xref: /spdk/lib/nvmf/rdma.c (revision 94a84ae98590bea46939eb1dcd7a9876bd393b54)
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright (c) Intel Corporation. All rights reserved.
5  *   Copyright (c) 2019, 2020 Mellanox Technologies LTD. All rights reserved.
6  *
7  *   Redistribution and use in source and binary forms, with or without
8  *   modification, are permitted provided that the following conditions
9  *   are met:
10  *
11  *     * Redistributions of source code must retain the above copyright
12  *       notice, this list of conditions and the following disclaimer.
13  *     * Redistributions in binary form must reproduce the above copyright
14  *       notice, this list of conditions and the following disclaimer in
15  *       the documentation and/or other materials provided with the
16  *       distribution.
17  *     * Neither the name of Intel Corporation nor the names of its
18  *       contributors may be used to endorse or promote products derived
19  *       from this software without specific prior written permission.
20  *
21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
32  */
33 
34 #include "spdk/stdinc.h"
35 
36 #include <infiniband/verbs.h>
37 #include <rdma/rdma_cma.h>
38 #include <rdma/rdma_verbs.h>
39 
40 #include "spdk/config.h"
41 #include "spdk/thread.h"
42 #include "spdk/likely.h"
43 #include "spdk/nvmf_transport.h"
44 #include "spdk/string.h"
45 #include "spdk/trace.h"
46 #include "spdk/util.h"
47 
48 #include "spdk_internal/assert.h"
49 #include "spdk_internal/log.h"
50 
51 struct spdk_nvme_rdma_hooks g_nvmf_hooks = {};
52 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma;
53 
54 /*
55  RDMA Connection Resource Defaults
56  */
57 #define NVMF_DEFAULT_TX_SGE		SPDK_NVMF_MAX_SGL_ENTRIES
58 #define NVMF_DEFAULT_RSP_SGE		1
59 #define NVMF_DEFAULT_RX_SGE		2
60 
61 /* The RDMA completion queue size */
62 #define DEFAULT_NVMF_RDMA_CQ_SIZE	4096
63 #define MAX_WR_PER_QP(queue_depth)	(queue_depth * 3 + 2)
64 
65 /* Timeout for destroying defunct rqpairs */
66 #define NVMF_RDMA_QPAIR_DESTROY_TIMEOUT_US	4000000
67 
68 static int g_spdk_nvmf_ibv_query_mask =
69 	IBV_QP_STATE |
70 	IBV_QP_PKEY_INDEX |
71 	IBV_QP_PORT |
72 	IBV_QP_ACCESS_FLAGS |
73 	IBV_QP_AV |
74 	IBV_QP_PATH_MTU |
75 	IBV_QP_DEST_QPN |
76 	IBV_QP_RQ_PSN |
77 	IBV_QP_MAX_DEST_RD_ATOMIC |
78 	IBV_QP_MIN_RNR_TIMER |
79 	IBV_QP_SQ_PSN |
80 	IBV_QP_TIMEOUT |
81 	IBV_QP_RETRY_CNT |
82 	IBV_QP_RNR_RETRY |
83 	IBV_QP_MAX_QP_RD_ATOMIC;
84 
85 enum spdk_nvmf_rdma_request_state {
86 	/* The request is not currently in use */
87 	RDMA_REQUEST_STATE_FREE = 0,
88 
89 	/* Initial state when request first received */
90 	RDMA_REQUEST_STATE_NEW,
91 
92 	/* The request is queued until a data buffer is available. */
93 	RDMA_REQUEST_STATE_NEED_BUFFER,
94 
95 	/* The request is waiting on RDMA queue depth availability
96 	 * to transfer data from the host to the controller.
97 	 */
98 	RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING,
99 
100 	/* The request is currently transferring data from the host to the controller. */
101 	RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER,
102 
103 	/* The request is ready to execute at the block device */
104 	RDMA_REQUEST_STATE_READY_TO_EXECUTE,
105 
106 	/* The request is currently executing at the block device */
107 	RDMA_REQUEST_STATE_EXECUTING,
108 
109 	/* The request finished executing at the block device */
110 	RDMA_REQUEST_STATE_EXECUTED,
111 
112 	/* The request is waiting on RDMA queue depth availability
113 	 * to transfer data from the controller to the host.
114 	 */
115 	RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING,
116 
117 	/* The request is ready to send a completion */
118 	RDMA_REQUEST_STATE_READY_TO_COMPLETE,
119 
120 	/* The request is currently transferring data from the controller to the host. */
121 	RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST,
122 
123 	/* The request currently has an outstanding completion without an
124 	 * associated data transfer.
125 	 */
126 	RDMA_REQUEST_STATE_COMPLETING,
127 
128 	/* The request completed and can be marked free. */
129 	RDMA_REQUEST_STATE_COMPLETED,
130 
131 	/* Terminator */
132 	RDMA_REQUEST_NUM_STATES,
133 };
134 
135 #define OBJECT_NVMF_RDMA_IO				0x40
136 
137 #define TRACE_GROUP_NVMF_RDMA				0x4
138 #define TRACE_RDMA_REQUEST_STATE_NEW					SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x0)
139 #define TRACE_RDMA_REQUEST_STATE_NEED_BUFFER				SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x1)
140 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING	SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x2)
141 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER	SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x3)
142 #define TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE			SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x4)
143 #define TRACE_RDMA_REQUEST_STATE_EXECUTING				SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x5)
144 #define TRACE_RDMA_REQUEST_STATE_EXECUTED				SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x6)
145 #define TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING		SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x7)
146 #define TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE			SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x8)
147 #define TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST	SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x9)
148 #define TRACE_RDMA_REQUEST_STATE_COMPLETING				SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xA)
149 #define TRACE_RDMA_REQUEST_STATE_COMPLETED				SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xB)
150 #define TRACE_RDMA_QP_CREATE						SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xC)
151 #define TRACE_RDMA_IBV_ASYNC_EVENT					SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xD)
152 #define TRACE_RDMA_CM_ASYNC_EVENT					SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xE)
153 #define TRACE_RDMA_QP_STATE_CHANGE					SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0xF)
154 #define TRACE_RDMA_QP_DISCONNECT					SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x10)
155 #define TRACE_RDMA_QP_DESTROY						SPDK_TPOINT_ID(TRACE_GROUP_NVMF_RDMA, 0x11)
156 
157 SPDK_TRACE_REGISTER_FN(nvmf_trace, "nvmf_rdma", TRACE_GROUP_NVMF_RDMA)
158 {
159 	spdk_trace_register_object(OBJECT_NVMF_RDMA_IO, 'r');
160 	spdk_trace_register_description("RDMA_REQ_NEW", TRACE_RDMA_REQUEST_STATE_NEW,
161 					OWNER_NONE, OBJECT_NVMF_RDMA_IO, 1, 1, "cmid:   ");
162 	spdk_trace_register_description("RDMA_REQ_NEED_BUFFER", TRACE_RDMA_REQUEST_STATE_NEED_BUFFER,
163 					OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid:   ");
164 	spdk_trace_register_description("RDMA_REQ_TX_PENDING_C2H",
165 					TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING,
166 					OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid:   ");
167 	spdk_trace_register_description("RDMA_REQ_TX_PENDING_H2C",
168 					TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING,
169 					OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid:   ");
170 	spdk_trace_register_description("RDMA_REQ_TX_H2C",
171 					TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER,
172 					OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid:   ");
173 	spdk_trace_register_description("RDMA_REQ_RDY_TO_EXECUTE",
174 					TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE,
175 					OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid:   ");
176 	spdk_trace_register_description("RDMA_REQ_EXECUTING",
177 					TRACE_RDMA_REQUEST_STATE_EXECUTING,
178 					OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid:   ");
179 	spdk_trace_register_description("RDMA_REQ_EXECUTED",
180 					TRACE_RDMA_REQUEST_STATE_EXECUTED,
181 					OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid:   ");
182 	spdk_trace_register_description("RDMA_REQ_RDY_TO_COMPL",
183 					TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE,
184 					OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid:   ");
185 	spdk_trace_register_description("RDMA_REQ_COMPLETING_C2H",
186 					TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST,
187 					OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid:   ");
188 	spdk_trace_register_description("RDMA_REQ_COMPLETING",
189 					TRACE_RDMA_REQUEST_STATE_COMPLETING,
190 					OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid:   ");
191 	spdk_trace_register_description("RDMA_REQ_COMPLETED",
192 					TRACE_RDMA_REQUEST_STATE_COMPLETED,
193 					OWNER_NONE, OBJECT_NVMF_RDMA_IO, 0, 1, "cmid:   ");
194 
195 	spdk_trace_register_description("RDMA_QP_CREATE", TRACE_RDMA_QP_CREATE,
196 					OWNER_NONE, OBJECT_NONE, 0, 0, "");
197 	spdk_trace_register_description("RDMA_IBV_ASYNC_EVENT", TRACE_RDMA_IBV_ASYNC_EVENT,
198 					OWNER_NONE, OBJECT_NONE, 0, 0, "type:   ");
199 	spdk_trace_register_description("RDMA_CM_ASYNC_EVENT", TRACE_RDMA_CM_ASYNC_EVENT,
200 					OWNER_NONE, OBJECT_NONE, 0, 0, "type:   ");
201 	spdk_trace_register_description("RDMA_QP_STATE_CHANGE", TRACE_RDMA_QP_STATE_CHANGE,
202 					OWNER_NONE, OBJECT_NONE, 0, 1, "state:  ");
203 	spdk_trace_register_description("RDMA_QP_DISCONNECT", TRACE_RDMA_QP_DISCONNECT,
204 					OWNER_NONE, OBJECT_NONE, 0, 0, "");
205 	spdk_trace_register_description("RDMA_QP_DESTROY", TRACE_RDMA_QP_DESTROY,
206 					OWNER_NONE, OBJECT_NONE, 0, 0, "");
207 }
208 
209 enum spdk_nvmf_rdma_wr_type {
210 	RDMA_WR_TYPE_RECV,
211 	RDMA_WR_TYPE_SEND,
212 	RDMA_WR_TYPE_DATA,
213 };
214 
215 struct spdk_nvmf_rdma_wr {
216 	enum spdk_nvmf_rdma_wr_type	type;
217 };
218 
219 /* This structure holds commands as they are received off the wire.
220  * It must be dynamically paired with a full request object
221  * (spdk_nvmf_rdma_request) to service a request. It is separate
222  * from the request because RDMA does not appear to order
223  * completions, so occasionally we'll get a new incoming
224  * command when there aren't any free request objects.
225  */
226 struct spdk_nvmf_rdma_recv {
227 	struct ibv_recv_wr			wr;
228 	struct ibv_sge				sgl[NVMF_DEFAULT_RX_SGE];
229 
230 	struct spdk_nvmf_rdma_qpair		*qpair;
231 
232 	/* In-capsule data buffer */
233 	uint8_t					*buf;
234 
235 	struct spdk_nvmf_rdma_wr		rdma_wr;
236 	uint64_t				receive_tsc;
237 
238 	STAILQ_ENTRY(spdk_nvmf_rdma_recv)	link;
239 };
240 
241 struct spdk_nvmf_rdma_request_data {
242 	struct spdk_nvmf_rdma_wr	rdma_wr;
243 	struct ibv_send_wr		wr;
244 	struct ibv_sge			sgl[SPDK_NVMF_MAX_SGL_ENTRIES];
245 };
246 
247 struct spdk_nvmf_rdma_request {
248 	struct spdk_nvmf_request		req;
249 
250 	enum spdk_nvmf_rdma_request_state	state;
251 
252 	struct spdk_nvmf_rdma_recv		*recv;
253 
254 	struct {
255 		struct spdk_nvmf_rdma_wr	rdma_wr;
256 		struct	ibv_send_wr		wr;
257 		struct	ibv_sge			sgl[NVMF_DEFAULT_RSP_SGE];
258 	} rsp;
259 
260 	struct spdk_nvmf_rdma_request_data	data;
261 
262 	uint32_t				iovpos;
263 
264 	uint32_t				num_outstanding_data_wr;
265 	uint64_t				receive_tsc;
266 
267 	STAILQ_ENTRY(spdk_nvmf_rdma_request)	state_link;
268 };
269 
270 enum spdk_nvmf_rdma_qpair_disconnect_flags {
271 	RDMA_QP_DISCONNECTING		= 1,
272 	RDMA_QP_RECV_DRAINED		= 1 << 1,
273 	RDMA_QP_SEND_DRAINED		= 1 << 2
274 };
275 
276 struct spdk_nvmf_rdma_resource_opts {
277 	struct spdk_nvmf_rdma_qpair	*qpair;
278 	/* qp points either to an ibv_qp object or an ibv_srq object depending on the value of shared. */
279 	void				*qp;
280 	struct ibv_pd			*pd;
281 	uint32_t			max_queue_depth;
282 	uint32_t			in_capsule_data_size;
283 	bool				shared;
284 };
285 
286 struct spdk_nvmf_send_wr_list {
287 	struct ibv_send_wr	*first;
288 	struct ibv_send_wr	*last;
289 };
290 
291 struct spdk_nvmf_recv_wr_list {
292 	struct ibv_recv_wr	*first;
293 	struct ibv_recv_wr	*last;
294 };
295 
296 struct spdk_nvmf_rdma_resources {
297 	/* Array of size "max_queue_depth" containing RDMA requests. */
298 	struct spdk_nvmf_rdma_request		*reqs;
299 
300 	/* Array of size "max_queue_depth" containing RDMA recvs. */
301 	struct spdk_nvmf_rdma_recv		*recvs;
302 
303 	/* Array of size "max_queue_depth" containing 64 byte capsules
304 	 * used for receive.
305 	 */
306 	union nvmf_h2c_msg			*cmds;
307 	struct ibv_mr				*cmds_mr;
308 
309 	/* Array of size "max_queue_depth" containing 16 byte completions
310 	 * to be sent back to the user.
311 	 */
312 	union nvmf_c2h_msg			*cpls;
313 	struct ibv_mr				*cpls_mr;
314 
315 	/* Array of size "max_queue_depth * InCapsuleDataSize" containing
316 	 * buffers to be used for in capsule data.
317 	 */
318 	void					*bufs;
319 	struct ibv_mr				*bufs_mr;
320 
321 	/* The list of pending recvs to transfer */
322 	struct spdk_nvmf_recv_wr_list		recvs_to_post;
323 
324 	/* Receives that are waiting for a request object */
325 	STAILQ_HEAD(, spdk_nvmf_rdma_recv)	incoming_queue;
326 
327 	/* Queue to track free requests */
328 	STAILQ_HEAD(, spdk_nvmf_rdma_request)	free_queue;
329 };
330 
331 typedef void (*spdk_nvmf_rdma_qpair_ibv_event)(struct spdk_nvmf_rdma_qpair *rqpair);
332 
333 struct spdk_nvmf_rdma_ibv_event_ctx {
334 	struct spdk_nvmf_rdma_qpair			*rqpair;
335 	spdk_nvmf_rdma_qpair_ibv_event			cb_fn;
336 	/* Link to other ibv events associated with this qpair */
337 	STAILQ_ENTRY(spdk_nvmf_rdma_ibv_event_ctx)	link;
338 };
339 
340 struct spdk_nvmf_rdma_qpair {
341 	struct spdk_nvmf_qpair			qpair;
342 
343 	struct spdk_nvmf_rdma_device		*device;
344 	struct spdk_nvmf_rdma_poller		*poller;
345 
346 	struct rdma_cm_id			*cm_id;
347 	struct ibv_srq				*srq;
348 	struct rdma_cm_id			*listen_id;
349 
350 	/* The maximum number of I/O outstanding on this connection at one time */
351 	uint16_t				max_queue_depth;
352 
353 	/* The maximum number of active RDMA READ and ATOMIC operations at one time */
354 	uint16_t				max_read_depth;
355 
356 	/* The maximum number of RDMA SEND operations at one time */
357 	uint32_t				max_send_depth;
358 
359 	/* The current number of outstanding WRs from this qpair's
360 	 * recv queue. Should not exceed device->attr.max_queue_depth.
361 	 */
362 	uint16_t				current_recv_depth;
363 
364 	/* The current number of active RDMA READ operations */
365 	uint16_t				current_read_depth;
366 
367 	/* The current number of posted WRs from this qpair's
368 	 * send queue. Should not exceed max_send_depth.
369 	 */
370 	uint32_t				current_send_depth;
371 
372 	/* The maximum number of SGEs per WR on the send queue */
373 	uint32_t				max_send_sge;
374 
375 	/* The maximum number of SGEs per WR on the recv queue */
376 	uint32_t				max_recv_sge;
377 
378 	/* The list of pending send requests for a transfer */
379 	struct spdk_nvmf_send_wr_list		sends_to_post;
380 
381 	struct spdk_nvmf_rdma_resources		*resources;
382 
383 	STAILQ_HEAD(, spdk_nvmf_rdma_request)	pending_rdma_read_queue;
384 
385 	STAILQ_HEAD(, spdk_nvmf_rdma_request)	pending_rdma_write_queue;
386 
387 	/* Number of requests not in the free state */
388 	uint32_t				qd;
389 
390 	TAILQ_ENTRY(spdk_nvmf_rdma_qpair)	link;
391 
392 	STAILQ_ENTRY(spdk_nvmf_rdma_qpair)	recv_link;
393 
394 	STAILQ_ENTRY(spdk_nvmf_rdma_qpair)	send_link;
395 
396 	/* IBV queue pair attributes: they are used to manage
397 	 * qp state and recover from errors.
398 	 */
399 	enum ibv_qp_state			ibv_state;
400 
401 	uint32_t				disconnect_flags;
402 
403 	/* Poller registered in case the qpair doesn't properly
404 	 * complete the qpair destruct process and becomes defunct.
405 	 */
406 
407 	struct spdk_poller			*destruct_poller;
408 
409 	/* List of ibv async events */
410 	STAILQ_HEAD(, spdk_nvmf_rdma_ibv_event_ctx)	ibv_events;
411 
412 	/* There are several ways a disconnect can start on a qpair
413 	 * and they are not all mutually exclusive. It is important
414 	 * that we only initialize one of these paths.
415 	 */
416 	bool					disconnect_started;
417 	/* Lets us know that we have received the last_wqe event. */
418 	bool					last_wqe_reached;
419 };
420 
421 struct spdk_nvmf_rdma_poller_stat {
422 	uint64_t				completions;
423 	uint64_t				polls;
424 	uint64_t				requests;
425 	uint64_t				request_latency;
426 	uint64_t				pending_free_request;
427 	uint64_t				pending_rdma_read;
428 	uint64_t				pending_rdma_write;
429 };
430 
431 struct spdk_nvmf_rdma_poller {
432 	struct spdk_nvmf_rdma_device		*device;
433 	struct spdk_nvmf_rdma_poll_group	*group;
434 
435 	int					num_cqe;
436 	int					required_num_wr;
437 	struct ibv_cq				*cq;
438 
439 	/* The maximum number of I/O outstanding on the shared receive queue at one time */
440 	uint16_t				max_srq_depth;
441 
442 	/* Shared receive queue */
443 	struct ibv_srq				*srq;
444 
445 	struct spdk_nvmf_rdma_resources		*resources;
446 	struct spdk_nvmf_rdma_poller_stat	stat;
447 
448 	TAILQ_HEAD(, spdk_nvmf_rdma_qpair)	qpairs;
449 
450 	STAILQ_HEAD(, spdk_nvmf_rdma_qpair)	qpairs_pending_recv;
451 
452 	STAILQ_HEAD(, spdk_nvmf_rdma_qpair)	qpairs_pending_send;
453 
454 	TAILQ_ENTRY(spdk_nvmf_rdma_poller)	link;
455 };
456 
457 struct spdk_nvmf_rdma_poll_group_stat {
458 	uint64_t				pending_data_buffer;
459 };
460 
461 struct spdk_nvmf_rdma_poll_group {
462 	struct spdk_nvmf_transport_poll_group		group;
463 	struct spdk_nvmf_rdma_poll_group_stat		stat;
464 	TAILQ_HEAD(, spdk_nvmf_rdma_poller)		pollers;
465 	TAILQ_ENTRY(spdk_nvmf_rdma_poll_group)		link;
466 	/*
467 	 * buffers which are split across multiple RDMA
468 	 * memory regions cannot be used by this transport.
469 	 */
470 	STAILQ_HEAD(, spdk_nvmf_transport_pg_cache_buf)	retired_bufs;
471 };
472 
473 struct spdk_nvmf_rdma_conn_sched {
474 	struct spdk_nvmf_rdma_poll_group *next_admin_pg;
475 	struct spdk_nvmf_rdma_poll_group *next_io_pg;
476 };
477 
478 /* Assuming rdma_cm uses just one protection domain per ibv_context. */
479 struct spdk_nvmf_rdma_device {
480 	struct ibv_device_attr			attr;
481 	struct ibv_context			*context;
482 
483 	struct spdk_mem_map			*map;
484 	struct ibv_pd				*pd;
485 
486 	int					num_srq;
487 
488 	TAILQ_ENTRY(spdk_nvmf_rdma_device)	link;
489 };
490 
491 struct spdk_nvmf_rdma_port {
492 	const struct spdk_nvme_transport_id	*trid;
493 	struct rdma_cm_id			*id;
494 	struct spdk_nvmf_rdma_device		*device;
495 	TAILQ_ENTRY(spdk_nvmf_rdma_port)	link;
496 };
497 
498 struct spdk_nvmf_rdma_transport {
499 	struct spdk_nvmf_transport	transport;
500 
501 	struct spdk_nvmf_rdma_conn_sched conn_sched;
502 
503 	struct rdma_event_channel	*event_channel;
504 
505 	struct spdk_mempool		*data_wr_pool;
506 
507 	pthread_mutex_t			lock;
508 
509 	/* fields used to poll RDMA/IB events */
510 	nfds_t			npoll_fds;
511 	struct pollfd		*poll_fds;
512 
513 	TAILQ_HEAD(, spdk_nvmf_rdma_device)	devices;
514 	TAILQ_HEAD(, spdk_nvmf_rdma_port)	ports;
515 	TAILQ_HEAD(, spdk_nvmf_rdma_poll_group)	poll_groups;
516 };
517 
518 static inline void
519 spdk_nvmf_rdma_start_disconnect(struct spdk_nvmf_rdma_qpair *rqpair);
520 
521 static bool
522 spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport,
523 			       struct spdk_nvmf_rdma_request *rdma_req);
524 
525 static inline int
526 spdk_nvmf_rdma_check_ibv_state(enum ibv_qp_state state)
527 {
528 	switch (state) {
529 	case IBV_QPS_RESET:
530 	case IBV_QPS_INIT:
531 	case IBV_QPS_RTR:
532 	case IBV_QPS_RTS:
533 	case IBV_QPS_SQD:
534 	case IBV_QPS_SQE:
535 	case IBV_QPS_ERR:
536 		return 0;
537 	default:
538 		return -1;
539 	}
540 }
541 
542 static inline enum spdk_nvme_media_error_status_code
543 spdk_nvmf_rdma_dif_error_to_compl_status(uint8_t err_type) {
544 	enum spdk_nvme_media_error_status_code result;
545 	switch (err_type)
546 	{
547 	case SPDK_DIF_REFTAG_ERROR:
548 		result = SPDK_NVME_SC_REFERENCE_TAG_CHECK_ERROR;
549 		break;
550 	case SPDK_DIF_APPTAG_ERROR:
551 		result = SPDK_NVME_SC_APPLICATION_TAG_CHECK_ERROR;
552 		break;
553 	case SPDK_DIF_GUARD_ERROR:
554 		result = SPDK_NVME_SC_GUARD_CHECK_ERROR;
555 		break;
556 	default:
557 		SPDK_UNREACHABLE();
558 	}
559 
560 	return result;
561 }
562 
563 static enum ibv_qp_state
564 spdk_nvmf_rdma_update_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair) {
565 	enum ibv_qp_state old_state, new_state;
566 	struct ibv_qp_attr qp_attr;
567 	struct ibv_qp_init_attr init_attr;
568 	int rc;
569 
570 	old_state = rqpair->ibv_state;
571 	rc = ibv_query_qp(rqpair->cm_id->qp, &qp_attr,
572 			  g_spdk_nvmf_ibv_query_mask, &init_attr);
573 
574 	if (rc)
575 	{
576 		SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n");
577 		return IBV_QPS_ERR + 1;
578 	}
579 
580 	new_state = qp_attr.qp_state;
581 	rqpair->ibv_state = new_state;
582 	qp_attr.ah_attr.port_num = qp_attr.port_num;
583 
584 	rc = spdk_nvmf_rdma_check_ibv_state(new_state);
585 	if (rc)
586 	{
587 		SPDK_ERRLOG("QP#%d: bad state updated: %u, maybe hardware issue\n", rqpair->qpair.qid, new_state);
588 		/*
589 		 * IBV_QPS_UNKNOWN undefined if lib version smaller than libibverbs-1.1.8
590 		 * IBV_QPS_UNKNOWN is the enum element after IBV_QPS_ERR
591 		 */
592 		return IBV_QPS_ERR + 1;
593 	}
594 
595 	if (old_state != new_state)
596 	{
597 		spdk_trace_record(TRACE_RDMA_QP_STATE_CHANGE, 0, 0,
598 				  (uintptr_t)rqpair->cm_id, new_state);
599 	}
600 	return new_state;
601 }
602 
603 static const char *str_ibv_qp_state[] = {
604 	"IBV_QPS_RESET",
605 	"IBV_QPS_INIT",
606 	"IBV_QPS_RTR",
607 	"IBV_QPS_RTS",
608 	"IBV_QPS_SQD",
609 	"IBV_QPS_SQE",
610 	"IBV_QPS_ERR",
611 	"IBV_QPS_UNKNOWN"
612 };
613 
614 static int
615 spdk_nvmf_rdma_set_ibv_state(struct spdk_nvmf_rdma_qpair *rqpair,
616 			     enum ibv_qp_state new_state)
617 {
618 	struct ibv_qp_attr qp_attr;
619 	struct ibv_qp_init_attr init_attr;
620 	int rc;
621 	enum ibv_qp_state state;
622 	static int attr_mask_rc[] = {
623 		[IBV_QPS_RESET] = IBV_QP_STATE,
624 		[IBV_QPS_INIT] = (IBV_QP_STATE |
625 				  IBV_QP_PKEY_INDEX |
626 				  IBV_QP_PORT |
627 				  IBV_QP_ACCESS_FLAGS),
628 		[IBV_QPS_RTR] = (IBV_QP_STATE |
629 				 IBV_QP_AV |
630 				 IBV_QP_PATH_MTU |
631 				 IBV_QP_DEST_QPN |
632 				 IBV_QP_RQ_PSN |
633 				 IBV_QP_MAX_DEST_RD_ATOMIC |
634 				 IBV_QP_MIN_RNR_TIMER),
635 		[IBV_QPS_RTS] = (IBV_QP_STATE |
636 				 IBV_QP_SQ_PSN |
637 				 IBV_QP_TIMEOUT |
638 				 IBV_QP_RETRY_CNT |
639 				 IBV_QP_RNR_RETRY |
640 				 IBV_QP_MAX_QP_RD_ATOMIC),
641 		[IBV_QPS_SQD] = IBV_QP_STATE,
642 		[IBV_QPS_SQE] = IBV_QP_STATE,
643 		[IBV_QPS_ERR] = IBV_QP_STATE,
644 	};
645 
646 	rc = spdk_nvmf_rdma_check_ibv_state(new_state);
647 	if (rc) {
648 		SPDK_ERRLOG("QP#%d: bad state requested: %u\n",
649 			    rqpair->qpair.qid, new_state);
650 		return rc;
651 	}
652 
653 	rc = ibv_query_qp(rqpair->cm_id->qp, &qp_attr,
654 			  g_spdk_nvmf_ibv_query_mask, &init_attr);
655 
656 	if (rc) {
657 		SPDK_ERRLOG("Failed to get updated RDMA queue pair state!\n");
658 	}
659 
660 	qp_attr.cur_qp_state = rqpair->ibv_state;
661 	qp_attr.qp_state = new_state;
662 
663 	rc = ibv_modify_qp(rqpair->cm_id->qp, &qp_attr,
664 			   attr_mask_rc[new_state]);
665 
666 	if (rc) {
667 		SPDK_ERRLOG("QP#%d: failed to set state to: %s, %d (%s)\n",
668 			    rqpair->qpair.qid, str_ibv_qp_state[new_state], errno, strerror(errno));
669 		return rc;
670 	}
671 
672 	state = spdk_nvmf_rdma_update_ibv_state(rqpair);
673 
674 	if (state != new_state) {
675 		SPDK_ERRLOG("QP#%d: expected state: %s, actual state: %s\n",
676 			    rqpair->qpair.qid, str_ibv_qp_state[new_state],
677 			    str_ibv_qp_state[state]);
678 		return -1;
679 	}
680 	SPDK_DEBUGLOG(SPDK_LOG_RDMA, "IBV QP#%u changed to: %s\n", rqpair->qpair.qid,
681 		      str_ibv_qp_state[state]);
682 	return 0;
683 }
684 
685 static void
686 nvmf_rdma_request_free_data(struct spdk_nvmf_rdma_request *rdma_req,
687 			    struct spdk_nvmf_rdma_transport *rtransport)
688 {
689 	struct spdk_nvmf_rdma_request_data	*data_wr;
690 	struct ibv_send_wr			*next_send_wr;
691 	uint64_t				req_wrid;
692 
693 	rdma_req->num_outstanding_data_wr = 0;
694 	data_wr = &rdma_req->data;
695 	req_wrid = data_wr->wr.wr_id;
696 	while (data_wr && data_wr->wr.wr_id == req_wrid) {
697 		memset(data_wr->sgl, 0, sizeof(data_wr->wr.sg_list[0]) * data_wr->wr.num_sge);
698 		data_wr->wr.num_sge = 0;
699 		next_send_wr = data_wr->wr.next;
700 		if (data_wr != &rdma_req->data) {
701 			spdk_mempool_put(rtransport->data_wr_pool, data_wr);
702 		}
703 		data_wr = (!next_send_wr || next_send_wr == &rdma_req->rsp.wr) ? NULL :
704 			  SPDK_CONTAINEROF(next_send_wr, struct spdk_nvmf_rdma_request_data, wr);
705 	}
706 }
707 
708 static void
709 nvmf_rdma_dump_request(struct spdk_nvmf_rdma_request *req)
710 {
711 	SPDK_ERRLOG("\t\tRequest Data From Pool: %d\n", req->req.data_from_pool);
712 	if (req->req.cmd) {
713 		SPDK_ERRLOG("\t\tRequest opcode: %d\n", req->req.cmd->nvmf_cmd.opcode);
714 	}
715 	if (req->recv) {
716 		SPDK_ERRLOG("\t\tRequest recv wr_id%lu\n", req->recv->wr.wr_id);
717 	}
718 }
719 
720 static void
721 nvmf_rdma_dump_qpair_contents(struct spdk_nvmf_rdma_qpair *rqpair)
722 {
723 	int i;
724 
725 	SPDK_ERRLOG("Dumping contents of queue pair (QID %d)\n", rqpair->qpair.qid);
726 	for (i = 0; i < rqpair->max_queue_depth; i++) {
727 		if (rqpair->resources->reqs[i].state != RDMA_REQUEST_STATE_FREE) {
728 			nvmf_rdma_dump_request(&rqpair->resources->reqs[i]);
729 		}
730 	}
731 }
732 
733 static void
734 nvmf_rdma_resources_destroy(struct spdk_nvmf_rdma_resources *resources)
735 {
736 	if (resources->cmds_mr) {
737 		ibv_dereg_mr(resources->cmds_mr);
738 	}
739 
740 	if (resources->cpls_mr) {
741 		ibv_dereg_mr(resources->cpls_mr);
742 	}
743 
744 	if (resources->bufs_mr) {
745 		ibv_dereg_mr(resources->bufs_mr);
746 	}
747 
748 	spdk_free(resources->cmds);
749 	spdk_free(resources->cpls);
750 	spdk_free(resources->bufs);
751 	free(resources->reqs);
752 	free(resources->recvs);
753 	free(resources);
754 }
755 
756 
757 static struct spdk_nvmf_rdma_resources *
758 nvmf_rdma_resources_create(struct spdk_nvmf_rdma_resource_opts *opts)
759 {
760 	struct spdk_nvmf_rdma_resources	*resources;
761 	struct spdk_nvmf_rdma_request	*rdma_req;
762 	struct spdk_nvmf_rdma_recv	*rdma_recv;
763 	struct ibv_qp			*qp;
764 	struct ibv_srq			*srq;
765 	uint32_t			i;
766 	int				rc;
767 
768 	resources = calloc(1, sizeof(struct spdk_nvmf_rdma_resources));
769 	if (!resources) {
770 		SPDK_ERRLOG("Unable to allocate resources for receive queue.\n");
771 		return NULL;
772 	}
773 
774 	resources->reqs = calloc(opts->max_queue_depth, sizeof(*resources->reqs));
775 	resources->recvs = calloc(opts->max_queue_depth, sizeof(*resources->recvs));
776 	resources->cmds = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cmds),
777 				       0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
778 	resources->cpls = spdk_zmalloc(opts->max_queue_depth * sizeof(*resources->cpls),
779 				       0x1000, NULL, SPDK_ENV_LCORE_ID_ANY, SPDK_MALLOC_DMA);
780 
781 	if (opts->in_capsule_data_size > 0) {
782 		resources->bufs = spdk_zmalloc(opts->max_queue_depth * opts->in_capsule_data_size,
783 					       0x1000, NULL, SPDK_ENV_LCORE_ID_ANY,
784 					       SPDK_MALLOC_DMA);
785 	}
786 
787 	if (!resources->reqs || !resources->recvs || !resources->cmds ||
788 	    !resources->cpls || (opts->in_capsule_data_size && !resources->bufs)) {
789 		SPDK_ERRLOG("Unable to allocate sufficient memory for RDMA queue.\n");
790 		goto cleanup;
791 	}
792 
793 	resources->cmds_mr = ibv_reg_mr(opts->pd, resources->cmds,
794 					opts->max_queue_depth * sizeof(*resources->cmds),
795 					IBV_ACCESS_LOCAL_WRITE);
796 	resources->cpls_mr = ibv_reg_mr(opts->pd, resources->cpls,
797 					opts->max_queue_depth * sizeof(*resources->cpls),
798 					0);
799 
800 	if (opts->in_capsule_data_size) {
801 		resources->bufs_mr = ibv_reg_mr(opts->pd, resources->bufs,
802 						opts->max_queue_depth *
803 						opts->in_capsule_data_size,
804 						IBV_ACCESS_LOCAL_WRITE | IBV_ACCESS_REMOTE_WRITE);
805 	}
806 
807 	if (!resources->cmds_mr || !resources->cpls_mr ||
808 	    (opts->in_capsule_data_size &&
809 	     !resources->bufs_mr)) {
810 		goto cleanup;
811 	}
812 	SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Command Array: %p Length: %lx LKey: %x\n",
813 		      resources->cmds, opts->max_queue_depth * sizeof(*resources->cmds),
814 		      resources->cmds_mr->lkey);
815 	SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Completion Array: %p Length: %lx LKey: %x\n",
816 		      resources->cpls, opts->max_queue_depth * sizeof(*resources->cpls),
817 		      resources->cpls_mr->lkey);
818 	if (resources->bufs && resources->bufs_mr) {
819 		SPDK_DEBUGLOG(SPDK_LOG_RDMA, "In Capsule Data Array: %p Length: %x LKey: %x\n",
820 			      resources->bufs, opts->max_queue_depth *
821 			      opts->in_capsule_data_size, resources->bufs_mr->lkey);
822 	}
823 
824 	/* Initialize queues */
825 	STAILQ_INIT(&resources->incoming_queue);
826 	STAILQ_INIT(&resources->free_queue);
827 
828 	for (i = 0; i < opts->max_queue_depth; i++) {
829 		struct ibv_recv_wr *bad_wr = NULL;
830 
831 		rdma_recv = &resources->recvs[i];
832 		rdma_recv->qpair = opts->qpair;
833 
834 		/* Set up memory to receive commands */
835 		if (resources->bufs) {
836 			rdma_recv->buf = (void *)((uintptr_t)resources->bufs + (i *
837 						  opts->in_capsule_data_size));
838 		}
839 
840 		rdma_recv->rdma_wr.type = RDMA_WR_TYPE_RECV;
841 
842 		rdma_recv->sgl[0].addr = (uintptr_t)&resources->cmds[i];
843 		rdma_recv->sgl[0].length = sizeof(resources->cmds[i]);
844 		rdma_recv->sgl[0].lkey = resources->cmds_mr->lkey;
845 		rdma_recv->wr.num_sge = 1;
846 
847 		if (rdma_recv->buf && resources->bufs_mr) {
848 			rdma_recv->sgl[1].addr = (uintptr_t)rdma_recv->buf;
849 			rdma_recv->sgl[1].length = opts->in_capsule_data_size;
850 			rdma_recv->sgl[1].lkey = resources->bufs_mr->lkey;
851 			rdma_recv->wr.num_sge++;
852 		}
853 
854 		rdma_recv->wr.wr_id = (uintptr_t)&rdma_recv->rdma_wr;
855 		rdma_recv->wr.sg_list = rdma_recv->sgl;
856 		if (opts->shared) {
857 			srq = (struct ibv_srq *)opts->qp;
858 			rc = ibv_post_srq_recv(srq, &rdma_recv->wr, &bad_wr);
859 		} else {
860 			qp = (struct ibv_qp *)opts->qp;
861 			rc = ibv_post_recv(qp, &rdma_recv->wr, &bad_wr);
862 		}
863 		if (rc) {
864 			goto cleanup;
865 		}
866 	}
867 
868 	for (i = 0; i < opts->max_queue_depth; i++) {
869 		rdma_req = &resources->reqs[i];
870 
871 		if (opts->qpair != NULL) {
872 			rdma_req->req.qpair = &opts->qpair->qpair;
873 		} else {
874 			rdma_req->req.qpair = NULL;
875 		}
876 		rdma_req->req.cmd = NULL;
877 
878 		/* Set up memory to send responses */
879 		rdma_req->req.rsp = &resources->cpls[i];
880 
881 		rdma_req->rsp.sgl[0].addr = (uintptr_t)&resources->cpls[i];
882 		rdma_req->rsp.sgl[0].length = sizeof(resources->cpls[i]);
883 		rdma_req->rsp.sgl[0].lkey = resources->cpls_mr->lkey;
884 
885 		rdma_req->rsp.rdma_wr.type = RDMA_WR_TYPE_SEND;
886 		rdma_req->rsp.wr.wr_id = (uintptr_t)&rdma_req->rsp.rdma_wr;
887 		rdma_req->rsp.wr.next = NULL;
888 		rdma_req->rsp.wr.opcode = IBV_WR_SEND;
889 		rdma_req->rsp.wr.send_flags = IBV_SEND_SIGNALED;
890 		rdma_req->rsp.wr.sg_list = rdma_req->rsp.sgl;
891 		rdma_req->rsp.wr.num_sge = SPDK_COUNTOF(rdma_req->rsp.sgl);
892 
893 		/* Set up memory for data buffers */
894 		rdma_req->data.rdma_wr.type = RDMA_WR_TYPE_DATA;
895 		rdma_req->data.wr.wr_id = (uintptr_t)&rdma_req->data.rdma_wr;
896 		rdma_req->data.wr.next = NULL;
897 		rdma_req->data.wr.send_flags = IBV_SEND_SIGNALED;
898 		rdma_req->data.wr.sg_list = rdma_req->data.sgl;
899 		rdma_req->data.wr.num_sge = SPDK_COUNTOF(rdma_req->data.sgl);
900 
901 		/* Initialize request state to FREE */
902 		rdma_req->state = RDMA_REQUEST_STATE_FREE;
903 		STAILQ_INSERT_TAIL(&resources->free_queue, rdma_req, state_link);
904 	}
905 
906 	return resources;
907 
908 cleanup:
909 	nvmf_rdma_resources_destroy(resources);
910 	return NULL;
911 }
912 
913 static void
914 spdk_nvmf_rdma_qpair_clean_ibv_events(struct spdk_nvmf_rdma_qpair *rqpair)
915 {
916 	struct spdk_nvmf_rdma_ibv_event_ctx *ctx, *tctx;
917 	STAILQ_FOREACH_SAFE(ctx, &rqpair->ibv_events, link, tctx) {
918 		ctx->rqpair = NULL;
919 		/* Memory allocated for ctx is freed in spdk_nvmf_rdma_qpair_process_ibv_event */
920 		STAILQ_REMOVE(&rqpair->ibv_events, ctx, spdk_nvmf_rdma_ibv_event_ctx, link);
921 	}
922 }
923 
924 static void
925 spdk_nvmf_rdma_qpair_destroy(struct spdk_nvmf_rdma_qpair *rqpair)
926 {
927 	struct spdk_nvmf_rdma_recv	*rdma_recv, *recv_tmp;
928 	struct ibv_recv_wr		*bad_recv_wr = NULL;
929 	int				rc;
930 
931 	spdk_trace_record(TRACE_RDMA_QP_DESTROY, 0, 0, (uintptr_t)rqpair->cm_id, 0);
932 
933 	spdk_poller_unregister(&rqpair->destruct_poller);
934 
935 	if (rqpair->qd != 0) {
936 		struct spdk_nvmf_qpair *qpair = &rqpair->qpair;
937 		struct spdk_nvmf_rdma_transport	*rtransport = SPDK_CONTAINEROF(qpair->transport,
938 				struct spdk_nvmf_rdma_transport, transport);
939 		struct spdk_nvmf_rdma_request *req;
940 		uint32_t i, max_req_count = 0;
941 
942 		SPDK_WARNLOG("Destroying qpair when queue depth is %d\n", rqpair->qd);
943 
944 		if (rqpair->srq == NULL) {
945 			nvmf_rdma_dump_qpair_contents(rqpair);
946 			max_req_count = rqpair->max_queue_depth;
947 		} else if (rqpair->poller && rqpair->resources) {
948 			max_req_count = rqpair->poller->max_srq_depth;
949 		}
950 
951 		SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Release incomplete requests\n");
952 		for (i = 0; i < max_req_count; i++) {
953 			req = &rqpair->resources->reqs[i];
954 			if (req->req.qpair == qpair && req->state != RDMA_REQUEST_STATE_FREE) {
955 				/* spdk_nvmf_rdma_request_process checks qpair ibv and internal state
956 				 * and completes a request */
957 				spdk_nvmf_rdma_request_process(rtransport, req);
958 			}
959 		}
960 		assert(rqpair->qd == 0);
961 	}
962 
963 	if (rqpair->poller) {
964 		TAILQ_REMOVE(&rqpair->poller->qpairs, rqpair, link);
965 
966 		if (rqpair->srq != NULL && rqpair->resources != NULL) {
967 			/* Drop all received but unprocessed commands for this queue and return them to SRQ */
968 			STAILQ_FOREACH_SAFE(rdma_recv, &rqpair->resources->incoming_queue, link, recv_tmp) {
969 				if (rqpair == rdma_recv->qpair) {
970 					STAILQ_REMOVE(&rqpair->resources->incoming_queue, rdma_recv, spdk_nvmf_rdma_recv, link);
971 					rc = ibv_post_srq_recv(rqpair->srq, &rdma_recv->wr, &bad_recv_wr);
972 					if (rc) {
973 						SPDK_ERRLOG("Unable to re-post rx descriptor\n");
974 					}
975 				}
976 			}
977 		}
978 	}
979 
980 	if (rqpair->cm_id) {
981 		if (rqpair->cm_id->qp != NULL) {
982 			rdma_destroy_qp(rqpair->cm_id);
983 		}
984 		rdma_destroy_id(rqpair->cm_id);
985 
986 		if (rqpair->poller != NULL && rqpair->srq == NULL) {
987 			rqpair->poller->required_num_wr -= MAX_WR_PER_QP(rqpair->max_queue_depth);
988 		}
989 	}
990 
991 	if (rqpair->srq == NULL && rqpair->resources != NULL) {
992 		nvmf_rdma_resources_destroy(rqpair->resources);
993 	}
994 
995 	spdk_nvmf_rdma_qpair_clean_ibv_events(rqpair);
996 
997 	free(rqpair);
998 }
999 
1000 static int
1001 nvmf_rdma_resize_cq(struct spdk_nvmf_rdma_qpair *rqpair, struct spdk_nvmf_rdma_device *device)
1002 {
1003 	struct spdk_nvmf_rdma_poller	*rpoller;
1004 	int				rc, num_cqe, required_num_wr;
1005 
1006 	/* Enlarge CQ size dynamically */
1007 	rpoller = rqpair->poller;
1008 	required_num_wr = rpoller->required_num_wr + MAX_WR_PER_QP(rqpair->max_queue_depth);
1009 	num_cqe = rpoller->num_cqe;
1010 	if (num_cqe < required_num_wr) {
1011 		num_cqe = spdk_max(num_cqe * 2, required_num_wr);
1012 		num_cqe = spdk_min(num_cqe, device->attr.max_cqe);
1013 	}
1014 
1015 	if (rpoller->num_cqe != num_cqe) {
1016 		if (required_num_wr > device->attr.max_cqe) {
1017 			SPDK_ERRLOG("RDMA CQE requirement (%d) exceeds device max_cqe limitation (%d)\n",
1018 				    required_num_wr, device->attr.max_cqe);
1019 			return -1;
1020 		}
1021 
1022 		SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Resize RDMA CQ from %d to %d\n", rpoller->num_cqe, num_cqe);
1023 		rc = ibv_resize_cq(rpoller->cq, num_cqe);
1024 		if (rc) {
1025 			SPDK_ERRLOG("RDMA CQ resize failed: errno %d: %s\n", errno, spdk_strerror(errno));
1026 			return -1;
1027 		}
1028 
1029 		rpoller->num_cqe = num_cqe;
1030 	}
1031 
1032 	rpoller->required_num_wr = required_num_wr;
1033 	return 0;
1034 }
1035 
1036 static int
1037 spdk_nvmf_rdma_qpair_initialize(struct spdk_nvmf_qpair *qpair)
1038 {
1039 	struct spdk_nvmf_rdma_qpair		*rqpair;
1040 	int					rc;
1041 	struct spdk_nvmf_rdma_transport		*rtransport;
1042 	struct spdk_nvmf_transport		*transport;
1043 	struct spdk_nvmf_rdma_resource_opts	opts;
1044 	struct spdk_nvmf_rdma_device		*device;
1045 	struct ibv_qp_init_attr			ibv_init_attr;
1046 
1047 	rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
1048 	device = rqpair->device;
1049 
1050 	memset(&ibv_init_attr, 0, sizeof(struct ibv_qp_init_attr));
1051 	ibv_init_attr.qp_context	= rqpair;
1052 	ibv_init_attr.qp_type		= IBV_QPT_RC;
1053 	ibv_init_attr.send_cq		= rqpair->poller->cq;
1054 	ibv_init_attr.recv_cq		= rqpair->poller->cq;
1055 
1056 	if (rqpair->srq) {
1057 		ibv_init_attr.srq		= rqpair->srq;
1058 	} else {
1059 		ibv_init_attr.cap.max_recv_wr	= rqpair->max_queue_depth;
1060 	}
1061 
1062 	ibv_init_attr.cap.max_send_wr	= (uint32_t)rqpair->max_queue_depth *
1063 					  2; /* SEND, READ or WRITE operations */
1064 	ibv_init_attr.cap.max_send_sge	= spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_TX_SGE);
1065 	ibv_init_attr.cap.max_recv_sge	= spdk_min((uint32_t)device->attr.max_sge, NVMF_DEFAULT_RX_SGE);
1066 
1067 	if (rqpair->srq == NULL && nvmf_rdma_resize_cq(rqpair, device) < 0) {
1068 		SPDK_ERRLOG("Failed to resize the completion queue. Cannot initialize qpair.\n");
1069 		goto error;
1070 	}
1071 
1072 	rc = rdma_create_qp(rqpair->cm_id, device->pd, &ibv_init_attr);
1073 	if (rc) {
1074 		SPDK_ERRLOG("rdma_create_qp failed: errno %d: %s\n", errno, spdk_strerror(errno));
1075 		goto error;
1076 	}
1077 
1078 	rqpair->max_send_depth = spdk_min((uint32_t)(rqpair->max_queue_depth * 2),
1079 					  ibv_init_attr.cap.max_send_wr);
1080 	rqpair->max_send_sge = spdk_min(NVMF_DEFAULT_TX_SGE, ibv_init_attr.cap.max_send_sge);
1081 	rqpair->max_recv_sge = spdk_min(NVMF_DEFAULT_RX_SGE, ibv_init_attr.cap.max_recv_sge);
1082 	spdk_trace_record(TRACE_RDMA_QP_CREATE, 0, 0, (uintptr_t)rqpair->cm_id, 0);
1083 	SPDK_DEBUGLOG(SPDK_LOG_RDMA, "New RDMA Connection: %p\n", qpair);
1084 
1085 	rqpair->sends_to_post.first = NULL;
1086 	rqpair->sends_to_post.last = NULL;
1087 
1088 	if (rqpair->poller->srq == NULL) {
1089 		rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
1090 		transport = &rtransport->transport;
1091 
1092 		opts.qp = rqpair->cm_id->qp;
1093 		opts.pd = rqpair->cm_id->pd;
1094 		opts.qpair = rqpair;
1095 		opts.shared = false;
1096 		opts.max_queue_depth = rqpair->max_queue_depth;
1097 		opts.in_capsule_data_size = transport->opts.in_capsule_data_size;
1098 
1099 		rqpair->resources = nvmf_rdma_resources_create(&opts);
1100 
1101 		if (!rqpair->resources) {
1102 			SPDK_ERRLOG("Unable to allocate resources for receive queue.\n");
1103 			rdma_destroy_qp(rqpair->cm_id);
1104 			goto error;
1105 		}
1106 	} else {
1107 		rqpair->resources = rqpair->poller->resources;
1108 	}
1109 
1110 	rqpair->current_recv_depth = 0;
1111 	STAILQ_INIT(&rqpair->pending_rdma_read_queue);
1112 	STAILQ_INIT(&rqpair->pending_rdma_write_queue);
1113 
1114 	return 0;
1115 
1116 error:
1117 	rdma_destroy_id(rqpair->cm_id);
1118 	rqpair->cm_id = NULL;
1119 	return -1;
1120 }
1121 
1122 /* Append the given recv wr structure to the resource structs outstanding recvs list. */
1123 /* This function accepts either a single wr or the first wr in a linked list. */
1124 static void
1125 nvmf_rdma_qpair_queue_recv_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *first)
1126 {
1127 	struct ibv_recv_wr *last;
1128 
1129 	last = first;
1130 	while (last->next != NULL) {
1131 		last = last->next;
1132 	}
1133 
1134 	if (rqpair->resources->recvs_to_post.first == NULL) {
1135 		rqpair->resources->recvs_to_post.first = first;
1136 		rqpair->resources->recvs_to_post.last = last;
1137 		if (rqpair->srq == NULL) {
1138 			STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_recv, rqpair, recv_link);
1139 		}
1140 	} else {
1141 		rqpair->resources->recvs_to_post.last->next = first;
1142 		rqpair->resources->recvs_to_post.last = last;
1143 	}
1144 }
1145 
1146 /* Append the given send wr structure to the qpair's outstanding sends list. */
1147 /* This function accepts either a single wr or the first wr in a linked list. */
1148 static void
1149 nvmf_rdma_qpair_queue_send_wrs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *first)
1150 {
1151 	struct ibv_send_wr *last;
1152 
1153 	last = first;
1154 	while (last->next != NULL) {
1155 		last = last->next;
1156 	}
1157 
1158 	if (rqpair->sends_to_post.first == NULL) {
1159 		rqpair->sends_to_post.first = first;
1160 		rqpair->sends_to_post.last = last;
1161 		STAILQ_INSERT_TAIL(&rqpair->poller->qpairs_pending_send, rqpair, send_link);
1162 	} else {
1163 		rqpair->sends_to_post.last->next = first;
1164 		rqpair->sends_to_post.last = last;
1165 	}
1166 }
1167 
1168 static int
1169 request_transfer_in(struct spdk_nvmf_request *req)
1170 {
1171 	struct spdk_nvmf_rdma_request	*rdma_req;
1172 	struct spdk_nvmf_qpair		*qpair;
1173 	struct spdk_nvmf_rdma_qpair	*rqpair;
1174 
1175 	qpair = req->qpair;
1176 	rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
1177 	rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
1178 
1179 	assert(req->xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER);
1180 	assert(rdma_req != NULL);
1181 
1182 	nvmf_rdma_qpair_queue_send_wrs(rqpair, &rdma_req->data.wr);
1183 	rqpair->current_read_depth += rdma_req->num_outstanding_data_wr;
1184 	rqpair->current_send_depth += rdma_req->num_outstanding_data_wr;
1185 	return 0;
1186 }
1187 
1188 static int
1189 request_transfer_out(struct spdk_nvmf_request *req, int *data_posted)
1190 {
1191 	int				num_outstanding_data_wr = 0;
1192 	struct spdk_nvmf_rdma_request	*rdma_req;
1193 	struct spdk_nvmf_qpair		*qpair;
1194 	struct spdk_nvmf_rdma_qpair	*rqpair;
1195 	struct spdk_nvme_cpl		*rsp;
1196 	struct ibv_send_wr		*first = NULL;
1197 
1198 	*data_posted = 0;
1199 	qpair = req->qpair;
1200 	rsp = &req->rsp->nvme_cpl;
1201 	rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
1202 	rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
1203 
1204 	/* Advance our sq_head pointer */
1205 	if (qpair->sq_head == qpair->sq_head_max) {
1206 		qpair->sq_head = 0;
1207 	} else {
1208 		qpair->sq_head++;
1209 	}
1210 	rsp->sqhd = qpair->sq_head;
1211 
1212 	/* queue the capsule for the recv buffer */
1213 	assert(rdma_req->recv != NULL);
1214 
1215 	nvmf_rdma_qpair_queue_recv_wrs(rqpair, &rdma_req->recv->wr);
1216 
1217 	rdma_req->recv = NULL;
1218 	assert(rqpair->current_recv_depth > 0);
1219 	rqpair->current_recv_depth--;
1220 
1221 	/* Build the response which consists of optional
1222 	 * RDMA WRITEs to transfer data, plus an RDMA SEND
1223 	 * containing the response.
1224 	 */
1225 	first = &rdma_req->rsp.wr;
1226 
1227 	if (rsp->status.sc == SPDK_NVME_SC_SUCCESS &&
1228 	    req->xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
1229 		first = &rdma_req->data.wr;
1230 		*data_posted = 1;
1231 		num_outstanding_data_wr = rdma_req->num_outstanding_data_wr;
1232 	}
1233 	nvmf_rdma_qpair_queue_send_wrs(rqpair, first);
1234 	/* +1 for the rsp wr */
1235 	rqpair->current_send_depth += num_outstanding_data_wr + 1;
1236 
1237 	return 0;
1238 }
1239 
1240 static int
1241 spdk_nvmf_rdma_event_accept(struct rdma_cm_id *id, struct spdk_nvmf_rdma_qpair *rqpair)
1242 {
1243 	struct spdk_nvmf_rdma_accept_private_data	accept_data;
1244 	struct rdma_conn_param				ctrlr_event_data = {};
1245 	int						rc;
1246 
1247 	accept_data.recfmt = 0;
1248 	accept_data.crqsize = rqpair->max_queue_depth;
1249 
1250 	ctrlr_event_data.private_data = &accept_data;
1251 	ctrlr_event_data.private_data_len = sizeof(accept_data);
1252 	if (id->ps == RDMA_PS_TCP) {
1253 		ctrlr_event_data.responder_resources = 0; /* We accept 0 reads from the host */
1254 		ctrlr_event_data.initiator_depth = rqpair->max_read_depth;
1255 	}
1256 
1257 	/* Configure infinite retries for the initiator side qpair.
1258 	 * When using a shared receive queue on the target side,
1259 	 * we need to pass this value to the initiator to prevent the
1260 	 * initiator side NIC from completing SEND requests back to the
1261 	 * initiator with status rnr_retry_count_exceeded. */
1262 	if (rqpair->srq != NULL) {
1263 		ctrlr_event_data.rnr_retry_count = 0x7;
1264 	}
1265 
1266 	rc = rdma_accept(id, &ctrlr_event_data);
1267 	if (rc) {
1268 		SPDK_ERRLOG("Error %d on rdma_accept\n", errno);
1269 	} else {
1270 		SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Sent back the accept\n");
1271 	}
1272 
1273 	return rc;
1274 }
1275 
1276 static void
1277 spdk_nvmf_rdma_event_reject(struct rdma_cm_id *id, enum spdk_nvmf_rdma_transport_error error)
1278 {
1279 	struct spdk_nvmf_rdma_reject_private_data	rej_data;
1280 
1281 	rej_data.recfmt = 0;
1282 	rej_data.sts = error;
1283 
1284 	rdma_reject(id, &rej_data, sizeof(rej_data));
1285 }
1286 
1287 static int
1288 nvmf_rdma_connect(struct spdk_nvmf_transport *transport, struct rdma_cm_event *event,
1289 		  new_qpair_fn cb_fn, void *cb_arg)
1290 {
1291 	struct spdk_nvmf_rdma_transport *rtransport;
1292 	struct spdk_nvmf_rdma_qpair	*rqpair = NULL;
1293 	struct spdk_nvmf_rdma_port	*port;
1294 	struct rdma_conn_param		*rdma_param = NULL;
1295 	const struct spdk_nvmf_rdma_request_private_data *private_data = NULL;
1296 	uint16_t			max_queue_depth;
1297 	uint16_t			max_read_depth;
1298 
1299 	rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
1300 
1301 	assert(event->id != NULL); /* Impossible. Can't even reject the connection. */
1302 	assert(event->id->verbs != NULL); /* Impossible. No way to handle this. */
1303 
1304 	rdma_param = &event->param.conn;
1305 	if (rdma_param->private_data == NULL ||
1306 	    rdma_param->private_data_len < sizeof(struct spdk_nvmf_rdma_request_private_data)) {
1307 		SPDK_ERRLOG("connect request: no private data provided\n");
1308 		spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_PRIVATE_DATA_LENGTH);
1309 		return -1;
1310 	}
1311 
1312 	private_data = rdma_param->private_data;
1313 	if (private_data->recfmt != 0) {
1314 		SPDK_ERRLOG("Received RDMA private data with RECFMT != 0\n");
1315 		spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_INVALID_RECFMT);
1316 		return -1;
1317 	}
1318 
1319 	SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Connect Recv on fabric intf name %s, dev_name %s\n",
1320 		      event->id->verbs->device->name, event->id->verbs->device->dev_name);
1321 
1322 	port = event->listen_id->context;
1323 	SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Listen Id was %p with verbs %p. ListenAddr: %p\n",
1324 		      event->listen_id, event->listen_id->verbs, port);
1325 
1326 	/* Figure out the supported queue depth. This is a multi-step process
1327 	 * that takes into account hardware maximums, host provided values,
1328 	 * and our target's internal memory limits */
1329 
1330 	SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Calculating Queue Depth\n");
1331 
1332 	/* Start with the maximum queue depth allowed by the target */
1333 	max_queue_depth = rtransport->transport.opts.max_queue_depth;
1334 	max_read_depth = rtransport->transport.opts.max_queue_depth;
1335 	SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Target Max Queue Depth: %d\n",
1336 		      rtransport->transport.opts.max_queue_depth);
1337 
1338 	/* Next check the local NIC's hardware limitations */
1339 	SPDK_DEBUGLOG(SPDK_LOG_RDMA,
1340 		      "Local NIC Max Send/Recv Queue Depth: %d Max Read/Write Queue Depth: %d\n",
1341 		      port->device->attr.max_qp_wr, port->device->attr.max_qp_rd_atom);
1342 	max_queue_depth = spdk_min(max_queue_depth, port->device->attr.max_qp_wr);
1343 	max_read_depth = spdk_min(max_read_depth, port->device->attr.max_qp_init_rd_atom);
1344 
1345 	/* Next check the remote NIC's hardware limitations */
1346 	SPDK_DEBUGLOG(SPDK_LOG_RDMA,
1347 		      "Host (Initiator) NIC Max Incoming RDMA R/W operations: %d Max Outgoing RDMA R/W operations: %d\n",
1348 		      rdma_param->initiator_depth, rdma_param->responder_resources);
1349 	if (rdma_param->initiator_depth > 0) {
1350 		max_read_depth = spdk_min(max_read_depth, rdma_param->initiator_depth);
1351 	}
1352 
1353 	/* Finally check for the host software requested values, which are
1354 	 * optional. */
1355 	if (rdma_param->private_data != NULL &&
1356 	    rdma_param->private_data_len >= sizeof(struct spdk_nvmf_rdma_request_private_data)) {
1357 		SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Receive Queue Size: %d\n", private_data->hrqsize);
1358 		SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Host Send Queue Size: %d\n", private_data->hsqsize);
1359 		max_queue_depth = spdk_min(max_queue_depth, private_data->hrqsize);
1360 		max_queue_depth = spdk_min(max_queue_depth, private_data->hsqsize + 1);
1361 	}
1362 
1363 	SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Final Negotiated Queue Depth: %d R/W Depth: %d\n",
1364 		      max_queue_depth, max_read_depth);
1365 
1366 	rqpair = calloc(1, sizeof(struct spdk_nvmf_rdma_qpair));
1367 	if (rqpair == NULL) {
1368 		SPDK_ERRLOG("Could not allocate new connection.\n");
1369 		spdk_nvmf_rdma_event_reject(event->id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES);
1370 		return -1;
1371 	}
1372 
1373 	rqpair->device = port->device;
1374 	rqpair->max_queue_depth = max_queue_depth;
1375 	rqpair->max_read_depth = max_read_depth;
1376 	rqpair->cm_id = event->id;
1377 	rqpair->listen_id = event->listen_id;
1378 	rqpair->qpair.transport = transport;
1379 	STAILQ_INIT(&rqpair->ibv_events);
1380 	/* use qid from the private data to determine the qpair type
1381 	   qid will be set to the appropriate value when the controller is created */
1382 	rqpair->qpair.qid = private_data->qid;
1383 
1384 	event->id->context = &rqpair->qpair;
1385 
1386 	cb_fn(&rqpair->qpair, cb_arg);
1387 
1388 	return 0;
1389 }
1390 
1391 static int
1392 spdk_nvmf_rdma_mem_notify(void *cb_ctx, struct spdk_mem_map *map,
1393 			  enum spdk_mem_map_notify_action action,
1394 			  void *vaddr, size_t size)
1395 {
1396 	struct ibv_pd *pd = cb_ctx;
1397 	struct ibv_mr *mr;
1398 	int rc;
1399 
1400 	switch (action) {
1401 	case SPDK_MEM_MAP_NOTIFY_REGISTER:
1402 		if (!g_nvmf_hooks.get_rkey) {
1403 			mr = ibv_reg_mr(pd, vaddr, size,
1404 					IBV_ACCESS_LOCAL_WRITE |
1405 					IBV_ACCESS_REMOTE_READ |
1406 					IBV_ACCESS_REMOTE_WRITE);
1407 			if (mr == NULL) {
1408 				SPDK_ERRLOG("ibv_reg_mr() failed\n");
1409 				return -1;
1410 			} else {
1411 				rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, size, (uint64_t)mr);
1412 			}
1413 		} else {
1414 			rc = spdk_mem_map_set_translation(map, (uint64_t)vaddr, size,
1415 							  g_nvmf_hooks.get_rkey(pd, vaddr, size));
1416 		}
1417 		break;
1418 	case SPDK_MEM_MAP_NOTIFY_UNREGISTER:
1419 		if (!g_nvmf_hooks.get_rkey) {
1420 			mr = (struct ibv_mr *)spdk_mem_map_translate(map, (uint64_t)vaddr, NULL);
1421 			if (mr) {
1422 				ibv_dereg_mr(mr);
1423 			}
1424 		}
1425 		rc = spdk_mem_map_clear_translation(map, (uint64_t)vaddr, size);
1426 		break;
1427 	default:
1428 		SPDK_UNREACHABLE();
1429 	}
1430 
1431 	return rc;
1432 }
1433 
1434 static int
1435 spdk_nvmf_rdma_check_contiguous_entries(uint64_t addr_1, uint64_t addr_2)
1436 {
1437 	/* Two contiguous mappings will point to the same address which is the start of the RDMA MR. */
1438 	return addr_1 == addr_2;
1439 }
1440 
1441 static inline void
1442 nvmf_rdma_setup_wr(struct ibv_send_wr *wr, struct ibv_send_wr *next,
1443 		   enum spdk_nvme_data_transfer xfer)
1444 {
1445 	if (xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
1446 		wr->opcode = IBV_WR_RDMA_WRITE;
1447 		wr->send_flags = 0;
1448 		wr->next = next;
1449 	} else if (xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) {
1450 		wr->opcode = IBV_WR_RDMA_READ;
1451 		wr->send_flags = IBV_SEND_SIGNALED;
1452 		wr->next = NULL;
1453 	} else {
1454 		assert(0);
1455 	}
1456 }
1457 
1458 static int
1459 nvmf_request_alloc_wrs(struct spdk_nvmf_rdma_transport *rtransport,
1460 		       struct spdk_nvmf_rdma_request *rdma_req,
1461 		       uint32_t num_sgl_descriptors)
1462 {
1463 	struct spdk_nvmf_rdma_request_data	*work_requests[SPDK_NVMF_MAX_SGL_ENTRIES];
1464 	struct spdk_nvmf_rdma_request_data	*current_data_wr;
1465 	uint32_t				i;
1466 
1467 	if (num_sgl_descriptors > SPDK_NVMF_MAX_SGL_ENTRIES) {
1468 		SPDK_ERRLOG("Requested too much entries (%u), the limit is %u\n",
1469 			    num_sgl_descriptors, SPDK_NVMF_MAX_SGL_ENTRIES);
1470 		return -EINVAL;
1471 	}
1472 
1473 	if (spdk_mempool_get_bulk(rtransport->data_wr_pool, (void **)work_requests, num_sgl_descriptors)) {
1474 		return -ENOMEM;
1475 	}
1476 
1477 	current_data_wr = &rdma_req->data;
1478 
1479 	for (i = 0; i < num_sgl_descriptors; i++) {
1480 		nvmf_rdma_setup_wr(&current_data_wr->wr, &work_requests[i]->wr, rdma_req->req.xfer);
1481 		current_data_wr->wr.next = &work_requests[i]->wr;
1482 		current_data_wr = work_requests[i];
1483 		current_data_wr->wr.sg_list = current_data_wr->sgl;
1484 		current_data_wr->wr.wr_id = rdma_req->data.wr.wr_id;
1485 	}
1486 
1487 	nvmf_rdma_setup_wr(&current_data_wr->wr, &rdma_req->rsp.wr, rdma_req->req.xfer);
1488 
1489 	return 0;
1490 }
1491 
1492 static inline void
1493 nvmf_rdma_setup_request(struct spdk_nvmf_rdma_request *rdma_req)
1494 {
1495 	struct ibv_send_wr		*wr = &rdma_req->data.wr;
1496 	struct spdk_nvme_sgl_descriptor	*sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1;
1497 
1498 	wr->wr.rdma.rkey = sgl->keyed.key;
1499 	wr->wr.rdma.remote_addr = sgl->address;
1500 	nvmf_rdma_setup_wr(wr, &rdma_req->rsp.wr, rdma_req->req.xfer);
1501 }
1502 
1503 static inline void
1504 nvmf_rdma_update_remote_addr(struct spdk_nvmf_rdma_request *rdma_req, uint32_t num_wrs)
1505 {
1506 	struct ibv_send_wr		*wr = &rdma_req->data.wr;
1507 	struct spdk_nvme_sgl_descriptor	*sgl = &rdma_req->req.cmd->nvme_cmd.dptr.sgl1;
1508 	uint32_t			i;
1509 	int				j;
1510 	uint64_t			remote_addr_offset = 0;
1511 
1512 	for (i = 0; i < num_wrs; ++i) {
1513 		wr->wr.rdma.rkey = sgl->keyed.key;
1514 		wr->wr.rdma.remote_addr = sgl->address + remote_addr_offset;
1515 		for (j = 0; j < wr->num_sge; ++j) {
1516 			remote_addr_offset += wr->sg_list[j].length;
1517 		}
1518 		wr = wr->next;
1519 	}
1520 }
1521 
1522 /* This function is used in the rare case that we have a buffer split over multiple memory regions. */
1523 static int
1524 nvmf_rdma_replace_buffer(struct spdk_nvmf_rdma_poll_group *rgroup, void **buf)
1525 {
1526 	struct spdk_nvmf_transport_poll_group	*group = &rgroup->group;
1527 	struct spdk_nvmf_transport		*transport = group->transport;
1528 	struct spdk_nvmf_transport_pg_cache_buf	*old_buf;
1529 	void					*new_buf;
1530 
1531 	if (!(STAILQ_EMPTY(&group->buf_cache))) {
1532 		group->buf_cache_count--;
1533 		new_buf = STAILQ_FIRST(&group->buf_cache);
1534 		STAILQ_REMOVE_HEAD(&group->buf_cache, link);
1535 		assert(*buf != NULL);
1536 	} else {
1537 		new_buf = spdk_mempool_get(transport->data_buf_pool);
1538 	}
1539 
1540 	if (*buf == NULL) {
1541 		return -ENOMEM;
1542 	}
1543 
1544 	old_buf = *buf;
1545 	STAILQ_INSERT_HEAD(&rgroup->retired_bufs, old_buf, link);
1546 	*buf = new_buf;
1547 	return 0;
1548 }
1549 
1550 static bool
1551 nvmf_rdma_get_lkey(struct spdk_nvmf_rdma_device *device, struct iovec *iov,
1552 		   uint32_t *_lkey)
1553 {
1554 	uint64_t	translation_len;
1555 	uint32_t	lkey;
1556 
1557 	translation_len = iov->iov_len;
1558 
1559 	if (!g_nvmf_hooks.get_rkey) {
1560 		lkey = ((struct ibv_mr *)spdk_mem_map_translate(device->map,
1561 				(uint64_t)iov->iov_base, &translation_len))->lkey;
1562 	} else {
1563 		lkey = spdk_mem_map_translate(device->map,
1564 					      (uint64_t)iov->iov_base, &translation_len);
1565 	}
1566 
1567 	if (spdk_unlikely(translation_len < iov->iov_len)) {
1568 		return false;
1569 	}
1570 
1571 	*_lkey = lkey;
1572 	return true;
1573 }
1574 
1575 static bool
1576 nvmf_rdma_fill_wr_sge(struct spdk_nvmf_rdma_device *device,
1577 		      struct iovec *iov, struct ibv_send_wr **_wr,
1578 		      uint32_t *_remaining_data_block, uint32_t *_offset,
1579 		      uint32_t *_num_extra_wrs,
1580 		      const struct spdk_dif_ctx *dif_ctx)
1581 {
1582 	struct ibv_send_wr *wr = *_wr;
1583 	struct ibv_sge	*sg_ele = &wr->sg_list[wr->num_sge];
1584 	uint32_t	lkey = 0;
1585 	uint32_t	remaining, data_block_size, md_size, sge_len;
1586 
1587 	if (spdk_unlikely(!nvmf_rdma_get_lkey(device, iov, &lkey))) {
1588 		/* This is a very rare case that can occur when using DPDK version < 19.05 */
1589 		SPDK_ERRLOG("Data buffer split over multiple RDMA Memory Regions. Removing it from circulation.\n");
1590 		return false;
1591 	}
1592 
1593 	if (spdk_likely(!dif_ctx)) {
1594 		sg_ele->lkey = lkey;
1595 		sg_ele->addr = (uintptr_t)(iov->iov_base);
1596 		sg_ele->length = iov->iov_len;
1597 		wr->num_sge++;
1598 	} else {
1599 		remaining = iov->iov_len - *_offset;
1600 		data_block_size = dif_ctx->block_size - dif_ctx->md_size;
1601 		md_size = dif_ctx->md_size;
1602 
1603 		while (remaining) {
1604 			if (wr->num_sge >= SPDK_NVMF_MAX_SGL_ENTRIES) {
1605 				if (*_num_extra_wrs > 0 && wr->next) {
1606 					*_wr = wr->next;
1607 					wr = *_wr;
1608 					wr->num_sge = 0;
1609 					sg_ele = &wr->sg_list[wr->num_sge];
1610 					(*_num_extra_wrs)--;
1611 				} else {
1612 					break;
1613 				}
1614 			}
1615 			sg_ele->lkey = lkey;
1616 			sg_ele->addr = (uintptr_t)((char *)iov->iov_base + *_offset);
1617 			sge_len = spdk_min(remaining, *_remaining_data_block);
1618 			sg_ele->length = sge_len;
1619 			remaining -= sge_len;
1620 			*_remaining_data_block -= sge_len;
1621 			*_offset += sge_len;
1622 
1623 			sg_ele++;
1624 			wr->num_sge++;
1625 
1626 			if (*_remaining_data_block == 0) {
1627 				/* skip metadata */
1628 				*_offset += md_size;
1629 				/* Metadata that do not fit this IO buffer will be included in the next IO buffer */
1630 				remaining -= spdk_min(remaining, md_size);
1631 				*_remaining_data_block = data_block_size;
1632 			}
1633 
1634 			if (remaining == 0) {
1635 				/* By subtracting the size of the last IOV from the offset, we ensure that we skip
1636 				   the remaining metadata bits at the beginning of the next buffer */
1637 				*_offset -= iov->iov_len;
1638 			}
1639 		}
1640 	}
1641 
1642 	return true;
1643 }
1644 
1645 static int
1646 nvmf_rdma_fill_wr_sgl(struct spdk_nvmf_rdma_poll_group *rgroup,
1647 		      struct spdk_nvmf_rdma_device *device,
1648 		      struct spdk_nvmf_rdma_request *rdma_req,
1649 		      struct ibv_send_wr *wr,
1650 		      uint32_t length,
1651 		      uint32_t num_extra_wrs)
1652 {
1653 	struct spdk_nvmf_request *req = &rdma_req->req;
1654 	struct spdk_dif_ctx *dif_ctx = NULL;
1655 	uint32_t remaining_data_block = 0;
1656 	uint32_t offset = 0;
1657 
1658 	if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) {
1659 		dif_ctx = &rdma_req->req.dif.dif_ctx;
1660 		remaining_data_block = dif_ctx->block_size - dif_ctx->md_size;
1661 	}
1662 
1663 	wr->num_sge = 0;
1664 
1665 	while (length && (num_extra_wrs || wr->num_sge < SPDK_NVMF_MAX_SGL_ENTRIES)) {
1666 		while (spdk_unlikely(!nvmf_rdma_fill_wr_sge(device, &req->iov[rdma_req->iovpos], &wr,
1667 				     &remaining_data_block, &offset, &num_extra_wrs, dif_ctx))) {
1668 			if (nvmf_rdma_replace_buffer(rgroup, &req->buffers[rdma_req->iovpos]) == -ENOMEM) {
1669 				return -ENOMEM;
1670 			}
1671 			req->iov[rdma_req->iovpos].iov_base = (void *)((uintptr_t)(req->buffers[rdma_req->iovpos] +
1672 							      NVMF_DATA_BUFFER_MASK) &
1673 							      ~NVMF_DATA_BUFFER_MASK);
1674 		}
1675 
1676 		length -= req->iov[rdma_req->iovpos].iov_len;
1677 		rdma_req->iovpos++;
1678 	}
1679 
1680 	if (length) {
1681 		SPDK_ERRLOG("Not enough SG entries to hold data buffer\n");
1682 		return -EINVAL;
1683 	}
1684 
1685 	return 0;
1686 }
1687 
1688 static inline uint32_t
1689 nvmf_rdma_calc_num_wrs(uint32_t length, uint32_t io_unit_size, uint32_t block_size)
1690 {
1691 	/* estimate the number of SG entries and WRs needed to process the request */
1692 	uint32_t num_sge = 0;
1693 	uint32_t i;
1694 	uint32_t num_buffers = SPDK_CEIL_DIV(length, io_unit_size);
1695 
1696 	for (i = 0; i < num_buffers && length > 0; i++) {
1697 		uint32_t buffer_len = spdk_min(length, io_unit_size);
1698 		uint32_t num_sge_in_block = SPDK_CEIL_DIV(buffer_len, block_size);
1699 
1700 		if (num_sge_in_block * block_size > buffer_len) {
1701 			++num_sge_in_block;
1702 		}
1703 		num_sge += num_sge_in_block;
1704 		length -= buffer_len;
1705 	}
1706 	return SPDK_CEIL_DIV(num_sge, SPDK_NVMF_MAX_SGL_ENTRIES);
1707 }
1708 
1709 static int
1710 spdk_nvmf_rdma_request_fill_iovs(struct spdk_nvmf_rdma_transport *rtransport,
1711 				 struct spdk_nvmf_rdma_device *device,
1712 				 struct spdk_nvmf_rdma_request *rdma_req,
1713 				 uint32_t length)
1714 {
1715 	struct spdk_nvmf_rdma_qpair		*rqpair;
1716 	struct spdk_nvmf_rdma_poll_group	*rgroup;
1717 	struct spdk_nvmf_request		*req = &rdma_req->req;
1718 	struct ibv_send_wr			*wr = &rdma_req->data.wr;
1719 	int					rc;
1720 	uint32_t				num_wrs = 1;
1721 
1722 	rqpair = SPDK_CONTAINEROF(req->qpair, struct spdk_nvmf_rdma_qpair, qpair);
1723 	rgroup = rqpair->poller->group;
1724 
1725 	/* rdma wr specifics */
1726 	nvmf_rdma_setup_request(rdma_req);
1727 
1728 	rc = spdk_nvmf_request_get_buffers(req, &rgroup->group, &rtransport->transport,
1729 					   length);
1730 	if (rc != 0) {
1731 		return rc;
1732 	}
1733 
1734 	assert(req->iovcnt <= rqpair->max_send_sge);
1735 
1736 	rdma_req->iovpos = 0;
1737 
1738 	if (spdk_unlikely(req->dif.dif_insert_or_strip)) {
1739 		num_wrs = nvmf_rdma_calc_num_wrs(length, rtransport->transport.opts.io_unit_size,
1740 						 req->dif.dif_ctx.block_size);
1741 		if (num_wrs > 1) {
1742 			rc = nvmf_request_alloc_wrs(rtransport, rdma_req, num_wrs - 1);
1743 			if (rc != 0) {
1744 				goto err_exit;
1745 			}
1746 		}
1747 	}
1748 
1749 	rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, wr, length, num_wrs - 1);
1750 	if (spdk_unlikely(rc != 0)) {
1751 		goto err_exit;
1752 	}
1753 
1754 	if (spdk_unlikely(num_wrs > 1)) {
1755 		nvmf_rdma_update_remote_addr(rdma_req, num_wrs);
1756 	}
1757 
1758 	/* set the number of outstanding data WRs for this request. */
1759 	rdma_req->num_outstanding_data_wr = num_wrs;
1760 
1761 	return rc;
1762 
1763 err_exit:
1764 	spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport);
1765 	nvmf_rdma_request_free_data(rdma_req, rtransport);
1766 	req->iovcnt = 0;
1767 	return rc;
1768 }
1769 
1770 static int
1771 nvmf_rdma_request_fill_iovs_multi_sgl(struct spdk_nvmf_rdma_transport *rtransport,
1772 				      struct spdk_nvmf_rdma_device *device,
1773 				      struct spdk_nvmf_rdma_request *rdma_req)
1774 {
1775 	struct spdk_nvmf_rdma_qpair		*rqpair;
1776 	struct spdk_nvmf_rdma_poll_group	*rgroup;
1777 	struct ibv_send_wr			*current_wr;
1778 	struct spdk_nvmf_request		*req = &rdma_req->req;
1779 	struct spdk_nvme_sgl_descriptor		*inline_segment, *desc;
1780 	uint32_t				num_sgl_descriptors;
1781 	uint32_t				lengths[SPDK_NVMF_MAX_SGL_ENTRIES];
1782 	uint32_t				i;
1783 	int					rc;
1784 
1785 	rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
1786 	rgroup = rqpair->poller->group;
1787 
1788 	inline_segment = &req->cmd->nvme_cmd.dptr.sgl1;
1789 	assert(inline_segment->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT);
1790 	assert(inline_segment->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET);
1791 
1792 	num_sgl_descriptors = inline_segment->unkeyed.length / sizeof(struct spdk_nvme_sgl_descriptor);
1793 	assert(num_sgl_descriptors <= SPDK_NVMF_MAX_SGL_ENTRIES);
1794 
1795 	if (nvmf_request_alloc_wrs(rtransport, rdma_req, num_sgl_descriptors - 1) != 0) {
1796 		return -ENOMEM;
1797 	}
1798 
1799 	desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address;
1800 	for (i = 0; i < num_sgl_descriptors; i++) {
1801 		if (spdk_likely(!req->dif.dif_insert_or_strip)) {
1802 			lengths[i] = desc->keyed.length;
1803 		} else {
1804 			req->dif.orig_length += desc->keyed.length;
1805 			lengths[i] = spdk_dif_get_length_with_md(desc->keyed.length, &req->dif.dif_ctx);
1806 			req->dif.elba_length += lengths[i];
1807 		}
1808 		desc++;
1809 	}
1810 
1811 	rc = spdk_nvmf_request_get_buffers_multi(req, &rgroup->group, &rtransport->transport,
1812 			lengths, num_sgl_descriptors);
1813 	if (rc != 0) {
1814 		nvmf_rdma_request_free_data(rdma_req, rtransport);
1815 		return rc;
1816 	}
1817 
1818 	/* The first WR must always be the embedded data WR. This is how we unwind them later. */
1819 	current_wr = &rdma_req->data.wr;
1820 	assert(current_wr != NULL);
1821 
1822 	req->length = 0;
1823 	rdma_req->iovpos = 0;
1824 	desc = (struct spdk_nvme_sgl_descriptor *)rdma_req->recv->buf + inline_segment->address;
1825 	for (i = 0; i < num_sgl_descriptors; i++) {
1826 		/* The descriptors must be keyed data block descriptors with an address, not an offset. */
1827 		if (spdk_unlikely(desc->generic.type != SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK ||
1828 				  desc->keyed.subtype != SPDK_NVME_SGL_SUBTYPE_ADDRESS)) {
1829 			rc = -EINVAL;
1830 			goto err_exit;
1831 		}
1832 
1833 		current_wr->num_sge = 0;
1834 
1835 		rc = nvmf_rdma_fill_wr_sgl(rgroup, device, rdma_req, current_wr, lengths[i], 0);
1836 		if (rc != 0) {
1837 			rc = -ENOMEM;
1838 			goto err_exit;
1839 		}
1840 
1841 		req->length += desc->keyed.length;
1842 		current_wr->wr.rdma.rkey = desc->keyed.key;
1843 		current_wr->wr.rdma.remote_addr = desc->address;
1844 		current_wr = current_wr->next;
1845 		desc++;
1846 	}
1847 
1848 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
1849 	/* Go back to the last descriptor in the list. */
1850 	desc--;
1851 	if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) {
1852 		if (desc->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) {
1853 			rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV;
1854 			rdma_req->rsp.wr.imm_data = desc->keyed.key;
1855 		}
1856 	}
1857 #endif
1858 
1859 	rdma_req->num_outstanding_data_wr = num_sgl_descriptors;
1860 
1861 	return 0;
1862 
1863 err_exit:
1864 	spdk_nvmf_request_free_buffers(req, &rgroup->group, &rtransport->transport);
1865 	nvmf_rdma_request_free_data(rdma_req, rtransport);
1866 	return rc;
1867 }
1868 
1869 static int
1870 spdk_nvmf_rdma_request_parse_sgl(struct spdk_nvmf_rdma_transport *rtransport,
1871 				 struct spdk_nvmf_rdma_device *device,
1872 				 struct spdk_nvmf_rdma_request *rdma_req)
1873 {
1874 	struct spdk_nvmf_request		*req = &rdma_req->req;
1875 	struct spdk_nvme_cpl			*rsp;
1876 	struct spdk_nvme_sgl_descriptor		*sgl;
1877 	int					rc;
1878 	uint32_t				length;
1879 
1880 	rsp = &req->rsp->nvme_cpl;
1881 	sgl = &req->cmd->nvme_cmd.dptr.sgl1;
1882 
1883 	if (sgl->generic.type == SPDK_NVME_SGL_TYPE_KEYED_DATA_BLOCK &&
1884 	    (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_ADDRESS ||
1885 	     sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY)) {
1886 
1887 		length = sgl->keyed.length;
1888 		if (length > rtransport->transport.opts.max_io_size) {
1889 			SPDK_ERRLOG("SGL length 0x%x exceeds max io size 0x%x\n",
1890 				    length, rtransport->transport.opts.max_io_size);
1891 			rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
1892 			return -1;
1893 		}
1894 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
1895 		if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) != 0) {
1896 			if (sgl->keyed.subtype == SPDK_NVME_SGL_SUBTYPE_INVALIDATE_KEY) {
1897 				rdma_req->rsp.wr.opcode = IBV_WR_SEND_WITH_INV;
1898 				rdma_req->rsp.wr.imm_data = sgl->keyed.key;
1899 			}
1900 		}
1901 #endif
1902 
1903 		/* fill request length and populate iovs */
1904 		req->length = length;
1905 
1906 		if (spdk_unlikely(req->dif.dif_insert_or_strip)) {
1907 			req->dif.orig_length = length;
1908 			length = spdk_dif_get_length_with_md(length, &req->dif.dif_ctx);
1909 			req->dif.elba_length = length;
1910 		}
1911 
1912 		rc = spdk_nvmf_rdma_request_fill_iovs(rtransport, device, rdma_req, length);
1913 		if (spdk_unlikely(rc < 0)) {
1914 			if (rc == -EINVAL) {
1915 				SPDK_ERRLOG("SGL length exceeds the max I/O size\n");
1916 				rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
1917 				return -1;
1918 			}
1919 			/* No available buffers. Queue this request up. */
1920 			SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req);
1921 			return 0;
1922 		}
1923 
1924 		/* backward compatible */
1925 		req->data = req->iov[0].iov_base;
1926 
1927 		SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req,
1928 			      req->iovcnt);
1929 
1930 		return 0;
1931 	} else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_DATA_BLOCK &&
1932 		   sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) {
1933 		uint64_t offset = sgl->address;
1934 		uint32_t max_len = rtransport->transport.opts.in_capsule_data_size;
1935 
1936 		SPDK_DEBUGLOG(SPDK_LOG_NVMF, "In-capsule data: offset 0x%" PRIx64 ", length 0x%x\n",
1937 			      offset, sgl->unkeyed.length);
1938 
1939 		if (offset > max_len) {
1940 			SPDK_ERRLOG("In-capsule offset 0x%" PRIx64 " exceeds capsule length 0x%x\n",
1941 				    offset, max_len);
1942 			rsp->status.sc = SPDK_NVME_SC_INVALID_SGL_OFFSET;
1943 			return -1;
1944 		}
1945 		max_len -= (uint32_t)offset;
1946 
1947 		if (sgl->unkeyed.length > max_len) {
1948 			SPDK_ERRLOG("In-capsule data length 0x%x exceeds capsule length 0x%x\n",
1949 				    sgl->unkeyed.length, max_len);
1950 			rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
1951 			return -1;
1952 		}
1953 
1954 		rdma_req->num_outstanding_data_wr = 0;
1955 		req->data = rdma_req->recv->buf + offset;
1956 		req->data_from_pool = false;
1957 		req->length = sgl->unkeyed.length;
1958 
1959 		req->iov[0].iov_base = req->data;
1960 		req->iov[0].iov_len = req->length;
1961 		req->iovcnt = 1;
1962 
1963 		return 0;
1964 	} else if (sgl->generic.type == SPDK_NVME_SGL_TYPE_LAST_SEGMENT &&
1965 		   sgl->unkeyed.subtype == SPDK_NVME_SGL_SUBTYPE_OFFSET) {
1966 
1967 		rc = nvmf_rdma_request_fill_iovs_multi_sgl(rtransport, device, rdma_req);
1968 		if (rc == -ENOMEM) {
1969 			SPDK_DEBUGLOG(SPDK_LOG_RDMA, "No available large data buffers. Queueing request %p\n", rdma_req);
1970 			return 0;
1971 		} else if (rc == -EINVAL) {
1972 			SPDK_ERRLOG("Multi SGL element request length exceeds the max I/O size\n");
1973 			rsp->status.sc = SPDK_NVME_SC_DATA_SGL_LENGTH_INVALID;
1974 			return -1;
1975 		}
1976 
1977 		/* backward compatible */
1978 		req->data = req->iov[0].iov_base;
1979 
1980 		SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p took %d buffer/s from central pool\n", rdma_req,
1981 			      req->iovcnt);
1982 
1983 		return 0;
1984 	}
1985 
1986 	SPDK_ERRLOG("Invalid NVMf I/O Command SGL:  Type 0x%x, Subtype 0x%x\n",
1987 		    sgl->generic.type, sgl->generic.subtype);
1988 	rsp->status.sc = SPDK_NVME_SC_SGL_DESCRIPTOR_TYPE_INVALID;
1989 	return -1;
1990 }
1991 
1992 static void
1993 nvmf_rdma_request_free(struct spdk_nvmf_rdma_request *rdma_req,
1994 		       struct spdk_nvmf_rdma_transport	*rtransport)
1995 {
1996 	struct spdk_nvmf_rdma_qpair		*rqpair;
1997 	struct spdk_nvmf_rdma_poll_group	*rgroup;
1998 
1999 	rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
2000 	if (rdma_req->req.data_from_pool) {
2001 		rgroup = rqpair->poller->group;
2002 
2003 		spdk_nvmf_request_free_buffers(&rdma_req->req, &rgroup->group, &rtransport->transport);
2004 	}
2005 	nvmf_rdma_request_free_data(rdma_req, rtransport);
2006 	rdma_req->req.length = 0;
2007 	rdma_req->req.iovcnt = 0;
2008 	rdma_req->req.data = NULL;
2009 	rdma_req->rsp.wr.next = NULL;
2010 	rdma_req->data.wr.next = NULL;
2011 	memset(&rdma_req->req.dif, 0, sizeof(rdma_req->req.dif));
2012 	rqpair->qd--;
2013 
2014 	STAILQ_INSERT_HEAD(&rqpair->resources->free_queue, rdma_req, state_link);
2015 	rdma_req->state = RDMA_REQUEST_STATE_FREE;
2016 }
2017 
2018 bool
2019 spdk_nvmf_rdma_request_process(struct spdk_nvmf_rdma_transport *rtransport,
2020 			       struct spdk_nvmf_rdma_request *rdma_req)
2021 {
2022 	struct spdk_nvmf_rdma_qpair	*rqpair;
2023 	struct spdk_nvmf_rdma_device	*device;
2024 	struct spdk_nvmf_rdma_poll_group *rgroup;
2025 	struct spdk_nvme_cpl		*rsp = &rdma_req->req.rsp->nvme_cpl;
2026 	int				rc;
2027 	struct spdk_nvmf_rdma_recv	*rdma_recv;
2028 	enum spdk_nvmf_rdma_request_state prev_state;
2029 	bool				progress = false;
2030 	int				data_posted;
2031 	uint32_t			num_blocks;
2032 
2033 	rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
2034 	device = rqpair->device;
2035 	rgroup = rqpair->poller->group;
2036 
2037 	assert(rdma_req->state != RDMA_REQUEST_STATE_FREE);
2038 
2039 	/* If the queue pair is in an error state, force the request to the completed state
2040 	 * to release resources. */
2041 	if (rqpair->ibv_state == IBV_QPS_ERR || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) {
2042 		if (rdma_req->state == RDMA_REQUEST_STATE_NEED_BUFFER) {
2043 			STAILQ_REMOVE(&rgroup->group.pending_buf_queue, &rdma_req->req, spdk_nvmf_request, buf_link);
2044 		} else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING) {
2045 			STAILQ_REMOVE(&rqpair->pending_rdma_read_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
2046 		} else if (rdma_req->state == RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING) {
2047 			STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
2048 		}
2049 		rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
2050 	}
2051 
2052 	/* The loop here is to allow for several back-to-back state changes. */
2053 	do {
2054 		prev_state = rdma_req->state;
2055 
2056 		SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Request %p entering state %d\n", rdma_req, prev_state);
2057 
2058 		switch (rdma_req->state) {
2059 		case RDMA_REQUEST_STATE_FREE:
2060 			/* Some external code must kick a request into RDMA_REQUEST_STATE_NEW
2061 			 * to escape this state. */
2062 			break;
2063 		case RDMA_REQUEST_STATE_NEW:
2064 			spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEW, 0, 0,
2065 					  (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id);
2066 			rdma_recv = rdma_req->recv;
2067 
2068 			/* The first element of the SGL is the NVMe command */
2069 			rdma_req->req.cmd = (union nvmf_h2c_msg *)rdma_recv->sgl[0].addr;
2070 			memset(rdma_req->req.rsp, 0, sizeof(*rdma_req->req.rsp));
2071 
2072 			if (rqpair->ibv_state == IBV_QPS_ERR  || rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) {
2073 				rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
2074 				break;
2075 			}
2076 
2077 			if (spdk_unlikely(spdk_nvmf_request_get_dif_ctx(&rdma_req->req, &rdma_req->req.dif.dif_ctx))) {
2078 				rdma_req->req.dif.dif_insert_or_strip = true;
2079 			}
2080 
2081 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
2082 			rdma_req->rsp.wr.opcode = IBV_WR_SEND;
2083 			rdma_req->rsp.wr.imm_data = 0;
2084 #endif
2085 
2086 			/* The next state transition depends on the data transfer needs of this request. */
2087 			rdma_req->req.xfer = spdk_nvmf_req_get_xfer(&rdma_req->req);
2088 
2089 			/* If no data to transfer, ready to execute. */
2090 			if (rdma_req->req.xfer == SPDK_NVME_DATA_NONE) {
2091 				rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
2092 				break;
2093 			}
2094 
2095 			rdma_req->state = RDMA_REQUEST_STATE_NEED_BUFFER;
2096 			STAILQ_INSERT_TAIL(&rgroup->group.pending_buf_queue, &rdma_req->req, buf_link);
2097 			break;
2098 		case RDMA_REQUEST_STATE_NEED_BUFFER:
2099 			spdk_trace_record(TRACE_RDMA_REQUEST_STATE_NEED_BUFFER, 0, 0,
2100 					  (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id);
2101 
2102 			assert(rdma_req->req.xfer != SPDK_NVME_DATA_NONE);
2103 
2104 			if (&rdma_req->req != STAILQ_FIRST(&rgroup->group.pending_buf_queue)) {
2105 				/* This request needs to wait in line to obtain a buffer */
2106 				break;
2107 			}
2108 
2109 			/* Try to get a data buffer */
2110 			rc = spdk_nvmf_rdma_request_parse_sgl(rtransport, device, rdma_req);
2111 			if (rc < 0) {
2112 				STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link);
2113 				rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE;
2114 				break;
2115 			}
2116 
2117 			if (!rdma_req->req.data) {
2118 				/* No buffers available. */
2119 				rgroup->stat.pending_data_buffer++;
2120 				break;
2121 			}
2122 
2123 			STAILQ_REMOVE_HEAD(&rgroup->group.pending_buf_queue, buf_link);
2124 
2125 			/* If data is transferring from host to controller and the data didn't
2126 			 * arrive using in capsule data, we need to do a transfer from the host.
2127 			 */
2128 			if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER &&
2129 			    rdma_req->req.data_from_pool) {
2130 				STAILQ_INSERT_TAIL(&rqpair->pending_rdma_read_queue, rdma_req, state_link);
2131 				rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING;
2132 				break;
2133 			}
2134 
2135 			rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
2136 			break;
2137 		case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING:
2138 			spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_CONTROLLER_PENDING, 0, 0,
2139 					  (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id);
2140 
2141 			if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_read_queue)) {
2142 				/* This request needs to wait in line to perform RDMA */
2143 				break;
2144 			}
2145 			if (rqpair->current_send_depth + rdma_req->num_outstanding_data_wr > rqpair->max_send_depth
2146 			    || rqpair->current_read_depth + rdma_req->num_outstanding_data_wr > rqpair->max_read_depth) {
2147 				/* We can only have so many WRs outstanding. we have to wait until some finish. */
2148 				rqpair->poller->stat.pending_rdma_read++;
2149 				break;
2150 			}
2151 
2152 			/* We have already verified that this request is the head of the queue. */
2153 			STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_read_queue, state_link);
2154 
2155 			rc = request_transfer_in(&rdma_req->req);
2156 			if (!rc) {
2157 				rdma_req->state = RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER;
2158 			} else {
2159 				rsp->status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR;
2160 				rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE;
2161 			}
2162 			break;
2163 		case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
2164 			spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER, 0, 0,
2165 					  (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id);
2166 			/* Some external code must kick a request into RDMA_REQUEST_STATE_READY_TO_EXECUTE
2167 			 * to escape this state. */
2168 			break;
2169 		case RDMA_REQUEST_STATE_READY_TO_EXECUTE:
2170 			spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_EXECUTE, 0, 0,
2171 					  (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id);
2172 
2173 			if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) {
2174 				if (rdma_req->req.xfer == SPDK_NVME_DATA_HOST_TO_CONTROLLER) {
2175 					/* generate DIF for write operation */
2176 					num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size);
2177 					assert(num_blocks > 0);
2178 
2179 					rc = spdk_dif_generate(rdma_req->req.iov, rdma_req->req.iovcnt,
2180 							       num_blocks, &rdma_req->req.dif.dif_ctx);
2181 					if (rc != 0) {
2182 						SPDK_ERRLOG("DIF generation failed\n");
2183 						rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
2184 						spdk_nvmf_rdma_start_disconnect(rqpair);
2185 						break;
2186 					}
2187 				}
2188 
2189 				assert(rdma_req->req.dif.elba_length >= rdma_req->req.length);
2190 				/* set extended length before IO operation */
2191 				rdma_req->req.length = rdma_req->req.dif.elba_length;
2192 			}
2193 
2194 			rdma_req->state = RDMA_REQUEST_STATE_EXECUTING;
2195 			spdk_nvmf_request_exec(&rdma_req->req);
2196 			break;
2197 		case RDMA_REQUEST_STATE_EXECUTING:
2198 			spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTING, 0, 0,
2199 					  (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id);
2200 			/* Some external code must kick a request into RDMA_REQUEST_STATE_EXECUTED
2201 			 * to escape this state. */
2202 			break;
2203 		case RDMA_REQUEST_STATE_EXECUTED:
2204 			spdk_trace_record(TRACE_RDMA_REQUEST_STATE_EXECUTED, 0, 0,
2205 					  (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id);
2206 			if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
2207 				STAILQ_INSERT_TAIL(&rqpair->pending_rdma_write_queue, rdma_req, state_link);
2208 				rdma_req->state = RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING;
2209 			} else {
2210 				rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE;
2211 			}
2212 			if (spdk_unlikely(rdma_req->req.dif.dif_insert_or_strip)) {
2213 				/* restore the original length */
2214 				rdma_req->req.length = rdma_req->req.dif.orig_length;
2215 
2216 				if (rdma_req->req.xfer == SPDK_NVME_DATA_CONTROLLER_TO_HOST) {
2217 					struct spdk_dif_error error_blk;
2218 
2219 					num_blocks = SPDK_CEIL_DIV(rdma_req->req.dif.elba_length, rdma_req->req.dif.dif_ctx.block_size);
2220 
2221 					rc = spdk_dif_verify(rdma_req->req.iov, rdma_req->req.iovcnt, num_blocks,
2222 							     &rdma_req->req.dif.dif_ctx, &error_blk);
2223 					if (rc) {
2224 						struct spdk_nvme_cpl *rsp = &rdma_req->req.rsp->nvme_cpl;
2225 
2226 						SPDK_ERRLOG("DIF error detected. type=%d, offset=%" PRIu32 "\n", error_blk.err_type,
2227 							    error_blk.err_offset);
2228 						rsp->status.sct = SPDK_NVME_SCT_MEDIA_ERROR;
2229 						rsp->status.sc = spdk_nvmf_rdma_dif_error_to_compl_status(error_blk.err_type);
2230 						rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE;
2231 						STAILQ_REMOVE(&rqpair->pending_rdma_write_queue, rdma_req, spdk_nvmf_rdma_request, state_link);
2232 					}
2233 				}
2234 			}
2235 			break;
2236 		case RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING:
2237 			spdk_trace_record(TRACE_RDMA_REQUEST_STATE_DATA_TRANSFER_TO_HOST_PENDING, 0, 0,
2238 					  (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id);
2239 
2240 			if (rdma_req != STAILQ_FIRST(&rqpair->pending_rdma_write_queue)) {
2241 				/* This request needs to wait in line to perform RDMA */
2242 				break;
2243 			}
2244 			if ((rqpair->current_send_depth + rdma_req->num_outstanding_data_wr + 1) >
2245 			    rqpair->max_send_depth) {
2246 				/* We can only have so many WRs outstanding. we have to wait until some finish.
2247 				 * +1 since each request has an additional wr in the resp. */
2248 				rqpair->poller->stat.pending_rdma_write++;
2249 				break;
2250 			}
2251 
2252 			/* We have already verified that this request is the head of the queue. */
2253 			STAILQ_REMOVE_HEAD(&rqpair->pending_rdma_write_queue, state_link);
2254 
2255 			/* The data transfer will be kicked off from
2256 			 * RDMA_REQUEST_STATE_READY_TO_COMPLETE state.
2257 			 */
2258 			rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE;
2259 			break;
2260 		case RDMA_REQUEST_STATE_READY_TO_COMPLETE:
2261 			spdk_trace_record(TRACE_RDMA_REQUEST_STATE_READY_TO_COMPLETE, 0, 0,
2262 					  (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id);
2263 			rc = request_transfer_out(&rdma_req->req, &data_posted);
2264 			assert(rc == 0); /* No good way to handle this currently */
2265 			if (rc) {
2266 				rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
2267 			} else {
2268 				rdma_req->state = data_posted ? RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST :
2269 						  RDMA_REQUEST_STATE_COMPLETING;
2270 			}
2271 			break;
2272 		case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST:
2273 			spdk_trace_record(TRACE_RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST, 0, 0,
2274 					  (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id);
2275 			/* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED
2276 			 * to escape this state. */
2277 			break;
2278 		case RDMA_REQUEST_STATE_COMPLETING:
2279 			spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETING, 0, 0,
2280 					  (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id);
2281 			/* Some external code must kick a request into RDMA_REQUEST_STATE_COMPLETED
2282 			 * to escape this state. */
2283 			break;
2284 		case RDMA_REQUEST_STATE_COMPLETED:
2285 			spdk_trace_record(TRACE_RDMA_REQUEST_STATE_COMPLETED, 0, 0,
2286 					  (uintptr_t)rdma_req, (uintptr_t)rqpair->cm_id);
2287 
2288 			rqpair->poller->stat.request_latency += spdk_get_ticks() - rdma_req->receive_tsc;
2289 			nvmf_rdma_request_free(rdma_req, rtransport);
2290 			break;
2291 		case RDMA_REQUEST_NUM_STATES:
2292 		default:
2293 			assert(0);
2294 			break;
2295 		}
2296 
2297 		if (rdma_req->state != prev_state) {
2298 			progress = true;
2299 		}
2300 	} while (rdma_req->state != prev_state);
2301 
2302 	return progress;
2303 }
2304 
2305 /* Public API callbacks begin here */
2306 
2307 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH 128
2308 #define SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH 128
2309 #define SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH 4096
2310 #define SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR 128
2311 #define SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE 4096
2312 #define SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE 131072
2313 #define SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE (SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE / SPDK_NVMF_MAX_SGL_ENTRIES)
2314 #define SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS 4095
2315 #define SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE 32
2316 #define SPDK_NVMF_RDMA_DEFAULT_NO_SRQ false
2317 #define SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP false
2318 
2319 static void
2320 spdk_nvmf_rdma_opts_init(struct spdk_nvmf_transport_opts *opts)
2321 {
2322 	opts->max_queue_depth =		SPDK_NVMF_RDMA_DEFAULT_MAX_QUEUE_DEPTH;
2323 	opts->max_qpairs_per_ctrlr =	SPDK_NVMF_RDMA_DEFAULT_MAX_QPAIRS_PER_CTRLR;
2324 	opts->in_capsule_data_size =	SPDK_NVMF_RDMA_DEFAULT_IN_CAPSULE_DATA_SIZE;
2325 	opts->max_io_size =		SPDK_NVMF_RDMA_DEFAULT_MAX_IO_SIZE;
2326 	opts->io_unit_size =		SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE;
2327 	opts->max_aq_depth =		SPDK_NVMF_RDMA_DEFAULT_AQ_DEPTH;
2328 	opts->num_shared_buffers =	SPDK_NVMF_RDMA_DEFAULT_NUM_SHARED_BUFFERS;
2329 	opts->buf_cache_size =		SPDK_NVMF_RDMA_DEFAULT_BUFFER_CACHE_SIZE;
2330 	opts->max_srq_depth =		SPDK_NVMF_RDMA_DEFAULT_SRQ_DEPTH;
2331 	opts->no_srq =			SPDK_NVMF_RDMA_DEFAULT_NO_SRQ;
2332 	opts->dif_insert_or_strip =	SPDK_NVMF_RDMA_DIF_INSERT_OR_STRIP;
2333 }
2334 
2335 const struct spdk_mem_map_ops g_nvmf_rdma_map_ops = {
2336 	.notify_cb = spdk_nvmf_rdma_mem_notify,
2337 	.are_contiguous = spdk_nvmf_rdma_check_contiguous_entries
2338 };
2339 
2340 static int spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport);
2341 
2342 static struct spdk_nvmf_transport *
2343 spdk_nvmf_rdma_create(struct spdk_nvmf_transport_opts *opts)
2344 {
2345 	int rc;
2346 	struct spdk_nvmf_rdma_transport *rtransport;
2347 	struct spdk_nvmf_rdma_device	*device, *tmp;
2348 	struct ibv_context		**contexts;
2349 	uint32_t			i;
2350 	int				flag;
2351 	uint32_t			sge_count;
2352 	uint32_t			min_shared_buffers;
2353 	int				max_device_sge = SPDK_NVMF_MAX_SGL_ENTRIES;
2354 	pthread_mutexattr_t		attr;
2355 
2356 	rtransport = calloc(1, sizeof(*rtransport));
2357 	if (!rtransport) {
2358 		return NULL;
2359 	}
2360 
2361 	if (pthread_mutexattr_init(&attr)) {
2362 		SPDK_ERRLOG("pthread_mutexattr_init() failed\n");
2363 		free(rtransport);
2364 		return NULL;
2365 	}
2366 
2367 	if (pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE)) {
2368 		SPDK_ERRLOG("pthread_mutexattr_settype() failed\n");
2369 		pthread_mutexattr_destroy(&attr);
2370 		free(rtransport);
2371 		return NULL;
2372 	}
2373 
2374 	if (pthread_mutex_init(&rtransport->lock, &attr)) {
2375 		SPDK_ERRLOG("pthread_mutex_init() failed\n");
2376 		pthread_mutexattr_destroy(&attr);
2377 		free(rtransport);
2378 		return NULL;
2379 	}
2380 
2381 	pthread_mutexattr_destroy(&attr);
2382 
2383 	TAILQ_INIT(&rtransport->devices);
2384 	TAILQ_INIT(&rtransport->ports);
2385 	TAILQ_INIT(&rtransport->poll_groups);
2386 
2387 	rtransport->transport.ops = &spdk_nvmf_transport_rdma;
2388 
2389 	SPDK_INFOLOG(SPDK_LOG_RDMA, "*** RDMA Transport Init ***\n"
2390 		     "  Transport opts:  max_ioq_depth=%d, max_io_size=%d,\n"
2391 		     "  max_qpairs_per_ctrlr=%d, io_unit_size=%d,\n"
2392 		     "  in_capsule_data_size=%d, max_aq_depth=%d,\n"
2393 		     "  num_shared_buffers=%d, max_srq_depth=%d, no_srq=%d\n",
2394 		     opts->max_queue_depth,
2395 		     opts->max_io_size,
2396 		     opts->max_qpairs_per_ctrlr,
2397 		     opts->io_unit_size,
2398 		     opts->in_capsule_data_size,
2399 		     opts->max_aq_depth,
2400 		     opts->num_shared_buffers,
2401 		     opts->max_srq_depth,
2402 		     opts->no_srq);
2403 
2404 	/* I/O unit size cannot be larger than max I/O size */
2405 	if (opts->io_unit_size > opts->max_io_size) {
2406 		opts->io_unit_size = opts->max_io_size;
2407 	}
2408 
2409 	if (opts->num_shared_buffers < (SPDK_NVMF_MAX_SGL_ENTRIES * 2)) {
2410 		SPDK_ERRLOG("The number of shared data buffers (%d) is less than"
2411 			    "the minimum number required to guarantee that forward progress can be made (%d)\n",
2412 			    opts->num_shared_buffers, (SPDK_NVMF_MAX_SGL_ENTRIES * 2));
2413 		spdk_nvmf_rdma_destroy(&rtransport->transport);
2414 		return NULL;
2415 	}
2416 
2417 	min_shared_buffers = spdk_thread_get_count() * opts->buf_cache_size;
2418 	if (min_shared_buffers > opts->num_shared_buffers) {
2419 		SPDK_ERRLOG("There are not enough buffers to satisfy"
2420 			    "per-poll group caches for each thread. (%" PRIu32 ")"
2421 			    "supplied. (%" PRIu32 ") required\n", opts->num_shared_buffers, min_shared_buffers);
2422 		SPDK_ERRLOG("Please specify a larger number of shared buffers\n");
2423 		spdk_nvmf_rdma_destroy(&rtransport->transport);
2424 		return NULL;
2425 	}
2426 
2427 	sge_count = opts->max_io_size / opts->io_unit_size;
2428 	if (sge_count > NVMF_DEFAULT_TX_SGE) {
2429 		SPDK_ERRLOG("Unsupported IO Unit size specified, %d bytes\n", opts->io_unit_size);
2430 		spdk_nvmf_rdma_destroy(&rtransport->transport);
2431 		return NULL;
2432 	}
2433 
2434 	rtransport->event_channel = rdma_create_event_channel();
2435 	if (rtransport->event_channel == NULL) {
2436 		SPDK_ERRLOG("rdma_create_event_channel() failed, %s\n", spdk_strerror(errno));
2437 		spdk_nvmf_rdma_destroy(&rtransport->transport);
2438 		return NULL;
2439 	}
2440 
2441 	flag = fcntl(rtransport->event_channel->fd, F_GETFL);
2442 	if (fcntl(rtransport->event_channel->fd, F_SETFL, flag | O_NONBLOCK) < 0) {
2443 		SPDK_ERRLOG("fcntl can't set nonblocking mode for socket, fd: %d (%s)\n",
2444 			    rtransport->event_channel->fd, spdk_strerror(errno));
2445 		spdk_nvmf_rdma_destroy(&rtransport->transport);
2446 		return NULL;
2447 	}
2448 
2449 	rtransport->data_wr_pool = spdk_mempool_create("spdk_nvmf_rdma_wr_data",
2450 				   opts->max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES,
2451 				   sizeof(struct spdk_nvmf_rdma_request_data),
2452 				   SPDK_MEMPOOL_DEFAULT_CACHE_SIZE,
2453 				   SPDK_ENV_SOCKET_ID_ANY);
2454 	if (!rtransport->data_wr_pool) {
2455 		SPDK_ERRLOG("Unable to allocate work request pool for poll group\n");
2456 		spdk_nvmf_rdma_destroy(&rtransport->transport);
2457 		return NULL;
2458 	}
2459 
2460 	contexts = rdma_get_devices(NULL);
2461 	if (contexts == NULL) {
2462 		SPDK_ERRLOG("rdma_get_devices() failed: %s (%d)\n", spdk_strerror(errno), errno);
2463 		spdk_nvmf_rdma_destroy(&rtransport->transport);
2464 		return NULL;
2465 	}
2466 
2467 	i = 0;
2468 	rc = 0;
2469 	while (contexts[i] != NULL) {
2470 		device = calloc(1, sizeof(*device));
2471 		if (!device) {
2472 			SPDK_ERRLOG("Unable to allocate memory for RDMA devices.\n");
2473 			rc = -ENOMEM;
2474 			break;
2475 		}
2476 		device->context = contexts[i];
2477 		rc = ibv_query_device(device->context, &device->attr);
2478 		if (rc < 0) {
2479 			SPDK_ERRLOG("Failed to query RDMA device attributes.\n");
2480 			free(device);
2481 			break;
2482 
2483 		}
2484 
2485 		max_device_sge = spdk_min(max_device_sge, device->attr.max_sge);
2486 
2487 #ifdef SPDK_CONFIG_RDMA_SEND_WITH_INVAL
2488 		if ((device->attr.device_cap_flags & IBV_DEVICE_MEM_MGT_EXTENSIONS) == 0) {
2489 			SPDK_WARNLOG("The libibverbs on this system supports SEND_WITH_INVALIDATE,");
2490 			SPDK_WARNLOG("but the device with vendor ID %u does not.\n", device->attr.vendor_id);
2491 		}
2492 
2493 		/**
2494 		 * The vendor ID is assigned by the IEEE and an ID of 0 implies Soft-RoCE.
2495 		 * The Soft-RoCE RXE driver does not currently support send with invalidate,
2496 		 * but incorrectly reports that it does. There are changes making their way
2497 		 * through the kernel now that will enable this feature. When they are merged,
2498 		 * we can conditionally enable this feature.
2499 		 *
2500 		 * TODO: enable this for versions of the kernel rxe driver that support it.
2501 		 */
2502 		if (device->attr.vendor_id == 0) {
2503 			device->attr.device_cap_flags &= ~(IBV_DEVICE_MEM_MGT_EXTENSIONS);
2504 		}
2505 #endif
2506 
2507 		/* set up device context async ev fd as NON_BLOCKING */
2508 		flag = fcntl(device->context->async_fd, F_GETFL);
2509 		rc = fcntl(device->context->async_fd, F_SETFL, flag | O_NONBLOCK);
2510 		if (rc < 0) {
2511 			SPDK_ERRLOG("Failed to set context async fd to NONBLOCK.\n");
2512 			free(device);
2513 			break;
2514 		}
2515 
2516 		TAILQ_INSERT_TAIL(&rtransport->devices, device, link);
2517 		i++;
2518 
2519 		if (g_nvmf_hooks.get_ibv_pd) {
2520 			device->pd = g_nvmf_hooks.get_ibv_pd(NULL, device->context);
2521 		} else {
2522 			device->pd = ibv_alloc_pd(device->context);
2523 		}
2524 
2525 		if (!device->pd) {
2526 			SPDK_ERRLOG("Unable to allocate protection domain.\n");
2527 			rc = -ENOMEM;
2528 			break;
2529 		}
2530 
2531 		assert(device->map == NULL);
2532 
2533 		device->map = spdk_mem_map_alloc(0, &g_nvmf_rdma_map_ops, device->pd);
2534 		if (!device->map) {
2535 			SPDK_ERRLOG("Unable to allocate memory map for listen address\n");
2536 			rc = -ENOMEM;
2537 			break;
2538 		}
2539 
2540 		assert(device->map != NULL);
2541 		assert(device->pd != NULL);
2542 	}
2543 	rdma_free_devices(contexts);
2544 
2545 	if (opts->io_unit_size * max_device_sge < opts->max_io_size) {
2546 		/* divide and round up. */
2547 		opts->io_unit_size = (opts->max_io_size + max_device_sge - 1) / max_device_sge;
2548 
2549 		/* round up to the nearest 4k. */
2550 		opts->io_unit_size = (opts->io_unit_size + NVMF_DATA_BUFFER_ALIGNMENT - 1) & ~NVMF_DATA_BUFFER_MASK;
2551 
2552 		opts->io_unit_size = spdk_max(opts->io_unit_size, SPDK_NVMF_RDMA_MIN_IO_BUFFER_SIZE);
2553 		SPDK_NOTICELOG("Adjusting the io unit size to fit the device's maximum I/O size. New I/O unit size %u\n",
2554 			       opts->io_unit_size);
2555 	}
2556 
2557 	if (rc < 0) {
2558 		spdk_nvmf_rdma_destroy(&rtransport->transport);
2559 		return NULL;
2560 	}
2561 
2562 	/* Set up poll descriptor array to monitor events from RDMA and IB
2563 	 * in a single poll syscall
2564 	 */
2565 	rtransport->npoll_fds = i + 1;
2566 	i = 0;
2567 	rtransport->poll_fds = calloc(rtransport->npoll_fds, sizeof(struct pollfd));
2568 	if (rtransport->poll_fds == NULL) {
2569 		SPDK_ERRLOG("poll_fds allocation failed\n");
2570 		spdk_nvmf_rdma_destroy(&rtransport->transport);
2571 		return NULL;
2572 	}
2573 
2574 	rtransport->poll_fds[i].fd = rtransport->event_channel->fd;
2575 	rtransport->poll_fds[i++].events = POLLIN;
2576 
2577 	TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) {
2578 		rtransport->poll_fds[i].fd = device->context->async_fd;
2579 		rtransport->poll_fds[i++].events = POLLIN;
2580 	}
2581 
2582 	return &rtransport->transport;
2583 }
2584 
2585 static int
2586 spdk_nvmf_rdma_destroy(struct spdk_nvmf_transport *transport)
2587 {
2588 	struct spdk_nvmf_rdma_transport	*rtransport;
2589 	struct spdk_nvmf_rdma_port	*port, *port_tmp;
2590 	struct spdk_nvmf_rdma_device	*device, *device_tmp;
2591 
2592 	rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
2593 
2594 	TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, port_tmp) {
2595 		TAILQ_REMOVE(&rtransport->ports, port, link);
2596 		rdma_destroy_id(port->id);
2597 		free(port);
2598 	}
2599 
2600 	if (rtransport->poll_fds != NULL) {
2601 		free(rtransport->poll_fds);
2602 	}
2603 
2604 	if (rtransport->event_channel != NULL) {
2605 		rdma_destroy_event_channel(rtransport->event_channel);
2606 	}
2607 
2608 	TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, device_tmp) {
2609 		TAILQ_REMOVE(&rtransport->devices, device, link);
2610 		if (device->map) {
2611 			spdk_mem_map_free(&device->map);
2612 		}
2613 		if (device->pd) {
2614 			if (!g_nvmf_hooks.get_ibv_pd) {
2615 				ibv_dealloc_pd(device->pd);
2616 			}
2617 		}
2618 		free(device);
2619 	}
2620 
2621 	if (rtransport->data_wr_pool != NULL) {
2622 		if (spdk_mempool_count(rtransport->data_wr_pool) !=
2623 		    (transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES)) {
2624 			SPDK_ERRLOG("transport wr pool count is %zu but should be %u\n",
2625 				    spdk_mempool_count(rtransport->data_wr_pool),
2626 				    transport->opts.max_queue_depth * SPDK_NVMF_MAX_SGL_ENTRIES);
2627 		}
2628 	}
2629 
2630 	spdk_mempool_free(rtransport->data_wr_pool);
2631 
2632 	pthread_mutex_destroy(&rtransport->lock);
2633 	free(rtransport);
2634 
2635 	return 0;
2636 }
2637 
2638 static int
2639 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id,
2640 			       struct spdk_nvme_transport_id *trid,
2641 			       bool peer);
2642 
2643 static int
2644 spdk_nvmf_rdma_listen(struct spdk_nvmf_transport *transport,
2645 		      const struct spdk_nvme_transport_id *trid)
2646 {
2647 	struct spdk_nvmf_rdma_transport	*rtransport;
2648 	struct spdk_nvmf_rdma_device	*device;
2649 	struct spdk_nvmf_rdma_port	*port;
2650 	struct addrinfo			*res;
2651 	struct addrinfo			hints;
2652 	int				family;
2653 	int				rc;
2654 
2655 	rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
2656 	assert(rtransport->event_channel != NULL);
2657 
2658 	pthread_mutex_lock(&rtransport->lock);
2659 	port = calloc(1, sizeof(*port));
2660 	if (!port) {
2661 		SPDK_ERRLOG("Port allocation failed\n");
2662 		pthread_mutex_unlock(&rtransport->lock);
2663 		return -ENOMEM;
2664 	}
2665 
2666 	port->trid = trid;
2667 
2668 	switch (trid->adrfam) {
2669 	case SPDK_NVMF_ADRFAM_IPV4:
2670 		family = AF_INET;
2671 		break;
2672 	case SPDK_NVMF_ADRFAM_IPV6:
2673 		family = AF_INET6;
2674 		break;
2675 	default:
2676 		SPDK_ERRLOG("Unhandled ADRFAM %d\n", trid->adrfam);
2677 		free(port);
2678 		pthread_mutex_unlock(&rtransport->lock);
2679 		return -EINVAL;
2680 	}
2681 
2682 	memset(&hints, 0, sizeof(hints));
2683 	hints.ai_family = family;
2684 	hints.ai_flags = AI_NUMERICSERV;
2685 	hints.ai_socktype = SOCK_STREAM;
2686 	hints.ai_protocol = 0;
2687 
2688 	rc = getaddrinfo(trid->traddr, trid->trsvcid, &hints, &res);
2689 	if (rc) {
2690 		SPDK_ERRLOG("getaddrinfo failed: %s (%d)\n", gai_strerror(rc), rc);
2691 		free(port);
2692 		pthread_mutex_unlock(&rtransport->lock);
2693 		return -EINVAL;
2694 	}
2695 
2696 	rc = rdma_create_id(rtransport->event_channel, &port->id, port, RDMA_PS_TCP);
2697 	if (rc < 0) {
2698 		SPDK_ERRLOG("rdma_create_id() failed\n");
2699 		freeaddrinfo(res);
2700 		free(port);
2701 		pthread_mutex_unlock(&rtransport->lock);
2702 		return rc;
2703 	}
2704 
2705 	rc = rdma_bind_addr(port->id, res->ai_addr);
2706 	freeaddrinfo(res);
2707 
2708 	if (rc < 0) {
2709 		SPDK_ERRLOG("rdma_bind_addr() failed\n");
2710 		rdma_destroy_id(port->id);
2711 		free(port);
2712 		pthread_mutex_unlock(&rtransport->lock);
2713 		return rc;
2714 	}
2715 
2716 	if (!port->id->verbs) {
2717 		SPDK_ERRLOG("ibv_context is null\n");
2718 		rdma_destroy_id(port->id);
2719 		free(port);
2720 		pthread_mutex_unlock(&rtransport->lock);
2721 		return -1;
2722 	}
2723 
2724 	rc = rdma_listen(port->id, 10); /* 10 = backlog */
2725 	if (rc < 0) {
2726 		SPDK_ERRLOG("rdma_listen() failed\n");
2727 		rdma_destroy_id(port->id);
2728 		free(port);
2729 		pthread_mutex_unlock(&rtransport->lock);
2730 		return rc;
2731 	}
2732 
2733 	TAILQ_FOREACH(device, &rtransport->devices, link) {
2734 		if (device->context == port->id->verbs) {
2735 			port->device = device;
2736 			break;
2737 		}
2738 	}
2739 	if (!port->device) {
2740 		SPDK_ERRLOG("Accepted a connection with verbs %p, but unable to find a corresponding device.\n",
2741 			    port->id->verbs);
2742 		rdma_destroy_id(port->id);
2743 		free(port);
2744 		pthread_mutex_unlock(&rtransport->lock);
2745 		return -EINVAL;
2746 	}
2747 
2748 	SPDK_NOTICELOG("*** NVMe/RDMA Target Listening on %s port %s ***\n",
2749 		       trid->traddr, trid->trsvcid);
2750 
2751 	TAILQ_INSERT_TAIL(&rtransport->ports, port, link);
2752 	pthread_mutex_unlock(&rtransport->lock);
2753 	return 0;
2754 }
2755 
2756 static void
2757 spdk_nvmf_rdma_stop_listen(struct spdk_nvmf_transport *transport,
2758 			   const struct spdk_nvme_transport_id *trid)
2759 {
2760 	struct spdk_nvmf_rdma_transport *rtransport;
2761 	struct spdk_nvmf_rdma_port *port, *tmp;
2762 
2763 	rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
2764 
2765 	pthread_mutex_lock(&rtransport->lock);
2766 	TAILQ_FOREACH_SAFE(port, &rtransport->ports, link, tmp) {
2767 		if (spdk_nvme_transport_id_compare(port->trid, trid) == 0) {
2768 			TAILQ_REMOVE(&rtransport->ports, port, link);
2769 			rdma_destroy_id(port->id);
2770 			free(port);
2771 			break;
2772 		}
2773 	}
2774 
2775 	pthread_mutex_unlock(&rtransport->lock);
2776 }
2777 
2778 static void
2779 spdk_nvmf_rdma_qpair_process_pending(struct spdk_nvmf_rdma_transport *rtransport,
2780 				     struct spdk_nvmf_rdma_qpair *rqpair, bool drain)
2781 {
2782 	struct spdk_nvmf_request *req, *tmp;
2783 	struct spdk_nvmf_rdma_request	*rdma_req, *req_tmp;
2784 	struct spdk_nvmf_rdma_resources *resources;
2785 
2786 	/* We process I/O in the data transfer pending queue at the highest priority. RDMA reads first */
2787 	STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_read_queue, state_link, req_tmp) {
2788 		if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
2789 			break;
2790 		}
2791 	}
2792 
2793 	/* Then RDMA writes since reads have stronger restrictions than writes */
2794 	STAILQ_FOREACH_SAFE(rdma_req, &rqpair->pending_rdma_write_queue, state_link, req_tmp) {
2795 		if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
2796 			break;
2797 		}
2798 	}
2799 
2800 	/* The second highest priority is I/O waiting on memory buffers. */
2801 	STAILQ_FOREACH_SAFE(req, &rqpair->poller->group->group.pending_buf_queue, buf_link, tmp) {
2802 		rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
2803 		if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false && drain == false) {
2804 			break;
2805 		}
2806 	}
2807 
2808 	resources = rqpair->resources;
2809 	while (!STAILQ_EMPTY(&resources->free_queue) && !STAILQ_EMPTY(&resources->incoming_queue)) {
2810 		rdma_req = STAILQ_FIRST(&resources->free_queue);
2811 		STAILQ_REMOVE_HEAD(&resources->free_queue, state_link);
2812 		rdma_req->recv = STAILQ_FIRST(&resources->incoming_queue);
2813 		STAILQ_REMOVE_HEAD(&resources->incoming_queue, link);
2814 
2815 		if (rqpair->srq != NULL) {
2816 			rdma_req->req.qpair = &rdma_req->recv->qpair->qpair;
2817 			rdma_req->recv->qpair->qd++;
2818 		} else {
2819 			rqpair->qd++;
2820 		}
2821 
2822 		rdma_req->receive_tsc = rdma_req->recv->receive_tsc;
2823 		rdma_req->state = RDMA_REQUEST_STATE_NEW;
2824 		if (spdk_nvmf_rdma_request_process(rtransport, rdma_req) == false) {
2825 			break;
2826 		}
2827 	}
2828 	if (!STAILQ_EMPTY(&resources->incoming_queue) && STAILQ_EMPTY(&resources->free_queue)) {
2829 		rqpair->poller->stat.pending_free_request++;
2830 	}
2831 }
2832 
2833 static void
2834 _nvmf_rdma_qpair_disconnect(void *ctx)
2835 {
2836 	struct spdk_nvmf_qpair *qpair = ctx;
2837 
2838 	spdk_nvmf_qpair_disconnect(qpair, NULL, NULL);
2839 }
2840 
2841 static void
2842 _nvmf_rdma_try_disconnect(void *ctx)
2843 {
2844 	struct spdk_nvmf_qpair *qpair = ctx;
2845 	struct spdk_nvmf_poll_group *group;
2846 
2847 	/* Read the group out of the qpair. This is normally set and accessed only from
2848 	 * the thread that created the group. Here, we're not on that thread necessarily.
2849 	 * The data member qpair->group begins it's life as NULL and then is assigned to
2850 	 * a pointer and never changes. So fortunately reading this and checking for
2851 	 * non-NULL is thread safe in the x86_64 memory model. */
2852 	group = qpair->group;
2853 
2854 	if (group == NULL) {
2855 		/* The qpair hasn't been assigned to a group yet, so we can't
2856 		 * process a disconnect. Send a message to ourself and try again. */
2857 		spdk_thread_send_msg(spdk_get_thread(), _nvmf_rdma_try_disconnect, qpair);
2858 		return;
2859 	}
2860 
2861 	spdk_thread_send_msg(group->thread, _nvmf_rdma_qpair_disconnect, qpair);
2862 }
2863 
2864 static inline void
2865 spdk_nvmf_rdma_start_disconnect(struct spdk_nvmf_rdma_qpair *rqpair)
2866 {
2867 	if (!__atomic_test_and_set(&rqpair->disconnect_started, __ATOMIC_RELAXED)) {
2868 		_nvmf_rdma_try_disconnect(&rqpair->qpair);
2869 	}
2870 }
2871 
2872 static void nvmf_rdma_destroy_drained_qpair(void *ctx)
2873 {
2874 	struct spdk_nvmf_rdma_qpair *rqpair = ctx;
2875 	struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
2876 			struct spdk_nvmf_rdma_transport, transport);
2877 
2878 	/* In non SRQ path, we will reach rqpair->max_queue_depth. In SRQ path, we will get the last_wqe event. */
2879 	if (rqpair->current_send_depth != 0) {
2880 		return;
2881 	}
2882 
2883 	if (rqpair->srq == NULL && rqpair->current_recv_depth != rqpair->max_queue_depth) {
2884 		return;
2885 	}
2886 
2887 	if (rqpair->srq != NULL && rqpair->last_wqe_reached == false) {
2888 		return;
2889 	}
2890 
2891 	spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true);
2892 
2893 	/* Qpair will be destroyed after nvmf layer closes this qpair */
2894 	if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ERROR) {
2895 		return;
2896 	}
2897 
2898 	spdk_nvmf_rdma_qpair_destroy(rqpair);
2899 }
2900 
2901 
2902 static int
2903 nvmf_rdma_disconnect(struct rdma_cm_event *evt)
2904 {
2905 	struct spdk_nvmf_qpair		*qpair;
2906 	struct spdk_nvmf_rdma_qpair	*rqpair;
2907 
2908 	if (evt->id == NULL) {
2909 		SPDK_ERRLOG("disconnect request: missing cm_id\n");
2910 		return -1;
2911 	}
2912 
2913 	qpair = evt->id->context;
2914 	if (qpair == NULL) {
2915 		SPDK_ERRLOG("disconnect request: no active connection\n");
2916 		return -1;
2917 	}
2918 
2919 	rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
2920 
2921 	spdk_trace_record(TRACE_RDMA_QP_DISCONNECT, 0, 0, (uintptr_t)rqpair->cm_id, 0);
2922 
2923 	spdk_nvmf_rdma_start_disconnect(rqpair);
2924 
2925 	return 0;
2926 }
2927 
2928 #ifdef DEBUG
2929 static const char *CM_EVENT_STR[] = {
2930 	"RDMA_CM_EVENT_ADDR_RESOLVED",
2931 	"RDMA_CM_EVENT_ADDR_ERROR",
2932 	"RDMA_CM_EVENT_ROUTE_RESOLVED",
2933 	"RDMA_CM_EVENT_ROUTE_ERROR",
2934 	"RDMA_CM_EVENT_CONNECT_REQUEST",
2935 	"RDMA_CM_EVENT_CONNECT_RESPONSE",
2936 	"RDMA_CM_EVENT_CONNECT_ERROR",
2937 	"RDMA_CM_EVENT_UNREACHABLE",
2938 	"RDMA_CM_EVENT_REJECTED",
2939 	"RDMA_CM_EVENT_ESTABLISHED",
2940 	"RDMA_CM_EVENT_DISCONNECTED",
2941 	"RDMA_CM_EVENT_DEVICE_REMOVAL",
2942 	"RDMA_CM_EVENT_MULTICAST_JOIN",
2943 	"RDMA_CM_EVENT_MULTICAST_ERROR",
2944 	"RDMA_CM_EVENT_ADDR_CHANGE",
2945 	"RDMA_CM_EVENT_TIMEWAIT_EXIT"
2946 };
2947 #endif /* DEBUG */
2948 
2949 static void
2950 nvmf_rdma_disconnect_qpairs_on_port(struct spdk_nvmf_rdma_transport *rtransport,
2951 				    struct spdk_nvmf_rdma_port *port)
2952 {
2953 	struct spdk_nvmf_rdma_poll_group	*rgroup;
2954 	struct spdk_nvmf_rdma_poller		*rpoller;
2955 	struct spdk_nvmf_rdma_qpair		*rqpair;
2956 
2957 	TAILQ_FOREACH(rgroup, &rtransport->poll_groups, link) {
2958 		TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
2959 			TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) {
2960 				if (rqpair->listen_id == port->id) {
2961 					spdk_nvmf_rdma_start_disconnect(rqpair);
2962 				}
2963 			}
2964 		}
2965 	}
2966 }
2967 
2968 static bool
2969 nvmf_rdma_handle_cm_event_addr_change(struct spdk_nvmf_transport *transport,
2970 				      struct rdma_cm_event *event)
2971 {
2972 	const struct spdk_nvme_transport_id	*trid;
2973 	struct spdk_nvmf_rdma_port		*port;
2974 	struct spdk_nvmf_rdma_transport		*rtransport;
2975 	bool					event_acked = false;
2976 
2977 	rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
2978 	TAILQ_FOREACH(port, &rtransport->ports, link) {
2979 		if (port->id == event->id) {
2980 			SPDK_ERRLOG("ADDR_CHANGE: IP %s:%s migrated\n", port->trid->traddr, port->trid->trsvcid);
2981 			rdma_ack_cm_event(event);
2982 			event_acked = true;
2983 			trid = port->trid;
2984 			break;
2985 		}
2986 	}
2987 
2988 	if (event_acked) {
2989 		nvmf_rdma_disconnect_qpairs_on_port(rtransport, port);
2990 
2991 		spdk_nvmf_rdma_stop_listen(transport, trid);
2992 		spdk_nvmf_rdma_listen(transport, trid);
2993 	}
2994 
2995 	return event_acked;
2996 }
2997 
2998 static void
2999 nvmf_rdma_handle_cm_event_port_removal(struct spdk_nvmf_transport *transport,
3000 				       struct rdma_cm_event *event)
3001 {
3002 	struct spdk_nvmf_rdma_port		*port;
3003 	struct spdk_nvmf_rdma_transport		*rtransport;
3004 
3005 	port = event->id->context;
3006 	rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3007 
3008 	SPDK_NOTICELOG("Port %s:%s is being removed\n", port->trid->traddr, port->trid->trsvcid);
3009 
3010 	nvmf_rdma_disconnect_qpairs_on_port(rtransport, port);
3011 
3012 	rdma_ack_cm_event(event);
3013 
3014 	while (spdk_nvmf_transport_stop_listen(transport, port->trid) == 0) {
3015 		;
3016 	}
3017 }
3018 
3019 static void
3020 spdk_nvmf_process_cm_event(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn, void *cb_arg)
3021 {
3022 	struct spdk_nvmf_rdma_transport *rtransport;
3023 	struct rdma_cm_event		*event;
3024 	int				rc;
3025 	bool				event_acked;
3026 
3027 	rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3028 
3029 	if (rtransport->event_channel == NULL) {
3030 		return;
3031 	}
3032 
3033 	while (1) {
3034 		event_acked = false;
3035 		rc = rdma_get_cm_event(rtransport->event_channel, &event);
3036 		if (rc) {
3037 			if (errno != EAGAIN && errno != EWOULDBLOCK) {
3038 				SPDK_ERRLOG("Acceptor Event Error: %s\n", spdk_strerror(errno));
3039 			}
3040 			break;
3041 		}
3042 
3043 		SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Acceptor Event: %s\n", CM_EVENT_STR[event->event]);
3044 
3045 		spdk_trace_record(TRACE_RDMA_CM_ASYNC_EVENT, 0, 0, 0, event->event);
3046 
3047 		switch (event->event) {
3048 		case RDMA_CM_EVENT_ADDR_RESOLVED:
3049 		case RDMA_CM_EVENT_ADDR_ERROR:
3050 		case RDMA_CM_EVENT_ROUTE_RESOLVED:
3051 		case RDMA_CM_EVENT_ROUTE_ERROR:
3052 			/* No action required. The target never attempts to resolve routes. */
3053 			break;
3054 		case RDMA_CM_EVENT_CONNECT_REQUEST:
3055 			rc = nvmf_rdma_connect(transport, event, cb_fn, cb_arg);
3056 			if (rc < 0) {
3057 				SPDK_ERRLOG("Unable to process connect event. rc: %d\n", rc);
3058 				break;
3059 			}
3060 			break;
3061 		case RDMA_CM_EVENT_CONNECT_RESPONSE:
3062 			/* The target never initiates a new connection. So this will not occur. */
3063 			break;
3064 		case RDMA_CM_EVENT_CONNECT_ERROR:
3065 			/* Can this happen? The docs say it can, but not sure what causes it. */
3066 			break;
3067 		case RDMA_CM_EVENT_UNREACHABLE:
3068 		case RDMA_CM_EVENT_REJECTED:
3069 			/* These only occur on the client side. */
3070 			break;
3071 		case RDMA_CM_EVENT_ESTABLISHED:
3072 			/* TODO: Should we be waiting for this event anywhere? */
3073 			break;
3074 		case RDMA_CM_EVENT_DISCONNECTED:
3075 			rc = nvmf_rdma_disconnect(event);
3076 			if (rc < 0) {
3077 				SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc);
3078 				break;
3079 			}
3080 			break;
3081 		case RDMA_CM_EVENT_DEVICE_REMOVAL:
3082 			/* In case of device removal, kernel IB part triggers IBV_EVENT_DEVICE_FATAL
3083 			 * which triggers RDMA_CM_EVENT_DEVICE_REMOVAL on all cma_id’s.
3084 			 * Once these events are sent to SPDK, we should release all IB resources and
3085 			 * don't make attempts to call any ibv_query/modify/create functions. We can only call
3086 			 * ibv_destory* functions to release user space memory allocated by IB. All kernel
3087 			 * resources are already cleaned. */
3088 			if (event->id->qp) {
3089 				/* If rdma_cm event has a valid `qp` pointer then the event refers to the
3090 				 * corresponding qpair. Otherwise the event refers to a listening device */
3091 				rc = nvmf_rdma_disconnect(event);
3092 				if (rc < 0) {
3093 					SPDK_ERRLOG("Unable to process disconnect event. rc: %d\n", rc);
3094 					break;
3095 				}
3096 			} else {
3097 				nvmf_rdma_handle_cm_event_port_removal(transport, event);
3098 				event_acked = true;
3099 			}
3100 			break;
3101 		case RDMA_CM_EVENT_MULTICAST_JOIN:
3102 		case RDMA_CM_EVENT_MULTICAST_ERROR:
3103 			/* Multicast is not used */
3104 			break;
3105 		case RDMA_CM_EVENT_ADDR_CHANGE:
3106 			event_acked = nvmf_rdma_handle_cm_event_addr_change(transport, event);
3107 			break;
3108 		case RDMA_CM_EVENT_TIMEWAIT_EXIT:
3109 			/* For now, do nothing. The target never re-uses queue pairs. */
3110 			break;
3111 		default:
3112 			SPDK_ERRLOG("Unexpected Acceptor Event [%d]\n", event->event);
3113 			break;
3114 		}
3115 		if (!event_acked) {
3116 			rdma_ack_cm_event(event);
3117 		}
3118 	}
3119 }
3120 
3121 static void
3122 nvmf_rdma_handle_qp_fatal(struct spdk_nvmf_rdma_qpair *rqpair)
3123 {
3124 	spdk_nvmf_rdma_update_ibv_state(rqpair);
3125 	spdk_nvmf_rdma_start_disconnect(rqpair);
3126 }
3127 
3128 static void
3129 nvmf_rdma_handle_last_wqe_reached(struct spdk_nvmf_rdma_qpair *rqpair)
3130 {
3131 	rqpair->last_wqe_reached = true;
3132 	nvmf_rdma_destroy_drained_qpair(rqpair);
3133 }
3134 
3135 static void
3136 nvmf_rdma_handle_sq_drained(struct spdk_nvmf_rdma_qpair *rqpair)
3137 {
3138 	spdk_nvmf_rdma_start_disconnect(rqpair);
3139 }
3140 
3141 static void
3142 spdk_nvmf_rdma_qpair_process_ibv_event(void *ctx)
3143 {
3144 	struct spdk_nvmf_rdma_ibv_event_ctx *event_ctx = ctx;
3145 
3146 	if (event_ctx->rqpair) {
3147 		STAILQ_REMOVE(&event_ctx->rqpair->ibv_events, event_ctx, spdk_nvmf_rdma_ibv_event_ctx, link);
3148 		if (event_ctx->cb_fn) {
3149 			event_ctx->cb_fn(event_ctx->rqpair);
3150 		}
3151 	}
3152 	free(event_ctx);
3153 }
3154 
3155 static int
3156 spdk_nvmf_rdma_send_qpair_async_event(struct spdk_nvmf_rdma_qpair *rqpair,
3157 				      spdk_nvmf_rdma_qpair_ibv_event fn)
3158 {
3159 	struct spdk_nvmf_rdma_ibv_event_ctx *ctx;
3160 
3161 	if (!rqpair->qpair.group) {
3162 		return EINVAL;
3163 	}
3164 
3165 	ctx = calloc(1, sizeof(*ctx));
3166 	if (!ctx) {
3167 		return ENOMEM;
3168 	}
3169 
3170 	ctx->rqpair = rqpair;
3171 	ctx->cb_fn = fn;
3172 	STAILQ_INSERT_TAIL(&rqpair->ibv_events, ctx, link);
3173 
3174 	return spdk_thread_send_msg(rqpair->qpair.group->thread, spdk_nvmf_rdma_qpair_process_ibv_event,
3175 				    ctx);
3176 }
3177 
3178 static void
3179 spdk_nvmf_process_ib_event(struct spdk_nvmf_rdma_device *device)
3180 {
3181 	int				rc;
3182 	struct spdk_nvmf_rdma_qpair	*rqpair = NULL;
3183 	struct ibv_async_event		event;
3184 
3185 	rc = ibv_get_async_event(device->context, &event);
3186 
3187 	if (rc) {
3188 		SPDK_ERRLOG("Failed to get async_event (%d): %s\n",
3189 			    errno, spdk_strerror(errno));
3190 		return;
3191 	}
3192 
3193 	switch (event.event_type) {
3194 	case IBV_EVENT_QP_FATAL:
3195 		rqpair = event.element.qp->qp_context;
3196 		SPDK_ERRLOG("Fatal event received for rqpair %p\n", rqpair);
3197 		spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0,
3198 				  (uintptr_t)rqpair->cm_id, event.event_type);
3199 		if (spdk_nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_qp_fatal)) {
3200 			SPDK_ERRLOG("Failed to send QP_FATAL event for rqpair %p\n", rqpair);
3201 			nvmf_rdma_handle_qp_fatal(rqpair);
3202 		}
3203 		break;
3204 	case IBV_EVENT_QP_LAST_WQE_REACHED:
3205 		/* This event only occurs for shared receive queues. */
3206 		rqpair = event.element.qp->qp_context;
3207 		SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Last WQE reached event received for rqpair %p\n", rqpair);
3208 		if (spdk_nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_last_wqe_reached)) {
3209 			SPDK_ERRLOG("Failed to send LAST_WQE_REACHED event for rqpair %p\n", rqpair);
3210 			rqpair->last_wqe_reached = true;
3211 		}
3212 		break;
3213 	case IBV_EVENT_SQ_DRAINED:
3214 		/* This event occurs frequently in both error and non-error states.
3215 		 * Check if the qpair is in an error state before sending a message. */
3216 		rqpair = event.element.qp->qp_context;
3217 		SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Last sq drained event received for rqpair %p\n", rqpair);
3218 		spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0,
3219 				  (uintptr_t)rqpair->cm_id, event.event_type);
3220 		if (spdk_nvmf_rdma_update_ibv_state(rqpair) == IBV_QPS_ERR) {
3221 			if (spdk_nvmf_rdma_send_qpair_async_event(rqpair, nvmf_rdma_handle_sq_drained)) {
3222 				SPDK_ERRLOG("Failed to send SQ_DRAINED event for rqpair %p\n", rqpair);
3223 				nvmf_rdma_handle_sq_drained(rqpair);
3224 			}
3225 		}
3226 		break;
3227 	case IBV_EVENT_QP_REQ_ERR:
3228 	case IBV_EVENT_QP_ACCESS_ERR:
3229 	case IBV_EVENT_COMM_EST:
3230 	case IBV_EVENT_PATH_MIG:
3231 	case IBV_EVENT_PATH_MIG_ERR:
3232 		SPDK_NOTICELOG("Async event: %s\n",
3233 			       ibv_event_type_str(event.event_type));
3234 		rqpair = event.element.qp->qp_context;
3235 		spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0,
3236 				  (uintptr_t)rqpair->cm_id, event.event_type);
3237 		spdk_nvmf_rdma_update_ibv_state(rqpair);
3238 		break;
3239 	case IBV_EVENT_CQ_ERR:
3240 	case IBV_EVENT_DEVICE_FATAL:
3241 	case IBV_EVENT_PORT_ACTIVE:
3242 	case IBV_EVENT_PORT_ERR:
3243 	case IBV_EVENT_LID_CHANGE:
3244 	case IBV_EVENT_PKEY_CHANGE:
3245 	case IBV_EVENT_SM_CHANGE:
3246 	case IBV_EVENT_SRQ_ERR:
3247 	case IBV_EVENT_SRQ_LIMIT_REACHED:
3248 	case IBV_EVENT_CLIENT_REREGISTER:
3249 	case IBV_EVENT_GID_CHANGE:
3250 	default:
3251 		SPDK_NOTICELOG("Async event: %s\n",
3252 			       ibv_event_type_str(event.event_type));
3253 		spdk_trace_record(TRACE_RDMA_IBV_ASYNC_EVENT, 0, 0, 0, event.event_type);
3254 		break;
3255 	}
3256 	ibv_ack_async_event(&event);
3257 }
3258 
3259 static void
3260 spdk_nvmf_rdma_accept(struct spdk_nvmf_transport *transport, new_qpair_fn cb_fn, void *cb_arg)
3261 {
3262 	int	nfds, i = 0;
3263 	struct spdk_nvmf_rdma_transport *rtransport;
3264 	struct spdk_nvmf_rdma_device *device, *tmp;
3265 
3266 	rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3267 	nfds = poll(rtransport->poll_fds, rtransport->npoll_fds, 0);
3268 
3269 	if (nfds <= 0) {
3270 		return;
3271 	}
3272 
3273 	/* The first poll descriptor is RDMA CM event */
3274 	if (rtransport->poll_fds[i++].revents & POLLIN) {
3275 		spdk_nvmf_process_cm_event(transport, cb_fn, cb_arg);
3276 		nfds--;
3277 	}
3278 
3279 	if (nfds == 0) {
3280 		return;
3281 	}
3282 
3283 	/* Second and subsequent poll descriptors are IB async events */
3284 	TAILQ_FOREACH_SAFE(device, &rtransport->devices, link, tmp) {
3285 		if (rtransport->poll_fds[i++].revents & POLLIN) {
3286 			spdk_nvmf_process_ib_event(device);
3287 			nfds--;
3288 		}
3289 	}
3290 	/* check all flagged fd's have been served */
3291 	assert(nfds == 0);
3292 }
3293 
3294 static void
3295 spdk_nvmf_rdma_discover(struct spdk_nvmf_transport *transport,
3296 			struct spdk_nvme_transport_id *trid,
3297 			struct spdk_nvmf_discovery_log_page_entry *entry)
3298 {
3299 	entry->trtype = SPDK_NVMF_TRTYPE_RDMA;
3300 	entry->adrfam = trid->adrfam;
3301 	entry->treq.secure_channel = SPDK_NVMF_TREQ_SECURE_CHANNEL_NOT_REQUIRED;
3302 
3303 	spdk_strcpy_pad(entry->trsvcid, trid->trsvcid, sizeof(entry->trsvcid), ' ');
3304 	spdk_strcpy_pad(entry->traddr, trid->traddr, sizeof(entry->traddr), ' ');
3305 
3306 	entry->tsas.rdma.rdma_qptype = SPDK_NVMF_RDMA_QPTYPE_RELIABLE_CONNECTED;
3307 	entry->tsas.rdma.rdma_prtype = SPDK_NVMF_RDMA_PRTYPE_NONE;
3308 	entry->tsas.rdma.rdma_cms = SPDK_NVMF_RDMA_CMS_RDMA_CM;
3309 }
3310 
3311 static void
3312 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group);
3313 
3314 static struct spdk_nvmf_transport_poll_group *
3315 spdk_nvmf_rdma_poll_group_create(struct spdk_nvmf_transport *transport)
3316 {
3317 	struct spdk_nvmf_rdma_transport		*rtransport;
3318 	struct spdk_nvmf_rdma_poll_group	*rgroup;
3319 	struct spdk_nvmf_rdma_poller		*poller;
3320 	struct spdk_nvmf_rdma_device		*device;
3321 	struct ibv_srq_init_attr		srq_init_attr;
3322 	struct spdk_nvmf_rdma_resource_opts	opts;
3323 	int					num_cqe;
3324 
3325 	rtransport = SPDK_CONTAINEROF(transport, struct spdk_nvmf_rdma_transport, transport);
3326 
3327 	rgroup = calloc(1, sizeof(*rgroup));
3328 	if (!rgroup) {
3329 		return NULL;
3330 	}
3331 
3332 	TAILQ_INIT(&rgroup->pollers);
3333 	STAILQ_INIT(&rgroup->retired_bufs);
3334 
3335 	pthread_mutex_lock(&rtransport->lock);
3336 	TAILQ_FOREACH(device, &rtransport->devices, link) {
3337 		poller = calloc(1, sizeof(*poller));
3338 		if (!poller) {
3339 			SPDK_ERRLOG("Unable to allocate memory for new RDMA poller\n");
3340 			spdk_nvmf_rdma_poll_group_destroy(&rgroup->group);
3341 			pthread_mutex_unlock(&rtransport->lock);
3342 			return NULL;
3343 		}
3344 
3345 		poller->device = device;
3346 		poller->group = rgroup;
3347 
3348 		TAILQ_INIT(&poller->qpairs);
3349 		STAILQ_INIT(&poller->qpairs_pending_send);
3350 		STAILQ_INIT(&poller->qpairs_pending_recv);
3351 
3352 		TAILQ_INSERT_TAIL(&rgroup->pollers, poller, link);
3353 		if (transport->opts.no_srq == false && device->num_srq < device->attr.max_srq) {
3354 			poller->max_srq_depth = transport->opts.max_srq_depth;
3355 
3356 			device->num_srq++;
3357 			memset(&srq_init_attr, 0, sizeof(struct ibv_srq_init_attr));
3358 			srq_init_attr.attr.max_wr = poller->max_srq_depth;
3359 			srq_init_attr.attr.max_sge = spdk_min(device->attr.max_sge, NVMF_DEFAULT_RX_SGE);
3360 			poller->srq = ibv_create_srq(device->pd, &srq_init_attr);
3361 			if (!poller->srq) {
3362 				SPDK_ERRLOG("Unable to create shared receive queue, errno %d\n", errno);
3363 				spdk_nvmf_rdma_poll_group_destroy(&rgroup->group);
3364 				pthread_mutex_unlock(&rtransport->lock);
3365 				return NULL;
3366 			}
3367 
3368 			opts.qp = poller->srq;
3369 			opts.pd = device->pd;
3370 			opts.qpair = NULL;
3371 			opts.shared = true;
3372 			opts.max_queue_depth = poller->max_srq_depth;
3373 			opts.in_capsule_data_size = transport->opts.in_capsule_data_size;
3374 
3375 			poller->resources = nvmf_rdma_resources_create(&opts);
3376 			if (!poller->resources) {
3377 				SPDK_ERRLOG("Unable to allocate resources for shared receive queue.\n");
3378 				spdk_nvmf_rdma_poll_group_destroy(&rgroup->group);
3379 				pthread_mutex_unlock(&rtransport->lock);
3380 				return NULL;
3381 			}
3382 		}
3383 
3384 		/*
3385 		 * When using an srq, we can limit the completion queue at startup.
3386 		 * The following formula represents the calculation:
3387 		 * num_cqe = num_recv + num_data_wr + num_send_wr.
3388 		 * where num_recv=num_data_wr=and num_send_wr=poller->max_srq_depth
3389 		 */
3390 		if (poller->srq) {
3391 			num_cqe = poller->max_srq_depth * 3;
3392 		} else {
3393 			num_cqe = DEFAULT_NVMF_RDMA_CQ_SIZE;
3394 		}
3395 
3396 		poller->cq = ibv_create_cq(device->context, num_cqe, poller, NULL, 0);
3397 		if (!poller->cq) {
3398 			SPDK_ERRLOG("Unable to create completion queue\n");
3399 			spdk_nvmf_rdma_poll_group_destroy(&rgroup->group);
3400 			pthread_mutex_unlock(&rtransport->lock);
3401 			return NULL;
3402 		}
3403 		poller->num_cqe = num_cqe;
3404 	}
3405 
3406 	TAILQ_INSERT_TAIL(&rtransport->poll_groups, rgroup, link);
3407 	if (rtransport->conn_sched.next_admin_pg == NULL) {
3408 		rtransport->conn_sched.next_admin_pg = rgroup;
3409 		rtransport->conn_sched.next_io_pg = rgroup;
3410 	}
3411 
3412 	pthread_mutex_unlock(&rtransport->lock);
3413 	return &rgroup->group;
3414 }
3415 
3416 static struct spdk_nvmf_transport_poll_group *
3417 spdk_nvmf_rdma_get_optimal_poll_group(struct spdk_nvmf_qpair *qpair)
3418 {
3419 	struct spdk_nvmf_rdma_transport *rtransport;
3420 	struct spdk_nvmf_rdma_poll_group **pg;
3421 	struct spdk_nvmf_transport_poll_group *result;
3422 
3423 	rtransport = SPDK_CONTAINEROF(qpair->transport, struct spdk_nvmf_rdma_transport, transport);
3424 
3425 	pthread_mutex_lock(&rtransport->lock);
3426 
3427 	if (TAILQ_EMPTY(&rtransport->poll_groups)) {
3428 		pthread_mutex_unlock(&rtransport->lock);
3429 		return NULL;
3430 	}
3431 
3432 	if (qpair->qid == 0) {
3433 		pg = &rtransport->conn_sched.next_admin_pg;
3434 	} else {
3435 		pg = &rtransport->conn_sched.next_io_pg;
3436 	}
3437 
3438 	assert(*pg != NULL);
3439 
3440 	result = &(*pg)->group;
3441 
3442 	*pg = TAILQ_NEXT(*pg, link);
3443 	if (*pg == NULL) {
3444 		*pg = TAILQ_FIRST(&rtransport->poll_groups);
3445 	}
3446 
3447 	pthread_mutex_unlock(&rtransport->lock);
3448 
3449 	return result;
3450 }
3451 
3452 static void
3453 spdk_nvmf_rdma_poll_group_destroy(struct spdk_nvmf_transport_poll_group *group)
3454 {
3455 	struct spdk_nvmf_rdma_poll_group	*rgroup, *next_rgroup;
3456 	struct spdk_nvmf_rdma_poller		*poller, *tmp;
3457 	struct spdk_nvmf_rdma_qpair		*qpair, *tmp_qpair;
3458 	struct spdk_nvmf_transport_pg_cache_buf	*buf, *tmp_buf;
3459 	struct spdk_nvmf_rdma_transport		*rtransport;
3460 
3461 	rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
3462 	if (!rgroup) {
3463 		return;
3464 	}
3465 
3466 	/* free all retired buffers back to the transport so we don't short the mempool. */
3467 	STAILQ_FOREACH_SAFE(buf, &rgroup->retired_bufs, link, tmp_buf) {
3468 		STAILQ_REMOVE(&rgroup->retired_bufs, buf, spdk_nvmf_transport_pg_cache_buf, link);
3469 		assert(group->transport != NULL);
3470 		spdk_mempool_put(group->transport->data_buf_pool, buf);
3471 	}
3472 
3473 	TAILQ_FOREACH_SAFE(poller, &rgroup->pollers, link, tmp) {
3474 		TAILQ_REMOVE(&rgroup->pollers, poller, link);
3475 
3476 		TAILQ_FOREACH_SAFE(qpair, &poller->qpairs, link, tmp_qpair) {
3477 			spdk_nvmf_rdma_qpair_destroy(qpair);
3478 		}
3479 
3480 		if (poller->srq) {
3481 			if (poller->resources) {
3482 				nvmf_rdma_resources_destroy(poller->resources);
3483 			}
3484 			ibv_destroy_srq(poller->srq);
3485 			SPDK_DEBUGLOG(SPDK_LOG_RDMA, "Destroyed RDMA shared queue %p\n", poller->srq);
3486 		}
3487 
3488 		if (poller->cq) {
3489 			ibv_destroy_cq(poller->cq);
3490 		}
3491 
3492 		free(poller);
3493 	}
3494 
3495 	if (rgroup->group.transport == NULL) {
3496 		/* Transport can be NULL when spdk_nvmf_rdma_poll_group_create()
3497 		 * calls this function directly in a failure path. */
3498 		free(rgroup);
3499 		return;
3500 	}
3501 
3502 	rtransport = SPDK_CONTAINEROF(rgroup->group.transport, struct spdk_nvmf_rdma_transport, transport);
3503 
3504 	pthread_mutex_lock(&rtransport->lock);
3505 	next_rgroup = TAILQ_NEXT(rgroup, link);
3506 	TAILQ_REMOVE(&rtransport->poll_groups, rgroup, link);
3507 	if (next_rgroup == NULL) {
3508 		next_rgroup = TAILQ_FIRST(&rtransport->poll_groups);
3509 	}
3510 	if (rtransport->conn_sched.next_admin_pg == rgroup) {
3511 		rtransport->conn_sched.next_admin_pg = next_rgroup;
3512 	}
3513 	if (rtransport->conn_sched.next_io_pg == rgroup) {
3514 		rtransport->conn_sched.next_io_pg = next_rgroup;
3515 	}
3516 	pthread_mutex_unlock(&rtransport->lock);
3517 
3518 	free(rgroup);
3519 }
3520 
3521 static void
3522 spdk_nvmf_rdma_qpair_reject_connection(struct spdk_nvmf_rdma_qpair *rqpair)
3523 {
3524 	if (rqpair->cm_id != NULL) {
3525 		spdk_nvmf_rdma_event_reject(rqpair->cm_id, SPDK_NVMF_RDMA_ERROR_NO_RESOURCES);
3526 	}
3527 	spdk_nvmf_rdma_qpair_destroy(rqpair);
3528 }
3529 
3530 static int
3531 spdk_nvmf_rdma_poll_group_add(struct spdk_nvmf_transport_poll_group *group,
3532 			      struct spdk_nvmf_qpair *qpair)
3533 {
3534 	struct spdk_nvmf_rdma_poll_group	*rgroup;
3535 	struct spdk_nvmf_rdma_qpair		*rqpair;
3536 	struct spdk_nvmf_rdma_device		*device;
3537 	struct spdk_nvmf_rdma_poller		*poller;
3538 	int					rc;
3539 
3540 	rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
3541 	rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
3542 
3543 	device = rqpair->device;
3544 
3545 	TAILQ_FOREACH(poller, &rgroup->pollers, link) {
3546 		if (poller->device == device) {
3547 			break;
3548 		}
3549 	}
3550 
3551 	if (!poller) {
3552 		SPDK_ERRLOG("No poller found for device.\n");
3553 		return -1;
3554 	}
3555 
3556 	TAILQ_INSERT_TAIL(&poller->qpairs, rqpair, link);
3557 	rqpair->poller = poller;
3558 	rqpair->srq = rqpair->poller->srq;
3559 
3560 	rc = spdk_nvmf_rdma_qpair_initialize(qpair);
3561 	if (rc < 0) {
3562 		SPDK_ERRLOG("Failed to initialize nvmf_rdma_qpair with qpair=%p\n", qpair);
3563 		return -1;
3564 	}
3565 
3566 	rc = spdk_nvmf_rdma_event_accept(rqpair->cm_id, rqpair);
3567 	if (rc) {
3568 		/* Try to reject, but we probably can't */
3569 		spdk_nvmf_rdma_qpair_reject_connection(rqpair);
3570 		return -1;
3571 	}
3572 
3573 	spdk_nvmf_rdma_update_ibv_state(rqpair);
3574 
3575 	return 0;
3576 }
3577 
3578 static int
3579 spdk_nvmf_rdma_request_free(struct spdk_nvmf_request *req)
3580 {
3581 	struct spdk_nvmf_rdma_request	*rdma_req = SPDK_CONTAINEROF(req, struct spdk_nvmf_rdma_request, req);
3582 	struct spdk_nvmf_rdma_transport	*rtransport = SPDK_CONTAINEROF(req->qpair->transport,
3583 			struct spdk_nvmf_rdma_transport, transport);
3584 
3585 	nvmf_rdma_request_free(rdma_req, rtransport);
3586 	return 0;
3587 }
3588 
3589 static int
3590 spdk_nvmf_rdma_request_complete(struct spdk_nvmf_request *req)
3591 {
3592 	struct spdk_nvmf_rdma_transport	*rtransport = SPDK_CONTAINEROF(req->qpair->transport,
3593 			struct spdk_nvmf_rdma_transport, transport);
3594 	struct spdk_nvmf_rdma_request	*rdma_req = SPDK_CONTAINEROF(req,
3595 			struct spdk_nvmf_rdma_request, req);
3596 	struct spdk_nvmf_rdma_qpair     *rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair,
3597 			struct spdk_nvmf_rdma_qpair, qpair);
3598 
3599 	if (rqpair->ibv_state != IBV_QPS_ERR) {
3600 		/* The connection is alive, so process the request as normal */
3601 		rdma_req->state = RDMA_REQUEST_STATE_EXECUTED;
3602 	} else {
3603 		/* The connection is dead. Move the request directly to the completed state. */
3604 		rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
3605 	}
3606 
3607 	spdk_nvmf_rdma_request_process(rtransport, rdma_req);
3608 
3609 	return 0;
3610 }
3611 
3612 static int
3613 spdk_nvmf_rdma_destroy_defunct_qpair(void *ctx)
3614 {
3615 	struct spdk_nvmf_rdma_qpair	*rqpair = ctx;
3616 	struct spdk_nvmf_rdma_transport *rtransport = SPDK_CONTAINEROF(rqpair->qpair.transport,
3617 			struct spdk_nvmf_rdma_transport, transport);
3618 
3619 	SPDK_INFOLOG(SPDK_LOG_RDMA, "QP#%d hasn't been drained as expected, manually destroy it\n",
3620 		     rqpair->qpair.qid);
3621 
3622 	spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, true);
3623 	spdk_nvmf_rdma_qpair_destroy(rqpair);
3624 
3625 	return 0;
3626 }
3627 
3628 static void
3629 spdk_nvmf_rdma_close_qpair(struct spdk_nvmf_qpair *qpair)
3630 {
3631 	struct spdk_nvmf_rdma_qpair *rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
3632 
3633 	if (rqpair->disconnect_flags & RDMA_QP_DISCONNECTING) {
3634 		return;
3635 	}
3636 
3637 	rqpair->disconnect_flags |= RDMA_QP_DISCONNECTING;
3638 
3639 	/* This happens only when the qpair is disconnected before
3640 	 * it is added to the poll group. Since there is no poll group,
3641 	 * the RDMA qp has not been initialized yet and the RDMA CM
3642 	 * event has not yet been acknowledged, so we need to reject it.
3643 	 */
3644 	if (rqpair->qpair.state == SPDK_NVMF_QPAIR_UNINITIALIZED) {
3645 		spdk_nvmf_rdma_qpair_reject_connection(rqpair);
3646 		return;
3647 	}
3648 
3649 	if (rqpair->ibv_state != IBV_QPS_ERR) {
3650 		spdk_nvmf_rdma_set_ibv_state(rqpair, IBV_QPS_ERR);
3651 	}
3652 
3653 	rqpair->destruct_poller = spdk_poller_register(spdk_nvmf_rdma_destroy_defunct_qpair, (void *)rqpair,
3654 				  NVMF_RDMA_QPAIR_DESTROY_TIMEOUT_US);
3655 }
3656 
3657 static struct spdk_nvmf_rdma_qpair *
3658 get_rdma_qpair_from_wc(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_wc *wc)
3659 {
3660 	struct spdk_nvmf_rdma_qpair *rqpair;
3661 	/* @todo: improve QP search */
3662 	TAILQ_FOREACH(rqpair, &rpoller->qpairs, link) {
3663 		if (wc->qp_num == rqpair->cm_id->qp->qp_num) {
3664 			return rqpair;
3665 		}
3666 	}
3667 	SPDK_ERRLOG("Didn't find QP with qp_num %u\n", wc->qp_num);
3668 	return NULL;
3669 }
3670 
3671 #ifdef DEBUG
3672 static int
3673 spdk_nvmf_rdma_req_is_completing(struct spdk_nvmf_rdma_request *rdma_req)
3674 {
3675 	return rdma_req->state == RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST ||
3676 	       rdma_req->state == RDMA_REQUEST_STATE_COMPLETING;
3677 }
3678 #endif
3679 
3680 static void
3681 _poller_reset_failed_recvs(struct spdk_nvmf_rdma_poller *rpoller, struct ibv_recv_wr *bad_recv_wr,
3682 			   int rc)
3683 {
3684 	struct spdk_nvmf_rdma_recv	*rdma_recv;
3685 	struct spdk_nvmf_rdma_wr	*bad_rdma_wr;
3686 
3687 	SPDK_ERRLOG("Failed to post a recv for the poller %p with errno %d\n", rpoller, -rc);
3688 	while (bad_recv_wr != NULL) {
3689 		bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_recv_wr->wr_id;
3690 		rdma_recv = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr);
3691 
3692 		rdma_recv->qpair->current_recv_depth++;
3693 		bad_recv_wr = bad_recv_wr->next;
3694 		SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rdma_recv->qpair, -rc);
3695 		spdk_nvmf_rdma_start_disconnect(rdma_recv->qpair);
3696 	}
3697 }
3698 
3699 static void
3700 _qp_reset_failed_recvs(struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_recv_wr *bad_recv_wr, int rc)
3701 {
3702 	SPDK_ERRLOG("Failed to post a recv for the qpair %p with errno %d\n", rqpair, -rc);
3703 	while (bad_recv_wr != NULL) {
3704 		bad_recv_wr = bad_recv_wr->next;
3705 		rqpair->current_recv_depth++;
3706 	}
3707 	spdk_nvmf_rdma_start_disconnect(rqpair);
3708 }
3709 
3710 static void
3711 _poller_submit_recvs(struct spdk_nvmf_rdma_transport *rtransport,
3712 		     struct spdk_nvmf_rdma_poller *rpoller)
3713 {
3714 	struct spdk_nvmf_rdma_qpair	*rqpair;
3715 	struct ibv_recv_wr		*bad_recv_wr;
3716 	int				rc;
3717 
3718 	if (rpoller->srq) {
3719 		if (rpoller->resources->recvs_to_post.first != NULL) {
3720 			rc = ibv_post_srq_recv(rpoller->srq, rpoller->resources->recvs_to_post.first, &bad_recv_wr);
3721 			if (rc) {
3722 				_poller_reset_failed_recvs(rpoller, bad_recv_wr, rc);
3723 			}
3724 			rpoller->resources->recvs_to_post.first = NULL;
3725 			rpoller->resources->recvs_to_post.last = NULL;
3726 		}
3727 	} else {
3728 		while (!STAILQ_EMPTY(&rpoller->qpairs_pending_recv)) {
3729 			rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_recv);
3730 			assert(rqpair->resources->recvs_to_post.first != NULL);
3731 			rc = ibv_post_recv(rqpair->cm_id->qp, rqpair->resources->recvs_to_post.first, &bad_recv_wr);
3732 			if (rc) {
3733 				_qp_reset_failed_recvs(rqpair, bad_recv_wr, rc);
3734 			}
3735 			rqpair->resources->recvs_to_post.first = NULL;
3736 			rqpair->resources->recvs_to_post.last = NULL;
3737 			STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_recv, recv_link);
3738 		}
3739 	}
3740 }
3741 
3742 static void
3743 _qp_reset_failed_sends(struct spdk_nvmf_rdma_transport *rtransport,
3744 		       struct spdk_nvmf_rdma_qpair *rqpair, struct ibv_send_wr *bad_wr, int rc)
3745 {
3746 	struct spdk_nvmf_rdma_wr	*bad_rdma_wr;
3747 	struct spdk_nvmf_rdma_request	*prev_rdma_req = NULL, *cur_rdma_req = NULL;
3748 
3749 	SPDK_ERRLOG("Failed to post a send for the qpair %p with errno %d\n", rqpair, -rc);
3750 	for (; bad_wr != NULL; bad_wr = bad_wr->next) {
3751 		bad_rdma_wr = (struct spdk_nvmf_rdma_wr *)bad_wr->wr_id;
3752 		assert(rqpair->current_send_depth > 0);
3753 		rqpair->current_send_depth--;
3754 		switch (bad_rdma_wr->type) {
3755 		case RDMA_WR_TYPE_DATA:
3756 			cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr);
3757 			if (bad_wr->opcode == IBV_WR_RDMA_READ) {
3758 				assert(rqpair->current_read_depth > 0);
3759 				rqpair->current_read_depth--;
3760 			}
3761 			break;
3762 		case RDMA_WR_TYPE_SEND:
3763 			cur_rdma_req = SPDK_CONTAINEROF(bad_rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr);
3764 			break;
3765 		default:
3766 			SPDK_ERRLOG("Found a RECV in the list of pending SEND requests for qpair %p\n", rqpair);
3767 			prev_rdma_req = cur_rdma_req;
3768 			continue;
3769 		}
3770 
3771 		if (prev_rdma_req == cur_rdma_req) {
3772 			/* this request was handled by an earlier wr. i.e. we were performing an nvme read. */
3773 			/* We only have to check against prev_wr since each requests wrs are contiguous in this list. */
3774 			continue;
3775 		}
3776 
3777 		switch (cur_rdma_req->state) {
3778 		case RDMA_REQUEST_STATE_TRANSFERRING_HOST_TO_CONTROLLER:
3779 			cur_rdma_req->req.rsp->nvme_cpl.status.sc = SPDK_NVME_SC_INTERNAL_DEVICE_ERROR;
3780 			cur_rdma_req->state = RDMA_REQUEST_STATE_READY_TO_COMPLETE;
3781 			break;
3782 		case RDMA_REQUEST_STATE_TRANSFERRING_CONTROLLER_TO_HOST:
3783 		case RDMA_REQUEST_STATE_COMPLETING:
3784 			cur_rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
3785 			break;
3786 		default:
3787 			SPDK_ERRLOG("Found a request in a bad state %d when draining pending SEND requests for qpair %p\n",
3788 				    cur_rdma_req->state, rqpair);
3789 			continue;
3790 		}
3791 
3792 		spdk_nvmf_rdma_request_process(rtransport, cur_rdma_req);
3793 		prev_rdma_req = cur_rdma_req;
3794 	}
3795 
3796 	if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) {
3797 		/* Disconnect the connection. */
3798 		spdk_nvmf_rdma_start_disconnect(rqpair);
3799 	}
3800 
3801 }
3802 
3803 static void
3804 _poller_submit_sends(struct spdk_nvmf_rdma_transport *rtransport,
3805 		     struct spdk_nvmf_rdma_poller *rpoller)
3806 {
3807 	struct spdk_nvmf_rdma_qpair	*rqpair;
3808 	struct ibv_send_wr		*bad_wr = NULL;
3809 	int				rc;
3810 
3811 	while (!STAILQ_EMPTY(&rpoller->qpairs_pending_send)) {
3812 		rqpair = STAILQ_FIRST(&rpoller->qpairs_pending_send);
3813 		assert(rqpair->sends_to_post.first != NULL);
3814 		rc = ibv_post_send(rqpair->cm_id->qp, rqpair->sends_to_post.first, &bad_wr);
3815 
3816 		/* bad wr always points to the first wr that failed. */
3817 		if (rc) {
3818 			_qp_reset_failed_sends(rtransport, rqpair, bad_wr, rc);
3819 		}
3820 		rqpair->sends_to_post.first = NULL;
3821 		rqpair->sends_to_post.last = NULL;
3822 		STAILQ_REMOVE_HEAD(&rpoller->qpairs_pending_send, send_link);
3823 	}
3824 }
3825 
3826 static int
3827 spdk_nvmf_rdma_poller_poll(struct spdk_nvmf_rdma_transport *rtransport,
3828 			   struct spdk_nvmf_rdma_poller *rpoller)
3829 {
3830 	struct ibv_wc wc[32];
3831 	struct spdk_nvmf_rdma_wr	*rdma_wr;
3832 	struct spdk_nvmf_rdma_request	*rdma_req;
3833 	struct spdk_nvmf_rdma_recv	*rdma_recv;
3834 	struct spdk_nvmf_rdma_qpair	*rqpair;
3835 	int reaped, i;
3836 	int count = 0;
3837 	bool error = false;
3838 	uint64_t poll_tsc = spdk_get_ticks();
3839 
3840 	/* Poll for completing operations. */
3841 	reaped = ibv_poll_cq(rpoller->cq, 32, wc);
3842 	if (reaped < 0) {
3843 		SPDK_ERRLOG("Error polling CQ! (%d): %s\n",
3844 			    errno, spdk_strerror(errno));
3845 		return -1;
3846 	}
3847 
3848 	rpoller->stat.polls++;
3849 	rpoller->stat.completions += reaped;
3850 
3851 	for (i = 0; i < reaped; i++) {
3852 
3853 		rdma_wr = (struct spdk_nvmf_rdma_wr *)wc[i].wr_id;
3854 
3855 		switch (rdma_wr->type) {
3856 		case RDMA_WR_TYPE_SEND:
3857 			rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, rsp.rdma_wr);
3858 			rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
3859 
3860 			if (!wc[i].status) {
3861 				count++;
3862 				assert(wc[i].opcode == IBV_WC_SEND);
3863 				assert(spdk_nvmf_rdma_req_is_completing(rdma_req));
3864 			}
3865 
3866 			rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
3867 			/* RDMA_WRITE operation completed. +1 since it was chained with rsp WR */
3868 			rqpair->current_send_depth -= rdma_req->num_outstanding_data_wr + 1;
3869 			rdma_req->num_outstanding_data_wr = 0;
3870 
3871 			spdk_nvmf_rdma_request_process(rtransport, rdma_req);
3872 			break;
3873 		case RDMA_WR_TYPE_RECV:
3874 			/* rdma_recv->qpair will be invalid if using an SRQ.  In that case we have to get the qpair from the wc. */
3875 			rdma_recv = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_recv, rdma_wr);
3876 			if (rpoller->srq != NULL) {
3877 				rdma_recv->qpair = get_rdma_qpair_from_wc(rpoller, &wc[i]);
3878 				/* It is possible that there are still some completions for destroyed QP
3879 				 * associated with SRQ. We just ignore these late completions and re-post
3880 				 * receive WRs back to SRQ.
3881 				 */
3882 				if (spdk_unlikely(NULL == rdma_recv->qpair)) {
3883 					struct ibv_recv_wr *bad_wr;
3884 					int rc;
3885 
3886 					rdma_recv->wr.next = NULL;
3887 					rc = ibv_post_srq_recv(rpoller->srq,
3888 							       &rdma_recv->wr,
3889 							       &bad_wr);
3890 					if (rc) {
3891 						SPDK_ERRLOG("Failed to re-post recv WR to SRQ, err %d\n", rc);
3892 					}
3893 					continue;
3894 				}
3895 			}
3896 			rqpair = rdma_recv->qpair;
3897 
3898 			assert(rqpair != NULL);
3899 			if (!wc[i].status) {
3900 				assert(wc[i].opcode == IBV_WC_RECV);
3901 				if (rqpair->current_recv_depth >= rqpair->max_queue_depth) {
3902 					spdk_nvmf_rdma_start_disconnect(rqpair);
3903 					break;
3904 				}
3905 			}
3906 
3907 			rdma_recv->wr.next = NULL;
3908 			rqpair->current_recv_depth++;
3909 			rdma_recv->receive_tsc = poll_tsc;
3910 			rpoller->stat.requests++;
3911 			STAILQ_INSERT_TAIL(&rqpair->resources->incoming_queue, rdma_recv, link);
3912 			break;
3913 		case RDMA_WR_TYPE_DATA:
3914 			rdma_req = SPDK_CONTAINEROF(rdma_wr, struct spdk_nvmf_rdma_request, data.rdma_wr);
3915 			rqpair = SPDK_CONTAINEROF(rdma_req->req.qpair, struct spdk_nvmf_rdma_qpair, qpair);
3916 
3917 			assert(rdma_req->num_outstanding_data_wr > 0);
3918 
3919 			rqpair->current_send_depth--;
3920 			rdma_req->num_outstanding_data_wr--;
3921 			if (!wc[i].status) {
3922 				assert(wc[i].opcode == IBV_WC_RDMA_READ);
3923 				rqpair->current_read_depth--;
3924 				/* wait for all outstanding reads associated with the same rdma_req to complete before proceeding. */
3925 				if (rdma_req->num_outstanding_data_wr == 0) {
3926 					rdma_req->state = RDMA_REQUEST_STATE_READY_TO_EXECUTE;
3927 					spdk_nvmf_rdma_request_process(rtransport, rdma_req);
3928 				}
3929 			} else {
3930 				/* If the data transfer fails still force the queue into the error state,
3931 				 * if we were performing an RDMA_READ, we need to force the request into a
3932 				 * completed state since it wasn't linked to a send. However, in the RDMA_WRITE
3933 				 * case, we should wait for the SEND to complete. */
3934 				if (rdma_req->data.wr.opcode == IBV_WR_RDMA_READ) {
3935 					rqpair->current_read_depth--;
3936 					if (rdma_req->num_outstanding_data_wr == 0) {
3937 						rdma_req->state = RDMA_REQUEST_STATE_COMPLETED;
3938 					}
3939 				}
3940 			}
3941 			break;
3942 		default:
3943 			SPDK_ERRLOG("Received an unknown opcode on the CQ: %d\n", wc[i].opcode);
3944 			continue;
3945 		}
3946 
3947 		/* Handle error conditions */
3948 		if (wc[i].status) {
3949 			SPDK_ERRLOG("Error on CQ %p, request 0x%lu, type %d, status: (%d): %s\n",
3950 				    rpoller->cq, wc[i].wr_id, rdma_wr->type, wc[i].status, ibv_wc_status_str(wc[i].status));
3951 
3952 			error = true;
3953 
3954 			if (rqpair->qpair.state == SPDK_NVMF_QPAIR_ACTIVE) {
3955 				/* Disconnect the connection. */
3956 				spdk_nvmf_rdma_start_disconnect(rqpair);
3957 			} else {
3958 				nvmf_rdma_destroy_drained_qpair(rqpair);
3959 			}
3960 			continue;
3961 		}
3962 
3963 		spdk_nvmf_rdma_qpair_process_pending(rtransport, rqpair, false);
3964 
3965 		if (rqpair->qpair.state != SPDK_NVMF_QPAIR_ACTIVE) {
3966 			nvmf_rdma_destroy_drained_qpair(rqpair);
3967 		}
3968 	}
3969 
3970 	if (error == true) {
3971 		return -1;
3972 	}
3973 
3974 	/* submit outstanding work requests. */
3975 	_poller_submit_recvs(rtransport, rpoller);
3976 	_poller_submit_sends(rtransport, rpoller);
3977 
3978 	return count;
3979 }
3980 
3981 static int
3982 spdk_nvmf_rdma_poll_group_poll(struct spdk_nvmf_transport_poll_group *group)
3983 {
3984 	struct spdk_nvmf_rdma_transport *rtransport;
3985 	struct spdk_nvmf_rdma_poll_group *rgroup;
3986 	struct spdk_nvmf_rdma_poller	*rpoller;
3987 	int				count, rc;
3988 
3989 	rtransport = SPDK_CONTAINEROF(group->transport, struct spdk_nvmf_rdma_transport, transport);
3990 	rgroup = SPDK_CONTAINEROF(group, struct spdk_nvmf_rdma_poll_group, group);
3991 
3992 	count = 0;
3993 	TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
3994 		rc = spdk_nvmf_rdma_poller_poll(rtransport, rpoller);
3995 		if (rc < 0) {
3996 			return rc;
3997 		}
3998 		count += rc;
3999 	}
4000 
4001 	return count;
4002 }
4003 
4004 static int
4005 spdk_nvmf_rdma_trid_from_cm_id(struct rdma_cm_id *id,
4006 			       struct spdk_nvme_transport_id *trid,
4007 			       bool peer)
4008 {
4009 	struct sockaddr *saddr;
4010 	uint16_t port;
4011 
4012 	spdk_nvme_trid_populate_transport(trid, SPDK_NVME_TRANSPORT_RDMA);
4013 
4014 	if (peer) {
4015 		saddr = rdma_get_peer_addr(id);
4016 	} else {
4017 		saddr = rdma_get_local_addr(id);
4018 	}
4019 	switch (saddr->sa_family) {
4020 	case AF_INET: {
4021 		struct sockaddr_in *saddr_in = (struct sockaddr_in *)saddr;
4022 
4023 		trid->adrfam = SPDK_NVMF_ADRFAM_IPV4;
4024 		inet_ntop(AF_INET, &saddr_in->sin_addr,
4025 			  trid->traddr, sizeof(trid->traddr));
4026 		if (peer) {
4027 			port = ntohs(rdma_get_dst_port(id));
4028 		} else {
4029 			port = ntohs(rdma_get_src_port(id));
4030 		}
4031 		snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port);
4032 		break;
4033 	}
4034 	case AF_INET6: {
4035 		struct sockaddr_in6 *saddr_in = (struct sockaddr_in6 *)saddr;
4036 		trid->adrfam = SPDK_NVMF_ADRFAM_IPV6;
4037 		inet_ntop(AF_INET6, &saddr_in->sin6_addr,
4038 			  trid->traddr, sizeof(trid->traddr));
4039 		if (peer) {
4040 			port = ntohs(rdma_get_dst_port(id));
4041 		} else {
4042 			port = ntohs(rdma_get_src_port(id));
4043 		}
4044 		snprintf(trid->trsvcid, sizeof(trid->trsvcid), "%u", port);
4045 		break;
4046 	}
4047 	default:
4048 		return -1;
4049 
4050 	}
4051 
4052 	return 0;
4053 }
4054 
4055 static int
4056 spdk_nvmf_rdma_qpair_get_peer_trid(struct spdk_nvmf_qpair *qpair,
4057 				   struct spdk_nvme_transport_id *trid)
4058 {
4059 	struct spdk_nvmf_rdma_qpair	*rqpair;
4060 
4061 	rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4062 
4063 	return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, true);
4064 }
4065 
4066 static int
4067 spdk_nvmf_rdma_qpair_get_local_trid(struct spdk_nvmf_qpair *qpair,
4068 				    struct spdk_nvme_transport_id *trid)
4069 {
4070 	struct spdk_nvmf_rdma_qpair	*rqpair;
4071 
4072 	rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4073 
4074 	return spdk_nvmf_rdma_trid_from_cm_id(rqpair->cm_id, trid, false);
4075 }
4076 
4077 static int
4078 spdk_nvmf_rdma_qpair_get_listen_trid(struct spdk_nvmf_qpair *qpair,
4079 				     struct spdk_nvme_transport_id *trid)
4080 {
4081 	struct spdk_nvmf_rdma_qpair	*rqpair;
4082 
4083 	rqpair = SPDK_CONTAINEROF(qpair, struct spdk_nvmf_rdma_qpair, qpair);
4084 
4085 	return spdk_nvmf_rdma_trid_from_cm_id(rqpair->listen_id, trid, false);
4086 }
4087 
4088 void
4089 spdk_nvmf_rdma_init_hooks(struct spdk_nvme_rdma_hooks *hooks)
4090 {
4091 	g_nvmf_hooks = *hooks;
4092 }
4093 
4094 static int
4095 spdk_nvmf_rdma_poll_group_get_stat(struct spdk_nvmf_tgt *tgt,
4096 				   struct spdk_nvmf_transport_poll_group_stat **stat)
4097 {
4098 	struct spdk_io_channel *ch;
4099 	struct spdk_nvmf_poll_group *group;
4100 	struct spdk_nvmf_transport_poll_group *tgroup;
4101 	struct spdk_nvmf_rdma_poll_group *rgroup;
4102 	struct spdk_nvmf_rdma_poller *rpoller;
4103 	struct spdk_nvmf_rdma_device_stat *device_stat;
4104 	uint64_t num_devices = 0;
4105 
4106 	if (tgt == NULL || stat == NULL) {
4107 		return -EINVAL;
4108 	}
4109 
4110 	ch = spdk_get_io_channel(tgt);
4111 	group = spdk_io_channel_get_ctx(ch);;
4112 	spdk_put_io_channel(ch);
4113 	TAILQ_FOREACH(tgroup, &group->tgroups, link) {
4114 		if (SPDK_NVME_TRANSPORT_RDMA == tgroup->transport->ops->type) {
4115 			*stat = calloc(1, sizeof(struct spdk_nvmf_transport_poll_group_stat));
4116 			if (!*stat) {
4117 				SPDK_ERRLOG("Failed to allocate memory for NVMf RDMA statistics\n");
4118 				return -ENOMEM;
4119 			}
4120 			(*stat)->trtype = SPDK_NVME_TRANSPORT_RDMA;
4121 
4122 			rgroup = SPDK_CONTAINEROF(tgroup, struct spdk_nvmf_rdma_poll_group, group);
4123 			/* Count devices to allocate enough memory */
4124 			TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
4125 				++num_devices;
4126 			}
4127 			(*stat)->rdma.devices = calloc(num_devices, sizeof(struct spdk_nvmf_rdma_device_stat));
4128 			if (!(*stat)->rdma.devices) {
4129 				SPDK_ERRLOG("Failed to allocate NVMf RDMA devices statistics\n");
4130 				free(*stat);
4131 				return -ENOMEM;
4132 			}
4133 
4134 			(*stat)->rdma.pending_data_buffer = rgroup->stat.pending_data_buffer;
4135 			(*stat)->rdma.num_devices = num_devices;
4136 			num_devices = 0;
4137 			TAILQ_FOREACH(rpoller, &rgroup->pollers, link) {
4138 				device_stat = &(*stat)->rdma.devices[num_devices++];
4139 				device_stat->name = ibv_get_device_name(rpoller->device->context->device);
4140 				device_stat->polls = rpoller->stat.polls;
4141 				device_stat->completions = rpoller->stat.completions;
4142 				device_stat->requests = rpoller->stat.requests;
4143 				device_stat->request_latency = rpoller->stat.request_latency;
4144 				device_stat->pending_free_request = rpoller->stat.pending_free_request;
4145 				device_stat->pending_rdma_read = rpoller->stat.pending_rdma_read;
4146 				device_stat->pending_rdma_write = rpoller->stat.pending_rdma_write;
4147 			}
4148 			return 0;
4149 		}
4150 	}
4151 	return -ENOENT;
4152 }
4153 
4154 static void
4155 spdk_nvmf_rdma_poll_group_free_stat(struct spdk_nvmf_transport_poll_group_stat *stat)
4156 {
4157 	if (stat) {
4158 		free(stat->rdma.devices);
4159 	}
4160 	free(stat);
4161 }
4162 
4163 const struct spdk_nvmf_transport_ops spdk_nvmf_transport_rdma = {
4164 	.name = "RDMA",
4165 	.type = SPDK_NVME_TRANSPORT_RDMA,
4166 	.opts_init = spdk_nvmf_rdma_opts_init,
4167 	.create = spdk_nvmf_rdma_create,
4168 	.destroy = spdk_nvmf_rdma_destroy,
4169 
4170 	.listen = spdk_nvmf_rdma_listen,
4171 	.stop_listen = spdk_nvmf_rdma_stop_listen,
4172 	.accept = spdk_nvmf_rdma_accept,
4173 
4174 	.listener_discover = spdk_nvmf_rdma_discover,
4175 
4176 	.poll_group_create = spdk_nvmf_rdma_poll_group_create,
4177 	.get_optimal_poll_group = spdk_nvmf_rdma_get_optimal_poll_group,
4178 	.poll_group_destroy = spdk_nvmf_rdma_poll_group_destroy,
4179 	.poll_group_add = spdk_nvmf_rdma_poll_group_add,
4180 	.poll_group_poll = spdk_nvmf_rdma_poll_group_poll,
4181 
4182 	.req_free = spdk_nvmf_rdma_request_free,
4183 	.req_complete = spdk_nvmf_rdma_request_complete,
4184 
4185 	.qpair_fini = spdk_nvmf_rdma_close_qpair,
4186 	.qpair_get_peer_trid = spdk_nvmf_rdma_qpair_get_peer_trid,
4187 	.qpair_get_local_trid = spdk_nvmf_rdma_qpair_get_local_trid,
4188 	.qpair_get_listen_trid = spdk_nvmf_rdma_qpair_get_listen_trid,
4189 
4190 	.poll_group_get_stat = spdk_nvmf_rdma_poll_group_get_stat,
4191 	.poll_group_free_stat = spdk_nvmf_rdma_poll_group_free_stat,
4192 };
4193 
4194 SPDK_NVMF_TRANSPORT_REGISTER(rdma, &spdk_nvmf_transport_rdma);
4195 SPDK_LOG_REGISTER_COMPONENT("rdma", SPDK_LOG_RDMA)
4196