
866 : :
867 [ # # # # : 0 : if (rqpair->srq == NULL) {
# # ]
868 : 0 : nvmf_rdma_dump_qpair_contents(rqpair);
869 [ # # # # ]: 0 : max_req_count = rqpair->max_queue_depth;
870 [ # # # # : 0 : } else if (rqpair->poller && rqpair->resources) {
# # # # #
# # # ]
871 [ # # # # : 0 : max_req_count = rqpair->poller->max_srq_depth;
# # # # ]
872 : 0 : }
873 : :
874 [ # # # # : 0 : SPDK_DEBUGLOG(rdma, "Release incomplete requests\n");
# # ]
875 [ # # ]: 0 : for (i = 0; i < max_req_count; i++) {
876 [ # # # # : 0 : req = &rqpair->resources->reqs[i];
# # # # #
# ]
877 [ # # # # : 0 : if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) {
# # # # #
# # # #
# ]
878 : : /* nvmf_rdma_request_process checks qpair ibv and internal state
879 : : * and completes a request */
880 : 0 : nvmf_rdma_request_process(rtransport, req);
881 : 0 : }
882 : 0 : }
883 [ # # # # : 0 : assert(rqpair->qd == 0);
# # # # ]
884 : 0 : }
885 : :
886 [ + - # # : 2575 : if (rqpair->poller) {
# # ]
887 [ # # # # : 2575 : RB_REMOVE(qpairs_tree, &rqpair->poller->qpairs, rqpair);
# # ]
888 : :
889 [ + - + - : 2575 : if (rqpair->srq != NULL && rqpair->resources != NULL) {
# # # # #
# # # ]
890 : : /* Drop all received but unprocessed commands for this queue and return them to SRQ */
891 [ - + # # : 2575 : STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) {
# # # # #
# # # # #
# # # # #
# ]
892 [ # # # # : 0 : if (rqpair == rdma_recv->qpair) {
# # ]
893 [ # # # # : 0 : STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # ]
894 [ # # # # : 0 : spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, &rdma_recv->wr);
# # ]
895 [ # # # # ]: 0 : rc = spdk_rdma_provider_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr);
896 [ # # ]: 0 : if (rc) {
897 : 0 : SPDK_ERRLOG("Unable to re-post rx descriptor\n");
898 : 0 : }
899 : 0 : }
900 : 0 : }
901 : 0 : }
902 : 0 : }
903 : :
904 [ + - # # : 2575 : if (rqpair->cm_id) {
# # ]
905 [ + - # # : 2575 : if (rqpair->rdma_qp != NULL) {
# # ]
906 [ # # # # ]: 2575 : spdk_rdma_provider_qp_destroy(rqpair->rdma_qp);
907 [ # # # # ]: 2575 : rqpair->rdma_qp = NULL;
908 : 0 : }
909 : :
910 [ + - - + : 2575 : if (rqpair->poller != NULL && rqpair->srq == NULL) {
# # # # #
# # # ]
911 [ # # # # : 0 : rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth);
# # # # #
# # # # #
# # # # ]
912 : 0 : }
913 : 0 : }
914 : :
915 [ - + - - : 2575 : if (rqpair->srq == NULL && rqpair->resources != NULL) {
# # # # #
# # # ]
916 [ # # # # ]: 0 : nvmf_rdma_resources_destroy(rqpair->resources);
917 : 0 : }
918 : :
919 : 2575 : nvmf_rdma_qpair_clean_ibv_events(rqpair);
920 : :
921 [ + - # # : 2575 : if (rqpair->destruct_channel) {
# # ]
922 [ # # # # ]: 2575 : spdk_put_io_channel(rqpair->destruct_channel);
923 [ # # # # ]: 2575 : rqpair->destruct_channel = NULL;
924 : 0 : }
925 : :
926 [ + - - + : 2575 : if (rqpair->poller && rqpair->poller->need_destroy && RB_EMPTY(&rqpair->poller->qpairs)) {
- + - - #
# # # # #
# # # # #
# # # # #
# # # # #
# ]
927 [ # # # # ]: 0 : nvmf_rdma_poller_destroy(rqpair->poller);
928 : 0 : }
929 : :
930 : : /* destroy cm_id last so cma device will not be freed before we destroy the cq. */
931 [ + - # # : 2575 : if (rqpair->cm_id) {
# # ]
932 [ # # # # ]: 2575 : rdma_destroy_id(rqpair->cm_id);
933 : 0 : }
934 : :
935 : 2575 : free(rqpair);
936 : 2575 : }
937 : :
938 : : static int
939 : 15 : nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device)
940 : : {
941 : : struct spdk_nvmf_rdma_poller *rpoller;
942 : : int rc, num_cqe, required_num_wr;
943 : :
944 : : /* Enlarge CQ size dynamically */
945 [ # # # # ]: 15 : rpoller = rqpair->poller;
946 [ # # # # : 15 : required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth);
# # # # #
# # # #
# ]
947 [ # # # # ]: 15 : num_cqe = rpoller->num_cqe;
948 [ + + ]: 15 : if (num_cqe < required_num_wr) {
949 [ # # # # : 12 : num_cqe = spdk_max(num_cqe * 2, required_num_wr);
# # ]
950 [ # # # # : 12 : num_cqe = spdk_min(num_cqe, device->attr.max_cqe);
# # # # #
# # # #
# ]
951 : 0 : }
952 : :
953 [ + + # # : 15 : if (rpoller->num_cqe != num_cqe) {
# # ]
954 [ + + # # : 12 : if (device->context->device->transport_type == IBV_TRANSPORT_IWARP) {
# # # # #
# # # #
# ]
955 [ # # # # ]: 3 : SPDK_ERRLOG("iWARP doesn't support CQ resize. Current capacity %u, required %u\n"
956 : : "Using CQ of insufficient size may lead to CQ overrun\n", rpoller->num_cqe, num_cqe);
957 : 3 : return -1;
958 : : }
959 [ + + # # : 9 : if (required_num_wr > device->attr.max_cqe) {
# # # # ]
960 [ # # # # : 3 : SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n",
# # ]
961 : : required_num_wr, device->attr.max_cqe);
962 : 3 : return -1;
963 : : }
964 : :
965 [ - + - + : 6 : SPDK_DEBUGLOG(rdma, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe);
# # # # #
# ]
966 [ # # # # ]: 6 : rc = ibv_resize_cq(rpoller->cq, num_cqe);
967 [ + + ]: 6 : if (rc) {
968 [ # # # # ]: 3 : SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno));
969 : 3 : return -1;
970 : : }
971 : :
972 [ # # # # ]: 3 : rpoller->num_cqe = num_cqe;
973 : 0 : }
974 : :
975 [ # # # # ]: 6 : rpoller->required_num_wr = required_num_wr;
976 : 6 : return 0;
977 : 0 : }
978 : :
979 : : static int
980 : 2575 : nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair)
981 : : {
982 : : struct spdk_nvmf_rdma_qpair *rqpair;
983 : : struct spdk_nvmf_rdma_transport *rtransport;
984 : : struct spdk_nvmf_transport *transport;
985 : 0 : struct spdk_nvmf_rdma_resource_opts opts;
986 : : struct spdk_nvmf_rdma_device *device;
987 : 2575 : struct spdk_rdma_provider_qp_init_attr qp_init_attr = {};
988 : :
989 : 2575 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
990 [ # # # # ]: 2575 : device = rqpair->device;
991 : :
992 : 2575 : qp_init_attr.qp_context = rqpair;
993 [ # # # # : 2575 : qp_init_attr.pd = device->pd;
# # ]
994 [ # # # # : 2575 : qp_init_attr.send_cq = rqpair->poller->cq;
# # # # #
# ]
995 [ # # # # : 2575 : qp_init_attr.recv_cq = rqpair->poller->cq;
# # # # #
# ]
996 : :
997 [ + - # # : 2575 : if (rqpair->srq) {
# # ]
998 [ # # # # : 2575 : qp_init_attr.srq = rqpair->srq->srq;
# # # # #
# ]
999 : 0 : } else {
1000 [ # # # # : 0 : qp_init_attr.cap.max_recv_wr = rqpair->max_queue_depth;
# # # # ]
1001 : : }
1002 : :
1003 : : /* SEND, READ, and WRITE operations */
1004 [ # # # # : 2575 : qp_init_attr.cap.max_send_wr = (uint32_t)rqpair->max_queue_depth * 2;
# # # # ]
1005 [ # # # # : 2575 : qp_init_attr.cap.max_send_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE);
# # # # #
# # # # #
# # # # ]
1006 [ # # # # : 2575 : qp_init_attr.cap.max_recv_sge = spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE);
# # # # #
# # # # #
# # # # ]
1007 [ # # # # : 2575 : qp_init_attr.stats = &rqpair->poller->stat.qp_stats;
# # # # #
# ]
1008 : :
1009 [ - + - - : 2575 : if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) {
# # # # ]
1010 : 0 : SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n");
1011 : 0 : goto error;
1012 : : }
1013 : :
1014 [ # # # # : 2575 : rqpair->rdma_qp = spdk_rdma_provider_qp_create(rqpair->cm_id, &qp_init_attr);
# # # # ]
1015 [ - + # # : 2575 : if (!rqpair->rdma_qp) {
# # ]
1016 : 0 : goto error;
1017 : : }
1018 : :
1019 [ # # # # : 2575 : rqpair->qp_num = rqpair->rdma_qp->qp->qp_num;
# # # # #
# # # # #
# # ]
1020 : :
1021 [ # # # # : 2575 : rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2),
# # # # #
# # # # #
# # # # #
# # # # #
# # ]
1022 : : qp_init_attr.cap.max_send_wr);
1023 [ # # # # : 2575 : rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, qp_init_attr.cap.max_send_sge);
# # # # #
# # # #
# ]
1024 [ # # # # : 2575 : rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, qp_init_attr.cap.max_recv_sge);
# # # # #
# # # #
# ]
1025 [ + + + + : 2575 : spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair);
# # # # #
# # # # #
# # ]
1026 [ - + - + : 2575 : SPDK_DEBUGLOG(rdma, "New RDMA Connection: %p\n", qpair);
# # ]
1027 : :
1028 [ - + # # : 2575 : if (rqpair->poller->srq == NULL) {
# # # # #
# ]
1029 [ # # # # ]: 0 : rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
1030 [ # # ]: 0 : transport = &rtransport->transport;
1031 : :
1032 [ # # # # : 0 : opts.qp = rqpair->rdma_qp;
# # ]
1033 [ # # # # : 0 : opts.map = device->map;
# # ]
1034 : 0 : opts.qpair = rqpair;
1035 [ # # ]: 0 : opts.shared = false;
1036 [ # # # # : 0 : opts.max_queue_depth = rqpair->max_queue_depth;
# # ]
1037 [ # # # # : 0 : opts.in_capsule_data_size = transport->opts.in_capsule_data_size;
# # # # ]
1038 : :
1039 [ # # # # ]: 0 : rqpair->resources = nvmf_rdma_resources_create(&opts);
1040 : :
1041 [ # # # # : 0 : if (!rqpair->resources) {
# # ]
1042 : 0 : SPDK_ERRLOG("Unable to allocate resources for receive queue.\n");
1043 [ # # # # ]: 0 : rdma_destroy_qp(rqpair->cm_id);
1044 : 0 : goto error;
1045 : : }
1046 : 0 : } else {
1047 [ # # # # : 2575 : rqpair->resources = rqpair->poller->resources;
# # # # #
# # # ]
1048 : : }
1049 : :
1050 [ # # # # ]: 2575 : rqpair->current_recv_depth = 0;
1051 [ # # # # : 2575 : STAILQ_INIT(&rqpair->pending_rdma_read_queue);
# # # # #
# # # # #
# # ]
1052 [ # # # # : 2575 : STAILQ_INIT(&rqpair->pending_rdma_write_queue);
# # # # #
# # # # #
# # ]
1053 [ # # # # : 2575 : STAILQ_INIT(&rqpair->pending_rdma_send_queue);
# # # # #
# # # # #
# # ]
1054 [ # # # # : 2575 : rqpair->qpair.queue_depth = 0;
# # ]
1055 : :
1056 : 2575 : return 0;
1057 : :
1058 : 0 : error:
1059 [ # # # # ]: 0 : rdma_destroy_id(rqpair->cm_id);
1060 [ # # # # ]: 0 : rqpair->cm_id = NULL;
1061 : 0 : return -1;
1062 : 0 : }
1063 : :
1064 : : /* Append the given recv wr structure to the resource structs outstanding recvs list. */
1065 : : /* This function accepts either a single wr or the first wr in a linked list. */
1066 : : static void
1067 : 7921540 : nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first)
1068 : : {
1069 [ # # # # : 7921540 : struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
# # ]
1070 : : struct spdk_nvmf_rdma_transport, transport);
1071 : :
1072 [ + + # # : 7921540 : if (rqpair->srq != NULL) {
# # ]
1073 [ # # # # ]: 7921522 : spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, first);
1074 : 0 : } else {
1075 [ + - # # : 18 : if (spdk_rdma_provider_qp_queue_recv_wrs(rqpair->rdma_qp, first)) {
# # ]
1076 [ # # # # : 18 : STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # ]
1077 : 0 : }
1078 : : }
1079 : :
1080 [ - + - + : 7921540 : if (rtransport->rdma_opts.no_wr_batching) {
# # # # #
# ]
1081 [ # # # # ]: 0 : _poller_submit_recvs(rtransport, rqpair->poller);
1082 : 0 : }
1083 : 7921540 : }
1084 : :
1085 : : static inline void
1086 : 784766 : request_transfer_in(struct spdk_nvmf_request *req)
1087 : : {
1088 : : struct spdk_nvmf_rdma_request *rdma_req;
1089 : : struct spdk_nvmf_qpair *qpair;
1090 : : struct spdk_nvmf_rdma_qpair *rqpair;
1091 : : struct spdk_nvmf_rdma_transport *rtransport;
1092 : :
1093 [ # # # # ]: 784766 : qpair = req->qpair;
1094 : 784766 : rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
1095 : 784766 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
1096 [ # # # # : 784766 : rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
# # ]
1097 : : struct spdk_nvmf_rdma_transport, transport);
1098 : :
1099 [ - + # # : 784766 : assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
# # # # ]
1100 [ - + # # ]: 784766 : assert(rdma_req != NULL);
1101 : :
1102 [ + + # # : 784766 : if (spdk_rdma_provider_qp_queue_send_wrs(rqpair->rdma_qp, rdma_req->transfer_wr)) {
# # # # #
# ]
1103 [ # # # # : 147432 : STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # ]
1104 : 0 : }
1105 [ - + - + : 784766 : if (rtransport->rdma_opts.no_wr_batching) {
# # # # #
# ]
1106 [ # # # # ]: 0 : _poller_submit_sends(rtransport, rqpair->poller);
1107 : 0 : }
1108 : :
1109 [ - + # # : 784766 : assert(rqpair->current_read_depth + rdma_req->num_outstanding_data_wr <= rqpair->max_read_depth);
# # # # #
# # # # #
# # ]
1110 [ # # # # : 784766 : rqpair->current_read_depth += rdma_req->num_outstanding_data_wr;
# # # # #
# ]
1111 [ - + # # : 784766 : assert(rqpair->current_send_depth + rdma_req->num_outstanding_data_wr <= rqpair->max_send_depth);
# # # # #
# # # # #
# # ]
1112 [ # # # # : 784766 : rqpair->current_send_depth += rdma_req->num_outstanding_data_wr;
# # # # ]
1113 : 784766 : }
1114 : :
1115 : : static inline void
1116 : 65838 : nvmf_rdma_request_reset_transfer_in(struct spdk_nvmf_rdma_request *rdma_req,
1117 : : struct spdk_nvmf_rdma_transport *rtransport)
1118 : : {
1119 : : /* Put completed WRs back to pool and move transfer_wr pointer */
1120 [ # # # # : 65838 : _nvmf_rdma_request_free_data(rdma_req, rdma_req->transfer_wr, rtransport->data_wr_pool);
# # # # ]
1121 [ # # # # : 65838 : rdma_req->transfer_wr = rdma_req->remaining_tranfer_in_wrs;
# # # # ]
1122 [ # # # # ]: 65838 : rdma_req->remaining_tranfer_in_wrs = NULL;
1123 [ # # # # : 65838 : rdma_req->num_outstanding_data_wr = rdma_req->num_remaining_data_wr;
# # # # ]
1124 [ # # # # ]: 65838 : rdma_req->num_remaining_data_wr = 0;
1125 : 65838 : }
1126 : :
1127 : : static inline int
1128 : 65838 : request_prepare_transfer_in_part(struct spdk_nvmf_request *req, uint32_t num_reads_available)
1129 : : {
1130 : : struct spdk_nvmf_rdma_request *rdma_req;
1131 : : struct ibv_send_wr *wr;
1132 : : uint32_t i;
1133 : :
1134 : 65838 : rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
1135 : :
1136 [ - + # # : 65838 : assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
# # # # ]
1137 [ - + # # ]: 65838 : assert(rdma_req != NULL);
1138 [ - + # # ]: 65838 : assert(num_reads_available > 0);
1139 [ - + # # : 65838 : assert(rdma_req->num_outstanding_data_wr > num_reads_available);
# # # # ]
1140 [ # # # # ]: 65838 : wr = rdma_req->transfer_wr;
1141 : :
1142 [ + + ]: 222920 : for (i = 0; i < num_reads_available - 1; i++) {
1143 [ # # # # ]: 157082 : wr = wr->next;
1144 : 0 : }
1145 : :
1146 [ # # # # : 65838 : rdma_req->remaining_tranfer_in_wrs = wr->next;
# # # # ]
1147 [ # # # # : 65838 : rdma_req->num_remaining_data_wr = rdma_req->num_outstanding_data_wr - num_reads_available;
# # # # ]
1148 [ # # # # ]: 65838 : rdma_req->num_outstanding_data_wr = num_reads_available;
1149 : : /* Break chain of WRs to send only part. Once this portion completes, we continue sending RDMA_READs */
1150 [ # # # # ]: 65838 : wr->next = NULL;
1151 : :
1152 : 65838 : return 0;
1153 : : }
1154 : :
1155 : : static int
1156 : 7921540 : request_transfer_out(struct spdk_nvmf_request *req, int *data_posted)
1157 : : {
1158 : 7921540 : int num_outstanding_data_wr = 0;
1159 : : struct spdk_nvmf_rdma_request *rdma_req;
1160 : : struct spdk_nvmf_qpair *qpair;
1161 : : struct spdk_nvmf_rdma_qpair *rqpair;
1162 : : struct spdk_nvme_cpl *rsp;
1163 : 7921540 : struct ibv_send_wr *first = NULL;
1164 : : struct spdk_nvmf_rdma_transport *rtransport;
1165 : :
1166 [ # # ]: 7921540 : *data_posted = 0;
1167 [ # # # # ]: 7921540 : qpair = req->qpair;
1168 [ # # # # : 7921540 : rsp = &req->rsp->nvme_cpl;
# # ]
1169 : 7921540 : rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
1170 : 7921540 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
1171 [ # # # # : 7921540 : rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
# # ]
1172 : : struct spdk_nvmf_rdma_transport, transport);
1173 : :
1174 : : /* Advance our sq_head pointer */
1175 [ + + # # : 7921540 : if (qpair->sq_head == qpair->sq_head_max) {
# # # # #
# ]
1176 [ # # # # ]: 65053 : qpair->sq_head = 0;
1177 : 0 : } else {
1178 [ # # ]: 7856487 : qpair->sq_head++;
1179 : : }
1180 [ # # # # : 7921540 : rsp->sqhd = qpair->sq_head;
# # # # ]
1181 : :
1182 : : /* queue the capsule for the recv buffer */
1183 [ - + # # : 7921540 : assert(rdma_req->recv != NULL);
# # # # ]
1184 : :
1185 [ # # # # : 7921540 : nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr);
# # ]
1186 : :
1187 [ # # # # ]: 7921540 : rdma_req->recv = NULL;
1188 [ - + # # : 7921540 : assert(rqpair->current_recv_depth > 0);
# # # # ]
1189 [ # # ]: 7921540 : rqpair->current_recv_depth--;
1190 : :
1191 : : /* Build the response which consists of optional
1192 : : * RDMA WRITEs to transfer data, plus an RDMA SEND
1193 : : * containing the response.
1194 : : */
1195 [ # # # # ]: 7921540 : first = &rdma_req->rsp.wr;
1196 : :
1197 [ + + # # : 7921540 : if (spdk_unlikely(rsp->status.sc != SPDK_NVME_SC_SUCCESS)) {
# # # # ]
1198 : : /* On failure, data was not read from the controller. So clear the
1199 : : * number of outstanding data WRs to zero.
1200 : : */
1201 [ # # # # ]: 1400459 : rdma_req->num_outstanding_data_wr = 0;
1202 [ + + # # : 6521081 : } else if (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
# # ]
1203 [ # # # # ]: 3841991 : first = rdma_req->transfer_wr;
1204 [ # # ]: 3841991 : *data_posted = 1;
1205 [ # # # # ]: 3841991 : num_outstanding_data_wr = rdma_req->num_outstanding_data_wr;
1206 : 0 : }
1207 [ + + # # : 7921540 : if (spdk_rdma_provider_qp_queue_send_wrs(rqpair->rdma_qp, first)) {
# # ]
1208 [ # # # # : 1202642 : STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # ]
1209 : 0 : }
1210 [ - + - + : 7921540 : if (rtransport->rdma_opts.no_wr_batching) {
# # # # #
# ]
1211 [ # # # # ]: 0 : _poller_submit_sends(rtransport, rqpair->poller);
1212 : 0 : }
1213 : :
1214 : : /* +1 for the rsp wr */
1215 [ - + # # : 7921540 : assert(rqpair->current_send_depth + num_outstanding_data_wr + 1 <= rqpair->max_send_depth);
# # # # #
# # # ]
1216 [ # # # # : 7921540 : rqpair->current_send_depth += num_outstanding_data_wr + 1;
# # ]
1217 : :
1218 : 7921540 : return 0;
1219 : : }
1220 : :
1221 : : static int
1222 : 2575 : nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair)
1223 : : {
1224 : 0 : struct spdk_nvmf_rdma_accept_private_data accept_data;
1225 : 2575 : struct rdma_conn_param ctrlr_event_data = {};
1226 : : int rc;
1227 : :
1228 : 2575 : accept_data.recfmt = 0;
1229 [ # # # # ]: 2575 : accept_data.crqsize = rqpair->max_queue_depth;
1230 : :
1231 : 2575 : ctrlr_event_data.private_data = &accept_data;
1232 [ # # ]: 2575 : ctrlr_event_data.private_data_len = sizeof(accept_data);
1233 [ + - # # : 2575 : if (id->ps == RDMA_PS_TCP) {
# # ]
1234 : 2575 : ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */
1235 [ # # # # : 2575 : ctrlr_event_data.initiator_depth = rqpair->max_read_depth;
# # ]
1236 : 0 : }
1237 : :
1238 : : /* Configure infinite retries for the initiator side qpair.
1239 : : * We need to pass this value to the initiator to prevent the
1240 : : * initiator side NIC from completing SEND requests back to the
1241 : : * initiator with status rnr_retry_count_exceeded. */
1242 : 2575 : ctrlr_event_data.rnr_retry_count = 0x7;
1243 : :
1244 : : /* When qpair is created without use of rdma cm API, an additional
1245 : : * information must be provided to initiator in the connection response:
1246 : : * whether qpair is using SRQ and its qp_num
1247 : : * Fields below are ignored by rdma cm if qpair has been
1248 : : * created using rdma cm API. */
1249 [ # # # # : 2575 : ctrlr_event_data.srq = rqpair->srq ? 1 : 0;
# # ]
1250 [ # # # # : 2575 : ctrlr_event_data.qp_num = rqpair->qp_num;
# # ]
1251 : :
1252 [ # # # # ]: 2575 : rc = spdk_rdma_provider_qp_accept(rqpair->rdma_qp, &ctrlr_event_data);
1253 [ - + ]: 2575 : if (rc) {
1254 [ # # ]: 0 : SPDK_ERRLOG("Error %d on spdk_rdma_provider_qp_accept\n", errno);
1255 : 0 : } else {
1256 [ - + - + : 2575 : SPDK_DEBUGLOG(rdma, "Sent back the accept\n");
# # ]
1257 : : }
1258 : :
1259 : 2575 : return rc;
1260 : : }
1261 : :
1262 : : static void
1263 : 0 : nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error)
1264 : : {
1265 : 0 : struct spdk_nvmf_rdma_reject_private_data rej_data;
1266 : :
1267 : 0 : rej_data.recfmt = 0;
1268 : 0 : rej_data.sts = error;
1269 : :
1270 : 0 : rdma_reject(id, &rej_data, sizeof(rej_data));
1271 : 0 : }
1272 : :
1273 : : static int
1274 : 2575 : nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event)
1275 : : {
1276 : : struct spdk_nvmf_rdma_transport *rtransport;
1277 : 2575 : struct spdk_nvmf_rdma_qpair *rqpair = NULL;
1278 : : struct spdk_nvmf_rdma_port *port;
1279 : 2575 : struct rdma_conn_param *rdma_param = NULL;
1280 : 2575 : const struct spdk_nvmf_rdma_request_private_data *private_data = NULL;
1281 : : uint16_t max_queue_depth;
1282 : : uint16_t max_read_depth;
1283 : :
1284 : 2575 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
1285 : :
1286 [ - + # # : 2575 : assert(event->id != NULL); /* Impossible. Can't even reject the connection. */
# # # # ]
1287 [ - + # # : 2575 : assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */
# # # # #
# # # ]
1288 : :
1289 [ # # # # ]: 2575 : rdma_param = &event->param.conn;
1290 [ + - # # : 2575 : if (rdma_param->private_data == NULL ||
# # # # ]
1291 [ - + # # ]: 2575 : rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) {
1292 : 0 : SPDK_ERRLOG("connect request: no private data provided\n");
1293 [ # # # # ]: 0 : nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH);
1294 : 0 : return -1;
1295 : : }
1296 : :
1297 [ # # # # ]: 2575 : private_data = rdma_param->private_data;
1298 [ - + # # : 2575 : if (private_data->recfmt != 0) {
# # ]
1299 : 0 : SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n");
1300 [ # # # # ]: 0 : nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT);
1301 : 0 : return -1;
1302 : : }
1303 : :
1304 [ - + - + : 2575 : SPDK_DEBUGLOG(rdma, "Connect Recv on fabric intf name %s, dev_name %s\n",
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # #
# ]
1305 : : event->id->verbs->device->name, event->id->verbs->device->dev_name);
1306 : :
1307 [ # # # # : 2575 : port = event->listen_id->context;
# # # # ]
1308 [ - + - + : 2575 : SPDK_DEBUGLOG(rdma, "Listen Id was %p with verbs %p. ListenAddr: %p\n",
# # # # #
# # # # #
# # # # ]
1309 : : event->listen_id, event->listen_id->verbs, port);
1310 : :
1311 : : /* Figure out the supported queue depth. This is a multi-step process
1312 : : * that takes into account hardware maximums, host provided values,
1313 : : * and our target's internal memory limits */
1314 : :
1315 [ - + - + : 2575 : SPDK_DEBUGLOG(rdma, "Calculating Queue Depth\n");
# # ]
1316 : :
1317 : : /* Start with the maximum queue depth allowed by the target */
1318 [ # # # # : 2575 : max_queue_depth = rtransport->transport.opts.max_queue_depth;
# # # # ]
1319 [ # # # # : 2575 : max_read_depth = rtransport->transport.opts.max_queue_depth;
# # # # ]
1320 [ - + - + : 2575 : SPDK_DEBUGLOG(rdma, "Target Max Queue Depth: %d\n",
# # # # #
# # # #
# ]
1321 : : rtransport->transport.opts.max_queue_depth);
1322 : :
1323 : : /* Next check the local NIC's hardware limitations */
1324 [ - + - + : 2575 : SPDK_DEBUGLOG(rdma,
# # # # #
# # # # #
# # # # #
# # # # #
# # ]
1325 : : "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n",
1326 : : port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom);
1327 [ # # # # : 2575 : max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr);
# # # # #
# # # # #
# # # # #
# # # ]
1328 [ # # # # : 2575 : max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom);
# # # # #
# # # # #
# # # # #
# # # ]
1329 : :
1330 : : /* Next check the remote NIC's hardware limitations */
1331 [ - + - + : 2575 : SPDK_DEBUGLOG(rdma,
# # # # #
# # # #
# ]
1332 : : "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n",
1333 : : rdma_param->initiator_depth, rdma_param->responder_resources);
1334 : : /* from man3 rdma_get_cm_event
1335 : : * responder_resources - Specifies the number of responder resources that is requested by the recipient.
1336 : : * The responder_resources field must match the initiator depth specified by the remote node when running
1337 : : * the rdma_connect and rdma_accept functions. */
1338 [ - + # # : 2575 : if (rdma_param->responder_resources != 0) {
# # ]
1339 [ # # # # : 0 : if (private_data->qid) {
# # ]
1340 [ # # # # : 0 : SPDK_DEBUGLOG(rdma, "Host (Initiator) is not allowed to use RDMA operations,"
# # # # #
# ]
1341 : : " responder_resources must be 0 but set to %u\n",
1342 : : rdma_param->responder_resources);
1343 : 0 : } else {
1344 [ # # # # ]: 0 : SPDK_WARNLOG("Host (Initiator) is not allowed to use RDMA operations,"
1345 : : " responder_resources must be 0 but set to %u\n",
1346 : : rdma_param->responder_resources);
1347 : : }
1348 : 0 : }
1349 : : /* from man3 rdma_get_cm_event
1350 : : * initiator_depth - Specifies the maximum number of outstanding RDMA read operations that the recipient holds.
1351 : : * The initiator_depth field must match the responder resources specified by the remote node when running
1352 : : * the rdma_connect and rdma_accept functions. */
1353 [ - + # # : 2575 : if (rdma_param->initiator_depth == 0) {
# # ]
1354 : 0 : SPDK_ERRLOG("Host (Initiator) doesn't support RDMA_READ or atomic operations\n");
1355 [ # # # # ]: 0 : nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_IRD);
1356 : 0 : return -1;
1357 : : }
1358 [ # # # # : 2575 : max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth);
# # # # #
# ]
1359 : :
1360 [ - + - + : 2575 : SPDK_DEBUGLOG(rdma, "Host Receive Queue Size: %d\n", private_data->hrqsize);
# # # # #
# ]
1361 [ - + - + : 2575 : SPDK_DEBUGLOG(rdma, "Host Send Queue Size: %d\n", private_data->hsqsize);
# # # # #
# ]
1362 [ # # # # : 2575 : max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize);
# # # # #
# ]
1363 [ + + # # : 2575 : max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1);
# # # # #
# # # #
# ]
1364 : :
1365 [ - + - + : 2575 : SPDK_DEBUGLOG(rdma, "Final Negotiated Queue Depth: %d R/W Depth: %d\n",
# # ]
1366 : : max_queue_depth, max_read_depth);
1367 : :
1368 : 2575 : rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair));
1369 [ - + ]: 2575 : if (rqpair == NULL) {
1370 : 0 : SPDK_ERRLOG("Could not allocate new connection.\n");
1371 [ # # # # ]: 0 : nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES);
1372 : 0 : return -1;
1373 : : }
1374 : :
1375 [ # # # # : 2575 : rqpair->device = port->device;
# # # # ]
1376 [ # # # # ]: 2575 : rqpair->max_queue_depth = max_queue_depth;
1377 [ # # # # ]: 2575 : rqpair->max_read_depth = max_read_depth;
1378 [ # # # # : 2575 : rqpair->cm_id = event->id;
# # # # ]
1379 [ # # # # : 2575 : rqpair->listen_id = event->listen_id;
# # # # ]
1380 [ # # # # : 2575 : rqpair->qpair.transport = transport;
# # ]
1381 : : /* use qid from the private data to determine the qpair type
1382 : : qid will be set to the appropriate value when the controller is created */
1383 [ # # # # : 2575 : rqpair->qpair.qid = private_data->qid;
# # # # #
# ]
1384 [ # # # # : 2575 : rqpair->qpair.numa.id_valid = 1;
# # ]
1385 [ # # # # : 2575 : rqpair->qpair.numa.id = spdk_rdma_cm_id_get_numa_id(rqpair->cm_id);
# # # # #
# ]
1386 : :
1387 [ # # # # : 2575 : event->id->context = &rqpair->qpair;
# # # # #
# ]
1388 : :
1389 [ # # # # : 2575 : spdk_nvmf_tgt_new_qpair(transport->tgt, &rqpair->qpair);
# # ]
1390 : :
1391 : 2575 : return 0;
1392 : 0 : }
1393 : :
1394 : : static inline void
1395 : 8117010 : nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next,
1396 : : enum spdk_nvme_data_transfer xfer)
1397 : : {
1398 [ + + ]: 8117010 : if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
1399 [ # # # # ]: 5502961 : wr->opcode = IBV_WR_RDMA_WRITE;
1400 [ # # # # ]: 5502961 : wr->send_flags = 0;
1401 [ # # # # ]: 5502961 : wr->next = next;
1402 [ + - ]: 2614049 : } else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) {
1403 [ # # # # ]: 2614049 : wr->opcode = IBV_WR_RDMA_READ;
1404 [ # # # # ]: 2614049 : wr->send_flags = IBV_SEND_SIGNALED;
1405 [ # # # # ]: 2614049 : wr->next = NULL;
1406 : 0 : } else {
1407 [ # # ]: 0 : assert(0);
1408 : : }
1409 : 8117010 : }
1410 : :
1411 : : static int
1412 : 1243419 : nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport,
1413 : : struct spdk_nvmf_rdma_request *rdma_req,
1414 : : uint32_t num_sgl_descriptors)
1415 : : {
1416 : 18 : struct spdk_nvmf_rdma_request_data *work_requests[SPDK_NVMF_MAX_SGL_ENTRIES];
1417 : : struct spdk_nvmf_rdma_request_data *current_data_wr;
1418 : : uint32_t i;
1419 : :
1420 [ - + ]: 1243419 : if (spdk_unlikely(num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES)) {
1421 : 0 : SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n",
1422 : : num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES);
1423 : 0 : return -EINVAL;
1424 : : }
1425 : :
1426 [ + + # # : 1243419 : if (spdk_unlikely(spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests,
# # ]
1427 : : num_sgl_descriptors))) {
1428 : 1075969 : return -ENOMEM;
1429 : : }
1430 : :
1431 [ # # ]: 167450 : current_data_wr = &rdma_req->data;
1432 : :
1433 [ + + ]: 1439886 : for (i = 0; i < num_sgl_descriptors; i++) {
1434 [ # # # # : 1272436 : nvmf_rdma_setup_wr(¤t_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer);
# # # # #
# # # # #
# # ]
1435 [ # # # # : 1272436 : current_data_wr->wr.next = &work_requests[i]->wr;
# # # # #
# # # #
# ]
1436 [ # # # # : 1272436 : current_data_wr = work_requests[i];
# # ]
1437 [ # # # # : 1272436 : current_data_wr->wr.sg_list = current_data_wr->sgl;
# # # # ]
1438 [ # # # # : 1272436 : current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id;
# # # # #
# # # #
# ]
1439 : 0 : }
1440 : :
1441 [ # # # # : 167450 : nvmf_rdma_setup_wr(¤t_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer);
# # # # #
# # # ]
1442 : :
1443 : 167450 : return 0;
1444 : 0 : }
1445 : :
1446 : : static inline void
1447 : 6677124 : nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req)
1448 : : {
1449 [ # # # # ]: 6677124 : struct ibv_send_wr *wr = &rdma_req->data.wr;
1450 [ # # # # : 6677124 : struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1;
# # # # #
# # # ]
1451 : :
1452 [ # # # # : 6677124 : wr->wr.rdma.rkey = sgl->keyed.key;
# # # # #
# # # #
# ]
1453 [ # # # # : 6677124 : wr->wr.rdma.remote_addr = sgl->address;
# # # # #
# # # ]
1454 [ # # # # : 6677124 : nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer);
# # # # #
# ]
1455 : 6677124 : }
1456 : :
1457 : : static inline void
1458 : 3 : nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs)
1459 : : {
1460 [ # # # # ]: 3 : struct ibv_send_wr *wr = &rdma_req->data.wr;
1461 [ # # # # : 3 : struct spdk_nvme_sgl_descriptor *sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1;
# # # # #
# # # ]
1462 : : uint32_t i;
1463 : : int j;
1464 : 3 : uint64_t remote_addr_offset = 0;
1465 : :
1466 [ + + ]: 9 : for (i = 0; i < num_wrs; ++i) {
1467 [ # # # # : 6 : wr->wr.rdma.rkey = sgl->keyed.key;
# # # # #
# # # #
# ]
1468 [ # # # # : 6 : wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset;
# # # # #
# # # ]
1469 [ + + # # : 57 : for (j = 0; j < wr->num_sge; ++j) {
# # # # ]
1470 [ # # # # : 51 : remote_addr_offset += wr->sg_list[j].length;
# # # # #
# ]
1471 : 0 : }
1472 [ # # # # ]: 6 : wr = wr->next;
1473 : 0 : }
1474 : 3 : }
1475 : :
1476 : : static int
1477 : 6003890 : nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_device *device,
1478 : : struct spdk_nvmf_rdma_request *rdma_req,
1479 : : struct ibv_send_wr *wr,
1480 : : uint32_t total_length)
1481 : : {
1482 : 45 : struct spdk_rdma_utils_memory_translation mem_translation;
1483 : : struct ibv_sge *sg_ele;
1484 : : struct iovec *iov;
1485 : : uint32_t lkey, remaining;
1486 : : int rc;
1487 : :
1488 [ # # # # ]: 6003890 : wr->num_sge = 0;
1489 : :
1490 [ + + + - : 29524604 : while (total_length && wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES) {
# # # # ]
1491 [ # # # # : 23520714 : iov = &rdma_req->req.iov[rdma_req->iovpos];
# # # # #
# # # ]
1492 [ # # # # : 23520714 : rc = spdk_rdma_utils_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation);
# # # # #
# # # ]
1493 [ - + ]: 23520714 : if (spdk_unlikely(rc)) {
1494 : 0 : return rc;
1495 : : }
1496 : :
1497 : 23520714 : lkey = spdk_rdma_utils_memory_translation_get_lkey(&mem_translation);
1498 [ # # # # : 23520714 : sg_ele = &wr->sg_list[wr->num_sge];
# # # # #
# ]
1499 [ # # # # : 23520714 : remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length);
# # # # #
# # # # #
# # # # ]
1500 : :
1501 [ # # # # ]: 23520714 : sg_ele->lkey = lkey;
1502 [ # # # # : 23520714 : sg_ele->addr = (uintptr_t)iov->iov_base + rdma_req->offset;
# # # # #
# # # ]
1503 [ # # # # ]: 23520714 : sg_ele->length = remaining;
1504 [ - + - + : 23520714 : SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele, sg_ele->addr,
# # # # #
# # # # #
# # # # ]
1505 : : sg_ele->length);
1506 [ # # # # : 23520714 : rdma_req->offset += sg_ele->length;
# # # # ]
1507 [ # # # # ]: 23520714 : total_length -= sg_ele->length;
1508 [ # # # # ]: 23520714 : wr->num_sge++;
1509 : :
1510 [ + + # # : 23520714 : if (rdma_req->offset == iov->iov_len) {
# # # # #
# ]
1511 [ # # # # ]: 22976017 : rdma_req->offset = 0;
1512 [ # # ]: 22976017 : rdma_req->iovpos++;
1513 : 0 : }
1514 : : }
1515 : :
1516 [ - + ]: 6003890 : if (spdk_unlikely(total_length)) {
1517 : 0 : SPDK_ERRLOG("Not enough SG entries to hold data buffer\n");
1518 : 0 : return -EINVAL;
1519 : : }
1520 : :
1521 : 6003890 : return 0;
1522 : 0 : }
1523 : :
1524 : : static int
1525 : 30 : nvmf_rdma_fill_wr_sgl_with_dif(struct spdk_nvmf_rdma_device *device,
1526 : : struct spdk_nvmf_rdma_request *rdma_req,
1527 : : struct ibv_send_wr *wr,
1528 : : uint32_t total_length,
1529 : : uint32_t num_extra_wrs)
1530 : : {
1531 : 30 : struct spdk_rdma_utils_memory_translation mem_translation;
1532 [ # # # # : 30 : struct spdk_dif_ctx *dif_ctx = &rdma_req->req.dif.dif_ctx;
# # ]
1533 : : struct ibv_sge *sg_ele;
1534 : : struct iovec *iov;
1535 : : struct iovec *rdma_iov;
1536 : : uint32_t lkey, remaining;
1537 : : uint32_t remaining_data_block, data_block_size, md_size;
1538 : : uint32_t sge_len;
1539 : : int rc;
1540 : :
1541 [ # # # # : 30 : data_block_size = dif_ctx->block_size - dif_ctx->md_size;
# # # # ]
1542 : :
1543 [ + + # # : 30 : if (spdk_likely(!rdma_req->req.stripped_data)) {
# # # # ]
1544 [ # # # # ]: 15 : rdma_iov = rdma_req->req.iov;
1545 : 15 : remaining_data_block = data_block_size;
1546 [ # # # # ]: 15 : md_size = dif_ctx->md_size;
1547 : 0 : } else {
1548 [ # # # # : 15 : rdma_iov = rdma_req->req.stripped_data->iov;
# # # # ]
1549 [ - + # # : 15 : total_length = total_length / dif_ctx->block_size * data_block_size;
# # ]
1550 : 15 : remaining_data_block = total_length;
1551 : 15 : md_size = 0;
1552 : : }
1553 : :
1554 [ # # # # ]: 30 : wr->num_sge = 0;
1555 : :
1556 [ + + + + : 75 : while (total_length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) {
+ - # # #
# ]
1557 [ # # # # : 45 : iov = rdma_iov + rdma_req->iovpos;
# # ]
1558 [ # # # # : 45 : rc = spdk_rdma_utils_get_translation(device->map, iov->iov_base, iov->iov_len, &mem_translation);
# # # # #
# # # ]
1559 [ - + ]: 45 : if (spdk_unlikely(rc)) {
1560 : 0 : return rc;
1561 : : }
1562 : :
1563 : 45 : lkey = spdk_rdma_utils_memory_translation_get_lkey(&mem_translation);
1564 [ # # # # : 45 : sg_ele = &wr->sg_list[wr->num_sge];
# # # # #
# ]
1565 [ # # # # : 45 : remaining = spdk_min((uint32_t)iov->iov_len - rdma_req->offset, total_length);
# # # # #
# # # # #
# # # # ]
1566 : :
1567 [ + + ]: 159 : while (remaining) {
1568 [ + + # # : 114 : if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) {
# # ]
1569 [ + - + - : 3 : if (num_extra_wrs > 0 && wr->next) {
# # # # ]
1570 [ # # # # ]: 3 : wr = wr->next;
1571 [ # # # # ]: 3 : wr->num_sge = 0;
1572 [ # # # # : 3 : sg_ele = &wr->sg_list[wr->num_sge];
# # # # #
# ]
1573 : 3 : num_extra_wrs--;
1574 : 0 : } else {
1575 : 0 : break;
1576 : : }
1577 : 0 : }
1578 [ # # # # ]: 114 : sg_ele->lkey = lkey;
1579 [ # # # # : 114 : sg_ele->addr = (uintptr_t)((char *)iov->iov_base + rdma_req->offset);
# # # # #
# # # #
# ]
1580 [ # # ]: 114 : sge_len = spdk_min(remaining, remaining_data_block);
1581 [ # # # # ]: 114 : sg_ele->length = sge_len;
1582 [ - + - + : 114 : SPDK_DEBUGLOG(rdma, "sge[%d] %p addr 0x%"PRIx64", len %u\n", wr->num_sge, sg_ele,
# # # # #
# # # # #
# # # # ]
1583 : : sg_ele->addr, sg_ele->length);
1584 : 114 : remaining -= sge_len;
1585 : 114 : remaining_data_block -= sge_len;
1586 [ # # # # ]: 114 : rdma_req->offset += sge_len;
1587 : 114 : total_length -= sge_len;
1588 : :
1589 [ # # ]: 114 : sg_ele++;
1590 [ # # # # ]: 114 : wr->num_sge++;
1591 : :
1592 [ + + ]: 114 : if (remaining_data_block == 0) {
1593 : : /* skip metadata */
1594 [ # # # # ]: 102 : rdma_req->offset += md_size;
1595 : 102 : total_length -= md_size;
1596 : : /* Metadata that do not fit this IO buffer will be included in the next IO buffer */
1597 [ # # ]: 102 : remaining -= spdk_min(remaining, md_size);
1598 : 102 : remaining_data_block = data_block_size;
1599 : 0 : }
1600 : :
1601 [ + + ]: 114 : if (remaining == 0) {
1602 : : /* By subtracting the size of the last IOV from the offset, we ensure that we skip
1603 : : the remaining metadata bits at the beginning of the next buffer */
1604 [ # # # # : 45 : rdma_req->offset -= spdk_min(iov->iov_len, rdma_req->offset);
# # # # #
# # # # #
# # # # #
# # # ]
1605 [ # # ]: 45 : rdma_req->iovpos++;
1606 : 0 : }
1607 : : }
1608 : : }
1609 : :
1610 [ - + ]: 30 : if (spdk_unlikely(total_length)) {
1611 : 0 : SPDK_ERRLOG("Not enough SG entries to hold data buffer\n");
1612 : 0 : return -EINVAL;
1613 : : }
1614 : :
1615 : 30 : return 0;
1616 : 0 : }
1617 : :
1618 : : static inline uint32_t
1619 : 24 : nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size)
1620 : : {
1621 : : /* estimate the number of SG entries and WRs needed to process the request */
1622 : 24 : uint32_t num_sge = 0;
1623 : : uint32_t i;
1624 [ - + ]: 24 : uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size);
1625 : :
1626 [ + + + - ]: 69 : for (i = 0; i < num_buffers && length > 0; i++) {
1627 [ # # ]: 45 : uint32_t buffer_len = spdk_min(length, io_unit_size);
1628 [ - + ]: 45 : uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size);
1629 : :
1630 [ + + ]: 45 : if (num_sge_in_block * block_size > buffer_len) {
1631 : 33 : ++num_sge_in_block;
1632 : 0 : }
1633 : 45 : num_sge += num_sge_in_block;
1634 : 45 : length -= buffer_len;
1635 : 0 : }
1636 [ # # ]: 24 : return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES);
1637 : : }
1638 : :
1639 : : static int
1640 : 6677124 : nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport,
1641 : : struct spdk_nvmf_rdma_device *device,
1642 : : struct spdk_nvmf_rdma_request *rdma_req)
1643 : : {
1644 : : struct spdk_nvmf_rdma_qpair *rqpair;
1645 : : struct spdk_nvmf_rdma_poll_group *rgroup;
1646 [ # # ]: 6677124 : struct spdk_nvmf_request *req = &rdma_req->req;
1647 [ # # # # ]: 6677124 : struct ibv_send_wr *wr = &rdma_req->data.wr;
1648 : : int rc;
1649 : 6677124 : uint32_t num_wrs = 1;
1650 : : uint32_t length;
1651 : :
1652 [ # # # # ]: 6677124 : rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair);
1653 [ # # # # : 6677124 : rgroup = rqpair->poller->group;
# # # # ]
1654 : :
1655 : : /* rdma wr specifics */
1656 : 6677124 : nvmf_rdma_setup_request(rdma_req);
1657 : :
1658 [ # # # # ]: 6677124 : length = req->length;
1659 [ + + # # : 6677124 : if (spdk_unlikely(req->dif_enabled)) {
# # # # ]
1660 [ # # # # : 24 : req->dif.orig_length = length;
# # ]
1661 [ # # # # ]: 24 : length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx);
1662 [ # # # # : 24 : req->dif.elba_length = length;
# # ]
1663 : 0 : }
1664 : :
1665 [ # # # # ]: 6677124 : rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport,
1666 : 0 : length);
1667 [ + + ]: 6677124 : if (spdk_unlikely(rc != 0)) {
1668 : 2113084 : return rc;
1669 : : }
1670 : :
1671 [ - + # # : 4564040 : assert(req->iovcnt <= rqpair->max_send_sge);
# # # # #
# # # ]
1672 : :
1673 : : /* When dif_insert_or_strip is true and the I/O data length is greater than one block,
1674 : : * the stripped_buffers are got for DIF stripping. */
1675 [ + + + - : 4564040 : if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST)
+ + + + #
# # # # #
# # # # #
# # # # #
# # # # #
# ]
1676 : : && (req->dif.elba_length > req->dif.dif_ctx.block_size))) {
1677 [ # # ]: 21 : rc = nvmf_request_get_stripped_buffers(req, &rgroup->group,
1678 [ # # # # : 0 : &rtransport->transport, req->dif.orig_length);
# # # # ]
1679 [ + + ]: 21 : if (rc != 0) {
1680 [ - + - + : 12 : SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc);
# # ]
1681 : 0 : }
1682 : 0 : }
1683 : :
1684 [ # # # # ]: 4564040 : rdma_req->iovpos = 0;
1685 : :
1686 [ + + # # : 4564040 : if (spdk_unlikely(req->dif_enabled)) {
# # # # ]
1687 [ # # # # : 24 : num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size,
# # # # ]
1688 [ # # # # : 0 : req->dif.dif_ctx.block_size);
# # # # ]
1689 [ + + ]: 24 : if (num_wrs > 1) {
1690 : 3 : rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1);
1691 [ - + ]: 3 : if (spdk_unlikely(rc != 0)) {
1692 : 0 : goto err_exit;
1693 : : }
1694 : 0 : }
1695 : :
1696 : 24 : rc = nvmf_rdma_fill_wr_sgl_with_dif(device, rdma_req, wr, length, num_wrs - 1);
1697 [ - + ]: 24 : if (spdk_unlikely(rc != 0)) {
1698 : 0 : goto err_exit;
1699 : : }
1700 : :
1701 [ + + ]: 24 : if (num_wrs > 1) {
1702 : 3 : nvmf_rdma_update_remote_addr(rdma_req, num_wrs);
1703 : 0 : }
1704 : 0 : } else {
1705 : 4564016 : rc = nvmf_rdma_fill_wr_sgl(device, rdma_req, wr, length);
1706 [ - + ]: 4564016 : if (spdk_unlikely(rc != 0)) {
1707 : 0 : goto err_exit;
1708 : : }
1709 : : }
1710 : :
1711 : : /* set the number of outstanding data WRs for this request. */
1712 [ # # # # ]: 4564040 : rdma_req->num_outstanding_data_wr = num_wrs;
1713 : :
1714 : 4564040 : return rc;
1715 : :
1716 : 0 : err_exit:
1717 [ # # # # ]: 0 : spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport);
1718 : 0 : nvmf_rdma_request_free_data(rdma_req, rtransport);
1719 [ # # # # ]: 0 : req->iovcnt = 0;
1720 : 0 : return rc;
1721 : 0 : }
1722 : :
1723 : : static int
1724 : 1243416 : nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport,
1725 : : struct spdk_nvmf_rdma_device *device,
1726 : : struct spdk_nvmf_rdma_request *rdma_req)
1727 : : {
1728 : : struct spdk_nvmf_rdma_qpair *rqpair;
1729 : : struct spdk_nvmf_rdma_poll_group *rgroup;
1730 : : struct ibv_send_wr *current_wr;
1731 [ # # ]: 1243416 : struct spdk_nvmf_request *req = &rdma_req->req;
1732 : : struct spdk_nvme_sgl_descriptor *inline_segment, *desc;
1733 : : uint32_t num_sgl_descriptors;
1734 : 1243416 : uint32_t lengths[SPDK_NVMF_MAX_SGL_ENTRIES], total_length = 0;
1735 : : uint32_t i;
1736 : : int rc;
1737 : :
1738 [ # # # # : 1243416 : rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
# # ]
1739 [ # # # # : 1243416 : rgroup = rqpair->poller->group;
# # # # ]
1740 : :
1741 [ # # # # : 1243416 : inline_segment = &req->cmd->nvme_cmd.dptr.sgl1;
# # # # #
# ]
1742 [ - + # # : 1243416 : assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT);
# # # # #
# ]
1743 [ - + # # : 1243416 : assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET);
# # # # #
# ]
1744 : :
1745 [ # # # # : 1243416 : num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor);
# # # # #
# ]
1746 [ - + # # ]: 1243416 : assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES);
1747 : :
1748 [ # # # # : 1243416 : desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address;
# # # # #
# # # #
# ]
1749 [ + + ]: 14518955 : for (i = 0; i < num_sgl_descriptors; i++) {
1750 [ + + # # : 13275539 : if (spdk_likely(!req->dif_enabled)) {
# # # # ]
1751 [ # # # # : 13275533 : lengths[i] = desc->keyed.length;
# # # # #
# # # ]
1752 : 0 : } else {
1753 [ # # # # : 6 : req->dif.orig_length += desc->keyed.length;
# # # # #
# # # ]
1754 [ # # # # : 6 : lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx);
# # # # #
# # # # #
# # ]
1755 [ # # # # : 6 : req->dif.elba_length += lengths[i];
# # # # #
# # # ]
1756 : : }
1757 [ # # # # : 13275539 : total_length += lengths[i];
# # ]
1758 [ # # ]: 13275539 : desc++;
1759 : 0 : }
1760 : :
1761 [ - + # # : 1243416 : if (spdk_unlikely(total_length > rtransport->transport.opts.max_io_size)) {
# # # # #
# ]
1762 [ # # # # : 0 : SPDK_ERRLOG("Multi SGL length 0x%x exceeds max io size 0x%x\n",
# # # # ]
1763 : : total_length, rtransport->transport.opts.max_io_size);
1764 [ # # # # : 0 : req->rsp->nvme_cpl.status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
# # # # #
# # # ]
1765 : 0 : return -EINVAL;
1766 : : }
1767 : :
1768 : 1243416 : rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1);
1769 [ + + ]: 1243416 : if (spdk_unlikely(rc != 0)) {
1770 : 1075969 : return -ENOMEM;
1771 : : }
1772 : :
1773 [ # # # # ]: 167447 : rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport, total_length);
1774 [ - + ]: 167447 : if (spdk_unlikely(rc != 0)) {
1775 : 0 : nvmf_rdma_request_free_data(rdma_req, rtransport);
1776 : 0 : return rc;
1777 : : }
1778 : :
1779 : : /* When dif_insert_or_strip is true and the I/O data length is greater than one block,
1780 : : * the stripped_buffers are got for DIF stripping. */
1781 [ + + + - : 167447 : if (spdk_unlikely(req->dif_enabled && (req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST)
+ + + - #
# # # # #
# # # # #
# # # # #
# # # # #
# ]
1782 : : && (req->dif.elba_length > req->dif.dif_ctx.block_size))) {
1783 [ # # ]: 3 : rc = nvmf_request_get_stripped_buffers(req, &rgroup->group,
1784 [ # # # # : 0 : &rtransport->transport, req->dif.orig_length);
# # # # ]
1785 [ - + ]: 3 : if (spdk_unlikely(rc != 0)) {
1786 [ # # # # : 0 : SPDK_INFOLOG(rdma, "Get stripped buffers fail %d, fallback to req.iov.\n", rc);
# # ]
1787 : 0 : }
1788 : 0 : }
1789 : :
1790 : : /* The first WR must always be the embedded data WR. This is how we unwind them later. */
1791 [ # # # # ]: 167447 : current_wr = &rdma_req->data.wr;
1792 [ - + # # ]: 167447 : assert(current_wr != NULL);
1793 : :
1794 [ # # # # ]: 167447 : req->length = 0;
1795 [ # # # # ]: 167447 : rdma_req->iovpos = 0;
1796 [ # # # # : 167447 : desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address;
# # # # #
# # # #
# ]
1797 [ + + ]: 1607327 : for (i = 0; i < num_sgl_descriptors; i++) {
1798 : : /* The descriptors must be keyed data block descriptors with an address, not an offset. */
1799 [ + - - + : 1439880 : if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK ||
# # # # #
# # # # #
# # ]
1800 : : desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) {
1801 : 0 : rc = -EINVAL;
1802 : 0 : goto err_exit;
1803 : : }
1804 : :
1805 [ + + # # : 1439880 : if (spdk_likely(!req->dif_enabled)) {
# # # # ]
1806 [ # # # # : 1439874 : rc = nvmf_rdma_fill_wr_sgl(device, rdma_req, current_wr, lengths[i]);
# # ]
1807 : 0 : } else {
1808 : 6 : rc = nvmf_rdma_fill_wr_sgl_with_dif(device, rdma_req, current_wr,
1809 [ # # # # : 4 : lengths[i], 0);
# # ]
1810 : : }
1811 [ - + ]: 1439880 : if (spdk_unlikely(rc != 0)) {
1812 : 0 : rc = -ENOMEM;
1813 : 0 : goto err_exit;
1814 : : }
1815 : :
1816 [ # # # # : 1439880 : req->length += desc->keyed.length;
# # # # #
# ]
1817 [ # # # # : 1439880 : current_wr->wr.rdma.rkey = desc->keyed.key;
# # # # #
# # # #
# ]
1818 [ # # # # : 1439880 : current_wr->wr.rdma.remote_addr = desc->address;
# # # # #
# # # ]
1819 [ # # # # ]: 1439880 : current_wr = current_wr->next;
1820 [ # # ]: 1439880 : desc++;
1821 : 0 : }
1822 : :
1823 : : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
1824 : : /* Go back to the last descriptor in the list. */
1825 [ # # ]: 167447 : desc--;
1826 [ + + # # : 167447 : if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) {
# # # # ]
1827 [ - + # # : 167432 : if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) {
# # # # ]
1828 [ # # # # : 0 : rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV;
# # # # ]
1829 [ # # # # : 0 : rdma_req->rsp.wr.imm_data = desc->keyed.key;
# # # # #
# # # # #
# # ]
1830 : 0 : }
1831 : 0 : }
1832 : : #endif
1833 : :
1834 [ # # # # ]: 167447 : rdma_req->num_outstanding_data_wr = num_sgl_descriptors;
1835 : :
1836 : 167447 : return 0;
1837 : :
1838 : 0 : err_exit:
1839 [ # # # # ]: 0 : spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport);
1840 : 0 : nvmf_rdma_request_free_data(rdma_req, rtransport);
1841 : 0 : return rc;
1842 : 0 : }
1843 : :
1844 : : static int
1845 : 9853140 : nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport,
1846 : : struct spdk_nvmf_rdma_device *device,
1847 : : struct spdk_nvmf_rdma_request *rdma_req)
1848 : : {
1849 [ # # ]: 9853140 : struct spdk_nvmf_request *req = &rdma_req->req;
1850 : : struct spdk_nvme_cpl *rsp;
1851 : : struct spdk_nvme_sgl_descriptor *sgl;
1852 : : int rc;
1853 : : uint32_t length;
1854 : :
1855 [ # # # # : 9853140 : rsp = &req->rsp->nvme_cpl;
# # ]
1856 [ # # # # : 9853140 : sgl = &req->cmd->nvme_cmd.dptr.sgl1;
# # # # #
# ]
1857 : :
1858 [ + + # # : 9853140 : if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK &&
# # # # #
# ]
1859 [ + + # # : 6677127 : (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS ||
# # # # ]
1860 [ + - # # : 3301153 : sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) {
# # ]
1861 : :
1862 [ # # # # : 6677127 : length = sgl->keyed.length;
# # ]
1863 [ + + # # : 6677127 : if (spdk_unlikely(length > rtransport->transport.opts.max_io_size)) {
# # # # #
# ]
1864 [ # # # # : 3 : SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n",
# # # # ]
1865 : : length, rtransport->transport.opts.max_io_size);
1866 [ # # # # : 3 : rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
# # ]
1867 : 3 : return -1;
1868 : : }
1869 : : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
1870 [ + + # # : 6677124 : if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) {
# # # # ]
1871 [ + + # # : 6677076 : if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) {
# # # # ]
1872 [ # # # # : 3301153 : rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV;
# # # # ]
1873 [ # # # # : 3301153 : rdma_req->rsp.wr.imm_data = sgl->keyed.key;
# # # # #
# # # # #
# # ]
1874 : 0 : }
1875 : 0 : }
1876 : : #endif
1877 : :
1878 : : /* fill request length and populate iovs */
1879 [ # # # # ]: 6677124 : req->length = length;
1880 : :
1881 : 6677124 : rc = nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req);
1882 [ + + ]: 6677124 : if (spdk_unlikely(rc < 0)) {
1883 [ - + ]: 2113084 : if (rc == -EINVAL) {
1884 : 0 : SPDK_ERRLOG("SGL length exceeds the max I/O size\n");
1885 [ # # # # : 0 : rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
# # ]
1886 : 0 : return -1;
1887 : : }
1888 : : /* No available buffers. Queue this request up. */
1889 [ - + - + : 2113084 : SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req);
# # ]
1890 : 2113084 : return 0;
1891 : : }
1892 : :
1893 [ - + - + : 4564040 : SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req,
# # # # #
# ]
1894 : : req->iovcnt);
1895 : :
1896 : 4564040 : return 0;
1897 [ + + # # : 3176013 : } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK &&
# # # # #
# ]
1898 [ + - # # : 1932597 : sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) {
# # ]
1899 [ # # # # ]: 1932597 : uint64_t offset = sgl->address;
1900 [ # # # # : 1932597 : uint32_t max_len = rtransport->transport.opts.in_capsule_data_size;
# # # # ]
1901 : :
1902 [ - + - + : 1932597 : SPDK_DEBUGLOG(nvmf, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n",
# # # # #
# # # #
# ]
1903 : : offset, sgl->unkeyed.length);
1904 : :
1905 [ - + ]: 1932597 : if (spdk_unlikely(offset > max_len)) {
1906 : 0 : SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n",
1907 : : offset, max_len);
1908 [ # # # # : 0 : rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET;
# # ]
1909 : 0 : return -1;
1910 : : }
1911 : 1932597 : max_len -= (uint32_t)offset;
1912 : :
1913 [ + + # # : 1932597 : if (spdk_unlikely(sgl->unkeyed.length > max_len)) {
# # # # #
# ]
1914 [ # # # # : 6 : SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n",
# # # # ]
1915 : : sgl->unkeyed.length, max_len);
1916 [ # # # # : 6 : rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
# # ]
1917 : 6 : return -1;
1918 : : }
1919 : :
1920 [ # # # # ]: 1932591 : rdma_req->num_outstanding_data_wr = 0;
1921 [ # # # # : 1932591 : req->data_from_pool = false;
# # ]
1922 [ # # # # : 1932591 : req->length = sgl->unkeyed.length;
# # # # #
# # # ]
1923 : :
1924 [ # # # # : 1932591 : req->iov[0].iov_base = rdma_req->recv->buf + offset;
# # # # #
# # # # #
# # # # #
# ]
1925 [ # # # # : 1932591 : req->iov[0].iov_len = req->length;
# # # # #
# # # #
# ]
1926 [ # # # # ]: 1932591 : req->iovcnt = 1;
1927 : :
1928 : 1932591 : return 0;
1929 [ + - # # : 1243416 : } else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT &&
# # # # #
# ]
1930 [ + - # # : 1243416 : sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) {
# # ]
1931 : :
1932 : 1243416 : rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req);
1933 [ + + ]: 1243416 : if (spdk_unlikely(rc == -ENOMEM)) {
1934 [ - + - + : 1075969 : SPDK_DEBUGLOG(rdma, "No available large data buffers. Queueing request %p\n", rdma_req);
# # ]
1935 : 1075969 : return 0;
1936 [ - + ]: 167447 : } else if (spdk_unlikely(rc == -EINVAL)) {
1937 : 0 : SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n");
1938 [ # # # # : 0 : rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
# # ]
1939 : 0 : return -1;
1940 : : }
1941 : :
1942 [ - + - + : 167447 : SPDK_DEBUGLOG(rdma, "Request %p took %d buffer/s from central pool\n", rdma_req,
# # # # #
# ]
1943 : : req->iovcnt);
1944 : :
1945 : 167447 : return 0;
1946 : : }
1947 : :
1948 [ # # # # : 0 : SPDK_ERRLOG("Invalid NVMf I/O Command SGL: Type 0x%x, Subtype 0x%x\n",
# # # # #
# # # ]
1949 : : sgl->generic.type, sgl->generic.subtype);
1950 [ # # # # : 0 : rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID;
# # ]
1951 : 0 : return -1;
1952 : 0 : }
1953 : :
1954 : : static void
1955 : 7924531 : _nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req,
1956 : : struct spdk_nvmf_rdma_transport *rtransport)
1957 : : {
1958 : : struct spdk_nvmf_rdma_qpair *rqpair;
1959 : : struct spdk_nvmf_rdma_poll_group *rgroup;
1960 : :
1961 [ # # # # : 7924531 : rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
# # ]
1962 [ + + # # : 7924531 : if (rdma_req->req.data_from_pool) {
# # # # #
# ]
1963 [ # # # # : 4731442 : rgroup = rqpair->poller->group;
# # # # ]
1964 : :
1965 [ # # # # : 4731442 : spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport);
# # ]
1966 : 0 : }
1967 [ - + # # : 7924531 : if (rdma_req->req.stripped_data) {
# # # # ]
1968 [ # # ]: 0 : nvmf_request_free_stripped_buffers(&rdma_req->req,
1969 [ # # # # : 0 : &rqpair->poller->group->group,
# # # # #
# ]
1970 [ # # ]: 0 : &rtransport->transport);
1971 : 0 : }
1972 : 7924531 : nvmf_rdma_request_free_data(rdma_req, rtransport);
1973 [ # # # # : 7924531 : rdma_req->req.length = 0;
# # ]
1974 [ # # # # : 7924531 : rdma_req->req.iovcnt = 0;
# # ]
1975 [ # # # # ]: 7924531 : rdma_req->offset = 0;
1976 [ # # # # : 7924531 : rdma_req->req.dif_enabled = false;
# # # # ]
1977 [ # # # # ]: 7924531 : rdma_req->fused_failed = false;
1978 [ # # # # ]: 7924531 : rdma_req->transfer_wr = NULL;
1979 [ - + # # : 7924531 : if (rdma_req->fused_pair) {
# # ]
1980 : : /* This req was part of a valid fused pair, but failed before it got to
1981 : : * READ_TO_EXECUTE state. This means we need to fail the other request
1982 : : * in the pair, because it is no longer part of a valid pair. If the pair
1983 : : * already reached READY_TO_EXECUTE state, we need to kick it.
1984 : : */
1985 [ # # # # : 0 : rdma_req->fused_pair->fused_failed = true;
# # # # ]
1986 [ # # # # : 0 : if (rdma_req->fused_pair->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) {
# # # # #
# ]
1987 [ # # # # ]: 0 : nvmf_rdma_request_process(rtransport, rdma_req->fused_pair);
1988 : 0 : }
1989 [ # # # # ]: 0 : rdma_req->fused_pair = NULL;
1990 : 0 : }
1991 [ - + # # : 7924531 : memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif));
# # ]
1992 [ # # ]: 7924531 : rqpair->qd--;
1993 : :
1994 [ + + # # : 7924531 : STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # ]
1995 [ # # # # ]: 7924531 : rqpair->qpair.queue_depth--;
1996 [ # # # # ]: 7924531 : rdma_req->state = RDMA_REQUEST_STATE_FREE;
1997 : 7924531 : }
1998 : :
1999 : : static void
2000 : 7924114 : nvmf_rdma_check_fused_ordering(struct spdk_nvmf_rdma_transport *rtransport,
2001 : : struct spdk_nvmf_rdma_qpair *rqpair,
2002 : : struct spdk_nvmf_rdma_request *rdma_req)
2003 : : {
2004 : : enum spdk_nvme_cmd_fuse last, next;
2005 : :
2006 [ + + # # : 7924114 : last = rqpair->fused_first ? rqpair->fused_first->req.cmd->nvme_cmd.fuse : SPDK_NVME_CMD_FUSE_NONE;
# # # # #
# # # # #
# # # # #
# ]
2007 [ # # # # : 7924114 : next = rdma_req->req.cmd->nvme_cmd.fuse;
# # # # #
# ]
2008 : :
2009 [ - + # # ]: 7924114 : assert(last != SPDK_NVME_CMD_FUSE_SECOND);
2010 : :
2011 [ + + + + ]: 7924114 : if (spdk_likely(last == SPDK_NVME_CMD_FUSE_NONE && next == SPDK_NVME_CMD_FUSE_NONE)) {
2012 : 7922056 : return;
2013 : : }
2014 : :
2015 [ + + ]: 2058 : if (last == SPDK_NVME_CMD_FUSE_FIRST) {
2016 [ + - ]: 1029 : if (next == SPDK_NVME_CMD_FUSE_SECOND) {
2017 : : /* This is a valid pair of fused commands. Point them at each other
2018 : : * so they can be submitted consecutively once ready to be executed.
2019 : : */
2020 [ # # # # : 1029 : rqpair->fused_first->fused_pair = rdma_req;
# # # # ]
2021 [ # # # # : 1029 : rdma_req->fused_pair = rqpair->fused_first;
# # # # ]
2022 [ # # # # ]: 1029 : rqpair->fused_first = NULL;
2023 : 1029 : return;
2024 : : } else {
2025 : : /* Mark the last req as failed since it wasn't followed by a SECOND. */
2026 [ # # # # : 0 : rqpair->fused_first->fused_failed = true;
# # # # ]
2027 : :
2028 : : /* If the last req is in READY_TO_EXECUTE state, then call
2029 : : * nvmf_rdma_request_process(), otherwise nothing else will kick it.
2030 : : */
2031 [ # # # # : 0 : if (rqpair->fused_first->state == RDMA_REQUEST_STATE_READY_TO_EXECUTE) {
# # # # #
# ]
2032 [ # # # # ]: 0 : nvmf_rdma_request_process(rtransport, rqpair->fused_first);
2033 : 0 : }
2034 : :
2035 [ # # # # ]: 0 : rqpair->fused_first = NULL;
2036 : : }
2037 : 0 : }
2038 : :
2039 [ + - ]: 1029 : if (next == SPDK_NVME_CMD_FUSE_FIRST) {
2040 : : /* Set rqpair->fused_first here so that we know to check that the next request
2041 : : * is a SECOND (and to fail this one if it isn't).
2042 : : */
2043 [ # # # # ]: 1029 : rqpair->fused_first = rdma_req;
2044 [ # # ]: 0 : } else if (next == SPDK_NVME_CMD_FUSE_SECOND) {
2045 : : /* Mark this req failed since it ia SECOND and the last one was not a FIRST. */
2046 [ # # # # ]: 0 : rdma_req->fused_failed = true;
2047 : 0 : }
2048 : 0 : }
2049 : :
2050 : : static void
2051 : 6664507 : nvmf_rdma_poll_group_insert_need_buffer_req(struct spdk_nvmf_rdma_poll_group *rgroup,
2052 : : struct spdk_nvmf_rdma_request *rdma_req)
2053 : : {
2054 : : struct spdk_nvmf_request *r;
2055 : :
2056 : : /* CONNECT commands have a timeout, so we need to avoid a CONNECT command
2057 : : * from getting buried behind a long list of other non-FABRIC requests
2058 : : * waiting for a buffer. Note that even though the CONNECT command's data is
2059 : : * in-capsule, the request still goes to this STAILQ.
2060 : : */
2061 [ + + # # : 6664507 : if (spdk_likely(rdma_req->req.cmd->nvme_cmd.opc != SPDK_NVME_OPC_FABRIC)) {
# # # # #
# # # ]
2062 : : /* This is the most likely case. */
2063 [ # # # # : 6660759 : STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # #
# ]
2064 : 6660759 : return;
2065 : : } else {
2066 : : /* STAILQ doesn't have INSERT_BEFORE, so we need to either INSERT_HEAD
2067 : : * or INSERT_AFTER. Put it after any other FABRIC commands that are
2068 : : * already in the queue.
2069 : : */
2070 [ # # # # : 3748 : r = STAILQ_FIRST(&rgroup->group.pending_buf_queue);
# # # # ]
2071 [ - + - - : 3748 : if (r == NULL || r->cmd->nvme_cmd.opc != SPDK_NVME_OPC_FABRIC) {
# # # # #
# # # ]
2072 [ + - # # : 3748 : STAILQ_INSERT_HEAD(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # ]
2073 : 3748 : return;
2074 : : }
2075 : 0 : while (true) {
2076 : : struct spdk_nvmf_request *next;
2077 : :
2078 [ # # # # : 0 : next = STAILQ_NEXT(r, buf_link);
# # ]
2079 [ # # # # : 0 : if (next == NULL || next->cmd->nvme_cmd.opc != SPDK_NVME_OPC_FABRIC) {
# # # # #
# # # ]
2080 [ # # # # : 0 : STAILQ_INSERT_AFTER(&rgroup->group.pending_buf_queue, r, &rdma_req->req, buf_link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # ]
2081 : 0 : return;
2082 : : }
2083 : 0 : r = next;
2084 : : }
2085 : : }
2086 : 0 : }
2087 : :
2088 : : bool
2089 : 32012967 : nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport,
2090 : : struct spdk_nvmf_rdma_request *rdma_req)
2091 : : {
2092 : : struct spdk_nvmf_rdma_qpair *rqpair;
2093 : : struct spdk_nvmf_rdma_device *device;
2094 : : struct spdk_nvmf_rdma_poll_group *rgroup;
2095 [ # # # # : 32012967 : struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl;
# # # # ]
2096 : : int rc;
2097 : : struct spdk_nvmf_rdma_recv *rdma_recv;
2098 : : enum spdk_nvmf_rdma_request_state prev_state;
2099 : 32012967 : bool progress = false;
2100 : 69 : int data_posted;
2101 : : uint32_t num_blocks, num_rdma_reads_available, qdepth;
2102 : :
2103 [ # # # # : 32012967 : rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
# # ]
2104 [ # # # # ]: 32012967 : device = rqpair->device;
2105 [ # # # # : 32012967 : rgroup = rqpair->poller->group;
# # # # ]
2106 : :
2107 [ - + # # : 32012967 : assert(rdma_req->state != RDMA_REQUEST_STATE_FREE);
# # # # ]
2108 : :
2109 : : /* If the queue pair is in an error state, force the request to the completed state
2110 : : * to release resources. */
2111 [ + + + + : 32012967 : if (spdk_unlikely(rqpair->ibv_in_error_state || !spdk_nvmf_qpair_is_active(&rqpair->qpair))) {
+ + # # #
# # # ]
2112 [ + + - - : 1923 : switch (rdma_req->state) {
- + # # #
# ]
2113 : 477 : case RDMA_REQUEST_STATE_NEED_BUFFER:
2114 [ + + + + : 52716 : STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link);
+ + + + #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # ]
2115 : 477 : break;
2116 : 478 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING:
2117 [ + + + + : 478 : STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
- + + + #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # ]
2118 : 478 : break;
2119 : 0 : case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
2120 [ # # # # : 0 : if (rdma_req->num_remaining_data_wr) {
# # ]
2121 : : /* Partially sent request is still in the pending_rdma_read_queue,
2122 : : * remove it before completing */
2123 [ # # # # ]: 0 : rdma_req->num_remaining_data_wr = 0;
2124 [ # # # # : 0 : STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # ]
2125 : 0 : }
2126 : 0 : break;
2127 : 0 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING:
2128 [ # # # # : 0 : STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # ]
2129 : 0 : break;
2130 : 0 : case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING:
2131 [ # # # # : 0 : STAILQ_REMOVE(&rqpair->pending_rdma_send_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # ]
2132 : 0 : break;
2133 : 968 : default:
2134 : 968 : break;
2135 : : }
2136 [ # # # # ]: 1923 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
2137 : 0 : }
2138 : :
2139 : : /* The loop here is to allow for several back-to-back state changes. */
2140 : 0 : do {
2141 [ # # # # ]: 86995764 : prev_state = rdma_req->state;
2142 : :
2143 [ - + - + : 86995764 : SPDK_DEBUGLOG(rdma, "Request %p entering state %d\n", rdma_req, prev_state);
# # ]
2144 : :
2145 [ + + + + : 86995764 : switch (rdma_req->state) {
+ + + + +
+ + + + +
- # # # #
# ]
2146 : 7923170 : case RDMA_REQUEST_STATE_FREE:
2147 : : /* Some external code must kick a request into RDMA_REQUEST_STATE_NEW
2148 : : * to escape this state. */
2149 : 7923170 : break;
2150 : 7924114 : case RDMA_REQUEST_STATE_NEW:
2151 [ + + + + : 7924114 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0,
# # # # #
# # # # #
# # # # #
# # # ]
2152 : : (uintptr_t)rdma_req, (uintptr_t)rqpair, rqpair->qpair.queue_depth);
2153 [ # # # # ]: 7924114 : rdma_recv = rdma_req->recv;
2154 : :
2155 : : /* The first element of the SGL is the NVMe command */
2156 [ # # # # : 7924114 : rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr;
# # # # #
# # # # #
# # ]
2157 [ - + # # : 7924114 : memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp));
# # # # ]
2158 [ # # # # : 7924114 : rdma_req->transfer_wr = &rdma_req->data.wr;
# # # # ]
2159 : :
2160 [ + + + + : 7924114 : if (spdk_unlikely(rqpair->ibv_in_error_state || !spdk_nvmf_qpair_is_active(&rqpair->qpair))) {
- + # # #
# # # ]
2161 [ # # # # ]: 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
2162 : 0 : break;
2163 : : }
2164 : :
2165 [ - + # # : 7924114 : if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) {
# # # # #
# ]
2166 [ # # # # : 0 : rdma_req->req.dif_enabled = true;
# # # # ]
2167 : 0 : }
2168 : :
2169 : 7924114 : nvmf_rdma_check_fused_ordering(rtransport, rqpair, rdma_req);
2170 : :
2171 : : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
2172 [ # # # # : 7924114 : rdma_req->rsp.wr.opcode = IBV_WR_SEND;
# # # # ]
2173 [ # # # # : 7924114 : rdma_req->rsp.wr.imm_data = 0;
# # # # #
# ]
2174 : : #endif
2175 : :
2176 : : /* The next state transition depends on the data transfer needs of this request. */
2177 [ # # # # : 7924114 : rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req);
# # # # ]
2178 : :
2179 [ + + # # : 7924114 : if (spdk_unlikely(rdma_req->req.xfer == SPDK_NVME_DATA_BIDIRECTIONAL)) {
# # # # ]
2180 [ # # # # : 3 : rsp->status.sct = SPDK_NVME_SCT_GENERIC;
# # ]
2181 [ # # # # : 3 : rsp->status.sc = SPDK_NVME_SC_INVALID_OPCODE;
# # ]
2182 [ # # # # : 3 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
# # # # #
# # # # #
# # # # #
# # # #
# ]
2183 [ # # # # ]: 3 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2184 [ - + - + : 3 : SPDK_DEBUGLOG(rdma, "Request %p: invalid xfer type (BIDIRECTIONAL)\n", rdma_req);
# # ]
2185 : 3 : break;
2186 : : }
2187 : :
2188 : : /* If no data to transfer, ready to execute. */
2189 [ + + # # : 7924111 : if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) {
# # # # ]
2190 [ # # # # ]: 1259604 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
2191 : 1259604 : break;
2192 : : }
2193 : :
2194 [ # # # # ]: 6664507 : rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER;
2195 : 6664507 : nvmf_rdma_poll_group_insert_need_buffer_req(rgroup, rdma_req);
2196 : 6664507 : break;
2197 : 10863401 : case RDMA_REQUEST_STATE_NEED_BUFFER:
2198 [ + + + + : 10863401 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0,
# # # # #
# # # # #
# # ]
2199 : : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2200 : :
2201 [ - + # # : 10863401 : assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE);
# # # # #
# ]
2202 : :
2203 [ + + # # : 10863401 : if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) {
# # # # #
# # # ]
2204 : : /* This request needs to wait in line to obtain a buffer */
2205 : 1010321 : break;
2206 : : }
2207 : :
2208 : : /* Try to get a data buffer */
2209 : 9853080 : rc = nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req);
2210 [ - + ]: 9853080 : if (spdk_unlikely(rc < 0)) {
2211 [ # # # # : 0 : STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # ]
2212 [ # # # # : 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
# # # # #
# # # # #
# # # # #
# # # #
# ]
2213 [ # # # # ]: 0 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2214 : 0 : break;
2215 : : }
2216 : :
2217 [ + + # # : 9853080 : if (rdma_req->req.iovcnt == 0) {
# # # # ]
2218 : : /* No buffers available. */
2219 [ # # # # ]: 3189050 : rgroup->stat.pending_data_buffer++;
2220 : 3189050 : break;
2221 : : }
2222 : :
2223 [ + + # # : 6664030 : STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # ]
2224 : :
2225 : : /* If data is transferring from host to controller and the data didn't
2226 : : * arrive using in capsule data, we need to do a transfer from the host.
2227 : : */
2228 [ + + + + : 6664030 : if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER &&
# # # # #
# ]
2229 [ # # # # : 0 : rdma_req->req.data_from_pool) {
# # # # ]
2230 [ # # # # : 719406 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link);
# # # # #
# # # # #
# # # # #
# # # #
# ]
2231 [ # # # # ]: 719406 : rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING;
2232 : 719406 : break;
2233 : : }
2234 : :
2235 [ # # # # ]: 5944624 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
2236 : 5944624 : break;
2237 : 4030651 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING:
2238 [ + + + + : 4030651 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0,
# # # # #
# # # # #
# # ]
2239 : : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2240 : :
2241 [ + + # # : 4030651 : if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) {
# # # # ]
2242 : : /* This request needs to wait in line to perform RDMA */
2243 : 1789530 : break;
2244 : : }
2245 [ - + # # : 2241121 : assert(rqpair->max_send_depth >= rqpair->current_send_depth);
# # # # #
# # # ]
2246 [ # # # # : 2241121 : qdepth = rqpair->max_send_depth - rqpair->current_send_depth;
# # # # ]
2247 [ - + # # : 2241121 : assert(rqpair->max_read_depth >= rqpair->current_read_depth);
# # # # #
# # # ]
2248 [ # # # # : 2241121 : num_rdma_reads_available = rqpair->max_read_depth - rqpair->current_read_depth;
# # # # ]
2249 [ + + # # : 2241121 : if (rdma_req->num_outstanding_data_wr > qdepth ||
# # # # ]
2250 [ + + # # ]: 2240698 : rdma_req->num_outstanding_data_wr > num_rdma_reads_available) {
2251 [ + + + + ]: 1522193 : if (num_rdma_reads_available && qdepth) {
2252 : : /* Send as much as we can */
2253 [ # # # # ]: 65838 : request_prepare_transfer_in_part(&rdma_req->req, spdk_min(num_rdma_reads_available, qdepth));
2254 : 0 : } else {
2255 : : /* We can only have so many WRs outstanding. we have to wait until some finish. */
2256 [ # # # # : 1456355 : rqpair->poller->stat.pending_rdma_read++;
# # # # ]
2257 : 1456355 : break;
2258 : : }
2259 : 0 : }
2260 : :
2261 : : /* We have already verified that this request is the head of the queue. */
2262 [ + + # # : 784766 : if (rdma_req->num_remaining_data_wr == 0) {
# # ]
2263 [ + + # # : 718928 : STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# ]
2264 : 0 : }
2265 : :
2266 [ # # ]: 784766 : request_transfer_in(&rdma_req->req);
2267 [ # # # # ]: 784766 : rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER;
2268 : :
2269 : 784766 : break;
2270 : 784766 : case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
2271 [ + + + + : 784766 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0,
# # # # #
# # # # #
# # ]
2272 : : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2273 : : /* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE
2274 : : * to escape this state. */
2275 : 784766 : break;
2276 : 7923129 : case RDMA_REQUEST_STATE_READY_TO_EXECUTE:
2277 [ + + + + : 7923129 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0,
# # # # #
# # # # #
# # ]
2278 : : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2279 : :
2280 [ - + # # : 7923129 : if (spdk_unlikely(rdma_req->req.dif_enabled)) {
# # # # #
# ]
2281 [ # # # # : 0 : if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) {
# # # # ]
2282 : : /* generate DIF for write operation */
2283 [ # # # # : 0 : num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# ]
2284 [ # # # # ]: 0 : assert(num_blocks > 0);
2285 : :
2286 [ # # # # : 0 : rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt,
# # # # #
# ]
2287 [ # # # # : 0 : num_blocks, &rdma_req->req.dif.dif_ctx);
# # ]
2288 [ # # ]: 0 : if (rc != 0) {
2289 : 0 : SPDK_ERRLOG("DIF generation failed\n");
2290 [ # # # # ]: 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
2291 [ # # ]: 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
2292 : 0 : break;
2293 : : }
2294 : 0 : }
2295 : :
2296 [ # # # # : 0 : assert(rdma_req->req.dif.elba_length >= rdma_req->req.length);
# # # # #
# # # # #
# # # # ]
2297 : : /* set extended length before IO operation */
2298 [ # # # # : 0 : rdma_req->req.length = rdma_req->req.dif.elba_length;
# # # # #
# # # #
# ]
2299 : 0 : }
2300 : :
2301 [ + + # # : 7923129 : if (rdma_req->req.cmd->nvme_cmd.fuse != SPDK_NVME_CMD_FUSE_NONE) {
# # # # #
# # # ]
2302 [ - + - + : 2058 : if (rdma_req->fused_failed) {
# # # # ]
2303 : : /* This request failed FUSED semantics. Fail it immediately, without
2304 : : * even sending it to the target layer.
2305 : : */
2306 [ # # # # : 0 : rsp->status.sct = SPDK_NVME_SCT_GENERIC;
# # ]
2307 [ # # # # : 0 : rsp->status.sc = SPDK_NVME_SC_ABORTED_MISSING_FUSED;
# # ]
2308 [ # # # # : 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
# # # # #
# # # # #
# # # # #
# # # #
# ]
2309 [ # # # # ]: 0 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2310 : 0 : break;
2311 : : }
2312 : :
2313 [ + + # # : 2058 : if (rdma_req->fused_pair == NULL ||
# # # # ]
2314 [ + - # # : 1029 : rdma_req->fused_pair->state != RDMA_REQUEST_STATE_READY_TO_EXECUTE) {
# # # # ]
2315 : : /* This request is ready to execute, but either we don't know yet if it's
2316 : : * valid - i.e. this is a FIRST but we haven't received the next
2317 : : * request yet or the other request of this fused pair isn't ready to
2318 : : * execute. So break here and this request will get processed later either
2319 : : * when the other request is ready or we find that this request isn't valid.
2320 : : */
2321 : 0 : break;
2322 : : }
2323 : 0 : }
2324 : :
2325 : : /* If we get to this point, and this request is a fused command, we know that
2326 : : * it is part of valid sequence (FIRST followed by a SECOND) and that both
2327 : : * requests are READY_TO_EXECUTE. So call spdk_nvmf_request_exec() both on this
2328 : : * request, and the other request of the fused pair, in the correct order.
2329 : : * Also clear the ->fused_pair pointers on both requests, since after this point
2330 : : * we no longer need to maintain the relationship between these two requests.
2331 : : */
2332 [ + + # # : 7922100 : if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_SECOND) {
# # # # #
# # # ]
2333 [ - + # # : 1029 : assert(rdma_req->fused_pair != NULL);
# # # # ]
2334 [ - + # # : 1029 : assert(rdma_req->fused_pair->fused_pair != NULL);
# # # # #
# # # ]
2335 [ # # # # : 1029 : rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING;
# # # # ]
2336 [ # # # # : 1029 : spdk_nvmf_request_exec(&rdma_req->fused_pair->req);
# # ]
2337 [ # # # # : 1029 : rdma_req->fused_pair->fused_pair = NULL;
# # # # ]
2338 [ # # # # ]: 1029 : rdma_req->fused_pair = NULL;
2339 : 0 : }
2340 [ # # # # ]: 7922100 : rdma_req->state = RDMA_REQUEST_STATE_EXECUTING;
2341 [ # # ]: 7922100 : spdk_nvmf_request_exec(&rdma_req->req);
2342 [ - + # # : 7922100 : if (rdma_req->req.cmd->nvme_cmd.fuse == SPDK_NVME_CMD_FUSE_FIRST) {
# # # # #
# # # ]
2343 [ # # # # : 0 : assert(rdma_req->fused_pair != NULL);
# # # # ]
2344 [ # # # # : 0 : assert(rdma_req->fused_pair->fused_pair != NULL);
# # # # #
# # # ]
2345 [ # # # # : 0 : rdma_req->fused_pair->state = RDMA_REQUEST_STATE_EXECUTING;
# # # # ]
2346 [ # # # # : 0 : spdk_nvmf_request_exec(&rdma_req->fused_pair->req);
# # ]
2347 [ # # # # : 0 : rdma_req->fused_pair->fused_pair = NULL;
# # # # ]
2348 [ # # # # ]: 0 : rdma_req->fused_pair = NULL;
2349 : 0 : }
2350 : 7922100 : break;
2351 : 6533925 : case RDMA_REQUEST_STATE_EXECUTING:
2352 [ + + + + : 6533925 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0,
# # # # #
# # # # #
# # ]
2353 : : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2354 : : /* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED
2355 : : * to escape this state. */
2356 : 6533925 : break;
2357 : 7921537 : case RDMA_REQUEST_STATE_EXECUTED:
2358 [ + + + + : 7921537 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0,
# # # # #
# # # # #
# # ]
2359 : : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2360 [ + + # # : 7921537 : if (rsp->status.sc == SPDK_NVME_SC_SUCCESS &&
# # # # #
# ]
2361 [ + + # # : 6521081 : rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
# # ]
2362 [ # # # # : 3841991 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link);
# # # # #
# # # # #
# # # # #
# # # #
# ]
2363 [ # # # # ]: 3841991 : rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING;
2364 : 0 : } else {
2365 [ # # # # : 4079546 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
# # # # #
# # # # #
# # # # #
# # # #
# ]
2366 [ # # # # ]: 4079546 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2367 : : }
2368 [ - + # # : 7921537 : if (spdk_unlikely(rdma_req->req.dif_enabled)) {
# # # # #
# ]
2369 : : /* restore the original length */
2370 [ # # # # : 0 : rdma_req->req.length = rdma_req->req.dif.orig_length;
# # # # #
# # # #
# ]
2371 : :
2372 [ # # # # : 0 : if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
# # # # ]
2373 : 0 : struct spdk_dif_error error_blk;
2374 : :
2375 [ # # # # : 0 : num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# ]
2376 [ # # # # : 0 : if (!rdma_req->req.stripped_data) {
# # # # ]
2377 [ # # # # : 0 : rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks,
# # # # #
# ]
2378 [ # # # # : 0 : &rdma_req->req.dif.dif_ctx, &error_blk);
# # ]
2379 : 0 : } else {
2380 [ # # # # : 0 : rc = spdk_dif_verify_copy(rdma_req->req.stripped_data->iov,
# # # # ]
2381 [ # # # # : 0 : rdma_req->req.stripped_data->iovcnt,
# # # # #
# ]
2382 [ # # # # : 0 : rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks,
# # # # #
# ]
2383 [ # # # # : 0 : &rdma_req->req.dif.dif_ctx, &error_blk);
# # ]
2384 : : }
2385 [ # # ]: 0 : if (rc) {
2386 [ # # # # : 0 : struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl;
# # # # ]
2387 : :
2388 [ # # ]: 0 : SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type,
2389 : : error_blk.err_offset);
2390 [ # # # # : 0 : rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR;
# # ]
2391 [ # # # # : 0 : rsp->status.sc = nvmf_rdma_dif_error_to_compl_status(error_blk.err_type);
# # ]
2392 [ # # # # : 0 : STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # ]
2393 [ # # # # : 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req, state_link);
# # # # #
# # # # #
# # # # #
# # # #
# ]
2394 [ # # # # ]: 0 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
2395 : 0 : }
2396 : 0 : }
2397 : 0 : }
2398 : 7921537 : break;
2399 : 3857002 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING:
2400 [ + + + + : 3857002 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0,
# # # # #
# # # # #
# # ]
2401 : : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2402 : :
2403 [ + + # # : 3857002 : if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) {
# # # # ]
2404 : : /* This request needs to wait in line to perform RDMA */
2405 : 4455 : break;
2406 : : }
2407 [ # # # # : 3852547 : if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) >
# # # # #
# ]
2408 [ + + # # ]: 3852547 : rqpair->max_send_depth) {
2409 : : /* We can only have so many WRs outstanding. we have to wait until some finish.
2410 : : * +1 since each request has an additional wr in the resp. */
2411 [ # # # # : 10556 : rqpair->poller->stat.pending_rdma_write++;
# # # # ]
2412 : 10556 : break;
2413 : : }
2414 : :
2415 : : /* We have already verified that this request is the head of the queue. */
2416 [ + + # # : 3841991 : STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# ]
2417 : :
2418 : : /* The data transfer will be kicked off from
2419 : : * RDMA_REQUEST_STATE_READY_TO_COMPLETE state.
2420 : : * We verified that data + response fit into send queue, so we can go to the next state directly
2421 : : */
2422 [ # # # # ]: 3841991 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE;
2423 : 3841991 : break;
2424 : 4079644 : case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING:
2425 [ + + + + : 4079644 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING, 0, 0,
# # # # #
# # # # #
# # ]
2426 : : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2427 : :
2428 [ + + # # : 4079644 : if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_send_queue)) {
# # # # ]
2429 : : /* This request needs to wait in line to send the completion */
2430 : 14 : break;
2431 : : }
2432 : :
2433 [ - + # # : 4079630 : assert(rqpair->current_send_depth <= rqpair->max_send_depth);
# # # # #
# # # ]
2434 [ + + # # : 4079630 : if (rqpair->current_send_depth == rqpair->max_send_depth) {
# # # # #
# ]
2435 : : /* We can only have so many WRs outstanding. we have to wait until some finish */
2436 [ # # # # : 81 : rqpair->poller->stat.pending_rdma_send++;
# # # # ]
2437 : 81 : break;
2438 : : }
2439 : :
2440 : : /* We have already verified that this request is the head of the queue. */
2441 [ + + # # : 4079549 : STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_send_queue, state_link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# ]
2442 : :
2443 : : /* The response sending will be kicked off from
2444 : : * RDMA_REQUEST_STATE_READY_TO_COMPLETE state.
2445 : : */
2446 [ # # # # ]: 4079549 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE;
2447 : 4079549 : break;
2448 : 7921540 : case RDMA_REQUEST_STATE_READY_TO_COMPLETE:
2449 [ + + + + : 7921540 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0,
# # # # #
# # # # #
# # ]
2450 : : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2451 [ # # ]: 7921540 : rc = request_transfer_out(&rdma_req->req, &data_posted);
2452 [ - + # # ]: 7921540 : assert(rc == 0); /* No good way to handle this currently */
2453 [ - + ]: 7921540 : if (spdk_unlikely(rc)) {
2454 [ # # # # ]: 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
2455 : 0 : } else {
2456 [ + + # # ]: 7921540 : rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST :
2457 : : RDMA_REQUEST_STATE_COMPLETING;
2458 : : }
2459 : 7921540 : break;
2460 : 3846472 : case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST:
2461 [ + + + + : 3846472 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0,
# # # # #
# # # # #
# # ]
2462 : : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2463 : : /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED
2464 : : * to escape this state. */
2465 : 3846472 : break;
2466 : 5463243 : case RDMA_REQUEST_STATE_COMPLETING:
2467 [ + + + + : 5463243 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0,
# # # # #
# # # # #
# # ]
2468 : : (uintptr_t)rdma_req, (uintptr_t)rqpair);
2469 : : /* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED
2470 : : * to escape this state. */
2471 : 5463243 : break;
2472 : 7923170 : case RDMA_REQUEST_STATE_COMPLETED:
2473 [ + + + + : 7923170 : spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0,
# # # # #
# # # # #
# # # # #
# # # ]
2474 : : (uintptr_t)rdma_req, (uintptr_t)rqpair, rqpair->qpair.queue_depth);
2475 : :
2476 [ # # # # : 7923170 : rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc;
# # # # #
# # # #
# ]
2477 : 7923170 : _nvmf_rdma_request_free(rdma_req, rtransport);
2478 : 7923170 : break;
2479 : 0 : case RDMA_REQUEST_NUM_STATES:
2480 : : default:
2481 [ # # ]: 0 : assert(0);
2482 : : break;
2483 : : }
2484 : :
2485 [ + + # # : 86995764 : if (rdma_req->state != prev_state) {
# # ]
2486 : 54982797 : progress = true;
2487 : 0 : }
2488 [ + + # # : 86995764 : } while (rdma_req->state != prev_state);
# # ]
2489 : :
2490 [ # # ]: 32012967 : return progress;
2491 : : }
2492 : :
2493 : : /* Public API callbacks begin here */
2494 : :
2495 : : #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128
2496 : : #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128
2497 : : #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096
2498 : : #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128
2499 : : #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096
2500 : : #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072
2501 : : #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES)
2502 : : #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095
2503 : : #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE UINT32_MAX
2504 : : #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false
2505 : : #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false
2506 : : #define SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG 100
2507 : : #define SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC 1
2508 : : #define SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING false
2509 : : #define SPDK_NVMF_RDMA_DEFAULT_DATA_WR_POOL_SIZE 4095
2510 : :
2511 : : static void
2512 : 50 : nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts)
2513 : : {
2514 [ # # # # ]: 50 : opts->max_queue_depth = SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH;
2515 [ # # # # ]: 50 : opts->max_qpairs_per_ctrlr = SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR;
2516 [ # # # # ]: 50 : opts->in_capsule_data_size = SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE;
2517 [ # # # # ]: 50 : opts->max_io_size = SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE;
2518 [ # # # # ]: 50 : opts->io_unit_size = SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE;
2519 [ # # # # ]: 50 : opts->max_aq_depth = SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH;
2520 [ # # # # ]: 50 : opts->num_shared_buffers = SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS;
2521 [ # # # # ]: 50 : opts->buf_cache_size = SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE;
2522 [ # # # # ]: 50 : opts->dif_insert_or_strip = SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP;
2523 [ # # # # ]: 50 : opts->abort_timeout_sec = SPDK_NVMF_RDMA_DEFAULT_ABORT_TIMEOUT_SEC;
2524 [ # # # # ]: 50 : opts->transport_specific = NULL;
2525 [ # # # # ]: 50 : opts->data_wr_pool_size = SPDK_NVMF_RDMA_DEFAULT_DATA_WR_POOL_SIZE;
2526 : 50 : }
2527 : :
2528 : : static int nvmf_rdma_destroy(struct spdk_nvmf_transport *transport,
2529 : : spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg);
2530 : :
2531 : : static inline bool
2532 : 2761 : nvmf_rdma_is_rxe_device(struct spdk_nvmf_rdma_device *device)
2533 : : {
2534 [ + - # # : 5522 : return device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_OLD ||
# # # # ]
2535 [ - + # # : 2761 : device->attr.vendor_id == SPDK_RDMA_RXE_VENDOR_ID_NEW;
# # ]
2536 : : }
2537 : :
2538 : : static int nvmf_rdma_accept(void *ctx);
2539 : : static bool nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport);
2540 : : static void destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport,
2541 : : struct spdk_nvmf_rdma_device *device);
2542 : :
2543 : : static int
2544 : 94 : create_ib_device(struct spdk_nvmf_rdma_transport *rtransport, struct ibv_context *context,
2545 : : struct spdk_nvmf_rdma_device **new_device)
2546 : : {
2547 : : struct spdk_nvmf_rdma_device *device;
2548 : 94 : int flag = 0;
2549 : 94 : int rc = 0;
2550 : :
2551 : 94 : device = calloc(1, sizeof(*device));
2552 [ - + ]: 94 : if (!device) {
2553 : 0 : SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n");
2554 : 0 : return -ENOMEM;
2555 : : }
2556 [ # # # # ]: 94 : device->context = context;
2557 [ # # # # : 94 : rc = ibv_query_device(device->context, &device->attr);
# # ]
2558 [ - + ]: 94 : if (rc < 0) {
2559 : 0 : SPDK_ERRLOG("Failed to query RDMA device attributes.\n");
2560 : 0 : free(device);
2561 : 0 : return rc;
2562 : : }
2563 : :
2564 : : #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
2565 [ - + # # : 94 : if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) {
# # # # ]
2566 : 0 : SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,");
2567 [ # # # # : 0 : SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id);
# # ]
2568 : 0 : }
2569 : :
2570 : : /**
2571 : : * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE.
2572 : : * The Soft-RoCE RXE driver does not currently support send with invalidate,
2573 : : * but incorrectly reports that it does. There are changes making their way
2574 : : * through the kernel now that will enable this feature. When they are merged,
2575 : : * we can conditionally enable this feature.
2576 : : *
2577 : : * TODO: enable this for versions of the kernel rxe driver that support it.
2578 : : */
2579 [ - + ]: 94 : if (nvmf_rdma_is_rxe_device(device)) {
2580 [ # # # # : 0 : device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS);
# # ]
2581 : 0 : }
2582 : : #endif
2583 : :
2584 : : /* set up device context async ev fd as NON_BLOCKING */
2585 [ # # # # : 94 : flag = fcntl(device->context->async_fd, F_GETFL);
# # # # ]
2586 [ # # # # : 94 : rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK);
# # # # ]
2587 [ - + ]: 94 : if (rc < 0) {
2588 : 0 : SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n");
2589 : 0 : free(device);
2590 : 0 : return rc;
2591 : : }
2592 : :
2593 [ # # # # : 94 : TAILQ_INSERT_TAIL(&rtransport->devices, device, link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # ]
2594 [ - + - + : 94 : SPDK_DEBUGLOG(rdma, "New device %p is added to RDMA transport\n", device);
# # ]
2595 : :
2596 [ - + ]: 94 : if (g_nvmf_hooks.get_ibv_pd) {
2597 [ # # # # : 0 : device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context);
# # # # #
# # # ]
2598 : 0 : } else {
2599 [ # # # # : 94 : device->pd = ibv_alloc_pd(device->context);
# # # # ]
2600 : : }
2601 : :
2602 [ - + # # : 94 : if (!device->pd) {
# # ]
2603 : 0 : SPDK_ERRLOG("Unable to allocate protection domain.\n");
2604 : 0 : destroy_ib_device(rtransport, device);
2605 : 0 : return -ENOMEM;
2606 : : }
2607 : :
2608 [ - + # # : 94 : assert(device->map == NULL);
# # # # ]
2609 : :
2610 [ # # # # : 94 : device->map = spdk_rdma_utils_create_mem_map(device->pd, &g_nvmf_hooks, IBV_ACCESS_LOCAL_WRITE);
# # # # ]
2611 [ - + # # : 94 : if (!device->map) {
# # ]
2612 : 0 : SPDK_ERRLOG("Unable to allocate memory map for listen address\n");
2613 : 0 : destroy_ib_device(rtransport, device);
2614 : 0 : return -ENOMEM;
2615 : : }
2616 : :
2617 [ - + # # : 94 : assert(device->map != NULL);
# # # # ]
2618 [ - + # # : 94 : assert(device->pd != NULL);
# # # # ]
2619 : :
2620 [ + - ]: 94 : if (new_device) {
2621 [ # # ]: 94 : *new_device = device;
2622 : 0 : }
2623 [ # # # # ]: 94 : SPDK_NOTICELOG("Create IB device %s(%p/%p) succeed.\n", ibv_get_device_name(context->device),
2624 : : device, context);
2625 : :
2626 : 94 : return 0;
2627 : 0 : }
2628 : :
2629 : : static void
2630 : 47 : free_poll_fds(struct spdk_nvmf_rdma_transport *rtransport)
2631 : : {
2632 [ + - # # : 47 : if (rtransport->poll_fds) {
# # ]
2633 [ # # # # ]: 47 : free(rtransport->poll_fds);
2634 [ # # # # ]: 47 : rtransport->poll_fds = NULL;
2635 : 0 : }
2636 [ # # # # ]: 47 : rtransport->npoll_fds = 0;
2637 : 47 : }
2638 : :
2639 : : static int
2640 : 47 : generate_poll_fds(struct spdk_nvmf_rdma_transport *rtransport)
2641 : : {
2642 : : /* Set up poll descriptor array to monitor events from RDMA and IB
2643 : : * in a single poll syscall
2644 : : */
2645 : 47 : int device_count = 0;
2646 : 47 : int i = 0;
2647 : : struct spdk_nvmf_rdma_device *device, *tmp;
2648 : :
2649 [ + + # # : 141 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) {
# # # # #
# # # # #
# # ]
2650 [ # # ]: 94 : device_count++;
2651 : 0 : }
2652 : :
2653 [ # # # # : 47 : rtransport->npoll_fds = device_count + 1;
# # ]
2654 : :
2655 [ # # # # : 47 : rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd));
# # # # ]
2656 [ - + # # : 47 : if (rtransport->poll_fds == NULL) {
# # ]
2657 : 0 : SPDK_ERRLOG("poll_fds allocation failed\n");
2658 : 0 : return -ENOMEM;
2659 : : }
2660 : :
2661 [ # # # # : 47 : rtransport->poll_fds[i].fd = rtransport->event_channel->fd;
# # # # #
# # # # #
# # # # ]
2662 [ # # # # : 47 : rtransport->poll_fds[i++].events = POLLIN;
# # # # #
# # # ]
2663 : :
2664 [ + + # # : 141 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) {
# # # # #
# # # # #
# # ]
2665 [ # # # # : 94 : rtransport->poll_fds[i].fd = device->context->async_fd;
# # # # #
# # # # #
# # # # ]
2666 [ # # # # : 94 : rtransport->poll_fds[i++].events = POLLIN;
# # # # #
# # # ]
2667 : 0 : }
2668 : :
2669 : 47 : return 0;
2670 : 0 : }
2671 : :
2672 : : static struct spdk_nvmf_transport *
2673 : 47 : nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts)
2674 : : {
2675 : : int rc;
2676 : : struct spdk_nvmf_rdma_transport *rtransport;
2677 : 0 : struct spdk_nvmf_rdma_device *device;
2678 : : struct ibv_context **contexts;
2679 : : size_t data_wr_pool_size;
2680 : : uint32_t i;
2681 : : int flag;
2682 : : uint32_t sge_count;
2683 : : uint32_t min_shared_buffers;
2684 : : uint32_t min_in_capsule_data_size;
2685 : 47 : int max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
2686 : :
2687 : 47 : rtransport = calloc(1, sizeof(*rtransport));
2688 [ - + ]: 47 : if (!rtransport) {
2689 : 0 : return NULL;
2690 : : }
2691 : :
2692 [ # # # # : 47 : TAILQ_INIT(&rtransport->devices);
# # # # #
# # # # #
# # ]
2693 [ # # # # : 47 : TAILQ_INIT(&rtransport->ports);
# # # # #
# # # # #
# # ]
2694 [ # # # # : 47 : TAILQ_INIT(&rtransport->poll_groups);
# # # # #
# # # # #
# # ]
2695 [ # # # # : 47 : TAILQ_INIT(&rtransport->retry_ports);
# # # # #
# # # # #
# # ]
2696 : :
2697 [ # # # # : 47 : rtransport->transport.ops = &spdk_nvmf_transport_rdma;
# # ]
2698 [ # # # # : 47 : rtransport->rdma_opts.num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE;
# # ]
2699 [ # # # # : 47 : rtransport->rdma_opts.max_srq_depth = SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH;
# # ]
2700 [ # # # # : 47 : rtransport->rdma_opts.no_srq = SPDK_NVMF_RDMA_DEFAULT_NO_SRQ;
# # ]
2701 [ # # # # : 47 : rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG;
# # ]
2702 [ # # # # : 47 : rtransport->rdma_opts.no_wr_batching = SPDK_NVMF_RDMA_DEFAULT_NO_WR_BATCHING;
# # ]
2703 [ + - - + : 94 : if (opts->transport_specific != NULL &&
# # # # ]
2704 [ # # # # ]: 47 : spdk_json_decode_object_relaxed(opts->transport_specific, rdma_transport_opts_decoder,
2705 : : SPDK_COUNTOF(rdma_transport_opts_decoder),
2706 [ # # ]: 47 : &rtransport->rdma_opts)) {
2707 : 0 : SPDK_ERRLOG("spdk_json_decode_object_relaxed failed\n");
2708 [ # # ]: 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2709 : 0 : return NULL;
2710 : : }
2711 : :
2712 [ - + - + : 47 : SPDK_INFOLOG(rdma, "*** RDMA Transport Init ***\n"
- - - - #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # #
# ]
2713 : : " Transport opts: max_ioq_depth=%d, max_io_size=%d,\n"
2714 : : " max_io_qpairs_per_ctrlr=%d, io_unit_size=%d,\n"
2715 : : " in_capsule_data_size=%d, max_aq_depth=%d,\n"
2716 : : " num_shared_buffers=%d, num_cqe=%d, max_srq_depth=%d, no_srq=%d,"
2717 : : " acceptor_backlog=%d, no_wr_batching=%d abort_timeout_sec=%d\n",
2718 : : opts->max_queue_depth,
2719 : : opts->max_io_size,
2720 : : opts->max_qpairs_per_ctrlr - 1,
2721 : : opts->io_unit_size,
2722 : : opts->in_capsule_data_size,
2723 : : opts->max_aq_depth,
2724 : : opts->num_shared_buffers,
2725 : : rtransport->rdma_opts.num_cqe,
2726 : : rtransport->rdma_opts.max_srq_depth,
2727 : : rtransport->rdma_opts.no_srq,
2728 : : rtransport->rdma_opts.acceptor_backlog,
2729 : : rtransport->rdma_opts.no_wr_batching,
2730 : : opts->abort_timeout_sec);
2731 : :
2732 : : /* I/O unit size cannot be larger than max I/O size */
2733 [ - + # # : 47 : if (opts->io_unit_size > opts->max_io_size) {
# # # # #
# ]
2734 [ # # # # : 0 : opts->io_unit_size = opts->max_io_size;
# # # # ]
2735 : 0 : }
2736 : :
2737 [ - + # # : 47 : if (rtransport->rdma_opts.acceptor_backlog <= 0) {
# # # # ]
2738 : 0 : SPDK_ERRLOG("The acceptor backlog cannot be less than 1, setting to the default value of (%d).\n",
2739 : : SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG);
2740 [ # # # # : 0 : rtransport->rdma_opts.acceptor_backlog = SPDK_NVMF_RDMA_ACCEPTOR_BACKLOG;
# # ]
2741 : 0 : }
2742 : :
2743 [ - + # # : 47 : if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) {
# # ]
2744 [ # # # # ]: 0 : SPDK_ERRLOG("The number of shared data buffers (%d) is less than"
2745 : : "the minimum number required to guarantee that forward progress can be made (%d)\n",
2746 : : opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2));
2747 [ # # ]: 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2748 : 0 : return NULL;
2749 : : }
2750 : :
2751 : : /* If buf_cache_size == UINT32_MAX, we will dynamically pick a cache size later that we know will fit. */
2752 [ - + # # : 47 : if (opts->buf_cache_size < UINT32_MAX) {
# # ]
2753 [ # # # # ]: 0 : min_shared_buffers = spdk_env_get_core_count() * opts->buf_cache_size;
2754 [ # # # # : 0 : if (min_shared_buffers > opts->num_shared_buffers) {
# # ]
2755 [ # # # # ]: 0 : SPDK_ERRLOG("There are not enough buffers to satisfy"
2756 : : "per-poll group caches for each thread. (%" PRIu32 ")"
2757 : : "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers);
2758 : 0 : SPDK_ERRLOG("Please specify a larger number of shared buffers\n");
2759 [ # # ]: 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2760 : 0 : return NULL;
2761 : : }
2762 : 0 : }
2763 : :
2764 [ - + # # : 47 : sge_count = opts->max_io_size / opts->io_unit_size;
# # # # #
# ]
2765 [ - + ]: 47 : if (sge_count > NVMF_DEFAULT_TX_SGE) {
2766 [ # # # # ]: 0 : SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size);
2767 [ # # ]: 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2768 : 0 : return NULL;
2769 : : }
2770 : :
2771 : 47 : min_in_capsule_data_size = sizeof(struct spdk_nvme_sgl_descriptor) * SPDK_NVMF_MAX_SGL_ENTRIES;
2772 [ + + # # : 47 : if (opts->in_capsule_data_size < min_in_capsule_data_size) {
# # ]
2773 : 4 : SPDK_WARNLOG("In capsule data size is set to %u, this is minimum size required to support msdbd=16\n",
2774 : : min_in_capsule_data_size);
2775 [ # # # # ]: 4 : opts->in_capsule_data_size = min_in_capsule_data_size;
2776 : 0 : }
2777 : :
2778 [ # # # # ]: 47 : rtransport->event_channel = rdma_create_event_channel();
2779 [ - + # # : 47 : if (rtransport->event_channel == NULL) {
# # ]
2780 [ # # ]: 0 : SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno));
2781 [ # # ]: 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2782 : 0 : return NULL;
2783 : : }
2784 : :
2785 [ # # # # : 47 : flag = fcntl(rtransport->event_channel->fd, F_GETFL);
# # # # ]
2786 [ - + # # : 47 : if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) {
# # # # #
# ]
2787 [ # # # # : 0 : SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n",
# # # # #
# ]
2788 : : rtransport->event_channel->fd, spdk_strerror(errno));
2789 [ # # ]: 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2790 : 0 : return NULL;
2791 : : }
2792 : :
2793 [ # # # # ]: 47 : data_wr_pool_size = opts->data_wr_pool_size;
2794 [ - + ]: 47 : if (data_wr_pool_size < SPDK_NVMF_MAX_SGL_ENTRIES * 2 * spdk_env_get_core_count()) {
2795 : 0 : data_wr_pool_size = SPDK_NVMF_MAX_SGL_ENTRIES * 2 * spdk_env_get_core_count();
2796 : 0 : SPDK_NOTICELOG("data_wr_pool_size is changed to %zu to guarantee enough cache for handling "
2797 : : "at least one IO in each core\n", data_wr_pool_size);
2798 : 0 : }
2799 [ # # # # ]: 47 : rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data", data_wr_pool_size,
2800 : : sizeof(struct spdk_nvmf_rdma_request_data), SPDK_MEMPOOL_DEFAULT_CACHE_SIZE,
2801 : : SPDK_ENV_NUMA_ID_ANY);
2802 [ - + # # : 47 : if (!rtransport->data_wr_pool) {
# # ]
2803 [ # # ]: 0 : if (spdk_mempool_lookup("spdk_nvmf_rdma_wr_data") != NULL) {
2804 : 0 : SPDK_ERRLOG("Unable to allocate work request pool for poll group: already exists\n");
2805 : 0 : SPDK_ERRLOG("Probably running in multiprocess environment, which is "
2806 : : "unsupported by the nvmf library\n");
2807 : 0 : } else {
2808 : 0 : SPDK_ERRLOG("Unable to allocate work request pool for poll group\n");
2809 : : }
2810 [ # # ]: 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2811 : 0 : return NULL;
2812 : : }
2813 : :
2814 : 47 : contexts = rdma_get_devices(NULL);
2815 [ - + ]: 47 : if (contexts == NULL) {
2816 [ # # # # ]: 0 : SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno);
2817 [ # # ]: 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2818 : 0 : return NULL;
2819 : : }
2820 : :
2821 : 47 : i = 0;
2822 : 47 : rc = 0;
2823 [ + + # # : 141 : while (contexts[i] != NULL) {
# # ]
2824 [ # # # # ]: 94 : rc = create_ib_device(rtransport, contexts[i], &device);
2825 [ - + ]: 94 : if (rc < 0) {
2826 : 0 : break;
2827 : : }
2828 : 94 : i++;
2829 [ # # # # : 94 : max_device_sge = spdk_min(max_device_sge, device->attr.max_sge);
# # # # #
# # # #
# ]
2830 [ # # # # ]: 94 : device->is_ready = true;
2831 : : }
2832 : 47 : rdma_free_devices(contexts);
2833 : :
2834 [ - + # # : 47 : if (opts->io_unit_size * max_device_sge < opts->max_io_size) {
# # # # #
# ]
2835 : : /* divide and round up. */
2836 [ # # # # : 0 : opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge;
# # # # #
# ]
2837 : :
2838 : : /* round up to the nearest 4k. */
2839 [ # # # # : 0 : opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK;
# # # # #
# # # ]
2840 : :
2841 [ # # # # : 0 : opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE);
# # # # #
# # # #
# ]
2842 [ # # # # ]: 0 : SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n",
2843 : : opts->io_unit_size);
2844 : 0 : }
2845 : :
2846 [ - + ]: 47 : if (rc < 0) {
2847 [ # # ]: 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2848 : 0 : return NULL;
2849 : : }
2850 : :
2851 : 47 : rc = generate_poll_fds(rtransport);
2852 [ - + ]: 47 : if (rc < 0) {
2853 [ # # ]: 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2854 : 0 : return NULL;
2855 : : }
2856 : :
2857 [ # # # # : 47 : rtransport->accept_poller = SPDK_POLLER_REGISTER(nvmf_rdma_accept, &rtransport->transport,
# # # # #
# ]
2858 : : opts->acceptor_poll_rate);
2859 [ - + # # : 47 : if (!rtransport->accept_poller) {
# # ]
2860 [ # # ]: 0 : nvmf_rdma_destroy(&rtransport->transport, NULL, NULL);
2861 : 0 : return NULL;
2862 : : }
2863 : :
2864 [ # # ]: 47 : return &rtransport->transport;
2865 : 0 : }
2866 : :
2867 : : static void
2868 : 94 : destroy_ib_device(struct spdk_nvmf_rdma_transport *rtransport,
2869 : : struct spdk_nvmf_rdma_device *device)
2870 : : {
2871 [ + + # # : 94 : TAILQ_REMOVE(&rtransport->devices, device, link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # ]
2872 [ # # ]: 94 : spdk_rdma_utils_free_mem_map(&device->map);
2873 [ + - # # : 94 : if (device->pd) {
# # ]
2874 [ + - ]: 94 : if (!g_nvmf_hooks.get_ibv_pd) {
2875 [ # # # # ]: 94 : ibv_dealloc_pd(device->pd);
2876 : 0 : }
2877 : 0 : }
2878 [ - + - + : 94 : SPDK_DEBUGLOG(rdma, "IB device [%p] is destroyed.\n", device);
# # ]
2879 : 94 : free(device);
2880 : 94 : }
2881 : :
2882 : : static void
2883 : 44 : nvmf_rdma_dump_opts(struct spdk_nvmf_transport *transport, struct spdk_json_write_ctx *w)
2884 : : {
2885 : : struct spdk_nvmf_rdma_transport *rtransport;
2886 [ - + # # ]: 44 : assert(w != NULL);
2887 : :
2888 : 44 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
2889 [ # # # # : 44 : spdk_json_write_named_uint32(w, "max_srq_depth", rtransport->rdma_opts.max_srq_depth);
# # ]
2890 [ - + # # : 44 : spdk_json_write_named_bool(w, "no_srq", rtransport->rdma_opts.no_srq);
# # # # ]
2891 [ - + - + : 44 : if (rtransport->rdma_opts.no_srq == true) {
# # # # #
# ]
2892 [ # # # # : 0 : spdk_json_write_named_int32(w, "num_cqe", rtransport->rdma_opts.num_cqe);
# # ]
2893 : 0 : }
2894 [ # # # # : 44 : spdk_json_write_named_int32(w, "acceptor_backlog", rtransport->rdma_opts.acceptor_backlog);
# # ]
2895 [ - + # # : 44 : spdk_json_write_named_bool(w, "no_wr_batching", rtransport->rdma_opts.no_wr_batching);
# # # # ]
2896 : 44 : }
2897 : :
2898 : : static int
2899 : 47 : nvmf_rdma_destroy(struct spdk_nvmf_transport *transport,
2900 : : spdk_nvmf_transport_destroy_done_cb cb_fn, void *cb_arg)
2901 : : {
2902 : : struct spdk_nvmf_rdma_transport *rtransport;
2903 : : struct spdk_nvmf_rdma_port *port, *port_tmp;
2904 : : struct spdk_nvmf_rdma_device *device, *device_tmp;
2905 : :
2906 : 47 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
2907 : :
2908 [ - + # # : 47 : TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, port_tmp) {
# # # # #
# # # # #
# # ]
2909 [ # # # # : 0 : TAILQ_REMOVE(&rtransport->retry_ports, port, link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # ]
2910 : 0 : free(port);
2911 : 0 : }
2912 : :
2913 [ - + # # : 47 : TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) {
# # # # #
# # # # #
# # ]
2914 [ # # # # : 0 : TAILQ_REMOVE(&rtransport->ports, port, link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # ]
2915 [ # # # # ]: 0 : rdma_destroy_id(port->id);
2916 : 0 : free(port);
2917 : 0 : }
2918 : :
2919 : 47 : free_poll_fds(rtransport);
2920 : :
2921 [ + - # # : 47 : if (rtransport->event_channel != NULL) {
# # ]
2922 [ # # # # ]: 47 : rdma_destroy_event_channel(rtransport->event_channel);
2923 : 0 : }
2924 : :
2925 [ + + # # : 141 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) {
# # # # #
# # # # #
# # ]
2926 : 94 : destroy_ib_device(rtransport, device);
2927 : 0 : }
2928 : :
2929 [ + - # # : 47 : if (rtransport->data_wr_pool != NULL) {
# # ]
2930 [ - + # # : 47 : if (spdk_mempool_count(rtransport->data_wr_pool) != transport->opts.data_wr_pool_size) {
# # # # #
# # # ]
2931 [ # # # # : 0 : SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n",
# # # # #
# # # ]
2932 : : spdk_mempool_count(rtransport->data_wr_pool),
2933 : : transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES);
2934 : 0 : }
2935 : 0 : }
2936 : :
2937 [ # # # # ]: 47 : spdk_mempool_free(rtransport->data_wr_pool);
2938 : :
2939 [ # # ]: 47 : spdk_poller_unregister(&rtransport->accept_poller);
2940 : 47 : free(rtransport);
2941 : :
2942 [ + - ]: 47 : if (cb_fn) {
2943 [ # # # # ]: 47 : cb_fn(cb_arg);
2944 : 0 : }
2945 : 47 : return 0;
2946 : : }
2947 : :
2948 : : static int nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id,
2949 : : struct spdk_nvme_transport_id *trid,
2950 : : bool peer);
2951 : :
2952 : : static bool nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport);
2953 : :
2954 : : static int
2955 : 67 : nvmf_rdma_listen(struct spdk_nvmf_transport *transport, const struct spdk_nvme_transport_id *trid,
2956 : : struct spdk_nvmf_listen_opts *listen_opts)
2957 : : {
2958 : : struct spdk_nvmf_rdma_transport *rtransport;
2959 : : struct spdk_nvmf_rdma_device *device;
2960 : : struct spdk_nvmf_rdma_port *port, *tmp_port;
2961 : 0 : struct addrinfo *res;
2962 : 0 : struct addrinfo hints;
2963 : : int family;
2964 : : int rc;
2965 : : long int port_val;
2966 : 67 : bool is_retry = false;
2967 : :
2968 [ - + # # : 67 : if (!strlen(trid->trsvcid)) {
# # ]
2969 : 0 : SPDK_ERRLOG("Service id is required\n");
2970 : 0 : return -EINVAL;
2971 : : }
2972 : :
2973 : 67 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
2974 [ - + # # : 67 : assert(rtransport->event_channel != NULL);
# # # # ]
2975 : :
2976 : 67 : port = calloc(1, sizeof(*port));
2977 [ - + ]: 67 : if (!port) {
2978 : 0 : SPDK_ERRLOG("Port allocation failed\n");
2979 : 0 : return -ENOMEM;
2980 : : }
2981 : :
2982 [ # # # # ]: 67 : port->trid = trid;
2983 : :
2984 [ + - - # : 67 : switch (trid->adrfam) {
# # # ]
2985 : 67 : case SPDK_NVMF_ADRFAM_IPV4:
2986 : 67 : family = AF_INET;
2987 : 67 : break;
2988 : 0 : case SPDK_NVMF_ADRFAM_IPV6:
2989 : 0 : family = AF_INET6;
2990 : 0 : break;
2991 : 0 : default:
2992 [ # # # # ]: 0 : SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam);
2993 : 0 : free(port);
2994 : 0 : return -EINVAL;
2995 : : }
2996 : :
2997 [ - + ]: 67 : memset(&hints, 0, sizeof(hints));
2998 [ # # ]: 67 : hints.ai_family = family;
2999 : 67 : hints.ai_flags = AI_NUMERICSERV;
3000 [ # # ]: 67 : hints.ai_socktype = SOCK_STREAM;
3001 [ # # ]: 67 : hints.ai_protocol = 0;
3002 : :
3003 : : /* Range check the trsvcid. Fail in 3 cases:
3004 : : * < 0: means that spdk_strtol hit an error
3005 : : * 0: this results in ephemeral port which we don't want
3006 : : * > 65535: port too high
3007 : : */
3008 [ # # ]: 67 : port_val = spdk_strtol(trid->trsvcid, 10);
3009 [ + - - + ]: 67 : if (port_val <= 0 || port_val > 65535) {
3010 [ # # ]: 0 : SPDK_ERRLOG("invalid trsvcid %s\n", trid->trsvcid);
3011 : 0 : free(port);
3012 : 0 : return -EINVAL;
3013 : : }
3014 : :
3015 [ # # # # ]: 67 : rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res);
3016 [ - + ]: 67 : if (rc) {
3017 : 0 : SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc);
3018 : 0 : free(port);
3019 [ # # ]: 0 : return -(abs(rc));
3020 : : }
3021 : :
3022 [ # # # # : 67 : rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP);
# # ]
3023 [ - + ]: 67 : if (rc < 0) {
3024 : 0 : SPDK_ERRLOG("rdma_create_id() failed\n");
3025 : 0 : freeaddrinfo(res);
3026 : 0 : free(port);
3027 : 0 : return rc;
3028 : : }
3029 : :
3030 [ # # # # : 67 : rc = rdma_bind_addr(port->id, res->ai_addr);
# # # # ]
3031 : 67 : freeaddrinfo(res);
3032 : :
3033 [ - + ]: 67 : if (rc < 0) {
3034 [ # # # # : 0 : TAILQ_FOREACH(tmp_port, &rtransport->retry_ports, link) {
# # # # #
# # # #
# ]
3035 [ # # # # : 0 : if (spdk_nvme_transport_id_compare(tmp_port->trid, trid) == 0) {
# # ]
3036 : 0 : is_retry = true;
3037 : 0 : break;
3038 : : }
3039 : 0 : }
3040 [ # # # # ]: 0 : if (!is_retry) {
3041 : 0 : SPDK_ERRLOG("rdma_bind_addr() failed\n");
3042 : 0 : }
3043 [ # # # # ]: 0 : rdma_destroy_id(port->id);
3044 : 0 : free(port);
3045 : 0 : return rc;
3046 : : }
3047 : :
3048 [ - + # # : 67 : if (!port->id->verbs) {
# # # # #
# ]
3049 : 0 : SPDK_ERRLOG("ibv_context is null\n");
3050 [ # # # # ]: 0 : rdma_destroy_id(port->id);
3051 : 0 : free(port);
3052 : 0 : return -1;
3053 : : }
3054 : :
3055 [ # # # # : 67 : rc = rdma_listen(port->id, rtransport->rdma_opts.acceptor_backlog);
# # # # #
# ]
3056 [ - + ]: 67 : if (rc < 0) {
3057 : 0 : SPDK_ERRLOG("rdma_listen() failed\n");
3058 [ # # # # ]: 0 : rdma_destroy_id(port->id);
3059 : 0 : free(port);
3060 : 0 : return rc;
3061 : : }
3062 : :
3063 [ + - # # : 68 : TAILQ_FOREACH(device, &rtransport->devices, link) {
# # # # #
# # # #
# ]
3064 [ + + - + : 68 : if (device->context == port->id->verbs && device->is_ready) {
+ - # # #
# # # # #
# # # # #
# # # ]
3065 [ # # # # ]: 67 : port->device = device;
3066 : 67 : break;
3067 : : }
3068 : 0 : }
3069 [ - + # # : 67 : if (!port->device) {
# # ]
3070 [ # # # # : 0 : SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n",
# # # # ]
3071 : : port->id->verbs);
3072 [ # # # # ]: 0 : rdma_destroy_id(port->id);
3073 : 0 : free(port);
3074 : 0 : nvmf_rdma_rescan_devices(rtransport);
3075 : 0 : return -EINVAL;
3076 : : }
3077 : :
3078 [ # # # # ]: 67 : SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n",
3079 : : trid->traddr, trid->trsvcid);
3080 : :
3081 [ # # # # : 67 : TAILQ_INSERT_TAIL(&rtransport->ports, port, link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # ]
3082 : 67 : return 0;
3083 : 0 : }
3084 : :
3085 : : static void
3086 : 67 : nvmf_rdma_stop_listen_ex(struct spdk_nvmf_transport *transport,
3087 : : const struct spdk_nvme_transport_id *trid, bool need_retry)
3088 : : {
3089 : : struct spdk_nvmf_rdma_transport *rtransport;
3090 : : struct spdk_nvmf_rdma_port *port, *tmp;
3091 : :
3092 : 67 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3093 : :
3094 [ + - # # ]: 67 : if (!need_retry) {
3095 [ - + # # : 67 : TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp) {
# # # # #
# # # # #
# # ]
3096 [ # # # # : 0 : if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) {
# # ]
3097 [ # # # # : 0 : TAILQ_REMOVE(&rtransport->retry_ports, port, link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # ]
3098 : 0 : free(port);
3099 : 0 : }
3100 : 0 : }
3101 : 0 : }
3102 : :
3103 [ + - # # : 76 : TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) {
# # # # #
# # # # #
# # ]
3104 [ + + # # : 76 : if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) {
# # ]
3105 [ - + - + : 67 : SPDK_DEBUGLOG(rdma, "Port %s:%s removed. need retry: %d\n",
# # # # #
# # # # #
# # # # #
# ]
3106 : : port->trid->traddr, port->trid->trsvcid, need_retry);
3107 [ + + # # : 67 : TAILQ_REMOVE(&rtransport->ports, port, link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # ]
3108 [ # # # # ]: 67 : rdma_destroy_id(port->id);
3109 [ # # # # ]: 67 : port->id = NULL;
3110 [ # # # # ]: 67 : port->device = NULL;
3111 [ - + # # ]: 67 : if (need_retry) {
3112 [ # # # # : 0 : TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # ]
3113 : 0 : } else {
3114 : 67 : free(port);
3115 : : }
3116 : 67 : break;
3117 : : }
3118 : 0 : }
3119 : 67 : }
3120 : :
3121 : : static void
3122 : 67 : nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport,
3123 : : const struct spdk_nvme_transport_id *trid)
3124 : : {
3125 : 67 : nvmf_rdma_stop_listen_ex(transport, trid, false);
3126 : 67 : }
3127 : :
3128 : : static void _nvmf_rdma_register_poller_in_group(void *c);
3129 : : static void _nvmf_rdma_remove_poller_in_group(void *c);
3130 : :
3131 : : static bool
3132 : 0 : nvmf_rdma_all_pollers_management_done(void *c)
3133 : : {
3134 : 0 : struct poller_manage_ctx *ctx = c;
3135 : : int counter;
3136 : :
3137 [ # # # # ]: 0 : counter = __atomic_sub_fetch(ctx->inflight_op_counter, 1, __ATOMIC_SEQ_CST);
3138 [ # # # # : 0 : SPDK_DEBUGLOG(rdma, "nvmf_rdma_all_pollers_management_done called. counter: %d, poller: %p\n",
# # # # #
# ]
3139 : : counter, ctx->rpoller);
3140 : :
3141 [ # # ]: 0 : if (counter == 0) {
3142 [ # # # # ]: 0 : free((void *)ctx->inflight_op_counter);
3143 : 0 : }
3144 : 0 : free(ctx);
3145 : :
3146 : 0 : return counter == 0;
3147 : : }
3148 : :
3149 : : static int
3150 : 0 : nvmf_rdma_manage_poller(struct spdk_nvmf_rdma_transport *rtransport,
3151 : : struct spdk_nvmf_rdma_device *device, bool *has_inflight, bool is_add)
3152 : : {
3153 : : struct spdk_nvmf_rdma_poll_group *rgroup;
3154 : : struct spdk_nvmf_rdma_poller *rpoller;
3155 : : struct spdk_nvmf_poll_group *poll_group;
3156 : : struct poller_manage_ctx *ctx;
3157 : : bool found;
3158 : : int *inflight_counter;
3159 : : spdk_msg_fn do_fn;
3160 : :
3161 [ # # ]: 0 : *has_inflight = false;
3162 [ # # ]: 0 : do_fn = is_add ? _nvmf_rdma_register_poller_in_group : _nvmf_rdma_remove_poller_in_group;
3163 : 0 : inflight_counter = calloc(1, sizeof(int));
3164 [ # # ]: 0 : if (!inflight_counter) {
3165 : 0 : SPDK_ERRLOG("Failed to allocate inflight counter when removing pollers\n");
3166 : 0 : return -ENOMEM;
3167 : : }
3168 : :
3169 [ # # # # : 0 : TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) {
# # # # #
# # # #
# ]
3170 [ # # ]: 0 : (*inflight_counter)++;
3171 : 0 : }
3172 : :
3173 [ # # # # : 0 : TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) {
# # # # #
# # # #
# ]
3174 : 0 : found = false;
3175 [ # # # # : 0 : TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
# # # # #
# # # #
# ]
3176 [ # # # # : 0 : if (rpoller->device == device) {
# # ]
3177 : 0 : found = true;
3178 : 0 : break;
3179 : : }
3180 : 0 : }
3181 [ # # # # : 0 : if (found == is_add) {
# # ]
3182 : 0 : __atomic_fetch_sub(inflight_counter, 1, __ATOMIC_SEQ_CST);
3183 : 0 : continue;
3184 : : }
3185 : :
3186 : 0 : ctx = calloc(1, sizeof(struct poller_manage_ctx));
3187 [ # # ]: 0 : if (!ctx) {
3188 : 0 : SPDK_ERRLOG("Failed to allocate poller_manage_ctx when removing pollers\n");
3189 [ # # # # : 0 : if (!*has_inflight) {
# # ]
3190 : 0 : free(inflight_counter);
3191 : 0 : }
3192 : 0 : return -ENOMEM;
3193 : : }
3194 : :
3195 [ # # # # ]: 0 : ctx->rtransport = rtransport;
3196 [ # # # # ]: 0 : ctx->rgroup = rgroup;
3197 [ # # # # ]: 0 : ctx->rpoller = rpoller;
3198 [ # # # # ]: 0 : ctx->device = device;
3199 [ # # # # ]: 0 : ctx->thread = spdk_get_thread();
3200 [ # # # # ]: 0 : ctx->inflight_op_counter = inflight_counter;
3201 [ # # ]: 0 : *has_inflight = true;
3202 : :
3203 [ # # # # : 0 : poll_group = rgroup->group.group;
# # ]
3204 [ # # # # : 0 : if (poll_group->thread != spdk_get_thread()) {
# # ]
3205 [ # # # # ]: 0 : spdk_thread_send_msg(poll_group->thread, do_fn, ctx);
3206 : 0 : } else {
3207 [ # # # # ]: 0 : do_fn(ctx);
3208 : : }
3209 : 0 : }
3210 : :
3211 [ # # # # : 0 : if (!*has_inflight) {
# # ]
3212 : 0 : free(inflight_counter);
3213 : 0 : }
3214 : :
3215 : 0 : return 0;
3216 : 0 : }
3217 : :
3218 : : static void nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport,
3219 : : struct spdk_nvmf_rdma_device *device);
3220 : :
3221 : : static struct spdk_nvmf_rdma_device *
3222 : 0 : nvmf_rdma_find_ib_device(struct spdk_nvmf_rdma_transport *rtransport,
3223 : : struct ibv_context *context)
3224 : : {
3225 : : struct spdk_nvmf_rdma_device *device, *tmp_device;
3226 : :
3227 [ # # # # : 0 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp_device) {
# # # # #
# # # # #
# # ]
3228 [ # # # # : 0 : if (device->need_destroy) {
# # # # ]
3229 : 0 : continue;
3230 : : }
3231 : :
3232 [ # # # # : 0 : if (strcmp(device->context->device->dev_name, context->device->dev_name) == 0) {
# # # # #
# # # # #
# # # # #
# # # ]
3233 : 0 : return device;
3234 : : }
3235 : 0 : }
3236 : :
3237 : 0 : return NULL;
3238 : 0 : }
3239 : :
3240 : : static bool
3241 : 0 : nvmf_rdma_check_devices_context(struct spdk_nvmf_rdma_transport *rtransport,
3242 : : struct ibv_context *context)
3243 : : {
3244 : 0 : struct spdk_nvmf_rdma_device *old_device, *new_device;
3245 : 0 : int rc = 0;
3246 : 0 : bool has_inflight;
3247 : :
3248 : 0 : old_device = nvmf_rdma_find_ib_device(rtransport, context);
3249 : :
3250 [ # # ]: 0 : if (old_device) {
3251 [ # # # # : 0 : if (old_device->context != context && !old_device->need_destroy && old_device->is_ready) {
# # # # #
# # # # #
# # # # #
# # # ]
3252 : : /* context may not have time to be cleaned when rescan. exactly one context
3253 : : * is valid for a device so this context must be invalid and just remove it. */
3254 [ # # # # ]: 0 : SPDK_WARNLOG("Device %p has a invalid context %p\n", old_device, old_device->context);
3255 [ # # # # ]: 0 : old_device->need_destroy = true;
3256 : 0 : nvmf_rdma_handle_device_removal(rtransport, old_device);
3257 : 0 : }
3258 : 0 : return false;
3259 : : }
3260 : :
3261 : 0 : rc = create_ib_device(rtransport, context, &new_device);
3262 : : /* TODO: update transport opts. */
3263 [ # # ]: 0 : if (rc < 0) {
3264 [ # # # # ]: 0 : SPDK_ERRLOG("Failed to create ib device for context: %s(%p)\n",
3265 : : ibv_get_device_name(context->device), context);
3266 : 0 : return false;
3267 : : }
3268 : :
3269 : 0 : rc = nvmf_rdma_manage_poller(rtransport, new_device, &has_inflight, true);
3270 [ # # ]: 0 : if (rc < 0) {
3271 [ # # # # ]: 0 : SPDK_ERRLOG("Failed to add poller for device context: %s(%p)\n",
3272 : : ibv_get_device_name(context->device), context);
3273 : 0 : return false;
3274 : : }
3275 : :
3276 [ # # # # ]: 0 : if (has_inflight) {
3277 [ # # # # ]: 0 : new_device->is_ready = true;
3278 : 0 : }
3279 : :
3280 : 0 : return true;
3281 : 0 : }
3282 : :
3283 : : static bool
3284 : 0 : nvmf_rdma_rescan_devices(struct spdk_nvmf_rdma_transport *rtransport)
3285 : : {
3286 : : struct spdk_nvmf_rdma_device *device;
3287 : 0 : struct ibv_device **ibv_device_list = NULL;
3288 : 0 : struct ibv_context **contexts = NULL;
3289 : 0 : int i = 0;
3290 : 0 : int num_dev = 0;
3291 : 0 : bool new_create = false, has_new_device = false;
3292 : 0 : struct ibv_context *tmp_verbs = NULL;
3293 : :
3294 : : /* do not rescan when any device is destroying, or context may be freed when
3295 : : * regenerating the poll fds.
3296 : : */
3297 [ # # # # : 0 : TAILQ_FOREACH(device, &rtransport->devices, link) {
# # # # #
# # # #
# ]
3298 [ # # # # : 0 : if (device->need_destroy) {
# # # # ]
3299 : 0 : return false;
3300 : : }
3301 : 0 : }
3302 : :
3303 : 0 : ibv_device_list = ibv_get_device_list(&num_dev);
3304 : :
3305 : : /* There is a bug in librdmacm. If verbs init failed in rdma_get_devices, it'll be
3306 : : * marked as dead verbs and never be init again. So we need to make sure the
3307 : : * verbs is available before we call rdma_get_devices. */
3308 [ # # ]: 0 : if (num_dev >= 0) {
3309 [ # # # # ]: 0 : for (i = 0; i < num_dev; i++) {
3310 [ # # # # ]: 0 : tmp_verbs = ibv_open_device(ibv_device_list[i]);
3311 [ # # ]: 0 : if (!tmp_verbs) {
3312 [ # # # # : 0 : SPDK_WARNLOG("Failed to init ibv device %p, err %d. Skip rescan.\n", ibv_device_list[i], errno);
# # ]
3313 : 0 : break;
3314 : : }
3315 [ # # ]: 0 : if (nvmf_rdma_find_ib_device(rtransport, tmp_verbs) == NULL) {
3316 [ # # # # : 0 : SPDK_DEBUGLOG(rdma, "Find new verbs init ibv device %p(%s).\n", ibv_device_list[i],
# # # # #
# # # # #
# # ]
3317 : : tmp_verbs->device->dev_name);
3318 : 0 : has_new_device = true;
3319 : 0 : }
3320 : 0 : ibv_close_device(tmp_verbs);
3321 : 0 : }
3322 : 0 : ibv_free_device_list(ibv_device_list);
3323 [ # # # # : 0 : if (!tmp_verbs || !has_new_device) {
# # ]
3324 : 0 : return false;
3325 : : }
3326 : 0 : }
3327 : :
3328 : 0 : contexts = rdma_get_devices(NULL);
3329 : :
3330 [ # # # # : 0 : for (i = 0; contexts && contexts[i] != NULL; i++) {
# # # # #
# ]
3331 [ # # # # : 0 : new_create |= nvmf_rdma_check_devices_context(rtransport, contexts[i]);
# # ]
3332 : 0 : }
3333 : :
3334 [ # # # # ]: 0 : if (new_create) {
3335 : 0 : free_poll_fds(rtransport);
3336 : 0 : generate_poll_fds(rtransport);
3337 : 0 : }
3338 : :
3339 [ # # ]: 0 : if (contexts) {
3340 : 0 : rdma_free_devices(contexts);
3341 : 0 : }
3342 : :
3343 [ # # ]: 0 : return new_create;
3344 : 0 : }
3345 : :
3346 : : static bool
3347 : 113395 : nvmf_rdma_retry_listen_port(struct spdk_nvmf_rdma_transport *rtransport)
3348 : : {
3349 : : struct spdk_nvmf_rdma_port *port, *tmp_port;
3350 : 113395 : int rc = 0;
3351 : 113395 : bool new_create = false;
3352 : :
3353 [ + - # # : 113395 : if (TAILQ_EMPTY(&rtransport->retry_ports)) {
# # # # ]
3354 : 113395 : return false;
3355 : : }
3356 : :
3357 : 0 : new_create = nvmf_rdma_rescan_devices(rtransport);
3358 : :
3359 [ # # # # : 0 : TAILQ_FOREACH_SAFE(port, &rtransport->retry_ports, link, tmp_port) {
# # # # #
# # # # #
# # ]
3360 [ # # # # : 0 : rc = nvmf_rdma_listen(&rtransport->transport, port->trid, NULL);
# # ]
3361 : :
3362 [ # # # # : 0 : TAILQ_REMOVE(&rtransport->retry_ports, port, link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # ]
3363 [ # # ]: 0 : if (rc) {
3364 [ # # # # ]: 0 : if (new_create) {
3365 [ # # # # : 0 : SPDK_ERRLOG("Found new IB device but port %s:%s is still failed(%d) to listen.\n",
# # # # #
# # # ]
3366 : : port->trid->traddr, port->trid->trsvcid, rc);
3367 : 0 : }
3368 [ # # # # : 0 : TAILQ_INSERT_TAIL(&rtransport->retry_ports, port, link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # ]
3369 : 0 : break;
3370 : : } else {
3371 [ # # # # : 0 : SPDK_NOTICELOG("Port %s:%s come back\n", port->trid->traddr, port->trid->trsvcid);
# # # # #
# # # ]
3372 : 0 : free(port);
3373 : : }
3374 : 0 : }
3375 : :
3376 : 0 : return true;
3377 : 0 : }
3378 : :
3379 : : static void
3380 : 17638301 : nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport,
3381 : : struct spdk_nvmf_rdma_qpair *rqpair, bool drain)
3382 : : {
3383 : : struct spdk_nvmf_request *req, *tmp;
3384 : : struct spdk_nvmf_rdma_request *rdma_req, *req_tmp;
3385 : : struct spdk_nvmf_rdma_resources *resources;
3386 : :
3387 : : /* First process requests which are waiting for response to be sent */
3388 [ + + # # : 17638377 : STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_send_queue, state_link, req_tmp) {
# # # # #
# # # # #
# # ]
3389 [ + + + - : 89 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
# # ]
3390 : 13 : break;
3391 : : }
3392 : 0 : }
3393 : :
3394 : : /* We process I/O in the data transfer pending queue at the highest priority. */
3395 [ + + # # : 19375790 : STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) {
# # # # #
# # # # #
# # ]
3396 [ + + # # : 4507454 : if (rdma_req->state != RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) {
# # ]
3397 : : /* Requests in this queue might be in state RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER,
3398 : : * they are transmitting data over network but we keep them in the list to guarantee
3399 : : * fair processing. */
3400 : 1261569 : continue;
3401 : : }
3402 [ + + + - : 3245885 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
# # ]
3403 : 2769965 : break;
3404 : : }
3405 : 0 : }
3406 : :
3407 : : /* Then RDMA writes since reads have stronger restrictions than writes */
3408 [ + + # # : 17643180 : STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) {
# # # # #
# # # # #
# # ]
3409 [ + + + - : 15011 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
# # ]
3410 : 10132 : break;
3411 : : }
3412 : 0 : }
3413 : :
3414 : : /* Then we handle request waiting on memory buffers. */
3415 [ + + # # : 18652641 : STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) {
# # # # #
# # # # #
# # # # #
# # # # #
# # ]
3416 : 4079352 : rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
3417 [ + + + + : 4079352 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
# # ]
3418 : 3065012 : break;
3419 : : }
3420 : 0 : }
3421 : :
3422 [ # # # # ]: 17638301 : resources = rqpair->resources;
3423 [ + + + + : 25562814 : while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) {
# # # # #
# # # # #
# # ]
3424 [ # # # # : 7924513 : rdma_req = STAILQ_FIRST(&resources->free_queue);
# # ]
3425 [ + + # # : 7924513 : STAILQ_REMOVE_HEAD(&resources->free_queue, state_link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# ]
3426 [ # # # # : 7924513 : rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue);
# # # # #
# ]
3427 [ + + # # : 7924513 : STAILQ_REMOVE_HEAD(&resources->incoming_queue, link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# ]
3428 : :
3429 [ + - # # : 7924513 : if (rqpair->srq != NULL) {
# # ]
3430 [ # # # # : 7924513 : rdma_req->req.qpair = &rdma_req->recv->qpair->qpair;
# # # # #
# # # # #
# # ]
3431 [ # # # # : 7924513 : rdma_req->recv->qpair->qd++;
# # # # #
# ]
3432 : 0 : } else {
3433 [ # # ]: 0 : rqpair->qd++;
3434 : : }
3435 : :
3436 [ # # # # : 7924513 : rdma_req->receive_tsc = rdma_req->recv->receive_tsc;
# # # # #
# # # ]
3437 [ # # # # ]: 7924513 : rdma_req->state = RDMA_REQUEST_STATE_NEW;
3438 [ - + ]: 7924513 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false) {
3439 : 0 : break;
3440 : : }
3441 : : }
3442 [ + + + - : 17638301 : if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) {
# # # # #
# # # # #
# # ]
3443 [ # # # # : 208 : rqpair->poller->stat.pending_free_request++;
# # # # ]
3444 : 0 : }
3445 : 17638301 : }
3446 : :
3447 : : static void
3448 : 898233522 : nvmf_rdma_poller_process_pending_buf_queue(struct spdk_nvmf_rdma_transport *rtransport,
3449 : : struct spdk_nvmf_rdma_poller *rpoller)
3450 : : {
3451 : : struct spdk_nvmf_request *req, *tmp;
3452 : : struct spdk_nvmf_rdma_request *rdma_req;
3453 : :
3454 [ + + # # : 898233526 : STAILQ_FOREACH_SAFE(req, &rpoller->group->group.pending_buf_queue, buf_link, tmp) {
# # # # #
# # # # #
# # # # #
# # # ]
3455 : 120019 : rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
3456 [ + + ]: 120019 : if (nvmf_rdma_request_process(rtransport, rdma_req) == false) {
3457 : 120015 : break;
3458 : : }
3459 : 0 : }
3460 : 898233522 : }
3461 : :
3462 : : static inline bool
3463 : 2667 : nvmf_rdma_device_supports_last_wqe_reached(struct spdk_nvmf_rdma_device *device)
3464 : : {
3465 : : /* iWARP transport and SoftRoCE driver don't support LAST_WQE_REACHED ibv async event */
3466 [ + - ]: 5334 : return !nvmf_rdma_is_rxe_device(device) &&
3467 [ + - # # : 2667 : device->context->device->transport_type != IBV_TRANSPORT_IWARP;
# # # # #
# # # ]
3468 : : }
3469 : :
3470 : : static void
3471 : 6142 : nvmf_rdma_destroy_drained_qpair(struct spdk_nvmf_rdma_qpair *rqpair)
3472 : : {
3473 [ # # # # : 6142 : struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
# # ]
3474 : : struct spdk_nvmf_rdma_transport, transport);
3475 : :
3476 : 6142 : nvmf_rdma_qpair_process_pending(rtransport, rqpair, true);
3477 : :
3478 : : /* nvmf_rdma_close_qpair is not called */
3479 [ - + + + : 6142 : if (!rqpair->to_close) {
# # # # ]
3480 : 46 : return;
3481 : : }
3482 : :
3483 : : /* device is already destroyed and we should force destroy this qpair. */
3484 [ + - - + : 6096 : if (rqpair->poller && rqpair->poller->need_destroy) {
- + # # #
# # # # #
# # # # ]
3485 : 0 : nvmf_rdma_qpair_destroy(rqpair);
3486 : 0 : return;
3487 : : }
3488 : :
3489 : : /* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */
3490 [ + + # # : 6096 : if (rqpair->current_send_depth != 0) {
# # ]
3491 : 854 : return;
3492 : : }
3493 : :
3494 [ - + - - : 5242 : if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) {
# # # # #
# # # # #
# # ]
3495 : 0 : return;
3496 : : }
3497 : :
3498 : : /* For devices that support LAST_WQE_REACHED with srq, we need to
3499 : : * wait to destroy the qpair until that event has been received.
3500 : : */
3501 [ + - - + : 7909 : if (rqpair->srq != NULL && rqpair->last_wqe_reached == false &&
+ + + - #
# # # # #
# # ]
3502 [ # # # # ]: 2667 : nvmf_rdma_device_supports_last_wqe_reached(rqpair->device)) {
3503 : 2667 : return;
3504 : : }
3505 : :
3506 [ - + # # : 2575 : assert(rqpair->qpair.state == SPDK_NVMF_QPAIR_ERROR);
# # # # #
# ]
3507 : :
3508 : 2575 : nvmf_rdma_qpair_destroy(rqpair);
3509 : 0 : }
3510 : :
3511 : : static int
3512 : 2494 : nvmf_rdma_disconnect(struct rdma_cm_event *evt, bool *event_acked)
3513 : : {
3514 : : struct spdk_nvmf_qpair *qpair;
3515 : : struct spdk_nvmf_rdma_qpair *rqpair;
3516 : :
3517 [ - + # # : 2494 : if (evt->id == NULL) {
# # ]
3518 : 0 : SPDK_ERRLOG("disconnect request: missing cm_id\n");
3519 : 0 : return -1;
3520 : : }
3521 : :
3522 [ # # # # : 2494 : qpair = evt->id->context;
# # # # ]
3523 [ - + ]: 2494 : if (qpair == NULL) {
3524 : 0 : SPDK_ERRLOG("disconnect request: no active connection\n");
3525 : 0 : return -1;
3526 : : }
3527 : :
3528 : 2494 : rdma_ack_cm_event(evt);
3529 [ # # ]: 2494 : *event_acked = true;
3530 : :
3531 : 2494 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
3532 : :
3533 [ + + + + : 2494 : spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair);
# # # # #
# # # # #
# # ]
3534 : :
3535 [ # # ]: 2494 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
3536 : :
3537 : 2494 : return 0;
3538 : 0 : }
3539 : :
3540 : : #ifdef DEBUG
3541 : : static const char *CM_EVENT_STR[] = {
3542 : : "RDMA_CM_EVENT_ADDR_RESOLVED",
3543 : : "RDMA_CM_EVENT_ADDR_ERROR",
3544 : : "RDMA_CM_EVENT_ROUTE_RESOLVED",
3545 : : "RDMA_CM_EVENT_ROUTE_ERROR",
3546 : : "RDMA_CM_EVENT_CONNECT_REQUEST",
3547 : : "RDMA_CM_EVENT_CONNECT_RESPONSE",
3548 : : "RDMA_CM_EVENT_CONNECT_ERROR",
3549 : : "RDMA_CM_EVENT_UNREACHABLE",
3550 : : "RDMA_CM_EVENT_REJECTED",
3551 : : "RDMA_CM_EVENT_ESTABLISHED",
3552 : : "RDMA_CM_EVENT_DISCONNECTED",
3553 : : "RDMA_CM_EVENT_DEVICE_REMOVAL",
3554 : : "RDMA_CM_EVENT_MULTICAST_JOIN",
3555 : : "RDMA_CM_EVENT_MULTICAST_ERROR",
3556 : : "RDMA_CM_EVENT_ADDR_CHANGE",
3557 : : "RDMA_CM_EVENT_TIMEWAIT_EXIT"
3558 : : };
3559 : : #endif /* DEBUG */
3560 : :
3561 : : static void
3562 : 0 : nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport,
3563 : : struct spdk_nvmf_rdma_port *port)
3564 : : {
3565 : : struct spdk_nvmf_rdma_poll_group *rgroup;
3566 : : struct spdk_nvmf_rdma_poller *rpoller;
3567 : : struct spdk_nvmf_rdma_qpair *rqpair;
3568 : :
3569 [ # # # # : 0 : TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) {
# # # # #
# # # #
# ]
3570 [ # # # # : 0 : TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
# # # # #
# # # #
# ]
3571 [ # # # # ]: 0 : RB_FOREACH(rqpair, qpairs_tree, &rpoller->qpairs) {
3572 [ # # # # : 0 : if (rqpair->listen_id == port->id) {
# # # # #
# ]
3573 [ # # ]: 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
3574 : 0 : }
3575 : 0 : }
3576 : 0 : }
3577 : 0 : }
3578 : 0 : }
3579 : :
3580 : : static bool
3581 : 0 : nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport,
3582 : : struct rdma_cm_event *event)
3583 : : {
3584 : : const struct spdk_nvme_transport_id *trid;
3585 : : struct spdk_nvmf_rdma_port *port;
3586 : : struct spdk_nvmf_rdma_transport *rtransport;
3587 : 0 : bool event_acked = false;
3588 : :
3589 : 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3590 [ # # # # : 0 : TAILQ_FOREACH(port, &rtransport->ports, link) {
# # # # #
# # # #
# ]
3591 [ # # # # : 0 : if (port->id == event->id) {
# # # # #
# ]
3592 [ # # # # : 0 : SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid);
# # # # #
# # # ]
3593 : 0 : rdma_ack_cm_event(event);
3594 : 0 : event_acked = true;
3595 [ # # # # ]: 0 : trid = port->trid;
3596 : 0 : break;
3597 : : }
3598 : 0 : }
3599 : :
3600 [ # # # # ]: 0 : if (event_acked) {
3601 : 0 : nvmf_rdma_disconnect_qpairs_on_port(rtransport, port);
3602 : :
3603 : 0 : nvmf_rdma_stop_listen(transport, trid);
3604 : 0 : nvmf_rdma_listen(transport, trid, NULL);
3605 : 0 : }
3606 : :
3607 [ # # ]: 0 : return event_acked;
3608 : : }
3609 : :
3610 : : static void
3611 : 0 : nvmf_rdma_handle_device_removal(struct spdk_nvmf_rdma_transport *rtransport,
3612 : : struct spdk_nvmf_rdma_device *device)
3613 : : {
3614 : : struct spdk_nvmf_rdma_port *port, *port_tmp;
3615 : : int rc;
3616 : 0 : bool has_inflight;
3617 : :
3618 : 0 : rc = nvmf_rdma_manage_poller(rtransport, device, &has_inflight, false);
3619 [ # # ]: 0 : if (rc) {
3620 : 0 : SPDK_ERRLOG("Failed to handle device removal, rc %d\n", rc);
3621 : 0 : return;
3622 : : }
3623 : :
3624 [ # # # # ]: 0 : if (!has_inflight) {
3625 : : /* no pollers, destroy the device */
3626 [ # # # # ]: 0 : device->ready_to_destroy = true;
3627 : 0 : spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_remove_destroyed_device, rtransport);
3628 : 0 : }
3629 : :
3630 [ # # # # : 0 : TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) {
# # # # #
# # # # #
# # ]
3631 [ # # # # : 0 : if (port->device == device) {
# # ]
3632 [ # # # # : 0 : SPDK_NOTICELOG("Port %s:%s on device %s is being removed.\n",
# # # # #
# # # # #
# # # # #
# # # #
# ]
3633 : : port->trid->traddr,
3634 : : port->trid->trsvcid,
3635 : : ibv_get_device_name(port->device->context->device));
3636 : :
3637 : : /* keep NVMF listener and only destroy structures of the
3638 : : * RDMA transport. when the device comes back we can retry listening
3639 : : * and the application's workflow will not be interrupted.
3640 : : */
3641 [ # # # # : 0 : nvmf_rdma_stop_listen_ex(&rtransport->transport, port->trid, true);
# # ]
3642 : 0 : }
3643 : 0 : }
3644 : 0 : }
3645 : :
3646 : : static void
3647 : 0 : nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport,
3648 : : struct rdma_cm_event *event)
3649 : : {
3650 : : struct spdk_nvmf_rdma_port *port, *tmp_port;
3651 : : struct spdk_nvmf_rdma_transport *rtransport;
3652 : :
3653 [ # # # # : 0 : port = event->id->context;
# # # # ]
3654 : 0 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3655 : :
3656 : 0 : rdma_ack_cm_event(event);
3657 : :
3658 : : /* if device removal happens during ctrl qpair disconnecting, it's possible that we receive
3659 : : * an DEVICE_REMOVAL event on qpair but the id->qp is just NULL. So we should make sure that
3660 : : * we are handling a port event here.
3661 : : */
3662 [ # # # # : 0 : TAILQ_FOREACH(tmp_port, &rtransport->ports, link) {
# # # # #
# # # #
# ]
3663 [ # # # # : 0 : if (port == tmp_port && port->device && !port->device->need_destroy) {
# # # # #
# # # # #
# # # # #
# ]
3664 [ # # # # : 0 : port->device->need_destroy = true;
# # # # ]
3665 [ # # # # ]: 0 : nvmf_rdma_handle_device_removal(rtransport, port->device);
3666 : 0 : }
3667 : 0 : }
3668 : 0 : }
3669 : :
3670 : : static void
3671 : 5949 : nvmf_process_cm_events(struct spdk_nvmf_transport *transport, uint32_t max_events)
3672 : : {
3673 : : struct spdk_nvmf_rdma_transport *rtransport;
3674 : 0 : struct rdma_cm_event *event;
3675 : : uint32_t i;
3676 : : int rc;
3677 : 0 : bool event_acked;
3678 : :
3679 : 5949 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3680 : :
3681 [ - + # # : 5949 : if (rtransport->event_channel == NULL) {
# # ]
3682 : 0 : return;
3683 : : }
3684 : :
3685 [ + - ]: 13593 : for (i = 0; i < max_events; i++) {
3686 : 13593 : event_acked = false;
3687 [ # # # # ]: 13593 : rc = rdma_get_cm_event(rtransport->event_channel, &event);
3688 [ + + ]: 13593 : if (rc) {
3689 [ - + - - : 5949 : if (errno != EAGAIN && errno != EWOULDBLOCK) {
# # # # ]
3690 [ # # ]: 0 : SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno));
3691 : 0 : }
3692 : 5949 : break;
3693 : : }
3694 : :
3695 [ - + - + : 7644 : SPDK_DEBUGLOG(rdma, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]);
# # # # #
# # # # #
# # ]
3696 : :
3697 [ + + + + : 7644 : spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event);
# # # # #
# # # # #
# # # # #
# ]
3698 : :
3699 [ - + - - : 7644 : switch (event->event) {
- + + - -
- - - # #
# # ]
3700 : 0 : case RDMA_CM_EVENT_ADDR_RESOLVED:
3701 : : case RDMA_CM_EVENT_ADDR_ERROR:
3702 : : case RDMA_CM_EVENT_ROUTE_RESOLVED:
3703 : : case RDMA_CM_EVENT_ROUTE_ERROR:
3704 : : /* No action required. The target never attempts to resolve routes. */
3705 : 0 : break;
3706 : 2575 : case RDMA_CM_EVENT_CONNECT_REQUEST:
3707 : 2575 : rc = nvmf_rdma_connect(transport, event);
3708 [ - + ]: 2575 : if (rc < 0) {
3709 : 0 : SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc);
3710 : 0 : break;
3711 : : }
3712 : 2575 : break;
3713 : 0 : case RDMA_CM_EVENT_CONNECT_RESPONSE:
3714 : : /* The target never initiates a new connection. So this will not occur. */
3715 : 0 : break;
3716 : 0 : case RDMA_CM_EVENT_CONNECT_ERROR:
3717 : : /* Can this happen? The docs say it can, but not sure what causes it. */
3718 : 0 : break;
3719 : 0 : case RDMA_CM_EVENT_UNREACHABLE:
3720 : : case RDMA_CM_EVENT_REJECTED:
3721 : : /* These only occur on the client side. */
3722 : 0 : break;
3723 : 2575 : case RDMA_CM_EVENT_ESTABLISHED:
3724 : : /* TODO: Should we be waiting for this event anywhere? */
3725 : 2575 : break;
3726 : 2494 : case RDMA_CM_EVENT_DISCONNECTED:
3727 : 2494 : rc = nvmf_rdma_disconnect(event, &event_acked);
3728 [ - + ]: 2494 : if (rc < 0) {
3729 : 0 : SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc);
3730 : 0 : break;
3731 : : }
3732 : 2494 : break;
3733 : 0 : case RDMA_CM_EVENT_DEVICE_REMOVAL:
3734 : : /* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL
3735 : : * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s.
3736 : : * Once these events are sent to SPDK, we should release all IB resources and
3737 : : * don't make attempts to call any ibv_query/modify/create functions. We can only call
3738 : : * ibv_destroy* functions to release user space memory allocated by IB. All kernel
3739 : : * resources are already cleaned. */
3740 [ # # # # : 0 : if (event->id->qp) {
# # # # #
# ]
3741 : : /* If rdma_cm event has a valid `qp` pointer then the event refers to the
3742 : : * corresponding qpair. Otherwise the event refers to a listening device. */
3743 : 0 : rc = nvmf_rdma_disconnect(event, &event_acked);
3744 [ # # ]: 0 : if (rc < 0) {
3745 : 0 : SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc);
3746 : 0 : break;
3747 : : }
3748 : 0 : } else {
3749 : 0 : nvmf_rdma_handle_cm_event_port_removal(transport, event);
3750 : 0 : event_acked = true;
3751 : : }
3752 : 0 : break;
3753 : 0 : case RDMA_CM_EVENT_MULTICAST_JOIN:
3754 : : case RDMA_CM_EVENT_MULTICAST_ERROR:
3755 : : /* Multicast is not used */
3756 : 0 : break;
3757 : 0 : case RDMA_CM_EVENT_ADDR_CHANGE:
3758 : 0 : event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event);
3759 : 0 : break;
3760 : 0 : case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3761 : : /* For now, do nothing. The target never re-uses queue pairs. */
3762 : 0 : break;
3763 : 0 : default:
3764 [ # # # # ]: 0 : SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event);
3765 : 0 : break;
3766 : : }
3767 [ - + + + ]: 7644 : if (!event_acked) {
3768 : 5150 : rdma_ack_cm_event(event);
3769 : 0 : }
3770 : 0 : }
3771 : 0 : }
3772 : :
3773 : : static void
3774 : 2575 : nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair)
3775 : : {
3776 [ # # # # ]: 2575 : rqpair->last_wqe_reached = true;
3777 : 2575 : nvmf_rdma_destroy_drained_qpair(rqpair);
3778 : 2575 : }
3779 : :
3780 : : static void
3781 : 2575 : nvmf_rdma_qpair_process_last_wqe_event(void *ctx)
3782 : : {
3783 : 2575 : struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx;
3784 : : struct spdk_nvmf_rdma_qpair *rqpair;
3785 : :
3786 [ # # # # ]: 2575 : rqpair = event_ctx->rqpair;
3787 : :
3788 [ + - ]: 2575 : if (rqpair) {
3789 [ - + # # : 2575 : assert(event_ctx == rqpair->last_wqe_reached_ctx);
# # # # ]
3790 [ # # # # ]: 2575 : rqpair->last_wqe_reached_ctx = NULL;
3791 : 2575 : nvmf_rdma_handle_last_wqe_reached(rqpair);
3792 : 0 : }
3793 : 2575 : free(event_ctx);
3794 : 2575 : }
3795 : :
3796 : : static int
3797 : 2575 : nvmf_rdma_send_qpair_last_wqe_event(struct spdk_nvmf_rdma_qpair *rqpair)
3798 : : {
3799 : : struct spdk_nvmf_rdma_ibv_event_ctx *ctx;
3800 : 2575 : struct spdk_thread *thr = NULL;
3801 : : int rc;
3802 : :
3803 [ - + # # : 2575 : if (rqpair->qpair.group) {
# # # # ]
3804 [ # # # # : 0 : thr = rqpair->qpair.group->thread;
# # # # #
# ]
3805 [ + - # # : 2575 : } else if (rqpair->destruct_channel) {
# # ]
3806 [ # # # # ]: 2575 : thr = spdk_io_channel_get_thread(rqpair->destruct_channel);
3807 : 0 : }
3808 : :
3809 [ - + ]: 2575 : if (!thr) {
3810 [ # # # # : 0 : SPDK_DEBUGLOG(rdma, "rqpair %p has no thread\n", rqpair);
# # ]
3811 : 0 : return -EINVAL;
3812 : : }
3813 : :
3814 [ - + + - : 2575 : if (rqpair->last_wqe_reached || rqpair->last_wqe_reached_ctx != NULL) {
- + # # #
# # # #
# ]
3815 : 0 : SPDK_ERRLOG("LAST_WQE_REACHED already received for rqpair %p\n", rqpair);
3816 : 0 : return -EALREADY;
3817 : : }
3818 : :
3819 : 2575 : ctx = calloc(1, sizeof(*ctx));
3820 [ - + ]: 2575 : if (!ctx) {
3821 : 0 : return -ENOMEM;
3822 : : }
3823 : :
3824 [ # # # # ]: 2575 : ctx->rqpair = rqpair;
3825 [ # # # # ]: 2575 : rqpair->last_wqe_reached_ctx = ctx;
3826 : :
3827 : 2575 : rc = spdk_thread_send_msg(thr, nvmf_rdma_qpair_process_last_wqe_event, ctx);
3828 [ - + ]: 2575 : if (rc) {
3829 [ # # # # ]: 0 : rqpair->last_wqe_reached_ctx = NULL;
3830 : 0 : free(ctx);
3831 : 0 : }
3832 : :
3833 : 2575 : return rc;
3834 : 0 : }
3835 : :
3836 : : static int
3837 : 3983 : nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device)
3838 : : {
3839 : : int rc;
3840 : 3983 : struct spdk_nvmf_rdma_qpair *rqpair = NULL;
3841 : 0 : struct ibv_async_event event;
3842 : :
3843 [ # # # # ]: 3983 : rc = ibv_get_async_event(device->context, &event);
3844 : :
3845 [ + + ]: 3983 : if (rc) {
3846 : : /* In non-blocking mode -1 means there are no events available */
3847 : 1408 : return rc;
3848 : : }
3849 : :
3850 [ + - - # : 2575 : switch (event.event_type) {
# # ]
3851 : 2575 : case IBV_EVENT_QP_FATAL:
3852 : : case IBV_EVENT_QP_LAST_WQE_REACHED:
3853 : : case IBV_EVENT_QP_REQ_ERR:
3854 : : case IBV_EVENT_QP_ACCESS_ERR:
3855 : : case IBV_EVENT_COMM_EST:
3856 : : case IBV_EVENT_PATH_MIG:
3857 : : case IBV_EVENT_PATH_MIG_ERR:
3858 [ # # # # ]: 2575 : rqpair = event.element.qp->qp_context;
3859 [ - + ]: 2575 : if (!rqpair) {
3860 : : /* Any QP event for NVMe-RDMA initiator may be returned. */
3861 [ # # ]: 0 : SPDK_NOTICELOG("Async QP event for unknown QP: %s\n",
3862 : : ibv_event_type_str(event.event_type));
3863 : 0 : break;
3864 : : }
3865 : :
3866 [ - + - - : 2575 : switch (event.event_type) {
# # ]
3867 : 0 : case IBV_EVENT_QP_FATAL:
3868 : 0 : SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair);
3869 [ # # # # : 0 : spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0,
# # # # #
# # # # #
# # # # ]
3870 : : (uintptr_t)rqpair, event.event_type);
3871 [ # # # # ]: 0 : rqpair->ibv_in_error_state = true;
3872 [ # # ]: 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
3873 : 0 : break;
3874 : 2575 : case IBV_EVENT_QP_LAST_WQE_REACHED:
3875 : : /* This event only occurs for shared receive queues. */
3876 [ - + - + : 2575 : SPDK_DEBUGLOG(rdma, "Last WQE reached event received for rqpair %p\n", rqpair);
# # ]
3877 : 2575 : rc = nvmf_rdma_send_qpair_last_wqe_event(rqpair);
3878 [ - + ]: 2575 : if (rc) {
3879 : 0 : SPDK_WARNLOG("Failed to send LAST_WQE_REACHED event. rqpair %p, err %d\n", rqpair, rc);
3880 [ # # # # ]: 0 : rqpair->last_wqe_reached = true;
3881 : 0 : }
3882 : 2575 : break;
3883 : 0 : case IBV_EVENT_QP_REQ_ERR:
3884 : : case IBV_EVENT_QP_ACCESS_ERR:
3885 : : case IBV_EVENT_COMM_EST:
3886 : : case IBV_EVENT_PATH_MIG:
3887 : : case IBV_EVENT_PATH_MIG_ERR:
3888 [ # # ]: 0 : SPDK_NOTICELOG("Async QP event: %s\n",
3889 : : ibv_event_type_str(event.event_type));
3890 [ # # # # : 0 : spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0,
# # # # #
# # # # #
# # # # ]
3891 : : (uintptr_t)rqpair, event.event_type);
3892 [ # # # # ]: 0 : rqpair->ibv_in_error_state = true;
3893 : 0 : break;
3894 : 0 : default:
3895 : 0 : break;
3896 : : }
3897 : 2575 : break;
3898 : 0 : case IBV_EVENT_DEVICE_FATAL:
3899 [ # # # # : 0 : SPDK_ERRLOG("Device Fatal event[%s] received on %s. device: %p\n",
# # # # #
# ]
3900 : : ibv_event_type_str(event.event_type), ibv_get_device_name(device->context->device), device);
3901 [ # # # # ]: 0 : device->need_destroy = true;
3902 : 0 : break;
3903 : 0 : case IBV_EVENT_CQ_ERR:
3904 : : case IBV_EVENT_PORT_ACTIVE:
3905 : : case IBV_EVENT_PORT_ERR:
3906 : : case IBV_EVENT_LID_CHANGE:
3907 : : case IBV_EVENT_PKEY_CHANGE:
3908 : : case IBV_EVENT_SM_CHANGE:
3909 : : case IBV_EVENT_SRQ_ERR:
3910 : : case IBV_EVENT_SRQ_LIMIT_REACHED:
3911 : : case IBV_EVENT_CLIENT_REREGISTER:
3912 : : case IBV_EVENT_GID_CHANGE:
3913 : 0 : case IBV_EVENT_SQ_DRAINED:
3914 : : default:
3915 [ # # ]: 0 : SPDK_NOTICELOG("Async event: %s\n",
3916 : : ibv_event_type_str(event.event_type));
3917 [ # # # # : 0 : spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type);
# # # # #
# # # # #
# # # # ]
3918 : 0 : break;
3919 : : }
3920 : 2575 : ibv_ack_async_event(&event);
3921 : :
3922 : 2575 : return 0;
3923 : 0 : }
3924 : :
3925 : : static void
3926 : 1408 : nvmf_process_ib_events(struct spdk_nvmf_rdma_device *device, uint32_t max_events)
3927 : : {
3928 : 1408 : int rc = 0;
3929 : 1408 : uint32_t i = 0;
3930 : :
3931 [ + - ]: 3983 : for (i = 0; i < max_events; i++) {
3932 : 3983 : rc = nvmf_process_ib_event(device);
3933 [ + + ]: 3983 : if (rc) {
3934 : 1408 : break;
3935 : : }
3936 : 0 : }
3937 : :
3938 [ - + - + : 1408 : SPDK_DEBUGLOG(rdma, "Device %s: %u events processed\n", device->context->device->name, i);
# # # # #
# # # # #
# # ]
3939 : 1408 : }
3940 : :
3941 : : static int
3942 : 113395 : nvmf_rdma_accept(void *ctx)
3943 : : {
3944 : 113395 : int nfds, i = 0;
3945 : 113395 : struct spdk_nvmf_transport *transport = ctx;
3946 : : struct spdk_nvmf_rdma_transport *rtransport;
3947 : : struct spdk_nvmf_rdma_device *device, *tmp;
3948 : : uint32_t count;
3949 : : short revents;
3950 : : bool do_retry;
3951 : :
3952 : 113395 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3953 : 113395 : do_retry = nvmf_rdma_retry_listen_port(rtransport);
3954 : :
3955 [ # # # # : 113395 : count = nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0);
# # # # ]
3956 : :
3957 [ + + ]: 113395 : if (nfds <= 0) {
3958 [ # # ]: 106702 : return do_retry ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE;
3959 : : }
3960 : :
3961 : : /* The first poll descriptor is RDMA CM event */
3962 [ + + # # : 6693 : if (rtransport->poll_fds[i++].revents & POLLIN) {
# # # # #
# # # #
# ]
3963 : 5949 : nvmf_process_cm_events(transport, NVMF_RDMA_MAX_EVENTS_PER_POLL);
3964 [ # # ]: 5949 : nfds--;
3965 : 0 : }
3966 : :
3967 [ + + ]: 6693 : if (nfds == 0) {
3968 : 5285 : return SPDK_POLLER_BUSY;
3969 : : }
3970 : :
3971 : : /* Second and subsequent poll descriptors are IB async events */
3972 [ + + # # : 4224 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) {
# # # # #
# # # # #
# # ]
3973 [ # # # # : 2816 : revents = rtransport->poll_fds[i++].revents;
# # # # #
# # # ]
3974 [ + + ]: 2816 : if (revents & POLLIN) {
3975 [ - + + - : 1408 : if (spdk_likely(!device->need_destroy)) {
# # # # ]
3976 : 1408 : nvmf_process_ib_events(device, NVMF_RDMA_MAX_EVENTS_PER_POLL);
3977 [ - + - + : 1408 : if (spdk_unlikely(device->need_destroy)) {
# # # # ]
3978 : 0 : nvmf_rdma_handle_device_removal(rtransport, device);
3979 : 0 : }
3980 : 0 : }
3981 [ # # ]: 1408 : nfds--;
3982 [ + - - + ]: 1408 : } else if (revents & POLLNVAL || revents & POLLHUP) {
3983 : 0 : SPDK_ERRLOG("Receive unknown revent %x on device %p\n", (int)revents, device);
3984 [ # # ]: 0 : nfds--;
3985 : 0 : }
3986 : 0 : }
3987 : : /* check all flagged fd's have been served */
3988 [ - + # # ]: 1408 : assert(nfds == 0);
3989 : :
3990 : 1408 : return count > 0 ? SPDK_POLLER_BUSY : SPDK_POLLER_IDLE;
3991 : 0 : }
3992 : :
3993 : : static void
3994 : 516 : nvmf_rdma_cdata_init(struct spdk_nvmf_transport *transport, struct spdk_nvmf_subsystem *subsystem,
3995 : : struct spdk_nvmf_ctrlr_data *cdata)
3996 : : {
3997 [ # # # # : 516 : cdata->nvmf_specific.msdbd = NVMF_DEFAULT_MSDBD;
# # ]
3998 : :
3999 : : /* Disable in-capsule data transfer for RDMA controller when dif_insert_or_strip is enabled
4000 : : since in-capsule data only works with NVME drives that support SGL memory layout */
4001 [ - + - + : 516 : if (transport->opts.dif_insert_or_strip) {
# # # # #
# ]
4002 [ # # # # : 0 : cdata->nvmf_specific.ioccsz = sizeof(struct spdk_nvme_cmd) / 16;
# # ]
4003 : 0 : }
4004 : :
4005 [ - + # # : 516 : if (cdata->nvmf_specific.ioccsz > ((sizeof(struct spdk_nvme_cmd) + 0x1000) / 16)) {
# # # # ]
4006 : 0 : SPDK_WARNLOG("RDMA is configured to support up to 16 SGL entries while in capsule"
4007 : : " data is greater than 4KiB.\n");
4008 : 0 : SPDK_WARNLOG("When used in conjunction with the NVMe-oF initiator from the Linux "
4009 : : "kernel between versions 5.4 and 5.12 data corruption may occur for "
4010 : : "writes that are not a multiple of 4KiB in size.\n");
4011 : 0 : }
4012 : 516 : }
4013 : :
4014 : : static void
4015 : 128 : nvmf_rdma_discover(struct spdk_nvmf_transport *transport,
4016 : : struct spdk_nvme_transport_id *trid,
4017 : : struct spdk_nvmf_discovery_log_page_entry *entry)
4018 : : {
4019 [ # # # # ]: 128 : entry->trtype = SPDK_NVMF_TRTYPE_RDMA;
4020 [ # # # # : 128 : entry->adrfam = trid->adrfam;
# # # # ]
4021 [ # # # # ]: 128 : entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED;
4022 : :
4023 [ # # # # ]: 128 : spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' ');
4024 [ # # # # ]: 128 : spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' ');
4025 : :
4026 [ # # # # : 128 : entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED;
# # # # ]
4027 [ # # # # : 128 : entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE;
# # # # ]
4028 [ # # # # : 128 : entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM;
# # # # ]
4029 : 128 : }
4030 : :
4031 : : static int
4032 : 282 : nvmf_rdma_poller_create(struct spdk_nvmf_rdma_transport *rtransport,
4033 : : struct spdk_nvmf_rdma_poll_group *rgroup, struct spdk_nvmf_rdma_device *device,
4034 : : struct spdk_nvmf_rdma_poller **out_poller)
4035 : : {
4036 : : struct spdk_nvmf_rdma_poller *poller;
4037 : 0 : struct spdk_rdma_provider_srq_init_attr srq_init_attr;
4038 : 0 : struct spdk_nvmf_rdma_resource_opts opts;
4039 : : int num_cqe;
4040 : :
4041 : 282 : poller = calloc(1, sizeof(*poller));
4042 [ - + ]: 282 : if (!poller) {
4043 : 0 : SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n");
4044 : 0 : return -1;
4045 : : }
4046 : :
4047 [ # # # # ]: 282 : poller->device = device;
4048 [ # # # # ]: 282 : poller->group = rgroup;
4049 [ # # ]: 282 : *out_poller = poller;
4050 : :
4051 [ # # # # : 282 : RB_INIT(&poller->qpairs);
# # ]
4052 [ # # # # : 282 : STAILQ_INIT(&poller->qpairs_pending_send);
# # # # #
# # # # #
# # ]
4053 [ # # # # : 282 : STAILQ_INIT(&poller->qpairs_pending_recv);
# # # # #
# # # # #
# # ]
4054 : :
4055 [ # # # # : 282 : TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # ]
4056 [ - + - + : 282 : SPDK_DEBUGLOG(rdma, "Create poller %p on device %p in poll group %p.\n", poller, device, rgroup);
# # ]
4057 [ - + + - : 282 : if (rtransport->rdma_opts.no_srq == false && device->num_srq < device->attr.max_srq) {
+ - # # #
# # # # #
# # # # #
# # # ]
4058 [ - + # # : 282 : if ((int)rtransport->rdma_opts.max_srq_depth > device->attr.max_srq_wr) {
# # # # #
# # # #
# ]
4059 [ # # # # : 0 : SPDK_WARNLOG("Requested SRQ depth %u, max supported by dev %s is %d\n",
# # # # #
# # # # #
# # # # #
# # # ]
4060 : : rtransport->rdma_opts.max_srq_depth, device->context->device->name, device->attr.max_srq_wr);
4061 : 0 : }
4062 [ # # # # : 282 : poller->max_srq_depth = spdk_min((int)rtransport->rdma_opts.max_srq_depth, device->attr.max_srq_wr);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# ]
4063 : :
4064 [ # # # # ]: 282 : device->num_srq++;
4065 [ # # ]: 282 : memset(&srq_init_attr, 0, sizeof(srq_init_attr));
4066 [ # # # # ]: 282 : srq_init_attr.pd = device->pd;
4067 [ # # # # : 282 : srq_init_attr.stats = &poller->stat.qp_stats.recv;
# # # # ]
4068 [ # # # # : 282 : srq_init_attr.srq_init_attr.attr.max_wr = poller->max_srq_depth;
# # # # #
# ]
4069 [ # # # # : 282 : srq_init_attr.srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE);
# # # # #
# # # # #
# # # # #
# ]
4070 [ # # # # ]: 282 : poller->srq = spdk_rdma_provider_srq_create(&srq_init_attr);
4071 [ - + # # : 282 : if (!poller->srq) {
# # ]
4072 [ # # ]: 0 : SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno);
4073 : 0 : return -1;
4074 : : }
4075 : :
4076 [ # # # # : 282 : opts.qp = poller->srq;
# # ]
4077 [ # # # # : 282 : opts.map = device->map;
# # ]
4078 : 282 : opts.qpair = NULL;
4079 [ # # ]: 282 : opts.shared = true;
4080 [ # # # # : 282 : opts.max_queue_depth = poller->max_srq_depth;
# # ]
4081 [ # # # # : 282 : opts.in_capsule_data_size = rtransport->transport.opts.in_capsule_data_size;
# # # # #
# ]
4082 : :
4083 [ # # # # ]: 282 : poller->resources = nvmf_rdma_resources_create(&opts);
4084 [ - + # # : 282 : if (!poller->resources) {
# # ]
4085 : 0 : SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n");
4086 : 0 : return -1;
4087 : : }
4088 : 0 : }
4089 : :
4090 : : /*
4091 : : * When using an srq, we can limit the completion queue at startup.
4092 : : * The following formula represents the calculation:
4093 : : * num_cqe = num_recv + num_data_wr + num_send_wr.
4094 : : * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth
4095 : : */
4096 [ + - # # : 282 : if (poller->srq) {
# # ]
4097 [ # # # # : 282 : num_cqe = poller->max_srq_depth * 3;
# # ]
4098 : 0 : } else {
4099 [ # # # # : 0 : num_cqe = rtransport->rdma_opts.num_cqe;
# # ]
4100 : : }
4101 : :
4102 [ # # # # : 282 : poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0);
# # # # ]
4103 [ - + # # : 282 : if (!poller->cq) {
# # ]
4104 : 0 : SPDK_ERRLOG("Unable to create completion queue\n");
4105 : 0 : return -1;
4106 : : }
4107 [ # # # # ]: 282 : poller->num_cqe = num_cqe;
4108 : 282 : return 0;
4109 : 0 : }
4110 : :
4111 : : static void
4112 : 0 : _nvmf_rdma_register_poller_in_group(void *c)
4113 : : {
4114 : 0 : struct spdk_nvmf_rdma_poller *poller;
4115 : 0 : struct poller_manage_ctx *ctx = c;
4116 : : struct spdk_nvmf_rdma_device *device;
4117 : : int rc;
4118 : :
4119 [ # # # # : 0 : rc = nvmf_rdma_poller_create(ctx->rtransport, ctx->rgroup, ctx->device, &poller);
# # # # #
# # # ]
4120 [ # # # # ]: 0 : if (rc < 0 && poller) {
4121 : 0 : nvmf_rdma_poller_destroy(poller);
4122 : 0 : }
4123 : :
4124 [ # # # # ]: 0 : device = ctx->device;
4125 [ # # ]: 0 : if (nvmf_rdma_all_pollers_management_done(ctx)) {
4126 [ # # # # ]: 0 : device->is_ready = true;
4127 : 0 : }
4128 : 0 : }
4129 : :
4130 : : static void nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group);
4131 : :
4132 : : static struct spdk_nvmf_transport_poll_group *
4133 : 156 : nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport,
4134 : : struct spdk_nvmf_poll_group *group)
4135 : : {
4136 : : struct spdk_nvmf_rdma_transport *rtransport;
4137 : : struct spdk_nvmf_rdma_poll_group *rgroup;
4138 : 15 : struct spdk_nvmf_rdma_poller *poller;
4139 : : struct spdk_nvmf_rdma_device *device;
4140 : : int rc;
4141 : :
4142 [ - + ]: 156 : if (spdk_interrupt_mode_is_enabled()) {
4143 : 0 : SPDK_ERRLOG("RDMA transport does not support interrupt mode\n");
4144 : 0 : return NULL;
4145 : : }
4146 : :
4147 : 156 : rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
4148 : :
4149 : 156 : rgroup = calloc(1, sizeof(*rgroup));
4150 [ - + ]: 156 : if (!rgroup) {
4151 : 0 : return NULL;
4152 : : }
4153 : :
4154 [ # # # # : 156 : TAILQ_INIT(&rgroup->pollers);
# # # # #
# # # # #
# # ]
4155 : :
4156 [ + + # # : 438 : TAILQ_FOREACH(device, &rtransport->devices, link) {
# # # # #
# # # #
# ]
4157 : 282 : rc = nvmf_rdma_poller_create(rtransport, rgroup, device, &poller);
4158 [ - + ]: 282 : if (rc < 0) {
4159 [ # # ]: 0 : nvmf_rdma_poll_group_destroy(&rgroup->group);
4160 : 0 : return NULL;
4161 : : }
4162 : 0 : }
4163 : :
4164 [ # # # # : 156 : TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # ]
4165 [ + + # # : 156 : if (rtransport->conn_sched.next_admin_pg == NULL) {
# # # # ]
4166 [ # # # # : 50 : rtransport->conn_sched.next_admin_pg = rgroup;
# # ]
4167 [ # # # # : 50 : rtransport->conn_sched.next_io_pg = rgroup;
# # ]
4168 : 0 : }
4169 : :
4170 [ # # ]: 156 : return &rgroup->group;
4171 : 0 : }
4172 : :
4173 : : static uint32_t
4174 : 7065 : nvmf_poll_group_get_io_qpair_count(struct spdk_nvmf_poll_group *pg)
4175 : : {
4176 : : uint32_t count;
4177 : :
4178 : : /* Just assume that unassociated qpairs will eventually be io
4179 : : * qpairs. This is close enough for the use cases for this
4180 : : * function.
4181 : : */
4182 [ - + # # ]: 7065 : pthread_mutex_lock(&pg->mutex);
4183 [ # # # # : 7065 : count = pg->stat.current_io_qpairs + pg->current_unassociated_qpairs;
# # # # #
# ]
4184 [ - + # # ]: 7065 : pthread_mutex_unlock(&pg->mutex);
4185 : :
4186 : 7065 : return count;
4187 : : }
4188 : :
4189 : : static struct spdk_nvmf_transport_poll_group *
4190 : 2617 : nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair)
4191 : : {
4192 : : struct spdk_nvmf_rdma_transport *rtransport;
4193 : : struct spdk_nvmf_rdma_poll_group **pg;
4194 : : struct spdk_nvmf_transport_poll_group *result;
4195 : : uint32_t count;
4196 : :
4197 [ # # # # ]: 2617 : rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
4198 : :
4199 [ + + # # : 2617 : if (TAILQ_EMPTY(&rtransport->poll_groups)) {
# # # # ]
4200 : 6 : return NULL;
4201 : : }
4202 : :
4203 [ + + # # : 2611 : if (qpair->qid == 0) {
# # ]
4204 [ # # # # ]: 536 : pg = &rtransport->conn_sched.next_admin_pg;
4205 : 0 : } else {
4206 : : struct spdk_nvmf_rdma_poll_group *pg_min, *pg_start, *pg_current;
4207 : : uint32_t min_value;
4208 : :
4209 [ # # # # ]: 2075 : pg = &rtransport->conn_sched.next_io_pg;
4210 [ # # ]: 2075 : pg_min = *pg;
4211 [ # # ]: 2075 : pg_start = *pg;
4212 [ # # ]: 2075 : pg_current = *pg;
4213 [ # # # # : 2075 : min_value = nvmf_poll_group_get_io_qpair_count(pg_current->group.group);
# # ]
4214 : :
4215 : 0 : while (1) {
4216 [ # # # # : 4990 : count = nvmf_poll_group_get_io_qpair_count(pg_current->group.group);
# # ]
4217 : :
4218 [ + + ]: 4990 : if (count < min_value) {
4219 : 1 : min_value = count;
4220 : 1 : pg_min = pg_current;
4221 : 0 : }
4222 : :
4223 [ # # # # : 4990 : pg_current = TAILQ_NEXT(pg_current, link);
# # ]
4224 [ + + ]: 4990 : if (pg_current == NULL) {
4225 [ # # # # : 1473 : pg_current = TAILQ_FIRST(&rtransport->poll_groups);
# # ]
4226 : 0 : }
4227 : :
4228 [ + + + + ]: 4990 : if (pg_current == pg_start || min_value == 0) {
4229 : 0 : break;
4230 : : }
4231 : : }
4232 [ # # ]: 2075 : *pg = pg_min;
4233 : : }
4234 : :
4235 [ - + # # : 2611 : assert(*pg != NULL);
# # ]
4236 : :
4237 [ # # # # ]: 2611 : result = &(*pg)->group;
4238 : :
4239 [ # # # # : 2611 : *pg = TAILQ_NEXT(*pg, link);
# # # # #
# ]
4240 [ + + # # ]: 2611 : if (*pg == NULL) {
4241 [ # # # # : 980 : *pg = TAILQ_FIRST(&rtransport->poll_groups);
# # # # ]
4242 : 0 : }
4243 : :
4244 : 2611 : return result;
4245 : 0 : }
4246 : :
4247 : : static void
4248 : 282 : nvmf_rdma_poller_destroy(struct spdk_nvmf_rdma_poller *poller)
4249 : : {
4250 : : struct spdk_nvmf_rdma_qpair *qpair, *tmp_qpair;
4251 : : int rc;
4252 : :
4253 [ + + # # : 282 : TAILQ_REMOVE(&poller->group->pollers, poller, link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # ]
4254 [ - + - - : 282 : RB_FOREACH_SAFE(qpair, qpairs_tree, &poller->qpairs, tmp_qpair) {
# # ]
4255 : 0 : nvmf_rdma_qpair_destroy(qpair);
4256 : 0 : }
4257 : :
4258 [ + - # # : 282 : if (poller->srq) {
# # ]
4259 [ + - # # : 282 : if (poller->resources) {
# # ]
4260 [ # # # # ]: 282 : nvmf_rdma_resources_destroy(poller->resources);
4261 : 0 : }
4262 [ # # # # ]: 282 : spdk_rdma_provider_srq_destroy(poller->srq);
4263 [ - + - + : 282 : SPDK_DEBUGLOG(rdma, "Destroyed RDMA shared queue %p\n", poller->srq);
# # # # #
# ]
4264 : 0 : }
4265 : :
4266 [ + - # # : 282 : if (poller->cq) {
# # ]
4267 [ # # # # ]: 282 : rc = ibv_destroy_cq(poller->cq);
4268 [ - + ]: 282 : if (rc != 0) {
4269 [ # # ]: 0 : SPDK_ERRLOG("Destroy cq return %d, error: %s\n", rc, strerror(errno));
4270 : 0 : }
4271 : 0 : }
4272 : :
4273 [ - + # # : 282 : if (poller->destroy_cb) {
# # ]
4274 [ # # # # : 0 : poller->destroy_cb(poller->destroy_cb_ctx);
# # # # #
# # # ]
4275 [ # # # # ]: 0 : poller->destroy_cb = NULL;
4276 : 0 : }
4277 : :
4278 : 282 : free(poller);
4279 : 282 : }
4280 : :
4281 : : static void
4282 : 156 : nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group)
4283 : : {
4284 : : struct spdk_nvmf_rdma_poll_group *rgroup, *next_rgroup;
4285 : : struct spdk_nvmf_rdma_poller *poller, *tmp;
4286 : : struct spdk_nvmf_rdma_transport *rtransport;
4287 : :
4288 : 156 : rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
4289 [ - + ]: 156 : if (!rgroup) {
4290 : 0 : return;
4291 : : }
4292 : :
4293 [ + + # # : 438 : TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) {
# # # # #
# # # # #
# # ]
4294 : 282 : nvmf_rdma_poller_destroy(poller);
4295 : 0 : }
4296 : :
4297 [ - + # # : 156 : if (rgroup->group.transport == NULL) {
# # # # ]
4298 : : /* Transport can be NULL when nvmf_rdma_poll_group_create()
4299 : : * calls this function directly in a failure path. */
4300 : 0 : free(rgroup);
4301 : 0 : return;
4302 : : }
4303 : :
4304 [ # # # # : 156 : rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport);
# # ]
4305 : :
4306 [ # # # # : 156 : next_rgroup = TAILQ_NEXT(rgroup, link);
# # ]
4307 [ + + # # : 156 : TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # ]
4308 [ + + ]: 156 : if (next_rgroup == NULL) {
4309 [ # # # # : 73 : next_rgroup = TAILQ_FIRST(&rtransport->poll_groups);
# # ]
4310 : 0 : }
4311 [ + + # # : 156 : if (rtransport->conn_sched.next_admin_pg == rgroup) {
# # # # ]
4312 [ # # # # : 97 : rtransport->conn_sched.next_admin_pg = next_rgroup;
# # ]
4313 : 0 : }
4314 [ + + # # : 156 : if (rtransport->conn_sched.next_io_pg == rgroup) {
# # # # ]
4315 [ # # # # : 94 : rtransport->conn_sched.next_io_pg = next_rgroup;
# # ]
4316 : 0 : }
4317 : :
4318 : 156 : free(rgroup);
4319 : 0 : }
4320 : :
4321 : : static void
4322 : 0 : nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair)
4323 : : {
4324 [ # # # # : 0 : if (rqpair->cm_id != NULL) {
# # ]
4325 [ # # # # ]: 0 : nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES);
4326 : 0 : }
4327 : 0 : }
4328 : :
4329 : : static int
4330 : 2575 : nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group,
4331 : : struct spdk_nvmf_qpair *qpair)
4332 : : {
4333 : : struct spdk_nvmf_rdma_poll_group *rgroup;
4334 : : struct spdk_nvmf_rdma_qpair *rqpair;
4335 : : struct spdk_nvmf_rdma_device *device;
4336 : : struct spdk_nvmf_rdma_poller *poller;
4337 : : int rc;
4338 : :
4339 : 2575 : rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
4340 : 2575 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4341 : :
4342 [ # # # # ]: 2575 : device = rqpair->device;
4343 : :
4344 [ + - # # : 2577 : TAILQ_FOREACH(poller, &rgroup->pollers, link) {
# # # # #
# # # #
# ]
4345 [ + + # # : 2577 : if (poller->device == device) {
# # ]
4346 : 2575 : break;
4347 : : }
4348 : 0 : }
4349 : :
4350 [ - + ]: 2575 : if (!poller) {
4351 : 0 : SPDK_ERRLOG("No poller found for device.\n");
4352 : 0 : return -1;
4353 : : }
4354 : :
4355 [ - + - + : 2575 : if (poller->need_destroy) {
# # # # ]
4356 : 0 : SPDK_ERRLOG("Poller is destroying.\n");
4357 : 0 : return -1;
4358 : : }
4359 : :
4360 [ # # # # ]: 2575 : rqpair->poller = poller;
4361 [ # # # # : 2575 : rqpair->srq = rqpair->poller->srq;
# # # # #
# # # ]
4362 : :
4363 : 2575 : rc = nvmf_rdma_qpair_initialize(qpair);
4364 [ - + ]: 2575 : if (rc < 0) {
4365 : 0 : SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair);
4366 [ # # # # ]: 0 : rqpair->poller = NULL;
4367 [ # # # # ]: 0 : rqpair->srq = NULL;
4368 : 0 : return -1;
4369 : : }
4370 : :
4371 [ # # ]: 2575 : RB_INSERT(qpairs_tree, &poller->qpairs, rqpair);
4372 : :
4373 [ # # # # ]: 2575 : rc = nvmf_rdma_event_accept(rqpair->cm_id, rqpair);
4374 [ - + ]: 2575 : if (rc) {
4375 : : /* Try to reject, but we probably can't */
4376 : 0 : nvmf_rdma_qpair_reject_connection(rqpair);
4377 : 0 : return -1;
4378 : : }
4379 : :
4380 : 2575 : return 0;
4381 : 0 : }
4382 : :
4383 : : static int
4384 : 2575 : nvmf_rdma_poll_group_remove(struct spdk_nvmf_transport_poll_group *group,
4385 : : struct spdk_nvmf_qpair *qpair)
4386 : : {
4387 : : struct spdk_nvmf_rdma_qpair *rqpair;
4388 : :
4389 : 2575 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4390 [ - + # # : 2575 : assert(group->transport->tgt != NULL);
# # # # #
# # # ]
4391 : :
4392 [ # # # # : 2575 : rqpair->destruct_channel = spdk_get_io_channel(group->transport->tgt);
# # # # #
# # # ]
4393 : :
4394 [ - + # # : 2575 : if (!rqpair->destruct_channel) {
# # ]
4395 : 0 : SPDK_WARNLOG("failed to get io_channel, qpair %p\n", qpair);
4396 : 0 : return 0;
4397 : : }
4398 : :
4399 : : /* Sanity check that we get io_channel on the correct thread */
4400 [ + - # # : 2575 : if (qpair->group) {
# # ]
4401 [ - + # # : 2575 : assert(qpair->group->thread == spdk_io_channel_get_thread(rqpair->destruct_channel));
# # # # #
# # # # #
# # ]
4402 : 0 : }
4403 : :
4404 : 2575 : return 0;
4405 : 0 : }
4406 : :
4407 : : static int
4408 : 1361 : nvmf_rdma_request_free(struct spdk_nvmf_request *req)
4409 : : {
4410 : 1361 : struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
4411 [ # # # # : 1361 : struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport,
# # # # ]
4412 : : struct spdk_nvmf_rdma_transport, transport);
4413 [ # # # # : 1361 : struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair,
# # ]
4414 : : struct spdk_nvmf_rdma_qpair, qpair);
4415 : :
4416 : : /*
4417 : : * AER requests are freed when a qpair is destroyed. The recv corresponding to that request
4418 : : * needs to be returned to the shared receive queue or the poll group will eventually be
4419 : : * starved of RECV structures.
4420 : : */
4421 [ + - + - : 1361 : if (rqpair->srq && rdma_req->recv) {
# # # # #
# # # ]
4422 : : int rc;
4423 : 0 : struct ibv_recv_wr *bad_recv_wr;
4424 : :
4425 [ # # # # : 1361 : spdk_rdma_provider_srq_queue_recv_wrs(rqpair->srq, &rdma_req->recv->wr);
# # # # #
# ]
4426 [ # # # # ]: 1361 : rc = spdk_rdma_provider_srq_flush_recv_wrs(rqpair->srq, &bad_recv_wr);
4427 [ - + ]: 1361 : if (rc) {
4428 : 0 : SPDK_ERRLOG("Unable to re-post rx descriptor\n");
4429 : 0 : }
4430 : 0 : }
4431 : :
4432 : 1361 : _nvmf_rdma_request_free(rdma_req, rtransport);
4433 : 1361 : return 0;
4434 : : }
4435 : :
4436 : : static int
4437 : 7921753 : nvmf_rdma_request_complete(struct spdk_nvmf_request *req)
4438 : : {
4439 [ # # # # : 7921753 : struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(req->qpair->transport,
# # # # ]
4440 : : struct spdk_nvmf_rdma_transport, transport);
4441 : 7921753 : struct spdk_nvmf_rdma_request *rdma_req = SPDK_CONTAINEROF(req,
4442 : : struct spdk_nvmf_rdma_request, req);
4443 [ # # # # : 7921753 : struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair,
# # ]
4444 : : struct spdk_nvmf_rdma_qpair, qpair);
4445 : :
4446 [ - + - + : 7921753 : if (spdk_unlikely(rqpair->ibv_in_error_state)) {
# # # # ]
4447 : : /* The connection is dead. Move the request directly to the completed state. */
4448 [ # # # # ]: 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
4449 : 0 : } else {
4450 : : /* The connection is alive, so process the request as normal */
4451 [ # # # # ]: 7921753 : rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
4452 : : }
4453 : :
4454 : 7921753 : nvmf_rdma_request_process(rtransport, rdma_req);
4455 : :
4456 : 7921753 : return 0;
4457 : : }
4458 : :
4459 : : static void
4460 : 2575 : nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair,
4461 : : spdk_nvmf_transport_qpair_fini_cb cb_fn, void *cb_arg)
4462 : : {
4463 : 2575 : struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4464 : :
4465 [ # # # # ]: 2575 : rqpair->to_close = true;
4466 : :
4467 : : /* This happens only when the qpair is disconnected before
4468 : : * it is added to the poll group. Since there is no poll group,
4469 : : * the RDMA qp has not been initialized yet and the RDMA CM
4470 : : * event has not yet been acknowledged, so we need to reject it.
4471 : : */
4472 [ - + # # : 2575 : if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) {
# # # # ]
4473 : 0 : nvmf_rdma_qpair_reject_connection(rqpair);
4474 : 0 : nvmf_rdma_qpair_destroy(rqpair);
4475 : 0 : return;
4476 : : }
4477 : :
4478 [ + - # # : 2575 : if (rqpair->rdma_qp) {
# # ]
4479 [ # # # # ]: 2575 : spdk_rdma_provider_qp_disconnect(rqpair->rdma_qp);
4480 : 0 : }
4481 : :
4482 : 2575 : nvmf_rdma_destroy_drained_qpair(rqpair);
4483 : :
4484 [ + - ]: 2575 : if (cb_fn) {
4485 [ # # # # ]: 2575 : cb_fn(cb_arg);
4486 : 0 : }
4487 : 0 : }
4488 : :
4489 : : static struct spdk_nvmf_rdma_qpair *
4490 : 7924518 : get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc)
4491 : : {
4492 : 0 : struct spdk_nvmf_rdma_qpair find;
4493 : :
4494 [ # # # # : 7924518 : find.qp_num = wc->qp_num;
# # ]
4495 : :
4496 [ # # ]: 7924518 : return RB_FIND(qpairs_tree, &rpoller->qpairs, &find);
4497 : : }
4498 : :
4499 : : #ifdef DEBUG
4500 : : static int
4501 : 7921522 : nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req)
4502 : : {
4503 [ + + # # : 12001056 : return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST ||
# # ]
4504 [ + - # # ]: 4079534 : rdma_req->state == RDMA_REQUEST_STATE_COMPLETING;
4505 : : }
4506 : : #endif
4507 : :
4508 : : static void
4509 : 0 : _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr,
4510 : : int rc)
4511 : : {
4512 : : struct spdk_nvmf_rdma_recv *rdma_recv;
4513 : : struct spdk_nvmf_rdma_wr *bad_rdma_wr;
4514 : :
4515 [ # # ]: 0 : SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc);
4516 [ # # ]: 0 : while (bad_recv_wr != NULL) {
4517 [ # # # # ]: 0 : bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id;
4518 : 0 : rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr);
4519 : :
4520 [ # # # # : 0 : rdma_recv->qpair->current_recv_depth++;
# # ]
4521 [ # # # # ]: 0 : bad_recv_wr = bad_recv_wr->next;
4522 [ # # # # : 0 : SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc);
# # ]
4523 [ # # # # : 0 : spdk_nvmf_qpair_disconnect(&rdma_recv->qpair->qpair);
# # ]
4524 : : }
4525 : 0 : }
4526 : :
4527 : : static void
4528 : 0 : _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc)
4529 : : {
4530 [ # # ]: 0 : SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc);
4531 [ # # ]: 0 : while (bad_recv_wr != NULL) {
4532 [ # # # # ]: 0 : bad_recv_wr = bad_recv_wr->next;
4533 [ # # ]: 0 : rqpair->current_recv_depth++;
4534 : : }
4535 [ # # ]: 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
4536 : 0 : }
4537 : :
4538 : : static void
4539 : 899882772 : _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport,
4540 : : struct spdk_nvmf_rdma_poller *rpoller)
4541 : : {
4542 : : struct spdk_nvmf_rdma_qpair *rqpair;
4543 : 0 : struct ibv_recv_wr *bad_recv_wr;
4544 : : int rc;
4545 : :
4546 [ + - # # : 899882772 : if (rpoller->srq) {
# # ]
4547 [ # # # # ]: 899882772 : rc = spdk_rdma_provider_srq_flush_recv_wrs(rpoller->srq, &bad_recv_wr);
4548 [ - + ]: 899882772 : if (spdk_unlikely(rc)) {
4549 : 0 : _poller_reset_failed_recvs(rpoller, bad_recv_wr, rc);
4550 : 0 : }
4551 : 0 : } else {
4552 [ # # # # : 0 : while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) {
# # # # ]
4553 [ # # # # : 0 : rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv);
# # ]
4554 [ # # # # ]: 0 : rc = spdk_rdma_provider_qp_flush_recv_wrs(rqpair->rdma_qp, &bad_recv_wr);
4555 [ # # ]: 0 : if (spdk_unlikely(rc)) {
4556 : 0 : _qp_reset_failed_recvs(rqpair, bad_recv_wr, rc);
4557 : 0 : }
4558 [ # # # # : 0 : STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# ]
4559 : : }
4560 : : }
4561 : 899882772 : }
4562 : :
4563 : : static void
4564 : 0 : _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport,
4565 : : struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc)
4566 : : {
4567 : : struct spdk_nvmf_rdma_wr *bad_rdma_wr;
4568 : 0 : struct spdk_nvmf_rdma_request *prev_rdma_req = NULL, *cur_rdma_req = NULL;
4569 : :
4570 [ # # ]: 0 : SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc);
4571 [ # # # # : 0 : for (; bad_wr != NULL; bad_wr = bad_wr->next) {
# # ]
4572 [ # # # # ]: 0 : bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id;
4573 [ # # # # : 0 : assert(rqpair->current_send_depth > 0);
# # # # ]
4574 [ # # ]: 0 : rqpair->current_send_depth--;
4575 [ # # # # : 0 : switch (bad_rdma_wr->type) {
# # # ]
4576 : 0 : case RDMA_WR_TYPE_DATA:
4577 : 0 : cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data_wr);
4578 [ # # # # : 0 : if (bad_wr->opcode == IBV_WR_RDMA_READ) {
# # ]
4579 [ # # # # : 0 : assert(rqpair->current_read_depth > 0);
# # # # ]
4580 [ # # ]: 0 : rqpair->current_read_depth--;
4581 : 0 : }
4582 : 0 : break;
4583 : 0 : case RDMA_WR_TYPE_SEND:
4584 : 0 : cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr);
4585 : 0 : break;
4586 : 0 : default:
4587 : 0 : SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair);
4588 : 0 : prev_rdma_req = cur_rdma_req;
4589 : 0 : continue;
4590 : : }
4591 : :
4592 [ # # ]: 0 : if (prev_rdma_req == cur_rdma_req) {
4593 : : /* this request was handled by an earlier wr. i.e. we were performing an nvme read. */
4594 : : /* We only have to check against prev_wr since each requests wrs are contiguous in this list. */
4595 : 0 : continue;
4596 : : }
4597 : :
4598 [ # # # # : 0 : switch (cur_rdma_req->state) {
# # # ]
4599 : 0 : case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
4600 [ # # # # : 0 : cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR;
# # # # #
# # # #
# ]
4601 [ # # # # : 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, cur_rdma_req, state_link);
# # # # #
# # # # #
# # # # #
# # # #
# ]
4602 [ # # # # ]: 0 : cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
4603 : 0 : break;
4604 : 0 : case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST:
4605 : : case RDMA_REQUEST_STATE_COMPLETING:
4606 [ # # # # ]: 0 : cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
4607 : 0 : break;
4608 : 0 : default:
4609 [ # # # # ]: 0 : SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n",
4610 : : cur_rdma_req->state, rqpair);
4611 : 0 : continue;
4612 : : }
4613 : :
4614 : 0 : nvmf_rdma_request_process(rtransport, cur_rdma_req);
4615 : 0 : prev_rdma_req = cur_rdma_req;
4616 : 0 : }
4617 : :
4618 [ # # # # ]: 0 : if (spdk_nvmf_qpair_is_active(&rqpair->qpair)) {
4619 : : /* Disconnect the connection. */
4620 [ # # ]: 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
4621 : 0 : }
4622 : :
4623 : 0 : }
4624 : :
4625 : : static void
4626 : 899882772 : _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport,
4627 : : struct spdk_nvmf_rdma_poller *rpoller)
4628 : : {
4629 : : struct spdk_nvmf_rdma_qpair *rqpair;
4630 : 899882772 : struct ibv_send_wr *bad_wr = NULL;
4631 : : int rc;
4632 : :
4633 [ + + # # : 1801115588 : while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) {
# # # # ]
4634 [ # # # # : 1350044 : rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send);
# # ]
4635 [ # # # # ]: 1350044 : rc = spdk_rdma_provider_qp_flush_send_wrs(rqpair->rdma_qp, &bad_wr);
4636 : :
4637 : : /* bad wr always points to the first wr that failed. */
4638 [ - + ]: 1350044 : if (spdk_unlikely(rc)) {
4639 : 0 : _qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc);
4640 : 0 : }
4641 [ + + # # : 1350044 : STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# ]
4642 : : }
4643 : 899882772 : }
4644 : :
4645 : : static const char *
4646 : 0 : nvmf_rdma_wr_type_str(enum spdk_nvmf_rdma_wr_type wr_type)
4647 : : {
4648 [ # # # # ]: 0 : switch (wr_type) {
4649 : 0 : case RDMA_WR_TYPE_RECV:
4650 : 0 : return "RECV";
4651 : 0 : case RDMA_WR_TYPE_SEND:
4652 : 0 : return "SEND";
4653 : 0 : case RDMA_WR_TYPE_DATA:
4654 : 0 : return "DATA";
4655 : 0 : default:
4656 : 0 : SPDK_ERRLOG("Unknown WR type %d\n", wr_type);
4657 [ # # ]: 0 : SPDK_UNREACHABLE();
4658 : : }
4659 : 0 : }
4660 : :
4661 : : static inline void
4662 : 0 : nvmf_rdma_log_wc_status(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_wc *wc)
4663 : : {
4664 [ # # # # : 0 : enum spdk_nvmf_rdma_wr_type wr_type = ((struct spdk_nvmf_rdma_wr *)wc->wr_id)->type;
# # # # ]
4665 : :
4666 [ # # # # : 0 : if (wc->status == IBV_WC_WR_FLUSH_ERR) {
# # ]
4667 : : /* If qpair is in ERR state, we will receive completions for all posted and not completed
4668 : : * Work Requests with IBV_WC_WR_FLUSH_ERR status. Don't log an error in that case */
4669 [ # # # # : 0 : SPDK_DEBUGLOG(rdma,
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # ]
4670 : : "Error on CQ %p, (qp state %d, in_error %d) request 0x%lu, type %s, status: (%d): %s\n",
4671 : : rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_in_error_state, wc->wr_id,
4672 : : nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status));
4673 : 0 : } else {
4674 [ # # # # : 0 : SPDK_ERRLOG("Error on CQ %p, (qp state %d, in_error %d) request 0x%lu, type %s, status: (%d): %s\n",
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # ]
4675 : : rqpair->poller->cq, rqpair->qpair.state, rqpair->ibv_in_error_state, wc->wr_id,
4676 : : nvmf_rdma_wr_type_str(wr_type), wc->status, ibv_wc_status_str(wc->status));
4677 : : }
4678 : 0 : }
4679 : :
4680 : : static int
4681 : 899882772 : nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport,
4682 : : struct spdk_nvmf_rdma_poller *rpoller)
4683 : : {
4684 : 0 : struct ibv_wc wc[32];
4685 : : struct spdk_nvmf_rdma_wr *rdma_wr;
4686 : : struct spdk_nvmf_rdma_request *rdma_req;
4687 : : struct spdk_nvmf_rdma_recv *rdma_recv;
4688 : : struct spdk_nvmf_rdma_qpair *rqpair, *tmp_rqpair;
4689 : : int reaped, i;
4690 : 899882772 : int count = 0;
4691 : : int rc;
4692 : 899882772 : bool error = false;
4693 : 899882772 : uint64_t poll_tsc = spdk_get_ticks();
4694 : :
4695 [ - + - + : 899882772 : if (spdk_unlikely(rpoller->need_destroy)) {
# # # # ]
4696 : : /* If qpair is closed before poller destroy, nvmf_rdma_destroy_drained_qpair may not
4697 : : * be called because we cannot poll anything from cq. So we call that here to force
4698 : : * destroy the qpair after to_close turning true.
4699 : : */
4700 [ # # # # : 0 : RB_FOREACH_SAFE(rqpair, qpairs_tree, &rpoller->qpairs, tmp_rqpair) {
# # ]
4701 : 0 : nvmf_rdma_destroy_drained_qpair(rqpair);
4702 : 0 : }
4703 : 0 : return 0;
4704 : : }
4705 : :
4706 : : /* Poll for completing operations. */
4707 [ # # # # ]: 899882772 : reaped = ibv_poll_cq(rpoller->cq, 32, wc);
4708 [ - + ]: 899882772 : if (spdk_unlikely(reaped < 0)) {
4709 [ # # # # ]: 0 : SPDK_ERRLOG("Error polling CQ! (%d): %s\n",
4710 : : errno, spdk_strerror(errno));
4711 : 0 : return -1;
4712 [ + + ]: 899882772 : } else if (reaped == 0) {
4713 [ # # # # ]: 898233522 : rpoller->stat.idle_polls++;
4714 : 0 : }
4715 : :
4716 [ # # # # ]: 899882772 : rpoller->stat.polls++;
4717 [ # # # # : 899882772 : rpoller->stat.completions += reaped;
# # ]
4718 : :
4719 [ + + # # ]: 917514936 : for (i = 0; i < reaped; i++) {
4720 : :
4721 [ # # # # : 17632164 : rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id;
# # # # ]
4722 : :
4723 [ + + + - : 17632164 : switch (rdma_wr->type) {
# # # # ]
4724 : 7921522 : case RDMA_WR_TYPE_SEND:
4725 : 7921522 : rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp_wr);
4726 [ # # # # : 7921522 : rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
# # ]
4727 : :
4728 [ + - # # : 7921522 : if (spdk_likely(!wc[i].status)) {
# # # # #
# ]
4729 [ # # ]: 7921522 : count++;
4730 [ - + # # : 7921522 : assert(wc[i].opcode == IBV_WC_SEND);
# # # # #
# # # ]
4731 [ - + # # ]: 7921522 : assert(nvmf_rdma_req_is_completing(rdma_req));
4732 : 0 : }
4733 : :
4734 [ # # # # ]: 7921522 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
4735 : : /* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */
4736 [ - + # # : 7921522 : assert(rqpair->current_send_depth >= (uint32_t)rdma_req->num_outstanding_data_wr + 1);
# # # # #
# # # ]
4737 [ # # # # : 7921522 : rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1;
# # # # #
# ]
4738 [ # # # # ]: 7921522 : rdma_req->num_outstanding_data_wr = 0;
4739 : :
4740 : 7921522 : nvmf_rdma_request_process(rtransport, rdma_req);
4741 : 7921522 : break;
4742 : 7924518 : case RDMA_WR_TYPE_RECV:
4743 : : /* rdma_recv->qpair will be invalid if using an SRQ. In that case we have to get the qpair from the wc. */
4744 : 7924518 : rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr);
4745 [ + - # # : 7924518 : if (rpoller->srq != NULL) {
# # ]
4746 [ # # # # : 7924518 : rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]);
# # # # ]
4747 : : /* It is possible that there are still some completions for destroyed QP
4748 : : * associated with SRQ. We just ignore these late completions and re-post
4749 : : * receive WRs back to SRQ.
4750 : : */
4751 [ + + # # : 7924518 : if (spdk_unlikely(NULL == rdma_recv->qpair)) {
# # ]
4752 : 0 : struct ibv_recv_wr *bad_wr;
4753 : :
4754 [ # # # # : 5 : rdma_recv->wr.next = NULL;
# # ]
4755 [ # # # # : 5 : spdk_rdma_provider_srq_queue_recv_wrs(rpoller->srq, &rdma_recv->wr);
# # ]
4756 [ # # # # ]: 5 : rc = spdk_rdma_provider_srq_flush_recv_wrs(rpoller->srq, &bad_wr);
4757 [ - + ]: 5 : if (rc) {
4758 : 0 : SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc);
4759 : 0 : }
4760 : 5 : continue;
4761 : : }
4762 : 0 : }
4763 [ # # # # ]: 7924513 : rqpair = rdma_recv->qpair;
4764 : :
4765 [ - + # # ]: 7924513 : assert(rqpair != NULL);
4766 [ + - # # : 7924513 : if (spdk_likely(!wc[i].status)) {
# # # # #
# ]
4767 [ - + # # : 7924513 : assert(wc[i].opcode == IBV_WC_RECV);
# # # # #
# # # ]
4768 [ - + # # : 7924513 : if (rqpair->current_recv_depth >= rqpair->max_queue_depth) {
# # # # #
# ]
4769 [ # # ]: 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
4770 : 0 : break;
4771 : : }
4772 : 0 : }
4773 : :
4774 [ # # # # : 7924513 : rdma_recv->wr.next = NULL;
# # ]
4775 [ # # ]: 7924513 : rqpair->current_recv_depth++;
4776 [ # # # # ]: 7924513 : rdma_recv->receive_tsc = poll_tsc;
4777 [ # # # # ]: 7924513 : rpoller->stat.requests++;
4778 [ + + # # : 7924513 : STAILQ_INSERT_HEAD(&rqpair->resources->incoming_queue, rdma_recv, link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # ]
4779 [ # # # # ]: 7924513 : rqpair->qpair.queue_depth++;
4780 : 7924513 : break;
4781 : 1786124 : case RDMA_WR_TYPE_DATA:
4782 : 1786124 : rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data_wr);
4783 [ # # # # : 1786124 : rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
# # ]
4784 : :
4785 [ - + # # : 1786124 : assert(rdma_req->num_outstanding_data_wr > 0);
# # # # ]
4786 : :
4787 [ # # ]: 1786124 : rqpair->current_send_depth--;
4788 [ # # ]: 1786124 : rdma_req->num_outstanding_data_wr--;
4789 [ + - # # : 1786124 : if (spdk_likely(!wc[i].status)) {
# # # # #
# ]
4790 [ - + # # : 1786124 : assert(wc[i].opcode == IBV_WC_RDMA_READ);
# # # # #
# # # ]
4791 [ # # ]: 1786124 : rqpair->current_read_depth--;
4792 : : /* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */
4793 [ + + # # : 1786124 : if (rdma_req->num_outstanding_data_wr == 0) {
# # ]
4794 [ + + # # : 784754 : if (rdma_req->num_remaining_data_wr) {
# # ]
4795 : : /* Only part of RDMA_READ operations was submitted, process the rest */
4796 : 65838 : nvmf_rdma_request_reset_transfer_in(rdma_req, rtransport);
4797 [ # # # # ]: 65838 : rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING;
4798 : 65838 : nvmf_rdma_request_process(rtransport, rdma_req);
4799 : 65838 : break;
4800 : : }
4801 [ # # # # ]: 718916 : rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
4802 : 718916 : nvmf_rdma_request_process(rtransport, rdma_req);
4803 : 0 : }
4804 : 0 : } else {
4805 : : /* If the data transfer fails still force the queue into the error state,
4806 : : * if we were performing an RDMA_READ, we need to force the request into a
4807 : : * completed state since it wasn't linked to a send. However, in the RDMA_WRITE
4808 : : * case, we should wait for the SEND to complete. */
4809 [ # # # # : 0 : if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) {
# # # # #
# ]
4810 [ # # ]: 0 : rqpair->current_read_depth--;
4811 [ # # # # : 0 : if (rdma_req->num_outstanding_data_wr == 0) {
# # ]
4812 [ # # # # : 0 : if (rdma_req->num_remaining_data_wr) {
# # ]
4813 : : /* Partially sent request is still in the pending_rdma_read_queue,
4814 : : * remove it now before completing */
4815 [ # # # # ]: 0 : rdma_req->num_remaining_data_wr = 0;
4816 [ # # # # : 0 : STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # ]
4817 : 0 : }
4818 [ # # # # ]: 0 : rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
4819 : 0 : nvmf_rdma_request_process(rtransport, rdma_req);
4820 : 0 : }
4821 : 0 : }
4822 : : }
4823 : 1720286 : break;
4824 : 0 : default:
4825 [ # # # # : 0 : SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode);
# # # # ]
4826 : 0 : continue;
4827 : : }
4828 : :
4829 : : /* Handle error conditions */
4830 [ - + # # : 17632159 : if (spdk_unlikely(wc[i].status)) {
# # # # #
# ]
4831 [ # # # # ]: 0 : rqpair->ibv_in_error_state = true;
4832 [ # # # # ]: 0 : nvmf_rdma_log_wc_status(rqpair, &wc[i]);
4833 : :
4834 : 0 : error = true;
4835 : :
4836 [ # # # # ]: 0 : if (spdk_nvmf_qpair_is_active(&rqpair->qpair)) {
4837 : : /* Disconnect the connection. */
4838 [ # # ]: 0 : spdk_nvmf_qpair_disconnect(&rqpair->qpair);
4839 : 0 : } else {
4840 : 0 : nvmf_rdma_destroy_drained_qpair(rqpair);
4841 : : }
4842 : 0 : continue;
4843 : : }
4844 : :
4845 : 17632159 : nvmf_rdma_qpair_process_pending(rtransport, rqpair, false);
4846 : :
4847 [ + + # # ]: 17632159 : if (spdk_unlikely(!spdk_nvmf_qpair_is_active(&rqpair->qpair))) {
4848 : 992 : nvmf_rdma_destroy_drained_qpair(rqpair);
4849 : 0 : }
4850 : 0 : }
4851 : :
4852 [ - + # # ]: 899882772 : if (spdk_unlikely(error == true)) {
4853 : 0 : return -1;
4854 : : }
4855 : :
4856 [ + + ]: 899882772 : if (reaped == 0) {
4857 : : /* In some cases we may not receive any CQE but we still may have pending IO requests waiting for
4858 : : * a resource (e.g. a WR from the data_wr_pool).
4859 : : * We need to start processing of such requests if no CQE reaped */
4860 : 898233522 : nvmf_rdma_poller_process_pending_buf_queue(rtransport, rpoller);
4861 : 0 : }
4862 : :
4863 : : /* submit outstanding work requests. */
4864 : 899882772 : _poller_submit_recvs(rtransport, rpoller);
4865 : 899882772 : _poller_submit_sends(rtransport, rpoller);
4866 : :
4867 : 899882772 : return count;
4868 : 0 : }
4869 : :
4870 : : static void
4871 : 0 : _nvmf_rdma_remove_destroyed_device(void *c)
4872 : : {
4873 : 0 : struct spdk_nvmf_rdma_transport *rtransport = c;
4874 : : struct spdk_nvmf_rdma_device *device, *device_tmp;
4875 : : int rc;
4876 : :
4877 [ # # # # : 0 : TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) {
# # # # #
# # # # #
# # ]
4878 [ # # # # : 0 : if (device->ready_to_destroy) {
# # # # ]
4879 : 0 : destroy_ib_device(rtransport, device);
4880 : 0 : }
4881 : 0 : }
4882 : :
4883 : 0 : free_poll_fds(rtransport);
4884 : 0 : rc = generate_poll_fds(rtransport);
4885 : : /* cannot handle fd allocation error here */
4886 [ # # ]: 0 : if (rc != 0) {
4887 : 0 : SPDK_ERRLOG("Failed to generate poll fds after remove ib device.\n");
4888 : 0 : }
4889 : 0 : }
4890 : :
4891 : : static void
4892 : 0 : _nvmf_rdma_remove_poller_in_group_cb(void *c)
4893 : : {
4894 : 0 : struct poller_manage_ctx *ctx = c;
4895 [ # # # # ]: 0 : struct spdk_nvmf_rdma_transport *rtransport = ctx->rtransport;
4896 [ # # # # ]: 0 : struct spdk_nvmf_rdma_device *device = ctx->device;
4897 [ # # # # ]: 0 : struct spdk_thread *thread = ctx->thread;
4898 : :
4899 [ # # ]: 0 : if (nvmf_rdma_all_pollers_management_done(c)) {
4900 : : /* destroy device when last poller is destroyed */
4901 [ # # # # ]: 0 : device->ready_to_destroy = true;
4902 : 0 : spdk_thread_send_msg(thread, _nvmf_rdma_remove_destroyed_device, rtransport);
4903 : 0 : }
4904 : 0 : }
4905 : :
4906 : : static void
4907 : 0 : _nvmf_rdma_remove_poller_in_group(void *c)
4908 : : {
4909 : 0 : struct poller_manage_ctx *ctx = c;
4910 : :
4911 [ # # # # : 0 : ctx->rpoller->need_destroy = true;
# # # # ]
4912 [ # # # # : 0 : ctx->rpoller->destroy_cb_ctx = ctx;
# # # # ]
4913 [ # # # # : 0 : ctx->rpoller->destroy_cb = _nvmf_rdma_remove_poller_in_group_cb;
# # # # ]
4914 : :
4915 : : /* qp will be disconnected after receiving a RDMA_CM_EVENT_DEVICE_REMOVAL event. */
4916 [ # # # # : 0 : if (RB_EMPTY(&ctx->rpoller->qpairs)) {
# # # # #
# # # ]
4917 [ # # # # ]: 0 : nvmf_rdma_poller_destroy(ctx->rpoller);
4918 : 0 : }
4919 : 0 : }
4920 : :
4921 : : static int
4922 : 449941386 : nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group)
4923 : : {
4924 : : struct spdk_nvmf_rdma_transport *rtransport;
4925 : : struct spdk_nvmf_rdma_poll_group *rgroup;
4926 : : struct spdk_nvmf_rdma_poller *rpoller, *tmp;
4927 : 449941386 : int count = 0, rc, rc2 = 0;
4928 : :
4929 [ # # # # ]: 449941386 : rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport);
4930 : 449941386 : rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
4931 : :
4932 [ + + # # : 1349824158 : TAILQ_FOREACH_SAFE(rpoller, &rgroup->pollers, link, tmp) {
# # # # #
# # # # #
# # ]
4933 : 899882772 : rc = nvmf_rdma_poller_poll(rtransport, rpoller);
4934 [ - + ]: 899882772 : if (spdk_unlikely(rc < 0)) {
4935 [ # # ]: 0 : if (rc2 == 0) {
4936 : 0 : rc2 = rc;
4937 : 0 : }
4938 : 0 : continue;
4939 : : }
4940 [ # # ]: 899882772 : count += rc;
4941 : 0 : }
4942 : :
4943 [ - + ]: 449941386 : return rc2 ? rc2 : count;
4944 : : }
4945 : :
4946 : : static int
4947 : 3273 : nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id,
4948 : : struct spdk_nvme_transport_id *trid,
4949 : : bool peer)
4950 : : {
4951 : : struct sockaddr *saddr;
4952 : : uint16_t port;
4953 : :
4954 : 3273 : spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA);
4955 : :
4956 [ + + # # ]: 3273 : if (peer) {
4957 : 74 : saddr = rdma_get_peer_addr(id);
4958 : 0 : } else {
4959 : 3199 : saddr = rdma_get_local_addr(id);
4960 : : }
4961 [ + - - # : 3273 : switch (saddr->sa_family) {
# # # ]
4962 : 3273 : case AF_INET: {
4963 : 3273 : struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr;
4964 : :
4965 [ # # # # ]: 3273 : trid->adrfam = SPDK_NVMF_ADRFAM_IPV4;
4966 [ # # ]: 3273 : inet_ntop(AF_INET, &saddr_in->sin_addr,
4967 [ # # ]: 3273 : trid->traddr, sizeof(trid->traddr));
4968 [ + + # # ]: 3273 : if (peer) {
4969 : 74 : port = ntohs(rdma_get_dst_port(id));
4970 : 0 : } else {
4971 : 3199 : port = ntohs(rdma_get_src_port(id));
4972 : : }
4973 [ # # ]: 3273 : snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port);
4974 : 3273 : break;
4975 : : }
4976 : 0 : case AF_INET6: {
4977 : 0 : struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr;
4978 [ # # # # ]: 0 : trid->adrfam = SPDK_NVMF_ADRFAM_IPV6;
4979 [ # # ]: 0 : inet_ntop(AF_INET6, &saddr_in->sin6_addr,
4980 [ # # ]: 0 : trid->traddr, sizeof(trid->traddr));
4981 [ # # # # ]: 0 : if (peer) {
4982 : 0 : port = ntohs(rdma_get_dst_port(id));
4983 : 0 : } else {
4984 : 0 : port = ntohs(rdma_get_src_port(id));
4985 : : }
4986 [ # # ]: 0 : snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port);
4987 : 0 : break;
4988 : : }
4989 : 0 : default:
4990 : 0 : return -1;
4991 : :
4992 : : }
4993 : :
4994 : 3273 : return 0;
4995 : 0 : }
4996 : :
4997 : : static int
4998 : 74 : nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair,
4999 : : struct spdk_nvme_transport_id *trid)
5000 : : {
5001 : : struct spdk_nvmf_rdma_qpair *rqpair;
5002 : :
5003 : 74 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
5004 : :
5005 [ # # # # ]: 74 : return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true);
5006 : : }
5007 : :
5008 : : static int
5009 : 0 : nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair,
5010 : : struct spdk_nvme_transport_id *trid)
5011 : : {
5012 : : struct spdk_nvmf_rdma_qpair *rqpair;
5013 : :
5014 : 0 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
5015 : :
5016 [ # # # # ]: 0 : return nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false);
5017 : : }
5018 : :
5019 : : static int
5020 : 3199 : nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair,
5021 : : struct spdk_nvme_transport_id *trid)
5022 : : {
5023 : : struct spdk_nvmf_rdma_qpair *rqpair;
5024 : :
5025 : 3199 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
5026 : :
5027 [ # # # # ]: 3199 : return nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false);
5028 : : }
5029 : :
5030 : : void
5031 : 0 : spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks)
5032 : : {
5033 : 0 : g_nvmf_hooks = *hooks;
5034 : 0 : }
5035 : :
5036 : : static void
5037 : 0 : nvmf_rdma_request_set_abort_status(struct spdk_nvmf_request *req,
5038 : : struct spdk_nvmf_rdma_request *rdma_req_to_abort,
5039 : : struct spdk_nvmf_rdma_qpair *rqpair)
5040 : : {
5041 [ # # # # : 0 : rdma_req_to_abort->req.rsp->nvme_cpl.status.sct = SPDK_NVME_SCT_GENERIC;
# # # # #
# # # #
# ]
5042 [ # # # # : 0 : rdma_req_to_abort->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_ABORTED_BY_REQUEST;
# # # # #
# # # #
# ]
5043 : :
5044 [ # # # # : 0 : STAILQ_INSERT_TAIL(&rqpair->pending_rdma_send_queue, rdma_req_to_abort, state_link);
# # # # #
# # # # #
# # # # #
# # # #
# ]
5045 [ # # # # ]: 0 : rdma_req_to_abort->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING;
5046 : :
5047 [ # # # # : 0 : req->rsp->nvme_cpl.cdw0 &= ~1U; /* Command was successfully aborted. */
# # # # #
# ]
5048 : 0 : }
5049 : :
5050 : : static int
5051 : 9528 : _nvmf_rdma_qpair_abort_request(void *ctx)
5052 : : {
5053 : 9528 : struct spdk_nvmf_request *req = ctx;
5054 [ # # # # ]: 9528 : struct spdk_nvmf_rdma_request *rdma_req_to_abort = SPDK_CONTAINEROF(
5055 : : req->req_to_abort, struct spdk_nvmf_rdma_request, req);
5056 [ # # # # : 9528 : struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(req->req_to_abort->qpair,
# # # # ]
5057 : : struct spdk_nvmf_rdma_qpair, qpair);
5058 : : int rc;
5059 : :
5060 [ # # ]: 9528 : spdk_poller_unregister(&req->poller);
5061 : :
5062 [ + - - - : 9528 : switch (rdma_req_to_abort->state) {
- - - # #
# # ]
5063 : 9528 : case RDMA_REQUEST_STATE_EXECUTING:
5064 : 9528 : rc = nvmf_ctrlr_abort_request(req);
5065 [ + - ]: 9528 : if (rc == SPDK_NVMF_REQUEST_EXEC_STATUS_ASYNCHRONOUS) {
5066 : 9528 : return SPDK_POLLER_BUSY;
5067 : : }
5068 : 0 : break;
5069 : :
5070 : 0 : case RDMA_REQUEST_STATE_NEED_BUFFER:
5071 [ # # # # : 0 : STAILQ_REMOVE(&rqpair->poller->group->group.pending_buf_queue,
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # ]
5072 : : &rdma_req_to_abort->req, spdk_nvmf_request, buf_link);
5073 : :
5074 : 0 : nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
5075 : 0 : break;
5076 : :
5077 : 0 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING:
5078 [ # # # # : 0 : STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req_to_abort,
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # ]
5079 : : spdk_nvmf_rdma_request, state_link);
5080 : :
5081 : 0 : nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
5082 : 0 : break;
5083 : :
5084 : 0 : case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING:
5085 [ # # # # : 0 : STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req_to_abort,
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # ]
5086 : : spdk_nvmf_rdma_request, state_link);
5087 : :
5088 : 0 : nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
5089 : 0 : break;
5090 : :
5091 : 0 : case RDMA_REQUEST_STATE_READY_TO_COMPLETE_PENDING:
5092 : : /* Remove req from the list here to re-use common function */
5093 [ # # # # : 0 : STAILQ_REMOVE(&rqpair->pending_rdma_send_queue, rdma_req_to_abort,
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # #
# # # # ]
5094 : : spdk_nvmf_rdma_request, state_link);
5095 : :
5096 : 0 : nvmf_rdma_request_set_abort_status(req, rdma_req_to_abort, rqpair);
5097 : 0 : break;
5098 : :
5099 : 0 : case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
5100 [ # # # # : 0 : if (spdk_get_ticks() < req->timeout_tsc) {
# # ]
5101 [ # # # # ]: 0 : req->poller = SPDK_POLLER_REGISTER(_nvmf_rdma_qpair_abort_request, req, 0);
5102 : 0 : return SPDK_POLLER_BUSY;
5103 : : }
5104 : 0 : break;
5105 : :
5106 : 0 : default:
5107 : 0 : break;
5108 : : }
5109 : :
5110 : 0 : spdk_nvmf_request_complete(req);
5111 : 0 : return SPDK_POLLER_BUSY;
5112 : 0 : }
5113 : :
5114 : : static void
5115 : 9588 : nvmf_rdma_qpair_abort_request(struct spdk_nvmf_qpair *qpair,
5116 : : struct spdk_nvmf_request *req)
5117 : : {
5118 : : struct spdk_nvmf_rdma_qpair *rqpair;
5119 : : struct spdk_nvmf_rdma_transport *rtransport;
5120 : : struct spdk_nvmf_transport *transport;
5121 : : uint16_t cid;
5122 : : uint32_t i, max_req_count;
5123 : 9588 : struct spdk_nvmf_rdma_request *rdma_req_to_abort = NULL, *rdma_req;
5124 : :
5125 : 9588 : rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
5126 [ # # # # ]: 9588 : rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
5127 [ # # ]: 9588 : transport = &rtransport->transport;
5128 : :
5129 [ # # # # : 9588 : cid = req->cmd->nvme_cmd.cdw10_bits.abort.cid;
# # # # #
# # # #
# ]
5130 [ - + # # : 9588 : max_req_count = rqpair->srq == NULL ? rqpair->max_queue_depth : rqpair->poller->max_srq_depth;
# # # # #
# # # # #
# # # # ]
5131 : :
5132 [ + + ]: 587912 : for (i = 0; i < max_req_count; i++) {
5133 [ # # # # : 587852 : rdma_req = &rqpair->resources->reqs[i];
# # # # #
# ]
5134 : : /* When SRQ == NULL, rqpair has its own requests and req.qpair pointer always points to the qpair
5135 : : * When SRQ != NULL all rqpairs share common requests and qpair pointer is assigned when we start to
5136 : : * process a request. So in both cases all requests which are not in FREE state have valid qpair ptr */
5137 [ + + + + : 587852 : if (rdma_req->state != RDMA_REQUEST_STATE_FREE && rdma_req->req.cmd->nvme_cmd.cid == cid &&
# # # # #
# # # # #
# # # # #
# # # ]
5138 [ + - # # : 9528 : rdma_req->req.qpair == qpair) {
# # ]
5139 : 9528 : rdma_req_to_abort = rdma_req;
5140 : 9528 : break;
5141 : : }
5142 : 0 : }
5143 : :
5144 [ + + ]: 9588 : if (rdma_req_to_abort == NULL) {
5145 : 60 : spdk_nvmf_request_complete(req);
5146 : 60 : return;
5147 : : }
5148 : :
5149 [ # # # # : 9528 : req->req_to_abort = &rdma_req_to_abort->req;
# # ]
5150 [ # # # # ]: 19056 : req->timeout_tsc = spdk_get_ticks() +
5151 [ # # # # : 9528 : transport->opts.abort_timeout_sec * spdk_get_ticks_hz();
# # ]
5152 [ # # # # ]: 9528 : req->poller = NULL;
5153 : :
5154 : 9528 : _nvmf_rdma_qpair_abort_request(req);
5155 : 0 : }
5156 : :
5157 : : static void
5158 : 8 : nvmf_rdma_poll_group_dump_stat(struct spdk_nvmf_transport_poll_group *group,
5159 : : struct spdk_json_write_ctx *w)
5160 : : {
5161 : : struct spdk_nvmf_rdma_poll_group *rgroup;
5162 : : struct spdk_nvmf_rdma_poller *rpoller;
5163 : :
5164 [ - + # # ]: 8 : assert(w != NULL);
5165 : :
5166 : 8 : rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
5167 : :
5168 [ # # # # : 8 : spdk_json_write_named_uint64(w, "pending_data_buffer", rgroup->stat.pending_data_buffer);
# # ]
5169 : :
5170 : 8 : spdk_json_write_named_array_begin(w, "devices");
5171 : :
5172 [ + + # # : 24 : TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
# # # # #
# # # #
# ]
5173 : 16 : spdk_json_write_object_begin(w);
5174 : 16 : spdk_json_write_named_string(w, "name",
5175 [ # # # # : 16 : ibv_get_device_name(rpoller->device->context->device));
# # # # #
# # # ]
5176 : 16 : spdk_json_write_named_uint64(w, "polls",
5177 [ # # # # : 0 : rpoller->stat.polls);
# # ]
5178 : 16 : spdk_json_write_named_uint64(w, "idle_polls",
5179 [ # # # # : 0 : rpoller->stat.idle_polls);
# # ]
5180 : 16 : spdk_json_write_named_uint64(w, "completions",
5181 [ # # # # : 0 : rpoller->stat.completions);
# # ]
5182 : 16 : spdk_json_write_named_uint64(w, "requests",
5183 [ # # # # : 0 : rpoller->stat.requests);
# # ]
5184 : 16 : spdk_json_write_named_uint64(w, "request_latency",
5185 [ # # # # : 0 : rpoller->stat.request_latency);
# # ]
5186 : 16 : spdk_json_write_named_uint64(w, "pending_free_request",
5187 [ # # # # : 0 : rpoller->stat.pending_free_request);
# # ]
5188 : 16 : spdk_json_write_named_uint64(w, "pending_rdma_read",
5189 [ # # # # : 0 : rpoller->stat.pending_rdma_read);
# # ]
5190 : 16 : spdk_json_write_named_uint64(w, "pending_rdma_write",
5191 [ # # # # : 0 : rpoller->stat.pending_rdma_write);
# # ]
5192 : 16 : spdk_json_write_named_uint64(w, "pending_rdma_send",
5193 [ # # # # : 0 : rpoller->stat.pending_rdma_send);
# # ]
5194 : 16 : spdk_json_write_named_uint64(w, "total_send_wrs",
5195 [ # # # # : 0 : rpoller->stat.qp_stats.send.num_submitted_wrs);
# # # # #
# ]
5196 : 16 : spdk_json_write_named_uint64(w, "send_doorbell_updates",
5197 [ # # # # : 0 : rpoller->stat.qp_stats.send.doorbell_updates);
# # # # #
# ]
5198 : 16 : spdk_json_write_named_uint64(w, "total_recv_wrs",
5199 [ # # # # : 0 : rpoller->stat.qp_stats.recv.num_submitted_wrs);
# # # # #
# ]
5200 : 16 : spdk_json_write_named_uint64(w, "recv_doorbell_updates",
5201 [ # # # # : 0 : rpoller->stat.qp_stats.recv.doorbell_updates);
# # # # #
# ]
5202 : 16 : spdk_json_write_object_end(w);
5203 : 0 : }
5204 : :
5205 : 8 : spdk_json_write_array_end(w);
5206 : 8 : }
5207 : :
5208 : : const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = {
5209 : : .name = "RDMA",
5210 : : .type = SPDK_NVME_TRANSPORT_RDMA,
5211 : : .opts_init = nvmf_rdma_opts_init,
5212 : : .create = nvmf_rdma_create,
5213 : : .dump_opts = nvmf_rdma_dump_opts,
5214 : : .destroy = nvmf_rdma_destroy,
5215 : :
5216 : : .listen = nvmf_rdma_listen,
5217 : : .stop_listen = nvmf_rdma_stop_listen,
5218 : : .cdata_init = nvmf_rdma_cdata_init,
5219 : :
5220 : : .listener_discover = nvmf_rdma_discover,
5221 : :
5222 : : .poll_group_create = nvmf_rdma_poll_group_create,
5223 : : .get_optimal_poll_group = nvmf_rdma_get_optimal_poll_group,
5224 : : .poll_group_destroy = nvmf_rdma_poll_group_destroy,
5225 : : .poll_group_add = nvmf_rdma_poll_group_add,
5226 : : .poll_group_remove = nvmf_rdma_poll_group_remove,
5227 : : .poll_group_poll = nvmf_rdma_poll_group_poll,
5228 : :
5229 : : .req_free = nvmf_rdma_request_free,
5230 : : .req_complete = nvmf_rdma_request_complete,
5231 : :
5232 : : .qpair_fini = nvmf_rdma_close_qpair,
5233 : : .qpair_get_peer_trid = nvmf_rdma_qpair_get_peer_trid,
5234 : : .qpair_get_local_trid = nvmf_rdma_qpair_get_local_trid,
5235 : : .qpair_get_listen_trid = nvmf_rdma_qpair_get_listen_trid,
5236 : : .qpair_abort_request = nvmf_rdma_qpair_abort_request,
5237 : :
5238 : : .poll_group_dump_stat = nvmf_rdma_poll_group_dump_stat,
5239 : : };
5240 : :
5241 : 987 : SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma);
5242 : 987 : SPDK_LOG_REGISTER_COMPONENT(rdma)
|