xref: /spdk/examples/nvme/reconnect/reconnect.c (revision b30d57cdad6d2bc75cc1e4e2ebbcebcb0d98dcfa)
1 /*-
2  *   BSD LICENSE
3  *
4  *   Copyright (c) Intel Corporation. All rights reserved.
5  *   Copyright (c) 2020 Mellanox Technologies LTD. All rights reserved.
6  *
7  *   Copyright (c) 2019 Mellanox Technologies LTD. All rights reserved.
8  *
9  *   Redistribution and use in source and binary forms, with or without
10  *   modification, are permitted provided that the following conditions
11  *   are met:
12  *
13  *     * Redistributions of source code must retain the above copyright
14  *       notice, this list of conditions and the following disclaimer.
15  *     * Redistributions in binary form must reproduce the above copyright
16  *       notice, this list of conditions and the following disclaimer in
17  *       the documentation and/or other materials provided with the
18  *       distribution.
19  *     * Neither the name of Intel Corporation nor the names of its
20  *       contributors may be used to endorse or promote products derived
21  *       from this software without specific prior written permission.
22  *
23  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
26  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
27  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
28  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
29  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
30  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
31  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
32  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
33  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
34  */
35 
36 #include "spdk/stdinc.h"
37 
38 #include "spdk/env.h"
39 #include "spdk/nvme.h"
40 #include "spdk/queue.h"
41 #include "spdk/string.h"
42 #include "spdk/util.h"
43 #include "spdk/log.h"
44 #include "spdk/likely.h"
45 
46 struct ctrlr_entry {
47 	struct spdk_nvme_ctrlr			*ctrlr;
48 	struct spdk_nvme_transport_id		failover_trid;
49 	enum spdk_nvme_transport_type		trtype;
50 	TAILQ_ENTRY(ctrlr_entry)		link;
51 	char					name[1024];
52 	int					num_resets;
53 };
54 
55 struct ns_entry {
56 	struct spdk_nvme_ctrlr	*ctrlr;
57 	struct spdk_nvme_ns	*ns;
58 
59 	TAILQ_ENTRY(ns_entry)	link;
60 	uint32_t		io_size_blocks;
61 	uint32_t		num_io_requests;
62 	uint64_t		size_in_ios;
63 	uint32_t		block_size;
64 	uint32_t		io_flags;
65 	char			name[1024];
66 };
67 
68 struct ns_worker_ctx {
69 	struct ns_entry			*entry;
70 	uint64_t			io_completed;
71 	uint64_t			current_queue_depth;
72 	uint64_t			offset_in_ios;
73 	bool				is_draining;
74 
75 	int				num_qpairs;
76 	struct spdk_nvme_qpair		**qpair;
77 	int				last_qpair;
78 
79 	TAILQ_ENTRY(ns_worker_ctx)	link;
80 };
81 
82 struct perf_task {
83 	struct ns_worker_ctx	*ns_ctx;
84 	struct iovec		iov;
85 	bool			is_read;
86 };
87 
88 struct worker_thread {
89 	TAILQ_HEAD(, ns_worker_ctx)	ns_ctx;
90 	TAILQ_ENTRY(worker_thread)	link;
91 	unsigned			lcore;
92 };
93 
94 /* For basic reset handling. */
95 static int g_max_ctrlr_resets = 15;
96 
97 static TAILQ_HEAD(, ctrlr_entry) g_controllers = TAILQ_HEAD_INITIALIZER(g_controllers);
98 static TAILQ_HEAD(, ns_entry) g_namespaces = TAILQ_HEAD_INITIALIZER(g_namespaces);
99 static int g_num_namespaces = 0;
100 static TAILQ_HEAD(, worker_thread) g_workers = TAILQ_HEAD_INITIALIZER(g_workers);
101 static int g_num_workers = 0;
102 
103 static uint64_t g_tsc_rate;
104 
105 static uint32_t g_io_align = 0x200;
106 static uint32_t g_io_size_bytes;
107 static uint32_t g_max_io_size_blocks;
108 static int g_rw_percentage;
109 static int g_is_random;
110 static int g_queue_depth;
111 static int g_time_in_sec;
112 static uint32_t g_max_completions;
113 static int g_dpdk_mem;
114 static bool g_warn;
115 static uint32_t g_keep_alive_timeout_in_ms = 0;
116 static uint8_t g_transport_retry_count = 4;
117 static uint8_t g_transport_ack_timeout = 0; /* disabled */
118 static bool g_dpdk_mem_single_seg = false;
119 
120 static const char *g_core_mask;
121 
122 struct trid_entry {
123 	struct spdk_nvme_transport_id	trid;
124 	struct spdk_nvme_transport_id	failover_trid;
125 	TAILQ_ENTRY(trid_entry)		tailq;
126 };
127 
128 static TAILQ_HEAD(, trid_entry) g_trid_list = TAILQ_HEAD_INITIALIZER(g_trid_list);
129 
130 static inline void
131 task_complete(struct perf_task *task);
132 static void submit_io(struct ns_worker_ctx *ns_ctx, int queue_depth);
133 
134 static void io_complete(void *ctx, const struct spdk_nvme_cpl *cpl);
135 
136 static void
137 nvme_setup_payload(struct perf_task *task)
138 {
139 	/* maximum extended lba format size from all active namespace,
140 	 * it's same with g_io_size_bytes for namespace without metadata.
141 	 */
142 	task->iov.iov_base = spdk_dma_zmalloc(g_io_size_bytes, g_io_align, NULL);
143 	task->iov.iov_len = g_io_size_bytes;
144 	if (task->iov.iov_base == NULL) {
145 		fprintf(stderr, "task->buf spdk_dma_zmalloc failed\n");
146 		exit(1);
147 	}
148 }
149 
150 static int
151 nvme_submit_io(struct perf_task *task, struct ns_worker_ctx *ns_ctx,
152 	       struct ns_entry *entry, uint64_t offset_in_ios)
153 {
154 	uint64_t lba;
155 	int qp_num;
156 
157 	lba = offset_in_ios * entry->io_size_blocks;
158 
159 	qp_num = ns_ctx->last_qpair;
160 	ns_ctx->last_qpair++;
161 	if (ns_ctx->last_qpair == ns_ctx->num_qpairs) {
162 		ns_ctx->last_qpair = 0;
163 	}
164 
165 	if (task->is_read) {
166 		return spdk_nvme_ns_cmd_read(entry->ns, ns_ctx->qpair[qp_num],
167 					     task->iov.iov_base, lba,
168 					     entry->io_size_blocks, io_complete,
169 					     task, entry->io_flags);
170 	}
171 
172 	return spdk_nvme_ns_cmd_write(entry->ns, ns_ctx->qpair[qp_num],
173 				      task->iov.iov_base, lba,
174 				      entry->io_size_blocks, io_complete,
175 				      task, entry->io_flags);
176 }
177 
178 static void
179 nvme_check_io(struct ns_worker_ctx *ns_ctx)
180 {
181 	int i, rc;
182 
183 	for (i = 0; i < ns_ctx->num_qpairs; i++) {
184 		rc = spdk_nvme_qpair_process_completions(ns_ctx->qpair[i], g_max_completions);
185 		/* The transport level qpair is failed and we need to reconnect it. */
186 		if (spdk_unlikely(rc == -ENXIO)) {
187 			rc = spdk_nvme_ctrlr_reconnect_io_qpair(ns_ctx->qpair[i]);
188 			/* successful reconnect */
189 			if (rc == 0) {
190 				continue;
191 			} else if (rc == -ENXIO) {
192 				/* This means the controller is failed. Defer to it to restore the qpair. */
193 				continue;
194 			} else {
195 				/*
196 				 * We were unable to restore the qpair on this attempt. We don't
197 				 * really know why. For naive handling, just keep trying.
198 				 * TODO: add a retry limit, and destroy the qpair after x iterations.
199 				 */
200 				fprintf(stderr, "qpair failed and we were unable to recover it.\n");
201 			}
202 		} else if (spdk_unlikely(rc < 0)) {
203 			fprintf(stderr, "Received an unknown error processing completions.\n");
204 			exit(1);
205 		}
206 	}
207 }
208 
209 /*
210  * TODO: If a controller has multiple namespaces, they could all use the same queue.
211  *  For now, give each namespace/thread combination its own queue.
212  */
213 static int
214 nvme_init_ns_worker_ctx(struct ns_worker_ctx *ns_ctx)
215 {
216 	struct spdk_nvme_io_qpair_opts opts;
217 	struct ns_entry *entry = ns_ctx->entry;
218 	int i;
219 
220 	ns_ctx->num_qpairs = 1;
221 	ns_ctx->qpair = calloc(ns_ctx->num_qpairs, sizeof(struct spdk_nvme_qpair *));
222 	if (!ns_ctx->qpair) {
223 		return -1;
224 	}
225 
226 	spdk_nvme_ctrlr_get_default_io_qpair_opts(entry->ctrlr, &opts, sizeof(opts));
227 	if (opts.io_queue_requests < entry->num_io_requests) {
228 		opts.io_queue_requests = entry->num_io_requests;
229 	}
230 
231 	for (i = 0; i < ns_ctx->num_qpairs; i++) {
232 		ns_ctx->qpair[i] = spdk_nvme_ctrlr_alloc_io_qpair(entry->ctrlr, &opts,
233 				   sizeof(opts));
234 		if (!ns_ctx->qpair[i]) {
235 			printf("ERROR: spdk_nvme_ctrlr_alloc_io_qpair failed\n");
236 			return -1;
237 		}
238 	}
239 
240 	return 0;
241 }
242 
243 static void
244 nvme_cleanup_ns_worker_ctx(struct ns_worker_ctx *ns_ctx)
245 {
246 	int i;
247 
248 	for (i = 0; i < ns_ctx->num_qpairs; i++) {
249 		spdk_nvme_ctrlr_free_io_qpair(ns_ctx->qpair[i]);
250 	}
251 
252 	free(ns_ctx->qpair);
253 }
254 
255 static void
256 build_nvme_name(char *name, size_t length, struct spdk_nvme_ctrlr *ctrlr)
257 {
258 	const struct spdk_nvme_transport_id *trid;
259 
260 	trid = spdk_nvme_ctrlr_get_transport_id(ctrlr);
261 
262 	switch (trid->trtype) {
263 	case SPDK_NVME_TRANSPORT_RDMA:
264 		snprintf(name, length, "RDMA (addr:%s subnqn:%s)", trid->traddr, trid->subnqn);
265 		break;
266 	case SPDK_NVME_TRANSPORT_TCP:
267 		snprintf(name, length, "TCP (addr:%s subnqn:%s)", trid->traddr, trid->subnqn);
268 		break;
269 	case SPDK_NVME_TRANSPORT_CUSTOM:
270 		snprintf(name, length, "CUSTOM (%s)", trid->traddr);
271 		break;
272 	default:
273 		fprintf(stderr, "Unknown transport type %d\n", trid->trtype);
274 		break;
275 	}
276 }
277 
278 static void
279 register_ns(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_ns *ns)
280 {
281 	struct ns_entry *entry;
282 	const struct spdk_nvme_ctrlr_data *cdata;
283 	uint32_t max_xfer_size, entries, sector_size;
284 	uint64_t ns_size;
285 	struct spdk_nvme_io_qpair_opts opts;
286 
287 	cdata = spdk_nvme_ctrlr_get_data(ctrlr);
288 
289 	if (!spdk_nvme_ns_is_active(ns)) {
290 		printf("Controller %-20.20s (%-20.20s): Skipping inactive NS %u\n",
291 		       cdata->mn, cdata->sn,
292 		       spdk_nvme_ns_get_id(ns));
293 		g_warn = true;
294 		return;
295 	}
296 
297 	ns_size = spdk_nvme_ns_get_size(ns);
298 	sector_size = spdk_nvme_ns_get_sector_size(ns);
299 
300 	if (ns_size < g_io_size_bytes || sector_size > g_io_size_bytes) {
301 		printf("WARNING: controller %-20.20s (%-20.20s) ns %u has invalid "
302 		       "ns size %" PRIu64 " / block size %u for I/O size %u\n",
303 		       cdata->mn, cdata->sn, spdk_nvme_ns_get_id(ns),
304 		       ns_size, spdk_nvme_ns_get_sector_size(ns), g_io_size_bytes);
305 		g_warn = true;
306 		return;
307 	}
308 
309 	max_xfer_size = spdk_nvme_ns_get_max_io_xfer_size(ns);
310 	spdk_nvme_ctrlr_get_default_io_qpair_opts(ctrlr, &opts, sizeof(opts));
311 	/* NVMe driver may add additional entries based on
312 	 * stripe size and maximum transfer size, we assume
313 	 * 1 more entry be used for stripe.
314 	 */
315 	entries = (g_io_size_bytes - 1) / max_xfer_size + 2;
316 	if ((g_queue_depth * entries) > opts.io_queue_size) {
317 		printf("controller IO queue size %u less than required\n",
318 		       opts.io_queue_size);
319 		printf("Consider using lower queue depth or small IO size because "
320 		       "IO requests may be queued at the NVMe driver.\n");
321 		g_warn = true;
322 	}
323 	/* For requests which have children requests, parent request itself
324 	 * will also occupy 1 entry.
325 	 */
326 	entries += 1;
327 
328 	entry = calloc(1, sizeof(struct ns_entry));
329 	if (entry == NULL) {
330 		perror("ns_entry malloc");
331 		exit(1);
332 	}
333 
334 	entry->ctrlr = ctrlr;
335 	entry->ns = ns;
336 	entry->num_io_requests = g_queue_depth * entries;
337 
338 	entry->size_in_ios = ns_size / g_io_size_bytes;
339 	entry->io_size_blocks = g_io_size_bytes / sector_size;
340 
341 	entry->block_size = spdk_nvme_ns_get_sector_size(ns);
342 
343 
344 	if (g_max_io_size_blocks < entry->io_size_blocks) {
345 		g_max_io_size_blocks = entry->io_size_blocks;
346 	}
347 
348 	build_nvme_name(entry->name, sizeof(entry->name), ctrlr);
349 
350 	g_num_namespaces++;
351 	TAILQ_INSERT_TAIL(&g_namespaces, entry, link);
352 }
353 
354 static void
355 unregister_namespaces(void)
356 {
357 	struct ns_entry *entry, *tmp;
358 
359 	TAILQ_FOREACH_SAFE(entry, &g_namespaces, link, tmp) {
360 		TAILQ_REMOVE(&g_namespaces, entry, link);
361 		free(entry);
362 	}
363 }
364 
365 static void
366 register_ctrlr(struct spdk_nvme_ctrlr *ctrlr, struct trid_entry *trid_entry)
367 {
368 	struct spdk_nvme_ns *ns;
369 	struct ctrlr_entry *entry = calloc(1, sizeof(struct ctrlr_entry));
370 	const struct spdk_nvme_transport_id *ctrlr_trid;
371 	uint32_t nsid;
372 
373 	if (entry == NULL) {
374 		perror("ctrlr_entry malloc");
375 		exit(1);
376 	}
377 
378 	ctrlr_trid = spdk_nvme_ctrlr_get_transport_id(ctrlr);
379 	assert(ctrlr_trid != NULL);
380 
381 	/* each controller needs a unique failover trid. */
382 	entry->failover_trid = trid_entry->failover_trid;
383 
384 	/*
385 	 * Users are allowed to leave the trid subnqn blank or specify a discovery controller subnqn.
386 	 * In those cases, the controller subnqn will not equal the trid_entry subnqn and, by association,
387 	 * the failover_trid subnqn.
388 	 * When we do failover, we want to reconnect to the same nqn so explicitly set the failover nqn to
389 	 * the ctrlr nqn here.
390 	 */
391 	snprintf(entry->failover_trid.subnqn, SPDK_NVMF_NQN_MAX_LEN + 1, "%s", ctrlr_trid->subnqn);
392 
393 
394 	build_nvme_name(entry->name, sizeof(entry->name), ctrlr);
395 
396 	entry->ctrlr = ctrlr;
397 	entry->trtype = trid_entry->trid.trtype;
398 	TAILQ_INSERT_TAIL(&g_controllers, entry, link);
399 
400 	for (nsid = spdk_nvme_ctrlr_get_first_active_ns(ctrlr);
401 	     nsid != 0; nsid = spdk_nvme_ctrlr_get_next_active_ns(ctrlr, nsid)) {
402 		ns = spdk_nvme_ctrlr_get_ns(ctrlr, nsid);
403 		if (ns == NULL) {
404 			continue;
405 		}
406 		register_ns(ctrlr, ns);
407 	}
408 }
409 
410 static __thread unsigned int seed = 0;
411 
412 static inline void
413 submit_single_io(struct perf_task *task)
414 {
415 	uint64_t		offset_in_ios;
416 	int			rc;
417 	struct ns_worker_ctx	*ns_ctx = task->ns_ctx;
418 	struct ns_entry		*entry = ns_ctx->entry;
419 
420 	if (g_is_random) {
421 		offset_in_ios = rand_r(&seed) % entry->size_in_ios;
422 	} else {
423 		offset_in_ios = ns_ctx->offset_in_ios++;
424 		if (ns_ctx->offset_in_ios == entry->size_in_ios) {
425 			ns_ctx->offset_in_ios = 0;
426 		}
427 	}
428 
429 	if ((g_rw_percentage == 100) ||
430 	    (g_rw_percentage != 0 && ((rand_r(&seed) % 100) < g_rw_percentage))) {
431 		task->is_read = true;
432 	} else {
433 		task->is_read = false;
434 	}
435 
436 	rc = nvme_submit_io(task, ns_ctx, entry, offset_in_ios);
437 
438 	if (spdk_unlikely(rc != 0)) {
439 		fprintf(stderr, "starting I/O failed\n");
440 	} else {
441 		ns_ctx->current_queue_depth++;
442 	}
443 }
444 
445 static inline void
446 task_complete(struct perf_task *task)
447 {
448 	struct ns_worker_ctx	*ns_ctx;
449 
450 	ns_ctx = task->ns_ctx;
451 	ns_ctx->current_queue_depth--;
452 	ns_ctx->io_completed++;
453 
454 	/*
455 	 * is_draining indicates when time has expired for the test run
456 	 * and we are just waiting for the previously submitted I/O
457 	 * to complete.  In this case, do not submit a new I/O to replace
458 	 * the one just completed.
459 	 */
460 	if (spdk_unlikely(ns_ctx->is_draining)) {
461 		spdk_dma_free(task->iov.iov_base);
462 		free(task);
463 	} else {
464 		submit_single_io(task);
465 	}
466 }
467 
468 static void
469 io_complete(void *ctx, const struct spdk_nvme_cpl *cpl)
470 {
471 	struct perf_task *task = ctx;
472 
473 	if (spdk_unlikely(spdk_nvme_cpl_is_error(cpl))) {
474 		fprintf(stderr, "%s completed with error (sct=%d, sc=%d)\n",
475 			task->is_read ? "Read" : "Write",
476 			cpl->status.sct, cpl->status.sc);
477 	}
478 
479 	task_complete(task);
480 }
481 
482 static void
483 check_io(struct ns_worker_ctx *ns_ctx)
484 {
485 	nvme_check_io(ns_ctx);
486 }
487 
488 static struct perf_task *
489 allocate_task(struct ns_worker_ctx *ns_ctx, int queue_depth)
490 {
491 	struct perf_task *task;
492 
493 	task = calloc(1, sizeof(*task));
494 	if (task == NULL) {
495 		fprintf(stderr, "Out of memory allocating tasks\n");
496 		exit(1);
497 	}
498 
499 	nvme_setup_payload(task);
500 
501 	task->ns_ctx = ns_ctx;
502 
503 	return task;
504 }
505 
506 static void
507 submit_io(struct ns_worker_ctx *ns_ctx, int queue_depth)
508 {
509 	struct perf_task *task;
510 
511 	while (queue_depth-- > 0) {
512 		task = allocate_task(ns_ctx, queue_depth);
513 		submit_single_io(task);
514 	}
515 }
516 
517 static int
518 work_fn(void *arg)
519 {
520 	uint64_t tsc_end;
521 	struct worker_thread *worker = (struct worker_thread *)arg;
522 	struct ns_worker_ctx *ns_ctx = NULL;
523 	uint32_t unfinished_ns_ctx;
524 
525 	printf("Starting thread on core %u\n", worker->lcore);
526 
527 	/* Allocate queue pairs for each namespace. */
528 	TAILQ_FOREACH(ns_ctx, &worker->ns_ctx, link) {
529 		if (nvme_init_ns_worker_ctx(ns_ctx) != 0) {
530 			printf("ERROR: init_ns_worker_ctx() failed\n");
531 			return 1;
532 		}
533 	}
534 
535 	tsc_end = spdk_get_ticks() + g_time_in_sec * g_tsc_rate;
536 
537 	/* Submit initial I/O for each namespace. */
538 	TAILQ_FOREACH(ns_ctx, &worker->ns_ctx, link) {
539 		submit_io(ns_ctx, g_queue_depth);
540 	}
541 
542 	while (1) {
543 		/*
544 		 * Check for completed I/O for each controller. A new
545 		 * I/O will be submitted in the io_complete callback
546 		 * to replace each I/O that is completed.
547 		 */
548 		TAILQ_FOREACH(ns_ctx, &worker->ns_ctx, link) {
549 			check_io(ns_ctx);
550 		}
551 
552 		if (spdk_get_ticks() > tsc_end) {
553 			break;
554 		}
555 	}
556 
557 	/* drain the io of each ns_ctx in round robin to make the fairness */
558 	do {
559 		unfinished_ns_ctx = 0;
560 		TAILQ_FOREACH(ns_ctx, &worker->ns_ctx, link) {
561 			/* first time will enter into this if case */
562 			if (!ns_ctx->is_draining) {
563 				ns_ctx->is_draining = true;
564 			}
565 
566 			if (ns_ctx->current_queue_depth > 0) {
567 				check_io(ns_ctx);
568 				if (ns_ctx->current_queue_depth == 0) {
569 					nvme_cleanup_ns_worker_ctx(ns_ctx);
570 				} else {
571 					unfinished_ns_ctx++;
572 				}
573 			}
574 		}
575 	} while (unfinished_ns_ctx > 0);
576 
577 	return 0;
578 }
579 
580 static void usage(char *program_name)
581 {
582 	printf("%s options", program_name);
583 	printf("\n");
584 	printf("\t[-q io depth]\n");
585 	printf("\t[-o io size in bytes]\n");
586 	printf("\t[-w io pattern type, must be one of\n");
587 	printf("\t\t(read, write, randread, randwrite, rw, randrw)]\n");
588 	printf("\t[-M rwmixread (100 for reads, 0 for writes)]\n");
589 	printf("\t[-t time in seconds]\n");
590 	printf("\t[-c core mask for I/O submission/completion.]\n");
591 	printf("\t\t(default: 1)\n");
592 	printf("\t[-r Transport ID for NVMeoF]\n");
593 	printf("\t Format: 'key:value [key:value] ...'\n");
594 	printf("\t Keys:\n");
595 	printf("\t  trtype      Transport type (e.g. RDMA)\n");
596 	printf("\t  adrfam      Address family (e.g. IPv4, IPv6)\n");
597 	printf("\t  traddr      Transport address (e.g. 192.168.100.8 for RDMA)\n");
598 	printf("\t  trsvcid     Transport service identifier (e.g. 4420)\n");
599 	printf("\t  subnqn      Subsystem NQN (default: %s)\n", SPDK_NVMF_DISCOVERY_NQN);
600 	printf("\t  alt_traddr  (Optional) Alternative Transport address for failover.\n");
601 	printf("\t Example: -r 'trtype:RDMA adrfam:IPv4 traddr:192.168.100.8 trsvcid:4420' for NVMeoF\n");
602 	printf("\t[-k keep alive timeout period in millisecond]\n");
603 	printf("\t[-s DPDK huge memory size in MB.]\n");
604 	printf("\t[-m max completions per poll]\n");
605 	printf("\t\t(default: 0 - unlimited)\n");
606 	printf("\t[-i shared memory group ID]\n");
607 	printf("\t[-A transport ACK timeout]\n");
608 	printf("\t[-R transport retry count]\n");
609 	printf("\t");
610 	spdk_log_usage(stdout, "-T");
611 #ifdef DEBUG
612 	printf("\t[-G enable debug logging]\n");
613 #else
614 	printf("\t[-G enable debug logging (flag disabled, must reconfigure with --enable-debug)\n");
615 #endif
616 }
617 
618 static void
619 unregister_trids(void)
620 {
621 	struct trid_entry *trid_entry, *tmp;
622 
623 	TAILQ_FOREACH_SAFE(trid_entry, &g_trid_list, tailq, tmp) {
624 		TAILQ_REMOVE(&g_trid_list, trid_entry, tailq);
625 		free(trid_entry);
626 	}
627 }
628 
629 static int
630 add_trid(const char *trid_str)
631 {
632 	struct trid_entry *trid_entry;
633 	struct spdk_nvme_transport_id *trid;
634 	char *alt_traddr;
635 	int len;
636 
637 	trid_entry = calloc(1, sizeof(*trid_entry));
638 	if (trid_entry == NULL) {
639 		return -1;
640 	}
641 
642 	trid = &trid_entry->trid;
643 	snprintf(trid->subnqn, sizeof(trid->subnqn), "%s", SPDK_NVMF_DISCOVERY_NQN);
644 
645 	if (spdk_nvme_transport_id_parse(trid, trid_str) != 0) {
646 		fprintf(stderr, "Invalid transport ID format '%s'\n", trid_str);
647 		free(trid_entry);
648 		return 1;
649 	}
650 
651 	trid_entry->failover_trid = trid_entry->trid;
652 
653 	alt_traddr = strcasestr(trid_str, "alt_traddr:");
654 	if (alt_traddr) {
655 		alt_traddr += strlen("alt_traddr:");
656 		len = strcspn(alt_traddr, " \t\n");
657 		if (len > SPDK_NVMF_TRADDR_MAX_LEN) {
658 			fprintf(stderr, "The failover traddr %s is too long.\n", alt_traddr);
659 			free(trid_entry);
660 			return -1;
661 		}
662 		snprintf(trid_entry->failover_trid.traddr, SPDK_NVMF_TRADDR_MAX_LEN + 1, "%s", alt_traddr);
663 	}
664 
665 	TAILQ_INSERT_TAIL(&g_trid_list, trid_entry, tailq);
666 	return 0;
667 }
668 
669 static int
670 parse_args(int argc, char **argv)
671 {
672 	struct trid_entry *trid_entry, *trid_entry_tmp;
673 	const char *workload_type;
674 	int op;
675 	bool mix_specified = false;
676 	long int val;
677 	int rc;
678 
679 	/* default value */
680 	g_queue_depth = 0;
681 	g_io_size_bytes = 0;
682 	workload_type = NULL;
683 	g_time_in_sec = 0;
684 	g_rw_percentage = -1;
685 	g_core_mask = NULL;
686 	g_max_completions = 0;
687 
688 	while ((op = getopt(argc, argv, "c:gm:o:q:r:k:s:t:w:A:GM:R:T:")) != -1) {
689 		switch (op) {
690 		case 'm':
691 		case 'o':
692 		case 'q':
693 		case 'k':
694 		case 's':
695 		case 't':
696 		case 'A':
697 		case 'M':
698 		case 'R':
699 			val = spdk_strtol(optarg, 10);
700 			if (val < 0) {
701 				fprintf(stderr, "Converting a string to integer failed\n");
702 				return val;
703 			}
704 			switch (op) {
705 			case 'm':
706 				g_max_completions = val;
707 				break;
708 			case 'o':
709 				g_io_size_bytes = val;
710 				break;
711 			case 'q':
712 				g_queue_depth = val;
713 				break;
714 			case 'k':
715 				g_keep_alive_timeout_in_ms = val;
716 				break;
717 			case 's':
718 				g_dpdk_mem = val;
719 				break;
720 			case 't':
721 				g_time_in_sec = val;
722 				break;
723 			case 'A':
724 				g_transport_ack_timeout = val;
725 				break;
726 			case 'M':
727 				g_rw_percentage = val;
728 				mix_specified = true;
729 				break;
730 			case 'R':
731 				g_transport_retry_count = val;
732 				break;
733 			}
734 			break;
735 		case 'c':
736 			g_core_mask = optarg;
737 			break;
738 		case 'g':
739 			g_dpdk_mem_single_seg = true;
740 			break;
741 		case 'r':
742 			if (add_trid(optarg)) {
743 				usage(argv[0]);
744 				return 1;
745 			}
746 			break;
747 		case 'w':
748 			workload_type = optarg;
749 			break;
750 		case 'G':
751 #ifndef DEBUG
752 			fprintf(stderr, "%s must be configured with --enable-debug for -G flag\n",
753 				argv[0]);
754 			usage(argv[0]);
755 			return 1;
756 #else
757 			spdk_log_set_flag("nvme");
758 			spdk_log_set_print_level(SPDK_LOG_DEBUG);
759 			break;
760 #endif
761 		case 'T':
762 			rc = spdk_log_set_flag(optarg);
763 			if (rc < 0) {
764 				fprintf(stderr, "unknown flag\n");
765 				usage(argv[0]);
766 				exit(EXIT_FAILURE);
767 			}
768 #ifdef DEBUG
769 			spdk_log_set_print_level(SPDK_LOG_DEBUG);
770 #endif
771 			break;
772 		default:
773 			usage(argv[0]);
774 			return 1;
775 		}
776 	}
777 
778 	if (!g_queue_depth) {
779 		usage(argv[0]);
780 		return 1;
781 	}
782 	if (!g_io_size_bytes) {
783 		usage(argv[0]);
784 		return 1;
785 	}
786 	if (!workload_type) {
787 		usage(argv[0]);
788 		return 1;
789 	}
790 	if (!g_time_in_sec) {
791 		usage(argv[0]);
792 		return 1;
793 	}
794 
795 	if (strcmp(workload_type, "read") &&
796 	    strcmp(workload_type, "write") &&
797 	    strcmp(workload_type, "randread") &&
798 	    strcmp(workload_type, "randwrite") &&
799 	    strcmp(workload_type, "rw") &&
800 	    strcmp(workload_type, "randrw")) {
801 		fprintf(stderr,
802 			"io pattern type must be one of\n"
803 			"(read, write, randread, randwrite, rw, randrw)\n");
804 		return 1;
805 	}
806 
807 	if (!strcmp(workload_type, "read") ||
808 	    !strcmp(workload_type, "randread")) {
809 		g_rw_percentage = 100;
810 	}
811 
812 	if (!strcmp(workload_type, "write") ||
813 	    !strcmp(workload_type, "randwrite")) {
814 		g_rw_percentage = 0;
815 	}
816 
817 	if (!strcmp(workload_type, "read") ||
818 	    !strcmp(workload_type, "randread") ||
819 	    !strcmp(workload_type, "write") ||
820 	    !strcmp(workload_type, "randwrite")) {
821 		if (mix_specified) {
822 			fprintf(stderr, "Ignoring -M option... Please use -M option"
823 				" only when using rw or randrw.\n");
824 		}
825 	}
826 
827 	if (!strcmp(workload_type, "rw") ||
828 	    !strcmp(workload_type, "randrw")) {
829 		if (g_rw_percentage < 0 || g_rw_percentage > 100) {
830 			fprintf(stderr,
831 				"-M must be specified to value from 0 to 100 "
832 				"for rw or randrw.\n");
833 			return 1;
834 		}
835 	}
836 
837 	if (!strcmp(workload_type, "read") ||
838 	    !strcmp(workload_type, "write") ||
839 	    !strcmp(workload_type, "rw")) {
840 		g_is_random = 0;
841 	} else {
842 		g_is_random = 1;
843 	}
844 
845 	if (TAILQ_EMPTY(&g_trid_list)) {
846 		fprintf(stderr, "You must specify at least one fabrics TRID.\n");
847 		return -1;
848 	}
849 
850 	/* check whether there is local PCIe type and fail. */
851 	TAILQ_FOREACH_SAFE(trid_entry, &g_trid_list, tailq, trid_entry_tmp) {
852 		if (trid_entry->trid.trtype == SPDK_NVME_TRANSPORT_PCIE) {
853 			fprintf(stderr, "This application was not intended to be run on PCIe controllers.\n");
854 			return 1;
855 		}
856 	}
857 
858 	return 0;
859 }
860 
861 static int
862 register_workers(void)
863 {
864 	uint32_t i;
865 	struct worker_thread *worker;
866 
867 	SPDK_ENV_FOREACH_CORE(i) {
868 		worker = calloc(1, sizeof(*worker));
869 		if (worker == NULL) {
870 			fprintf(stderr, "Unable to allocate worker\n");
871 			return -1;
872 		}
873 
874 		TAILQ_INIT(&worker->ns_ctx);
875 		worker->lcore = i;
876 		TAILQ_INSERT_TAIL(&g_workers, worker, link);
877 		g_num_workers++;
878 	}
879 
880 	return 0;
881 }
882 
883 static void
884 unregister_workers(void)
885 {
886 	struct worker_thread *worker, *tmp_worker;
887 	struct ns_worker_ctx *ns_ctx, *tmp_ns_ctx;
888 
889 	/* Free namespace context and worker thread */
890 	TAILQ_FOREACH_SAFE(worker, &g_workers, link, tmp_worker) {
891 		TAILQ_REMOVE(&g_workers, worker, link);
892 		TAILQ_FOREACH_SAFE(ns_ctx, &worker->ns_ctx, link, tmp_ns_ctx) {
893 			TAILQ_REMOVE(&worker->ns_ctx, ns_ctx, link);
894 			free(ns_ctx);
895 		}
896 
897 		free(worker);
898 	}
899 }
900 
901 static bool
902 probe_cb(void *cb_ctx, const struct spdk_nvme_transport_id *trid,
903 	 struct spdk_nvme_ctrlr_opts *opts)
904 {
905 	/* These should have been weeded out earlier. */
906 	assert(trid->trtype != SPDK_NVME_TRANSPORT_PCIE);
907 
908 	printf("Attaching to NVMe over Fabrics controller at %s:%s: %s\n",
909 	       trid->traddr, trid->trsvcid,
910 	       trid->subnqn);
911 
912 	/* Set io_queue_size to UINT16_MAX, NVMe driver
913 	 * will then reduce this to MQES to maximize
914 	 * the io_queue_size as much as possible.
915 	 */
916 	opts->io_queue_size = UINT16_MAX;
917 
918 	opts->keep_alive_timeout_ms = spdk_max(opts->keep_alive_timeout_ms,
919 					       g_keep_alive_timeout_in_ms);
920 
921 	opts->transport_retry_count = g_transport_retry_count;
922 	opts->transport_ack_timeout = g_transport_ack_timeout;
923 
924 	return true;
925 }
926 
927 static void
928 attach_cb(void *cb_ctx, const struct spdk_nvme_transport_id *trid,
929 	  struct spdk_nvme_ctrlr *ctrlr, const struct spdk_nvme_ctrlr_opts *opts)
930 {
931 	struct trid_entry	*trid_entry = cb_ctx;
932 
933 	printf("Attached to NVMe over Fabrics controller at %s:%s: %s\n",
934 	       trid->traddr, trid->trsvcid,
935 	       trid->subnqn);
936 
937 	register_ctrlr(ctrlr, trid_entry);
938 }
939 
940 static int
941 register_controllers(void)
942 {
943 	struct trid_entry *trid_entry;
944 
945 	printf("Initializing NVMe Controllers\n");
946 
947 	TAILQ_FOREACH(trid_entry, &g_trid_list, tailq) {
948 		if (spdk_nvme_probe(&trid_entry->trid, trid_entry, probe_cb, attach_cb, NULL) != 0) {
949 			fprintf(stderr, "spdk_nvme_probe() failed for transport address '%s'\n",
950 				trid_entry->trid.traddr);
951 			return -1;
952 		}
953 	}
954 
955 	return 0;
956 }
957 
958 static void
959 unregister_controllers(void)
960 {
961 	struct ctrlr_entry *entry, *tmp;
962 	struct spdk_nvme_detach_ctx *detach_ctx = NULL;
963 
964 	TAILQ_FOREACH_SAFE(entry, &g_controllers, link, tmp) {
965 		TAILQ_REMOVE(&g_controllers, entry, link);
966 		spdk_nvme_detach_async(entry->ctrlr, &detach_ctx);
967 		free(entry);
968 	}
969 
970 	while (detach_ctx && spdk_nvme_detach_poll_async(detach_ctx) == -EAGAIN) {
971 		;
972 	}
973 }
974 
975 static int
976 associate_workers_with_ns(void)
977 {
978 	struct ns_entry		*entry = TAILQ_FIRST(&g_namespaces);
979 	struct worker_thread	*worker = TAILQ_FIRST(&g_workers);
980 	struct ns_worker_ctx	*ns_ctx;
981 	int			i, count;
982 
983 	count = g_num_namespaces > g_num_workers ? g_num_namespaces : g_num_workers;
984 
985 	for (i = 0; i < count; i++) {
986 		if (entry == NULL) {
987 			break;
988 		}
989 
990 		ns_ctx = calloc(1, sizeof(struct ns_worker_ctx));
991 		if (!ns_ctx) {
992 			return -1;
993 		}
994 
995 		printf("Associating %s with lcore %d\n", entry->name, worker->lcore);
996 		ns_ctx->entry = entry;
997 
998 		TAILQ_INSERT_TAIL(&worker->ns_ctx, ns_ctx, link);
999 
1000 		worker = TAILQ_NEXT(worker, link);
1001 		if (worker == NULL) {
1002 			worker = TAILQ_FIRST(&g_workers);
1003 		}
1004 
1005 		entry = TAILQ_NEXT(entry, link);
1006 		if (entry == NULL) {
1007 			entry = TAILQ_FIRST(&g_namespaces);
1008 		}
1009 
1010 	}
1011 
1012 	return 0;
1013 }
1014 
1015 static void *
1016 nvme_poll_ctrlrs(void *arg)
1017 {
1018 	struct ctrlr_entry			*entry;
1019 	const struct spdk_nvme_transport_id	*old_trid;
1020 	int					oldstate;
1021 	int					rc;
1022 
1023 
1024 	spdk_unaffinitize_thread();
1025 
1026 	while (true) {
1027 		pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate);
1028 
1029 		TAILQ_FOREACH(entry, &g_controllers, link) {
1030 			rc = spdk_nvme_ctrlr_process_admin_completions(entry->ctrlr);
1031 			/* This controller has encountered a failure at the transport level. reset it. */
1032 			if (rc == -ENXIO) {
1033 				if (entry->num_resets == 0) {
1034 					old_trid = spdk_nvme_ctrlr_get_transport_id(entry->ctrlr);
1035 					fprintf(stderr, "A controller has encountered a failure and is being reset.\n");
1036 					if (spdk_nvme_transport_id_compare(old_trid, &entry->failover_trid)) {
1037 						fprintf(stderr, "Resorting to new failover address %s\n", entry->failover_trid.traddr);
1038 						spdk_nvme_ctrlr_fail(entry->ctrlr);
1039 						rc = spdk_nvme_ctrlr_set_trid(entry->ctrlr, &entry->failover_trid);
1040 						if (rc != 0) {
1041 							fprintf(stderr, "Unable to fail over to back up trid.\n");
1042 						}
1043 					}
1044 				}
1045 
1046 				rc = spdk_nvme_ctrlr_reset(entry->ctrlr);
1047 				if (rc != 0) {
1048 					entry->num_resets++;
1049 					fprintf(stderr, "Unable to reset the controller.\n");
1050 
1051 					if (entry->num_resets > g_max_ctrlr_resets) {
1052 						fprintf(stderr, "Controller cannot be recovered. Exiting.\n");
1053 						exit(1);
1054 					}
1055 				} else {
1056 					fprintf(stderr, "Controller properly reset.\n");
1057 				}
1058 			}
1059 		}
1060 
1061 		pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &oldstate);
1062 
1063 		/* This is a pthread cancellation point and cannot be removed. */
1064 		sleep(1);
1065 	}
1066 
1067 	return NULL;
1068 }
1069 
1070 int main(int argc, char **argv)
1071 {
1072 	int rc;
1073 	struct worker_thread *worker, *main_worker;
1074 	unsigned main_core;
1075 	struct spdk_env_opts opts;
1076 	pthread_t thread_id = 0;
1077 
1078 	rc = parse_args(argc, argv);
1079 	if (rc != 0) {
1080 		return rc;
1081 	}
1082 
1083 	spdk_env_opts_init(&opts);
1084 	opts.name = "reconnect";
1085 	if (g_core_mask) {
1086 		opts.core_mask = g_core_mask;
1087 	}
1088 
1089 	if (g_dpdk_mem) {
1090 		opts.mem_size = g_dpdk_mem;
1091 	}
1092 	opts.hugepage_single_segments = g_dpdk_mem_single_seg;
1093 	if (spdk_env_init(&opts) < 0) {
1094 		fprintf(stderr, "Unable to initialize SPDK env\n");
1095 		rc = 1;
1096 		goto cleanup;
1097 	}
1098 
1099 	g_tsc_rate = spdk_get_ticks_hz();
1100 
1101 	if (register_workers() != 0) {
1102 		rc = 1;
1103 		goto cleanup;
1104 	}
1105 
1106 	if (register_controllers() != 0) {
1107 		rc = 1;
1108 		goto cleanup;
1109 	}
1110 
1111 	if (g_warn) {
1112 		printf("WARNING: Some requested NVMe devices were skipped\n");
1113 	}
1114 
1115 	if (g_num_namespaces == 0) {
1116 		fprintf(stderr, "No valid NVMe controllers found\n");
1117 		goto cleanup;
1118 	}
1119 
1120 	rc = pthread_create(&thread_id, NULL, &nvme_poll_ctrlrs, NULL);
1121 	if (rc != 0) {
1122 		fprintf(stderr, "Unable to spawn a thread to poll admin queues.\n");
1123 		goto cleanup;
1124 	}
1125 
1126 	if (associate_workers_with_ns() != 0) {
1127 		rc = 1;
1128 		goto cleanup;
1129 	}
1130 
1131 	printf("Initialization complete. Launching workers.\n");
1132 
1133 	/* Launch all of the secondary workers */
1134 	main_core = spdk_env_get_current_core();
1135 	main_worker = NULL;
1136 	TAILQ_FOREACH(worker, &g_workers, link) {
1137 		if (worker->lcore != main_core) {
1138 			spdk_env_thread_launch_pinned(worker->lcore, work_fn, worker);
1139 		} else {
1140 			assert(main_worker == NULL);
1141 			main_worker = worker;
1142 		}
1143 	}
1144 
1145 	assert(main_worker != NULL);
1146 	rc = work_fn(main_worker);
1147 
1148 	spdk_env_thread_wait_all();
1149 
1150 cleanup:
1151 	if (thread_id && pthread_cancel(thread_id) == 0) {
1152 		pthread_join(thread_id, NULL);
1153 	}
1154 	unregister_trids();
1155 	unregister_namespaces();
1156 	unregister_controllers();
1157 	unregister_workers();
1158 
1159 	if (rc != 0) {
1160 		fprintf(stderr, "%s: errors occured\n", argv[0]);
1161 		/*
1162 		 * return a generic error to the caller. This allows us to
1163 		 * distinguish between a failure in the script and something
1164 		 * like a segfault or an invalid access which causes the program
1165 		 * to crash.
1166 		 */
1167 		rc = 1;
1168 	}
1169 
1170 	return rc;
1171 }
1172