]>
Commit | Line | Data |
---|---|---|
7c673cae FG |
1 | /*- |
2 | * BSD LICENSE | |
3 | * | |
4 | * Copyright (c) Intel Corporation. | |
5 | * All rights reserved. | |
6 | * | |
9f95a23c TL |
7 | * Copyright (c) 2019 Mellanox Technologies LTD. All rights reserved. |
8 | * | |
7c673cae FG |
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 | ||
11fdf7f2 | 36 | #include "spdk/stdinc.h" |
7c673cae | 37 | |
11fdf7f2 | 38 | #include "spdk/env.h" |
7c673cae FG |
39 | #include "spdk/fd.h" |
40 | #include "spdk/nvme.h" | |
7c673cae FG |
41 | #include "spdk/queue.h" |
42 | #include "spdk/string.h" | |
43 | #include "spdk/nvme_intel.h" | |
11fdf7f2 TL |
44 | #include "spdk/histogram_data.h" |
45 | #include "spdk/endian.h" | |
9f95a23c TL |
46 | #include "spdk/dif.h" |
47 | #include "spdk/util.h" | |
48 | #include "spdk/log.h" | |
49 | #include "spdk/likely.h" | |
7c673cae FG |
50 | |
51 | #if HAVE_LIBAIO | |
52 | #include <libaio.h> | |
7c673cae FG |
53 | #endif |
54 | ||
55 | struct ctrlr_entry { | |
56 | struct spdk_nvme_ctrlr *ctrlr; | |
9f95a23c | 57 | enum spdk_nvme_transport_type trtype; |
7c673cae | 58 | struct spdk_nvme_intel_rw_latency_page *latency_page; |
9f95a23c TL |
59 | |
60 | struct spdk_nvme_qpair **unused_qpairs; | |
61 | ||
7c673cae FG |
62 | struct ctrlr_entry *next; |
63 | char name[1024]; | |
64 | }; | |
65 | ||
66 | enum entry_type { | |
67 | ENTRY_TYPE_NVME_NS, | |
68 | ENTRY_TYPE_AIO_FILE, | |
69 | }; | |
70 | ||
9f95a23c TL |
71 | struct ns_fn_table; |
72 | ||
7c673cae FG |
73 | struct ns_entry { |
74 | enum entry_type type; | |
9f95a23c | 75 | const struct ns_fn_table *fn_table; |
7c673cae FG |
76 | |
77 | union { | |
78 | struct { | |
79 | struct spdk_nvme_ctrlr *ctrlr; | |
80 | struct spdk_nvme_ns *ns; | |
81 | } nvme; | |
82 | #if HAVE_LIBAIO | |
83 | struct { | |
84 | int fd; | |
85 | } aio; | |
86 | #endif | |
87 | } u; | |
88 | ||
89 | struct ns_entry *next; | |
90 | uint32_t io_size_blocks; | |
11fdf7f2 | 91 | uint32_t num_io_requests; |
7c673cae | 92 | uint64_t size_in_ios; |
9f95a23c TL |
93 | uint32_t block_size; |
94 | uint32_t md_size; | |
95 | bool md_interleave; | |
96 | bool pi_loc; | |
97 | enum spdk_nvme_pi_type pi_type; | |
11fdf7f2 | 98 | uint32_t io_flags; |
7c673cae | 99 | char name[1024]; |
11fdf7f2 TL |
100 | }; |
101 | ||
102 | static const double g_latency_cutoffs[] = { | |
103 | 0.01, | |
104 | 0.10, | |
105 | 0.25, | |
106 | 0.50, | |
107 | 0.75, | |
108 | 0.90, | |
109 | 0.95, | |
110 | 0.98, | |
111 | 0.99, | |
112 | 0.995, | |
113 | 0.999, | |
114 | 0.9999, | |
115 | 0.99999, | |
116 | 0.999999, | |
117 | 0.9999999, | |
118 | -1, | |
7c673cae FG |
119 | }; |
120 | ||
121 | struct ns_worker_ctx { | |
122 | struct ns_entry *entry; | |
123 | uint64_t io_completed; | |
124 | uint64_t total_tsc; | |
125 | uint64_t min_tsc; | |
126 | uint64_t max_tsc; | |
127 | uint64_t current_queue_depth; | |
128 | uint64_t offset_in_ios; | |
129 | bool is_draining; | |
130 | ||
131 | union { | |
132 | struct { | |
9f95a23c TL |
133 | int num_qpairs; |
134 | struct spdk_nvme_qpair **qpair; | |
135 | int last_qpair; | |
7c673cae FG |
136 | } nvme; |
137 | ||
138 | #if HAVE_LIBAIO | |
139 | struct { | |
140 | struct io_event *events; | |
141 | io_context_t ctx; | |
142 | } aio; | |
143 | #endif | |
144 | } u; | |
145 | ||
146 | struct ns_worker_ctx *next; | |
11fdf7f2 TL |
147 | |
148 | struct spdk_histogram_data *histogram; | |
7c673cae FG |
149 | }; |
150 | ||
151 | struct perf_task { | |
152 | struct ns_worker_ctx *ns_ctx; | |
9f95a23c TL |
153 | struct iovec iov; |
154 | struct iovec md_iov; | |
7c673cae | 155 | uint64_t submit_tsc; |
11fdf7f2 | 156 | bool is_read; |
9f95a23c | 157 | struct spdk_dif_ctx dif_ctx; |
7c673cae FG |
158 | #if HAVE_LIBAIO |
159 | struct iocb iocb; | |
160 | #endif | |
161 | }; | |
162 | ||
163 | struct worker_thread { | |
11fdf7f2 | 164 | struct ns_worker_ctx *ns_ctx; |
7c673cae FG |
165 | struct worker_thread *next; |
166 | unsigned lcore; | |
167 | }; | |
168 | ||
9f95a23c TL |
169 | struct ns_fn_table { |
170 | void (*setup_payload)(struct perf_task *task, uint8_t pattern); | |
171 | ||
172 | int (*submit_io)(struct perf_task *task, struct ns_worker_ctx *ns_ctx, | |
173 | struct ns_entry *entry, uint64_t offset_in_ios); | |
174 | ||
175 | void (*check_io)(struct ns_worker_ctx *ns_ctx); | |
176 | ||
177 | void (*verify_io)(struct perf_task *task, struct ns_entry *entry); | |
178 | ||
179 | int (*init_ns_worker_ctx)(struct ns_worker_ctx *ns_ctx); | |
180 | ||
181 | void (*cleanup_ns_worker_ctx)(struct ns_worker_ctx *ns_ctx); | |
182 | }; | |
183 | ||
7c673cae FG |
184 | static int g_outstanding_commands; |
185 | ||
11fdf7f2 TL |
186 | static bool g_latency_ssd_tracking_enable = false; |
187 | static int g_latency_sw_tracking_level = 0; | |
7c673cae FG |
188 | |
189 | static struct ctrlr_entry *g_controllers = NULL; | |
11fdf7f2 | 190 | static int g_controllers_found = 0; |
7c673cae FG |
191 | static struct ns_entry *g_namespaces = NULL; |
192 | static int g_num_namespaces = 0; | |
193 | static struct worker_thread *g_workers = NULL; | |
194 | static int g_num_workers = 0; | |
195 | ||
196 | static uint64_t g_tsc_rate; | |
197 | ||
198 | static uint32_t g_io_align = 0x200; | |
199 | static uint32_t g_io_size_bytes; | |
11fdf7f2 TL |
200 | static uint32_t g_max_io_md_size; |
201 | static uint32_t g_max_io_size_blocks; | |
202 | static uint32_t g_metacfg_pract_flag; | |
203 | static uint32_t g_metacfg_prchk_flags; | |
7c673cae FG |
204 | static int g_rw_percentage; |
205 | static int g_is_random; | |
206 | static int g_queue_depth; | |
9f95a23c TL |
207 | static int g_nr_io_queues_per_ns = 1; |
208 | static int g_nr_unused_io_queues = 0; | |
7c673cae FG |
209 | static int g_time_in_sec; |
210 | static uint32_t g_max_completions; | |
211 | static int g_dpdk_mem; | |
212 | static int g_shm_id = -1; | |
11fdf7f2 TL |
213 | static uint32_t g_disable_sq_cmb; |
214 | static bool g_no_pci; | |
215 | static bool g_warn; | |
9f95a23c TL |
216 | static bool g_header_digest; |
217 | static bool g_data_digest; | |
218 | static uint32_t g_keep_alive_timeout_in_ms = 0; | |
7c673cae FG |
219 | |
220 | static const char *g_core_mask; | |
221 | ||
222 | struct trid_entry { | |
223 | struct spdk_nvme_transport_id trid; | |
11fdf7f2 | 224 | uint16_t nsid; |
7c673cae FG |
225 | TAILQ_ENTRY(trid_entry) tailq; |
226 | }; | |
227 | ||
228 | static TAILQ_HEAD(, trid_entry) g_trid_list = TAILQ_HEAD_INITIALIZER(g_trid_list); | |
229 | ||
230 | static int g_aio_optind; /* Index of first AIO filename in argv */ | |
231 | ||
9f95a23c | 232 | static inline void |
7c673cae FG |
233 | task_complete(struct perf_task *task); |
234 | ||
9f95a23c | 235 | #if HAVE_LIBAIO |
7c673cae | 236 | static void |
9f95a23c | 237 | aio_setup_payload(struct perf_task *task, uint8_t pattern) |
7c673cae | 238 | { |
9f95a23c TL |
239 | task->iov.iov_base = spdk_dma_zmalloc(g_io_size_bytes, g_io_align, NULL); |
240 | task->iov.iov_len = g_io_size_bytes; | |
241 | if (task->iov.iov_base == NULL) { | |
242 | fprintf(stderr, "spdk_dma_zmalloc() for task->buf failed\n"); | |
7c673cae FG |
243 | exit(1); |
244 | } | |
9f95a23c | 245 | memset(task->iov.iov_base, pattern, task->iov.iov_len); |
7c673cae FG |
246 | } |
247 | ||
9f95a23c TL |
248 | static int |
249 | aio_submit(io_context_t aio_ctx, struct iocb *iocb, int fd, enum io_iocb_cmd cmd, | |
250 | struct iovec *iov, uint64_t offset, void *cb_ctx) | |
7c673cae | 251 | { |
9f95a23c TL |
252 | iocb->aio_fildes = fd; |
253 | iocb->aio_reqprio = 0; | |
254 | iocb->aio_lio_opcode = cmd; | |
255 | iocb->u.c.buf = iov->iov_base; | |
256 | iocb->u.c.nbytes = iov->iov_len; | |
257 | iocb->u.c.offset = offset * iov->iov_len; | |
258 | iocb->data = cb_ctx; | |
7c673cae | 259 | |
9f95a23c TL |
260 | if (io_submit(aio_ctx, 1, &iocb) < 0) { |
261 | printf("io_submit"); | |
262 | return -1; | |
7c673cae | 263 | } |
9f95a23c TL |
264 | |
265 | return 0; | |
7c673cae FG |
266 | } |
267 | ||
9f95a23c TL |
268 | static int |
269 | aio_submit_io(struct perf_task *task, struct ns_worker_ctx *ns_ctx, | |
270 | struct ns_entry *entry, uint64_t offset_in_ios) | |
7c673cae | 271 | { |
9f95a23c TL |
272 | if (task->is_read) { |
273 | return aio_submit(ns_ctx->u.aio.ctx, &task->iocb, entry->u.aio.fd, IO_CMD_PREAD, | |
274 | &task->iov, offset_in_ios, task); | |
275 | } else { | |
276 | return aio_submit(ns_ctx->u.aio.ctx, &task->iocb, entry->u.aio.fd, IO_CMD_PWRITE, | |
277 | &task->iov, offset_in_ios, task); | |
7c673cae | 278 | } |
7c673cae FG |
279 | } |
280 | ||
281 | static void | |
9f95a23c | 282 | aio_check_io(struct ns_worker_ctx *ns_ctx) |
7c673cae | 283 | { |
9f95a23c TL |
284 | int count, i; |
285 | struct timespec timeout; | |
7c673cae | 286 | |
9f95a23c TL |
287 | timeout.tv_sec = 0; |
288 | timeout.tv_nsec = 0; | |
7c673cae | 289 | |
9f95a23c TL |
290 | count = io_getevents(ns_ctx->u.aio.ctx, 1, g_queue_depth, ns_ctx->u.aio.events, &timeout); |
291 | if (count < 0) { | |
292 | fprintf(stderr, "io_getevents error\n"); | |
293 | exit(1); | |
7c673cae | 294 | } |
7c673cae | 295 | |
9f95a23c TL |
296 | for (i = 0; i < count; i++) { |
297 | task_complete(ns_ctx->u.aio.events[i].data); | |
7c673cae FG |
298 | } |
299 | } | |
300 | ||
301 | static void | |
9f95a23c | 302 | aio_verify_io(struct perf_task *task, struct ns_entry *entry) |
7c673cae | 303 | { |
9f95a23c | 304 | } |
7c673cae | 305 | |
9f95a23c TL |
306 | static int |
307 | aio_init_ns_worker_ctx(struct ns_worker_ctx *ns_ctx) | |
308 | { | |
309 | ns_ctx->u.aio.events = calloc(g_queue_depth, sizeof(struct io_event)); | |
310 | if (!ns_ctx->u.aio.events) { | |
311 | return -1; | |
11fdf7f2 | 312 | } |
9f95a23c TL |
313 | ns_ctx->u.aio.ctx = 0; |
314 | if (io_setup(g_queue_depth, &ns_ctx->u.aio.ctx) < 0) { | |
315 | free(ns_ctx->u.aio.events); | |
316 | perror("io_setup"); | |
317 | return -1; | |
7c673cae | 318 | } |
9f95a23c TL |
319 | return 0; |
320 | } | |
7c673cae | 321 | |
9f95a23c TL |
322 | static void |
323 | aio_cleanup_ns_worker_ctx(struct ns_worker_ctx *ns_ctx) | |
324 | { | |
325 | io_destroy(ns_ctx->u.aio.ctx); | |
326 | free(ns_ctx->u.aio.events); | |
7c673cae FG |
327 | } |
328 | ||
9f95a23c TL |
329 | static const struct ns_fn_table aio_fn_table = { |
330 | .setup_payload = aio_setup_payload, | |
331 | .submit_io = aio_submit_io, | |
332 | .check_io = aio_check_io, | |
333 | .verify_io = aio_verify_io, | |
334 | .init_ns_worker_ctx = aio_init_ns_worker_ctx, | |
335 | .cleanup_ns_worker_ctx = aio_cleanup_ns_worker_ctx, | |
336 | }; | |
337 | ||
7c673cae FG |
338 | static int |
339 | register_aio_file(const char *path) | |
340 | { | |
341 | struct ns_entry *entry; | |
342 | ||
343 | int flags, fd; | |
344 | uint64_t size; | |
345 | uint32_t blklen; | |
346 | ||
347 | if (g_rw_percentage == 100) { | |
348 | flags = O_RDONLY; | |
349 | } else if (g_rw_percentage == 0) { | |
350 | flags = O_WRONLY; | |
351 | } else { | |
352 | flags = O_RDWR; | |
353 | } | |
354 | ||
355 | flags |= O_DIRECT; | |
356 | ||
357 | fd = open(path, flags); | |
358 | if (fd < 0) { | |
359 | fprintf(stderr, "Could not open AIO device %s: %s\n", path, strerror(errno)); | |
360 | return -1; | |
361 | } | |
362 | ||
363 | size = spdk_fd_get_size(fd); | |
364 | if (size == 0) { | |
365 | fprintf(stderr, "Could not determine size of AIO device %s\n", path); | |
366 | close(fd); | |
367 | return -1; | |
368 | } | |
369 | ||
370 | blklen = spdk_fd_get_blocklen(fd); | |
371 | if (blklen == 0) { | |
372 | fprintf(stderr, "Could not determine block size of AIO device %s\n", path); | |
373 | close(fd); | |
374 | return -1; | |
375 | } | |
376 | ||
377 | /* | |
378 | * TODO: This should really calculate the LCM of the current g_io_align and blklen. | |
379 | * For now, it's fairly safe to just assume all block sizes are powers of 2. | |
380 | */ | |
381 | if (g_io_align < blklen) { | |
382 | g_io_align = blklen; | |
383 | } | |
384 | ||
385 | entry = malloc(sizeof(struct ns_entry)); | |
386 | if (entry == NULL) { | |
387 | close(fd); | |
388 | perror("aio ns_entry malloc"); | |
389 | return -1; | |
390 | } | |
391 | ||
392 | entry->type = ENTRY_TYPE_AIO_FILE; | |
9f95a23c | 393 | entry->fn_table = &aio_fn_table; |
7c673cae FG |
394 | entry->u.aio.fd = fd; |
395 | entry->size_in_ios = size / g_io_size_bytes; | |
396 | entry->io_size_blocks = g_io_size_bytes / blklen; | |
397 | ||
398 | snprintf(entry->name, sizeof(entry->name), "%s", path); | |
399 | ||
400 | g_num_namespaces++; | |
401 | entry->next = g_namespaces; | |
402 | g_namespaces = entry; | |
403 | ||
404 | return 0; | |
405 | } | |
406 | ||
407 | static int | |
9f95a23c | 408 | register_aio_files(int argc, char **argv) |
7c673cae | 409 | { |
9f95a23c | 410 | int i; |
7c673cae | 411 | |
9f95a23c TL |
412 | /* Treat everything after the options as files for AIO */ |
413 | for (i = g_aio_optind; i < argc; i++) { | |
414 | if (register_aio_file(argv[i]) != 0) { | |
415 | return 1; | |
416 | } | |
417 | } | |
418 | ||
419 | return 0; | |
420 | } | |
421 | #endif /* HAVE_LIBAIO */ | |
422 | ||
423 | static void io_complete(void *ctx, const struct spdk_nvme_cpl *cpl); | |
424 | ||
425 | static void | |
426 | nvme_setup_payload(struct perf_task *task, uint8_t pattern) | |
427 | { | |
428 | uint32_t max_io_size_bytes, max_io_md_size; | |
429 | ||
430 | /* maximum extended lba format size from all active namespace, | |
431 | * it's same with g_io_size_bytes for namespace without metadata. | |
432 | */ | |
433 | max_io_size_bytes = g_io_size_bytes + g_max_io_md_size * g_max_io_size_blocks; | |
434 | task->iov.iov_base = spdk_dma_zmalloc(max_io_size_bytes, g_io_align, NULL); | |
435 | task->iov.iov_len = max_io_size_bytes; | |
436 | if (task->iov.iov_base == NULL) { | |
437 | fprintf(stderr, "task->buf spdk_dma_zmalloc failed\n"); | |
438 | exit(1); | |
439 | } | |
440 | memset(task->iov.iov_base, pattern, task->iov.iov_len); | |
441 | ||
442 | max_io_md_size = g_max_io_md_size * g_max_io_size_blocks; | |
443 | if (max_io_md_size != 0) { | |
444 | task->md_iov.iov_base = spdk_dma_zmalloc(max_io_md_size, g_io_align, NULL); | |
445 | task->md_iov.iov_len = max_io_md_size; | |
446 | if (task->md_iov.iov_base == NULL) { | |
447 | fprintf(stderr, "task->md_buf spdk_dma_zmalloc failed\n"); | |
448 | spdk_dma_free(task->iov.iov_base); | |
449 | exit(1); | |
450 | } | |
451 | } | |
452 | } | |
453 | ||
454 | static int | |
455 | nvme_submit_io(struct perf_task *task, struct ns_worker_ctx *ns_ctx, | |
456 | struct ns_entry *entry, uint64_t offset_in_ios) | |
457 | { | |
458 | uint64_t lba; | |
459 | int rc; | |
460 | int qp_num; | |
461 | ||
462 | enum dif_mode { | |
463 | DIF_MODE_NONE = 0, | |
464 | DIF_MODE_DIF = 1, | |
465 | DIF_MODE_DIX = 2, | |
466 | } mode = DIF_MODE_NONE; | |
467 | ||
468 | lba = offset_in_ios * entry->io_size_blocks; | |
469 | ||
470 | if (entry->md_size != 0 && !(entry->io_flags & SPDK_NVME_IO_FLAGS_PRACT)) { | |
471 | if (entry->md_interleave) { | |
472 | mode = DIF_MODE_DIF; | |
473 | } else { | |
474 | mode = DIF_MODE_DIX; | |
475 | } | |
476 | } | |
477 | ||
478 | qp_num = ns_ctx->u.nvme.last_qpair; | |
479 | ns_ctx->u.nvme.last_qpair++; | |
480 | if (ns_ctx->u.nvme.last_qpair == ns_ctx->u.nvme.num_qpairs) { | |
481 | ns_ctx->u.nvme.last_qpair = 0; | |
482 | } | |
483 | ||
484 | if (mode != DIF_MODE_NONE) { | |
485 | rc = spdk_dif_ctx_init(&task->dif_ctx, entry->block_size, entry->md_size, | |
486 | entry->md_interleave, entry->pi_loc, | |
487 | (enum spdk_dif_type)entry->pi_type, entry->io_flags, | |
488 | lba, 0xFFFF, (uint16_t)entry->io_size_blocks, 0, 0); | |
489 | if (rc != 0) { | |
490 | fprintf(stderr, "Initialization of DIF context failed\n"); | |
491 | exit(1); | |
492 | } | |
493 | } | |
494 | ||
495 | if (task->is_read) { | |
496 | return spdk_nvme_ns_cmd_read_with_md(entry->u.nvme.ns, ns_ctx->u.nvme.qpair[qp_num], | |
497 | task->iov.iov_base, task->md_iov.iov_base, | |
498 | lba, | |
499 | entry->io_size_blocks, io_complete, | |
500 | task, entry->io_flags, | |
501 | task->dif_ctx.apptag_mask, task->dif_ctx.app_tag); | |
502 | } else { | |
503 | switch (mode) { | |
504 | case DIF_MODE_DIF: | |
505 | rc = spdk_dif_generate(&task->iov, 1, entry->io_size_blocks, &task->dif_ctx); | |
506 | if (rc != 0) { | |
507 | fprintf(stderr, "Generation of DIF failed\n"); | |
508 | return rc; | |
509 | } | |
510 | break; | |
511 | case DIF_MODE_DIX: | |
512 | rc = spdk_dix_generate(&task->iov, 1, &task->md_iov, entry->io_size_blocks, | |
513 | &task->dif_ctx); | |
514 | if (rc != 0) { | |
515 | fprintf(stderr, "Generation of DIX failed\n"); | |
516 | return rc; | |
517 | } | |
518 | break; | |
519 | default: | |
520 | break; | |
521 | } | |
522 | ||
523 | return spdk_nvme_ns_cmd_write_with_md(entry->u.nvme.ns, ns_ctx->u.nvme.qpair[qp_num], | |
524 | task->iov.iov_base, task->md_iov.iov_base, | |
525 | lba, | |
526 | entry->io_size_blocks, io_complete, | |
527 | task, entry->io_flags, | |
528 | task->dif_ctx.apptag_mask, task->dif_ctx.app_tag); | |
529 | } | |
530 | } | |
531 | ||
532 | static void | |
533 | nvme_check_io(struct ns_worker_ctx *ns_ctx) | |
534 | { | |
535 | int i, rc; | |
536 | ||
537 | for (i = 0; i < ns_ctx->u.nvme.num_qpairs; i++) { | |
538 | rc = spdk_nvme_qpair_process_completions(ns_ctx->u.nvme.qpair[i], g_max_completions); | |
539 | if (rc < 0) { | |
540 | fprintf(stderr, "NVMe io qpair process completion error\n"); | |
541 | exit(1); | |
542 | } | |
543 | } | |
544 | } | |
545 | ||
546 | static void | |
547 | nvme_verify_io(struct perf_task *task, struct ns_entry *entry) | |
548 | { | |
549 | struct spdk_dif_error err_blk = {}; | |
550 | int rc; | |
551 | ||
552 | if (!task->is_read || (entry->io_flags & SPDK_NVME_IO_FLAGS_PRACT)) { | |
553 | return; | |
554 | } | |
555 | ||
556 | if (entry->md_interleave) { | |
557 | rc = spdk_dif_verify(&task->iov, 1, entry->io_size_blocks, &task->dif_ctx, | |
558 | &err_blk); | |
559 | if (rc != 0) { | |
560 | fprintf(stderr, "DIF error detected. type=%d, offset=%" PRIu32 "\n", | |
561 | err_blk.err_type, err_blk.err_offset); | |
562 | } | |
563 | } else { | |
564 | rc = spdk_dix_verify(&task->iov, 1, &task->md_iov, entry->io_size_blocks, | |
565 | &task->dif_ctx, &err_blk); | |
566 | if (rc != 0) { | |
567 | fprintf(stderr, "DIX error detected. type=%d, offset=%" PRIu32 "\n", | |
568 | err_blk.err_type, err_blk.err_offset); | |
569 | } | |
570 | } | |
571 | } | |
572 | ||
573 | /* | |
574 | * TODO: If a controller has multiple namespaces, they could all use the same queue. | |
575 | * For now, give each namespace/thread combination its own queue. | |
576 | */ | |
577 | static int | |
578 | nvme_init_ns_worker_ctx(struct ns_worker_ctx *ns_ctx) | |
579 | { | |
580 | struct spdk_nvme_io_qpair_opts opts; | |
581 | struct ns_entry *entry = ns_ctx->entry; | |
582 | int i; | |
583 | ||
584 | ns_ctx->u.nvme.num_qpairs = g_nr_io_queues_per_ns; | |
585 | ns_ctx->u.nvme.qpair = calloc(ns_ctx->u.nvme.num_qpairs, sizeof(struct spdk_nvme_qpair *)); | |
586 | if (!ns_ctx->u.nvme.qpair) { | |
7c673cae FG |
587 | return -1; |
588 | } | |
589 | ||
9f95a23c TL |
590 | spdk_nvme_ctrlr_get_default_io_qpair_opts(entry->u.nvme.ctrlr, &opts, sizeof(opts)); |
591 | if (opts.io_queue_requests < entry->num_io_requests) { | |
592 | opts.io_queue_requests = entry->num_io_requests; | |
593 | } | |
594 | opts.delay_pcie_doorbell = true; | |
595 | ||
596 | for (i = 0; i < ns_ctx->u.nvme.num_qpairs; i++) { | |
597 | ns_ctx->u.nvme.qpair[i] = spdk_nvme_ctrlr_alloc_io_qpair(entry->u.nvme.ctrlr, &opts, | |
598 | sizeof(opts)); | |
599 | if (!ns_ctx->u.nvme.qpair[i]) { | |
600 | printf("ERROR: spdk_nvme_ctrlr_alloc_io_qpair failed\n"); | |
601 | return -1; | |
602 | } | |
603 | } | |
604 | ||
7c673cae FG |
605 | return 0; |
606 | } | |
607 | ||
608 | static void | |
9f95a23c | 609 | nvme_cleanup_ns_worker_ctx(struct ns_worker_ctx *ns_ctx) |
7c673cae | 610 | { |
9f95a23c | 611 | int i; |
7c673cae | 612 | |
9f95a23c TL |
613 | for (i = 0; i < ns_ctx->u.nvme.num_qpairs; i++) { |
614 | spdk_nvme_ctrlr_free_io_qpair(ns_ctx->u.nvme.qpair[i]); | |
615 | } | |
7c673cae | 616 | |
9f95a23c TL |
617 | free(ns_ctx->u.nvme.qpair); |
618 | } | |
619 | ||
620 | static const struct ns_fn_table nvme_fn_table = { | |
621 | .setup_payload = nvme_setup_payload, | |
622 | .submit_io = nvme_submit_io, | |
623 | .check_io = nvme_check_io, | |
624 | .verify_io = nvme_verify_io, | |
625 | .init_ns_worker_ctx = nvme_init_ns_worker_ctx, | |
626 | .cleanup_ns_worker_ctx = nvme_cleanup_ns_worker_ctx, | |
627 | }; | |
628 | ||
629 | static void | |
630 | build_nvme_name(char *name, size_t length, struct spdk_nvme_ctrlr *ctrlr) | |
631 | { | |
632 | const struct spdk_nvme_transport_id *trid; | |
633 | ||
634 | trid = spdk_nvme_ctrlr_get_transport_id(ctrlr); | |
635 | ||
636 | switch (trid->trtype) { | |
637 | case SPDK_NVME_TRANSPORT_PCIE: | |
638 | snprintf(name, length, "PCIE (%s)", trid->traddr); | |
639 | break; | |
640 | case SPDK_NVME_TRANSPORT_RDMA: | |
641 | snprintf(name, length, "RDMA (addr:%s subnqn:%s)", trid->traddr, trid->subnqn); | |
642 | break; | |
643 | case SPDK_NVME_TRANSPORT_TCP: | |
644 | snprintf(name, length, "TCP (addr:%s subnqn:%s)", trid->traddr, trid->subnqn); | |
645 | break; | |
646 | default: | |
647 | fprintf(stderr, "Unknown transport type %d\n", trid->trtype); | |
648 | break; | |
649 | } | |
650 | } | |
651 | ||
652 | static void | |
653 | register_ns(struct spdk_nvme_ctrlr *ctrlr, struct spdk_nvme_ns *ns) | |
654 | { | |
655 | struct ns_entry *entry; | |
656 | const struct spdk_nvme_ctrlr_data *cdata; | |
657 | uint32_t max_xfer_size, entries, sector_size; | |
658 | uint64_t ns_size; | |
659 | struct spdk_nvme_io_qpair_opts opts; | |
660 | ||
661 | cdata = spdk_nvme_ctrlr_get_data(ctrlr); | |
662 | ||
663 | if (!spdk_nvme_ns_is_active(ns)) { | |
664 | printf("Controller %-20.20s (%-20.20s): Skipping inactive NS %u\n", | |
665 | cdata->mn, cdata->sn, | |
666 | spdk_nvme_ns_get_id(ns)); | |
667 | g_warn = true; | |
668 | return; | |
669 | } | |
670 | ||
671 | ns_size = spdk_nvme_ns_get_size(ns); | |
672 | sector_size = spdk_nvme_ns_get_sector_size(ns); | |
673 | ||
674 | if (ns_size < g_io_size_bytes || sector_size > g_io_size_bytes) { | |
675 | printf("WARNING: controller %-20.20s (%-20.20s) ns %u has invalid " | |
676 | "ns size %" PRIu64 " / block size %u for I/O size %u\n", | |
677 | cdata->mn, cdata->sn, spdk_nvme_ns_get_id(ns), | |
678 | ns_size, spdk_nvme_ns_get_sector_size(ns), g_io_size_bytes); | |
679 | g_warn = true; | |
680 | return; | |
681 | } | |
682 | ||
683 | max_xfer_size = spdk_nvme_ns_get_max_io_xfer_size(ns); | |
684 | spdk_nvme_ctrlr_get_default_io_qpair_opts(ctrlr, &opts, sizeof(opts)); | |
685 | /* NVMe driver may add additional entries based on | |
686 | * stripe size and maximum transfer size, we assume | |
687 | * 1 more entry be used for stripe. | |
688 | */ | |
689 | entries = (g_io_size_bytes - 1) / max_xfer_size + 2; | |
690 | if ((g_queue_depth * entries) > opts.io_queue_size) { | |
691 | printf("controller IO queue size %u less than required\n", | |
692 | opts.io_queue_size); | |
693 | printf("Consider using lower queue depth or small IO size because " | |
694 | "IO requests may be queued at the NVMe driver.\n"); | |
695 | g_warn = true; | |
696 | } | |
697 | /* For requests which have children requests, parent request itself | |
698 | * will also occupy 1 entry. | |
699 | */ | |
700 | entries += 1; | |
701 | ||
702 | entry = calloc(1, sizeof(struct ns_entry)); | |
703 | if (entry == NULL) { | |
704 | perror("ns_entry malloc"); | |
7c673cae FG |
705 | exit(1); |
706 | } | |
707 | ||
9f95a23c TL |
708 | entry->type = ENTRY_TYPE_NVME_NS; |
709 | entry->fn_table = &nvme_fn_table; | |
710 | entry->u.nvme.ctrlr = ctrlr; | |
711 | entry->u.nvme.ns = ns; | |
712 | entry->num_io_requests = g_queue_depth * entries; | |
713 | ||
714 | entry->size_in_ios = ns_size / g_io_size_bytes; | |
715 | entry->io_size_blocks = g_io_size_bytes / sector_size; | |
716 | ||
717 | entry->block_size = spdk_nvme_ns_get_extended_sector_size(ns); | |
718 | entry->md_size = spdk_nvme_ns_get_md_size(ns); | |
719 | entry->md_interleave = spdk_nvme_ns_supports_extended_lba(ns); | |
720 | entry->pi_loc = spdk_nvme_ns_get_data(ns)->dps.md_start; | |
721 | entry->pi_type = spdk_nvme_ns_get_pi_type(ns); | |
722 | ||
723 | if (spdk_nvme_ns_get_flags(ns) & SPDK_NVME_NS_DPS_PI_SUPPORTED) { | |
724 | entry->io_flags = g_metacfg_pract_flag | g_metacfg_prchk_flags; | |
725 | } | |
726 | ||
727 | if (g_max_io_md_size < entry->md_size) { | |
728 | g_max_io_md_size = entry->md_size; | |
7c673cae | 729 | } |
9f95a23c TL |
730 | |
731 | if (g_max_io_size_blocks < entry->io_size_blocks) { | |
732 | g_max_io_size_blocks = entry->io_size_blocks; | |
733 | } | |
734 | ||
735 | build_nvme_name(entry->name, sizeof(entry->name), ctrlr); | |
736 | ||
737 | g_num_namespaces++; | |
738 | entry->next = g_namespaces; | |
739 | g_namespaces = entry; | |
7c673cae | 740 | } |
7c673cae | 741 | |
11fdf7f2 | 742 | static void |
9f95a23c | 743 | unregister_namespaces(void) |
7c673cae | 744 | { |
9f95a23c | 745 | struct ns_entry *entry = g_namespaces; |
11fdf7f2 | 746 | |
9f95a23c TL |
747 | while (entry) { |
748 | struct ns_entry *next = entry->next; | |
749 | free(entry); | |
750 | entry = next; | |
751 | } | |
752 | } | |
11fdf7f2 | 753 | |
9f95a23c TL |
754 | static void |
755 | enable_latency_tracking_complete(void *cb_arg, const struct spdk_nvme_cpl *cpl) | |
756 | { | |
757 | if (spdk_nvme_cpl_is_error(cpl)) { | |
758 | printf("enable_latency_tracking_complete failed\n"); | |
759 | } | |
760 | g_outstanding_commands--; | |
761 | } | |
762 | ||
763 | static void | |
764 | set_latency_tracking_feature(struct spdk_nvme_ctrlr *ctrlr, bool enable) | |
765 | { | |
766 | int res; | |
767 | union spdk_nvme_intel_feat_latency_tracking latency_tracking; | |
768 | ||
769 | if (enable) { | |
770 | latency_tracking.bits.enable = 0x01; | |
771 | } else { | |
772 | latency_tracking.bits.enable = 0x00; | |
11fdf7f2 TL |
773 | } |
774 | ||
9f95a23c TL |
775 | res = spdk_nvme_ctrlr_cmd_set_feature(ctrlr, SPDK_NVME_INTEL_FEAT_LATENCY_TRACKING, |
776 | latency_tracking.raw, 0, NULL, 0, enable_latency_tracking_complete, NULL); | |
777 | if (res) { | |
778 | printf("fail to allocate nvme request.\n"); | |
11fdf7f2 TL |
779 | return; |
780 | } | |
9f95a23c TL |
781 | g_outstanding_commands++; |
782 | ||
783 | while (g_outstanding_commands) { | |
784 | spdk_nvme_ctrlr_process_admin_completions(ctrlr); | |
785 | } | |
786 | } | |
787 | ||
788 | static void | |
789 | register_ctrlr(struct spdk_nvme_ctrlr *ctrlr, struct trid_entry *trid_entry) | |
790 | { | |
791 | struct spdk_nvme_ns *ns; | |
792 | struct ctrlr_entry *entry = malloc(sizeof(struct ctrlr_entry)); | |
793 | uint32_t nsid; | |
11fdf7f2 | 794 | |
9f95a23c TL |
795 | if (entry == NULL) { |
796 | perror("ctrlr_entry malloc"); | |
797 | exit(1); | |
11fdf7f2 TL |
798 | } |
799 | ||
9f95a23c TL |
800 | entry->latency_page = spdk_dma_zmalloc(sizeof(struct spdk_nvme_intel_rw_latency_page), |
801 | 4096, NULL); | |
802 | if (entry->latency_page == NULL) { | |
803 | printf("Allocation error (latency page)\n"); | |
804 | exit(1); | |
11fdf7f2 TL |
805 | } |
806 | ||
9f95a23c | 807 | build_nvme_name(entry->name, sizeof(entry->name), ctrlr); |
11fdf7f2 | 808 | |
9f95a23c TL |
809 | entry->ctrlr = ctrlr; |
810 | entry->trtype = trid_entry->trid.trtype; | |
811 | entry->next = g_controllers; | |
812 | g_controllers = entry; | |
11fdf7f2 | 813 | |
9f95a23c TL |
814 | if (g_latency_ssd_tracking_enable && |
815 | spdk_nvme_ctrlr_is_feature_supported(ctrlr, SPDK_NVME_INTEL_FEAT_LATENCY_TRACKING)) { | |
816 | set_latency_tracking_feature(ctrlr, true); | |
11fdf7f2 TL |
817 | } |
818 | ||
9f95a23c TL |
819 | if (trid_entry->nsid == 0) { |
820 | for (nsid = spdk_nvme_ctrlr_get_first_active_ns(ctrlr); | |
821 | nsid != 0; nsid = spdk_nvme_ctrlr_get_next_active_ns(ctrlr, nsid)) { | |
822 | ns = spdk_nvme_ctrlr_get_ns(ctrlr, nsid); | |
823 | if (ns == NULL) { | |
824 | continue; | |
11fdf7f2 | 825 | } |
9f95a23c TL |
826 | register_ns(ctrlr, ns); |
827 | } | |
828 | } else { | |
829 | ns = spdk_nvme_ctrlr_get_ns(ctrlr, trid_entry->nsid); | |
830 | if (!ns) { | |
831 | perror("Namespace does not exist."); | |
832 | exit(1); | |
11fdf7f2 | 833 | } |
11fdf7f2 | 834 | |
9f95a23c | 835 | register_ns(ctrlr, ns); |
11fdf7f2 TL |
836 | } |
837 | ||
9f95a23c TL |
838 | if (g_nr_unused_io_queues) { |
839 | int i; | |
11fdf7f2 | 840 | |
9f95a23c | 841 | printf("Creating %u unused qpairs for controller %s\n", g_nr_unused_io_queues, entry->name); |
11fdf7f2 | 842 | |
9f95a23c TL |
843 | entry->unused_qpairs = calloc(g_nr_unused_io_queues, sizeof(struct spdk_nvme_qpair *)); |
844 | if (!entry->unused_qpairs) { | |
845 | fprintf(stderr, "Unable to allocate memory for qpair array\n"); | |
846 | exit(1); | |
11fdf7f2 | 847 | } |
9f95a23c TL |
848 | |
849 | for (i = 0; i < g_nr_unused_io_queues; i++) { | |
850 | entry->unused_qpairs[i] = spdk_nvme_ctrlr_alloc_io_qpair(ctrlr, NULL, 0); | |
851 | if (!entry->unused_qpairs[i]) { | |
852 | fprintf(stderr, "Unable to allocate unused qpair. Did you request too many?\n"); | |
853 | exit(1); | |
11fdf7f2 TL |
854 | } |
855 | } | |
7c673cae | 856 | } |
7c673cae | 857 | |
9f95a23c | 858 | } |
7c673cae FG |
859 | |
860 | static __thread unsigned int seed = 0; | |
861 | ||
9f95a23c | 862 | static inline void |
11fdf7f2 | 863 | submit_single_io(struct perf_task *task) |
7c673cae | 864 | { |
7c673cae FG |
865 | uint64_t offset_in_ios; |
866 | int rc; | |
11fdf7f2 | 867 | struct ns_worker_ctx *ns_ctx = task->ns_ctx; |
7c673cae FG |
868 | struct ns_entry *entry = ns_ctx->entry; |
869 | ||
7c673cae FG |
870 | if (g_is_random) { |
871 | offset_in_ios = rand_r(&seed) % entry->size_in_ios; | |
872 | } else { | |
873 | offset_in_ios = ns_ctx->offset_in_ios++; | |
874 | if (ns_ctx->offset_in_ios == entry->size_in_ios) { | |
875 | ns_ctx->offset_in_ios = 0; | |
876 | } | |
877 | } | |
878 | ||
879 | task->submit_tsc = spdk_get_ticks(); | |
880 | ||
881 | if ((g_rw_percentage == 100) || | |
882 | (g_rw_percentage != 0 && ((rand_r(&seed) % 100) < g_rw_percentage))) { | |
9f95a23c | 883 | task->is_read = true; |
7c673cae | 884 | } else { |
9f95a23c | 885 | task->is_read = false; |
7c673cae FG |
886 | } |
887 | ||
9f95a23c TL |
888 | rc = entry->fn_table->submit_io(task, ns_ctx, entry, offset_in_ios); |
889 | ||
890 | if (spdk_unlikely(rc != 0)) { | |
7c673cae | 891 | fprintf(stderr, "starting I/O failed\n"); |
11fdf7f2 TL |
892 | } else { |
893 | ns_ctx->current_queue_depth++; | |
7c673cae | 894 | } |
7c673cae FG |
895 | } |
896 | ||
9f95a23c | 897 | static inline void |
7c673cae FG |
898 | task_complete(struct perf_task *task) |
899 | { | |
900 | struct ns_worker_ctx *ns_ctx; | |
901 | uint64_t tsc_diff; | |
11fdf7f2 | 902 | struct ns_entry *entry; |
7c673cae FG |
903 | |
904 | ns_ctx = task->ns_ctx; | |
11fdf7f2 | 905 | entry = ns_ctx->entry; |
7c673cae FG |
906 | ns_ctx->current_queue_depth--; |
907 | ns_ctx->io_completed++; | |
908 | tsc_diff = spdk_get_ticks() - task->submit_tsc; | |
909 | ns_ctx->total_tsc += tsc_diff; | |
9f95a23c | 910 | if (spdk_unlikely(ns_ctx->min_tsc > tsc_diff)) { |
7c673cae FG |
911 | ns_ctx->min_tsc = tsc_diff; |
912 | } | |
9f95a23c | 913 | if (spdk_unlikely(ns_ctx->max_tsc < tsc_diff)) { |
7c673cae FG |
914 | ns_ctx->max_tsc = tsc_diff; |
915 | } | |
9f95a23c | 916 | if (spdk_unlikely(g_latency_sw_tracking_level > 0)) { |
11fdf7f2 TL |
917 | spdk_histogram_data_tally(ns_ctx->histogram, tsc_diff); |
918 | } | |
7c673cae | 919 | |
9f95a23c TL |
920 | if (spdk_unlikely(entry->md_size > 0)) { |
921 | /* add application level verification for end-to-end data protection */ | |
922 | entry->fn_table->verify_io(task, entry); | |
11fdf7f2 | 923 | } |
7c673cae FG |
924 | |
925 | /* | |
926 | * is_draining indicates when time has expired for the test run | |
927 | * and we are just waiting for the previously submitted I/O | |
928 | * to complete. In this case, do not submit a new I/O to replace | |
929 | * the one just completed. | |
930 | */ | |
9f95a23c TL |
931 | if (spdk_unlikely(ns_ctx->is_draining)) { |
932 | spdk_dma_free(task->iov.iov_base); | |
933 | spdk_dma_free(task->md_iov.iov_base); | |
11fdf7f2 TL |
934 | free(task); |
935 | } else { | |
936 | submit_single_io(task); | |
7c673cae FG |
937 | } |
938 | } | |
939 | ||
940 | static void | |
9f95a23c | 941 | io_complete(void *ctx, const struct spdk_nvme_cpl *cpl) |
7c673cae | 942 | { |
9f95a23c TL |
943 | struct perf_task *task = ctx; |
944 | ||
945 | if (spdk_unlikely(spdk_nvme_cpl_is_error(cpl))) { | |
946 | fprintf(stderr, "%s completed with error (sct=%d, sc=%d)\n", | |
947 | task->is_read ? "Read" : "Write", | |
948 | cpl->status.sct, cpl->status.sc); | |
949 | } | |
950 | ||
951 | task_complete(task); | |
7c673cae FG |
952 | } |
953 | ||
954 | static void | |
955 | check_io(struct ns_worker_ctx *ns_ctx) | |
956 | { | |
9f95a23c | 957 | ns_ctx->entry->fn_table->check_io(ns_ctx); |
7c673cae FG |
958 | } |
959 | ||
9f95a23c TL |
960 | static struct perf_task * |
961 | allocate_task(struct ns_worker_ctx *ns_ctx, int queue_depth) | |
7c673cae | 962 | { |
11fdf7f2 | 963 | struct perf_task *task; |
11fdf7f2 | 964 | |
9f95a23c TL |
965 | task = calloc(1, sizeof(*task)); |
966 | if (task == NULL) { | |
967 | fprintf(stderr, "Out of memory allocating tasks\n"); | |
968 | exit(1); | |
969 | } | |
11fdf7f2 | 970 | |
9f95a23c | 971 | ns_ctx->entry->fn_table->setup_payload(task, queue_depth % 8 + 1); |
11fdf7f2 | 972 | |
9f95a23c | 973 | task->ns_ctx = ns_ctx; |
11fdf7f2 | 974 | |
9f95a23c | 975 | return task; |
7c673cae FG |
976 | } |
977 | ||
978 | static void | |
9f95a23c | 979 | submit_io(struct ns_worker_ctx *ns_ctx, int queue_depth) |
7c673cae | 980 | { |
9f95a23c TL |
981 | struct perf_task *task; |
982 | ||
983 | while (queue_depth-- > 0) { | |
984 | task = allocate_task(ns_ctx, queue_depth); | |
985 | submit_single_io(task); | |
7c673cae FG |
986 | } |
987 | } | |
988 | ||
989 | static int | |
990 | init_ns_worker_ctx(struct ns_worker_ctx *ns_ctx) | |
991 | { | |
9f95a23c | 992 | return ns_ctx->entry->fn_table->init_ns_worker_ctx(ns_ctx); |
7c673cae FG |
993 | } |
994 | ||
995 | static void | |
996 | cleanup_ns_worker_ctx(struct ns_worker_ctx *ns_ctx) | |
997 | { | |
9f95a23c | 998 | ns_ctx->entry->fn_table->cleanup_ns_worker_ctx(ns_ctx); |
7c673cae FG |
999 | } |
1000 | ||
1001 | static int | |
1002 | work_fn(void *arg) | |
1003 | { | |
1004 | uint64_t tsc_end; | |
1005 | struct worker_thread *worker = (struct worker_thread *)arg; | |
1006 | struct ns_worker_ctx *ns_ctx = NULL; | |
9f95a23c | 1007 | uint32_t unfinished_ns_ctx; |
7c673cae FG |
1008 | |
1009 | printf("Starting thread on core %u\n", worker->lcore); | |
1010 | ||
9f95a23c | 1011 | /* Allocate queue pairs for each namespace. */ |
7c673cae FG |
1012 | ns_ctx = worker->ns_ctx; |
1013 | while (ns_ctx != NULL) { | |
1014 | if (init_ns_worker_ctx(ns_ctx) != 0) { | |
1015 | printf("ERROR: init_ns_worker_ctx() failed\n"); | |
1016 | return 1; | |
1017 | } | |
1018 | ns_ctx = ns_ctx->next; | |
1019 | } | |
1020 | ||
1021 | tsc_end = spdk_get_ticks() + g_time_in_sec * g_tsc_rate; | |
1022 | ||
1023 | /* Submit initial I/O for each namespace. */ | |
1024 | ns_ctx = worker->ns_ctx; | |
1025 | while (ns_ctx != NULL) { | |
1026 | submit_io(ns_ctx, g_queue_depth); | |
1027 | ns_ctx = ns_ctx->next; | |
1028 | } | |
1029 | ||
1030 | while (1) { | |
1031 | /* | |
1032 | * Check for completed I/O for each controller. A new | |
1033 | * I/O will be submitted in the io_complete callback | |
1034 | * to replace each I/O that is completed. | |
1035 | */ | |
1036 | ns_ctx = worker->ns_ctx; | |
1037 | while (ns_ctx != NULL) { | |
1038 | check_io(ns_ctx); | |
1039 | ns_ctx = ns_ctx->next; | |
1040 | } | |
1041 | ||
1042 | if (spdk_get_ticks() > tsc_end) { | |
1043 | break; | |
1044 | } | |
1045 | } | |
1046 | ||
9f95a23c TL |
1047 | /* drain the io of each ns_ctx in round robin to make the fairness */ |
1048 | do { | |
1049 | unfinished_ns_ctx = 0; | |
1050 | ns_ctx = worker->ns_ctx; | |
1051 | while (ns_ctx != NULL) { | |
1052 | /* first time will enter into this if case */ | |
1053 | if (!ns_ctx->is_draining) { | |
1054 | ns_ctx->is_draining = true; | |
1055 | } | |
1056 | ||
1057 | if (ns_ctx->current_queue_depth > 0) { | |
1058 | check_io(ns_ctx); | |
1059 | if (ns_ctx->current_queue_depth == 0) { | |
1060 | cleanup_ns_worker_ctx(ns_ctx); | |
1061 | } else { | |
1062 | unfinished_ns_ctx++; | |
1063 | } | |
1064 | } | |
1065 | ns_ctx = ns_ctx->next; | |
1066 | } | |
1067 | } while (unfinished_ns_ctx > 0); | |
7c673cae FG |
1068 | |
1069 | return 0; | |
1070 | } | |
1071 | ||
1072 | static void usage(char *program_name) | |
1073 | { | |
1074 | printf("%s options", program_name); | |
1075 | #if HAVE_LIBAIO | |
1076 | printf(" [AIO device(s)]..."); | |
1077 | #endif | |
1078 | printf("\n"); | |
1079 | printf("\t[-q io depth]\n"); | |
11fdf7f2 | 1080 | printf("\t[-o io size in bytes]\n"); |
9f95a23c TL |
1081 | printf("\t[-n number of io queues per namespace. default: 1]\n"); |
1082 | printf("\t[-U number of unused io queues per controller. default: 0]\n"); | |
7c673cae FG |
1083 | printf("\t[-w io pattern type, must be one of\n"); |
1084 | printf("\t\t(read, write, randread, randwrite, rw, randrw)]\n"); | |
1085 | printf("\t[-M rwmixread (100 for reads, 0 for writes)]\n"); | |
11fdf7f2 TL |
1086 | printf("\t[-L enable latency tracking via sw, default: disabled]\n"); |
1087 | printf("\t\t-L for latency summary, -LL for detailed histogram\n"); | |
1088 | printf("\t[-l enable latency tracking via ssd (if supported), default: disabled]\n"); | |
7c673cae FG |
1089 | printf("\t[-t time in seconds]\n"); |
1090 | printf("\t[-c core mask for I/O submission/completion.]\n"); | |
9f95a23c | 1091 | printf("\t\t(default: 1)\n"); |
11fdf7f2 | 1092 | printf("\t[-D disable submission queue in controller memory buffer, default: enabled]\n"); |
9f95a23c TL |
1093 | printf("\t[-H enable header digest for TCP transport, default: disabled]\n"); |
1094 | printf("\t[-I enable data digest for TCP transport, default: disabled]\n"); | |
7c673cae FG |
1095 | printf("\t[-r Transport ID for local PCIe NVMe or NVMeoF]\n"); |
1096 | printf("\t Format: 'key:value [key:value] ...'\n"); | |
1097 | printf("\t Keys:\n"); | |
1098 | printf("\t trtype Transport type (e.g. PCIe, RDMA)\n"); | |
1099 | printf("\t adrfam Address family (e.g. IPv4, IPv6)\n"); | |
1100 | printf("\t traddr Transport address (e.g. 0000:04:00.0 for PCIe or 192.168.100.8 for RDMA)\n"); | |
1101 | printf("\t trsvcid Transport service identifier (e.g. 4420)\n"); | |
1102 | printf("\t subnqn Subsystem NQN (default: %s)\n", SPDK_NVMF_DISCOVERY_NQN); | |
1103 | printf("\t Example: -r 'trtype:PCIe traddr:0000:04:00.0' for PCIe or\n"); | |
1104 | printf("\t -r 'trtype:RDMA adrfam:IPv4 traddr:192.168.100.8 trsvcid:4420' for NVMeoF\n"); | |
11fdf7f2 TL |
1105 | printf("\t[-e metadata configuration]\n"); |
1106 | printf("\t Keys:\n"); | |
1107 | printf("\t PRACT Protection Information Action bit (PRACT=1 or PRACT=0)\n"); | |
1108 | printf("\t PRCHK Control of Protection Information Checking (PRCHK=GUARD|REFTAG|APPTAG)\n"); | |
1109 | printf("\t Example: -e 'PRACT=0,PRCHK=GUARD|REFTAG|APPTAG'\n"); | |
1110 | printf("\t -e 'PRACT=1,PRCHK=GUARD'\n"); | |
9f95a23c | 1111 | printf("\t[-k keep alive timeout period in millisecond]\n"); |
11fdf7f2 | 1112 | printf("\t[-s DPDK huge memory size in MB.]\n"); |
7c673cae FG |
1113 | printf("\t[-m max completions per poll]\n"); |
1114 | printf("\t\t(default: 0 - unlimited)\n"); | |
1115 | printf("\t[-i shared memory group ID]\n"); | |
9f95a23c TL |
1116 | #ifdef DEBUG |
1117 | printf("\t[-G enable debug logging]\n"); | |
1118 | #else | |
1119 | printf("\t[-G enable debug logging (flag disabled, must reconfigure with --enable-debug)\n"); | |
1120 | #endif | |
7c673cae FG |
1121 | } |
1122 | ||
11fdf7f2 TL |
1123 | static void |
1124 | check_cutoff(void *ctx, uint64_t start, uint64_t end, uint64_t count, | |
1125 | uint64_t total, uint64_t so_far) | |
1126 | { | |
1127 | double so_far_pct; | |
1128 | double **cutoff = ctx; | |
1129 | ||
1130 | if (count == 0) { | |
1131 | return; | |
1132 | } | |
1133 | ||
1134 | so_far_pct = (double)so_far / total; | |
1135 | while (so_far_pct >= **cutoff && **cutoff > 0) { | |
1136 | printf("%9.5f%% : %9.3fus\n", **cutoff * 100, (double)end * 1000 * 1000 / g_tsc_rate); | |
1137 | (*cutoff)++; | |
1138 | } | |
1139 | } | |
1140 | ||
1141 | static void | |
1142 | print_bucket(void *ctx, uint64_t start, uint64_t end, uint64_t count, | |
1143 | uint64_t total, uint64_t so_far) | |
1144 | { | |
1145 | double so_far_pct; | |
1146 | ||
1147 | if (count == 0) { | |
1148 | return; | |
1149 | } | |
1150 | ||
1151 | so_far_pct = (double)so_far * 100 / total; | |
1152 | printf("%9.3f - %9.3f: %9.4f%% (%9ju)\n", | |
1153 | (double)start * 1000 * 1000 / g_tsc_rate, | |
1154 | (double)end * 1000 * 1000 / g_tsc_rate, | |
1155 | so_far_pct, count); | |
1156 | } | |
1157 | ||
7c673cae FG |
1158 | static void |
1159 | print_performance(void) | |
1160 | { | |
11fdf7f2 TL |
1161 | uint64_t total_io_completed, total_io_tsc; |
1162 | double io_per_second, mb_per_second, average_latency, min_latency, max_latency; | |
1163 | double sum_ave_latency, min_latency_so_far, max_latency_so_far; | |
1164 | double total_io_per_second, total_mb_per_second; | |
7c673cae FG |
1165 | int ns_count; |
1166 | struct worker_thread *worker; | |
1167 | struct ns_worker_ctx *ns_ctx; | |
9f95a23c | 1168 | uint32_t max_strlen; |
7c673cae FG |
1169 | |
1170 | total_io_per_second = 0; | |
1171 | total_mb_per_second = 0; | |
1172 | total_io_completed = 0; | |
11fdf7f2 TL |
1173 | total_io_tsc = 0; |
1174 | min_latency_so_far = (double)UINT64_MAX; | |
1175 | max_latency_so_far = 0; | |
7c673cae FG |
1176 | ns_count = 0; |
1177 | ||
9f95a23c TL |
1178 | max_strlen = 0; |
1179 | worker = g_workers; | |
1180 | while (worker) { | |
1181 | ns_ctx = worker->ns_ctx; | |
1182 | while (ns_ctx) { | |
1183 | max_strlen = spdk_max(strlen(ns_ctx->entry->name), max_strlen); | |
1184 | ns_ctx = ns_ctx->next; | |
1185 | } | |
1186 | worker = worker->next; | |
1187 | } | |
1188 | ||
7c673cae | 1189 | printf("========================================================\n"); |
9f95a23c TL |
1190 | printf("%*s\n", max_strlen + 60, "Latency(us)"); |
1191 | printf("%-*s: %10s %10s %10s %10s %10s\n", | |
1192 | max_strlen + 12, "Device Information", "IOPS", "MiB/s", "Average", "min", "max"); | |
7c673cae FG |
1193 | |
1194 | worker = g_workers; | |
1195 | while (worker) { | |
1196 | ns_ctx = worker->ns_ctx; | |
1197 | while (ns_ctx) { | |
11fdf7f2 TL |
1198 | if (ns_ctx->io_completed != 0) { |
1199 | io_per_second = (double)ns_ctx->io_completed / g_time_in_sec; | |
1200 | mb_per_second = io_per_second * g_io_size_bytes / (1024 * 1024); | |
1201 | average_latency = ((double)ns_ctx->total_tsc / ns_ctx->io_completed) * 1000 * 1000 / g_tsc_rate; | |
1202 | min_latency = (double)ns_ctx->min_tsc * 1000 * 1000 / g_tsc_rate; | |
1203 | if (min_latency < min_latency_so_far) { | |
1204 | min_latency_so_far = min_latency; | |
1205 | } | |
1206 | ||
1207 | max_latency = (double)ns_ctx->max_tsc * 1000 * 1000 / g_tsc_rate; | |
1208 | if (max_latency > max_latency_so_far) { | |
1209 | max_latency_so_far = max_latency; | |
1210 | } | |
1211 | ||
9f95a23c TL |
1212 | printf("%-*.*s from core %u: %10.2f %10.2f %10.2f %10.2f %10.2f\n", |
1213 | max_strlen, max_strlen, ns_ctx->entry->name, worker->lcore, | |
11fdf7f2 TL |
1214 | io_per_second, mb_per_second, |
1215 | average_latency, min_latency, max_latency); | |
1216 | total_io_per_second += io_per_second; | |
1217 | total_mb_per_second += mb_per_second; | |
1218 | total_io_completed += ns_ctx->io_completed; | |
1219 | total_io_tsc += ns_ctx->total_tsc; | |
1220 | ns_count++; | |
1221 | } | |
1222 | ns_ctx = ns_ctx->next; | |
1223 | } | |
1224 | worker = worker->next; | |
1225 | } | |
1226 | ||
1227 | if (ns_count != 0 && total_io_completed) { | |
1228 | sum_ave_latency = ((double)total_io_tsc / total_io_completed) * 1000 * 1000 / g_tsc_rate; | |
1229 | printf("========================================================\n"); | |
9f95a23c TL |
1230 | printf("%-*s: %10.2f %10.2f %10.2f %10.2f %10.2f\n", |
1231 | max_strlen + 12, "Total", total_io_per_second, total_mb_per_second, | |
11fdf7f2 TL |
1232 | sum_ave_latency, min_latency_so_far, max_latency_so_far); |
1233 | printf("\n"); | |
1234 | } | |
1235 | ||
1236 | if (g_latency_sw_tracking_level == 0 || total_io_completed == 0) { | |
1237 | return; | |
1238 | } | |
1239 | ||
1240 | worker = g_workers; | |
1241 | while (worker) { | |
1242 | ns_ctx = worker->ns_ctx; | |
1243 | while (ns_ctx) { | |
1244 | const double *cutoff = g_latency_cutoffs; | |
1245 | ||
1246 | printf("Summary latency data for %-43.43s from core %u:\n", ns_ctx->entry->name, worker->lcore); | |
1247 | printf("=================================================================================\n"); | |
1248 | ||
1249 | spdk_histogram_data_iterate(ns_ctx->histogram, check_cutoff, &cutoff); | |
1250 | ||
1251 | printf("\n"); | |
1252 | ns_ctx = ns_ctx->next; | |
1253 | } | |
1254 | worker = worker->next; | |
1255 | } | |
1256 | ||
1257 | if (g_latency_sw_tracking_level == 1) { | |
1258 | return; | |
1259 | } | |
1260 | ||
1261 | worker = g_workers; | |
1262 | while (worker) { | |
1263 | ns_ctx = worker->ns_ctx; | |
1264 | while (ns_ctx) { | |
1265 | printf("Latency histogram for %-43.43s from core %u:\n", ns_ctx->entry->name, worker->lcore); | |
1266 | printf("==============================================================================\n"); | |
1267 | printf(" Range in us Cumulative IO count\n"); | |
1268 | ||
1269 | spdk_histogram_data_iterate(ns_ctx->histogram, print_bucket, NULL); | |
1270 | printf("\n"); | |
7c673cae FG |
1271 | ns_ctx = ns_ctx->next; |
1272 | } | |
1273 | worker = worker->next; | |
1274 | } | |
1275 | ||
7c673cae FG |
1276 | } |
1277 | ||
1278 | static void | |
1279 | print_latency_page(struct ctrlr_entry *entry) | |
1280 | { | |
1281 | int i; | |
1282 | ||
1283 | printf("\n"); | |
1284 | printf("%s\n", entry->name); | |
1285 | printf("--------------------------------------------------------\n"); | |
1286 | ||
1287 | for (i = 0; i < 32; i++) { | |
11fdf7f2 | 1288 | if (entry->latency_page->buckets_32us[i]) { |
7c673cae | 1289 | printf("Bucket %dus - %dus: %d\n", i * 32, (i + 1) * 32, entry->latency_page->buckets_32us[i]); |
11fdf7f2 | 1290 | } |
7c673cae FG |
1291 | } |
1292 | for (i = 0; i < 31; i++) { | |
11fdf7f2 | 1293 | if (entry->latency_page->buckets_1ms[i]) { |
7c673cae | 1294 | printf("Bucket %dms - %dms: %d\n", i + 1, i + 2, entry->latency_page->buckets_1ms[i]); |
11fdf7f2 | 1295 | } |
7c673cae FG |
1296 | } |
1297 | for (i = 0; i < 31; i++) { | |
1298 | if (entry->latency_page->buckets_32ms[i]) | |
1299 | printf("Bucket %dms - %dms: %d\n", (i + 1) * 32, (i + 2) * 32, | |
1300 | entry->latency_page->buckets_32ms[i]); | |
1301 | } | |
1302 | } | |
1303 | ||
1304 | static void | |
1305 | print_latency_statistics(const char *op_name, enum spdk_nvme_intel_log_page log_page) | |
1306 | { | |
1307 | struct ctrlr_entry *ctrlr; | |
1308 | ||
1309 | printf("%s Latency Statistics:\n", op_name); | |
1310 | printf("========================================================\n"); | |
1311 | ctrlr = g_controllers; | |
1312 | while (ctrlr) { | |
1313 | if (spdk_nvme_ctrlr_is_log_page_supported(ctrlr->ctrlr, log_page)) { | |
1314 | if (spdk_nvme_ctrlr_cmd_get_log_page(ctrlr->ctrlr, log_page, SPDK_NVME_GLOBAL_NS_TAG, | |
1315 | ctrlr->latency_page, sizeof(struct spdk_nvme_intel_rw_latency_page), 0, | |
1316 | enable_latency_tracking_complete, | |
1317 | NULL)) { | |
1318 | printf("nvme_ctrlr_cmd_get_log_page() failed\n"); | |
1319 | exit(1); | |
1320 | } | |
1321 | ||
1322 | g_outstanding_commands++; | |
1323 | } else { | |
1324 | printf("Controller %s: %s latency statistics not supported\n", ctrlr->name, op_name); | |
1325 | } | |
1326 | ctrlr = ctrlr->next; | |
1327 | } | |
1328 | ||
1329 | while (g_outstanding_commands) { | |
1330 | ctrlr = g_controllers; | |
1331 | while (ctrlr) { | |
1332 | spdk_nvme_ctrlr_process_admin_completions(ctrlr->ctrlr); | |
1333 | ctrlr = ctrlr->next; | |
1334 | } | |
1335 | } | |
1336 | ||
1337 | ctrlr = g_controllers; | |
1338 | while (ctrlr) { | |
1339 | if (spdk_nvme_ctrlr_is_log_page_supported(ctrlr->ctrlr, log_page)) { | |
1340 | print_latency_page(ctrlr); | |
1341 | } | |
1342 | ctrlr = ctrlr->next; | |
1343 | } | |
1344 | printf("\n"); | |
1345 | } | |
1346 | ||
1347 | static void | |
1348 | print_stats(void) | |
1349 | { | |
1350 | print_performance(); | |
11fdf7f2 | 1351 | if (g_latency_ssd_tracking_enable) { |
7c673cae FG |
1352 | if (g_rw_percentage != 0) { |
1353 | print_latency_statistics("Read", SPDK_NVME_INTEL_LOG_READ_CMD_LATENCY); | |
1354 | } | |
1355 | if (g_rw_percentage != 100) { | |
1356 | print_latency_statistics("Write", SPDK_NVME_INTEL_LOG_WRITE_CMD_LATENCY); | |
1357 | } | |
1358 | } | |
1359 | } | |
1360 | ||
1361 | static void | |
1362 | unregister_trids(void) | |
1363 | { | |
1364 | struct trid_entry *trid_entry, *tmp; | |
1365 | ||
1366 | TAILQ_FOREACH_SAFE(trid_entry, &g_trid_list, tailq, tmp) { | |
9f95a23c | 1367 | TAILQ_REMOVE(&g_trid_list, trid_entry, tailq); |
7c673cae FG |
1368 | free(trid_entry); |
1369 | } | |
1370 | } | |
1371 | ||
1372 | static int | |
1373 | add_trid(const char *trid_str) | |
1374 | { | |
1375 | struct trid_entry *trid_entry; | |
1376 | struct spdk_nvme_transport_id *trid; | |
11fdf7f2 | 1377 | char *ns; |
7c673cae FG |
1378 | |
1379 | trid_entry = calloc(1, sizeof(*trid_entry)); | |
1380 | if (trid_entry == NULL) { | |
1381 | return -1; | |
1382 | } | |
1383 | ||
1384 | trid = &trid_entry->trid; | |
7c673cae FG |
1385 | trid->trtype = SPDK_NVME_TRANSPORT_PCIE; |
1386 | snprintf(trid->subnqn, sizeof(trid->subnqn), "%s", SPDK_NVMF_DISCOVERY_NQN); | |
1387 | ||
1388 | if (spdk_nvme_transport_id_parse(trid, trid_str) != 0) { | |
1389 | fprintf(stderr, "Invalid transport ID format '%s'\n", trid_str); | |
1390 | free(trid_entry); | |
1391 | return 1; | |
1392 | } | |
1393 | ||
11fdf7f2 TL |
1394 | ns = strcasestr(trid_str, "ns:"); |
1395 | if (ns) { | |
1396 | char nsid_str[6]; /* 5 digits maximum in an nsid */ | |
1397 | int len; | |
1398 | int nsid; | |
1399 | ||
1400 | ns += 3; | |
1401 | ||
1402 | len = strcspn(ns, " \t\n"); | |
1403 | if (len > 5) { | |
1404 | fprintf(stderr, "NVMe namespace IDs must be 5 digits or less\n"); | |
1405 | free(trid_entry); | |
1406 | return 1; | |
1407 | } | |
1408 | ||
1409 | memcpy(nsid_str, ns, len); | |
1410 | nsid_str[len] = '\0'; | |
1411 | ||
9f95a23c | 1412 | nsid = spdk_strtol(nsid_str, 10); |
11fdf7f2 TL |
1413 | if (nsid <= 0 || nsid > 65535) { |
1414 | fprintf(stderr, "NVMe namespace IDs must be less than 65536 and greater than 0\n"); | |
1415 | free(trid_entry); | |
1416 | return 1; | |
1417 | } | |
1418 | ||
1419 | trid_entry->nsid = (uint16_t)nsid; | |
1420 | } | |
1421 | ||
7c673cae FG |
1422 | TAILQ_INSERT_TAIL(&g_trid_list, trid_entry, tailq); |
1423 | return 0; | |
1424 | } | |
1425 | ||
9f95a23c TL |
1426 | static size_t |
1427 | parse_next_key(const char **str, char *key, char *val, size_t key_buf_size, | |
1428 | size_t val_buf_size) | |
11fdf7f2 TL |
1429 | { |
1430 | const char *sep; | |
9f95a23c TL |
1431 | const char *separator = ", \t\n"; |
1432 | size_t key_len, val_len; | |
11fdf7f2 | 1433 | |
9f95a23c | 1434 | *str += strspn(*str, separator); |
11fdf7f2 | 1435 | |
9f95a23c | 1436 | sep = strchr(*str, '='); |
11fdf7f2 | 1437 | if (!sep) { |
9f95a23c | 1438 | fprintf(stderr, "Key without '=' separator\n"); |
11fdf7f2 TL |
1439 | return 0; |
1440 | } | |
1441 | ||
9f95a23c TL |
1442 | key_len = sep - *str; |
1443 | if (key_len >= key_buf_size) { | |
1444 | fprintf(stderr, "Key length %zu is greater than maximum allowed %zu\n", | |
1445 | key_len, key_buf_size - 1); | |
1446 | return 0; | |
1447 | } | |
1448 | ||
1449 | memcpy(key, *str, key_len); | |
1450 | key[key_len] = '\0'; | |
1451 | ||
1452 | *str += key_len + 1; /* Skip key */ | |
1453 | val_len = strcspn(*str, separator); | |
1454 | if (val_len == 0) { | |
1455 | fprintf(stderr, "Key without value\n"); | |
1456 | return 0; | |
1457 | } | |
1458 | ||
1459 | if (val_len >= val_buf_size) { | |
1460 | fprintf(stderr, "Value length %zu is greater than maximum allowed %zu\n", | |
1461 | val_len, val_buf_size - 1); | |
1462 | return 0; | |
1463 | } | |
1464 | ||
1465 | memcpy(val, *str, val_len); | |
1466 | val[val_len] = '\0'; | |
1467 | ||
1468 | *str += val_len; | |
1469 | ||
1470 | return val_len; | |
1471 | } | |
1472 | ||
1473 | static int | |
1474 | parse_metadata(const char *metacfg_str) | |
1475 | { | |
1476 | const char *str; | |
1477 | size_t val_len; | |
1478 | char key[32]; | |
1479 | char val[1024]; | |
1480 | ||
1481 | if (metacfg_str == NULL) { | |
1482 | return -EINVAL; | |
1483 | } | |
1484 | ||
1485 | str = metacfg_str; | |
1486 | ||
1487 | while (*str != '\0') { | |
1488 | val_len = parse_next_key(&str, key, val, sizeof(key), sizeof(val)); | |
1489 | if (val_len == 0) { | |
1490 | fprintf(stderr, "Failed to parse metadata\n"); | |
1491 | return -EINVAL; | |
11fdf7f2 | 1492 | } |
9f95a23c TL |
1493 | |
1494 | if (strcmp(key, "PRACT") == 0) { | |
1495 | if (*val == '1') { | |
1496 | g_metacfg_prchk_flags = SPDK_NVME_IO_FLAGS_PRACT; | |
1497 | } | |
1498 | } else if (strcmp(key, "PRCHK") == 0) { | |
1499 | if (strstr(val, "GUARD") != NULL) { | |
1500 | g_metacfg_prchk_flags |= SPDK_NVME_IO_FLAGS_PRCHK_GUARD; | |
1501 | } | |
1502 | if (strstr(val, "REFTAG") != NULL) { | |
1503 | g_metacfg_prchk_flags |= SPDK_NVME_IO_FLAGS_PRCHK_REFTAG; | |
1504 | } | |
1505 | if (strstr(val, "APPTAG") != NULL) { | |
1506 | g_metacfg_prchk_flags |= SPDK_NVME_IO_FLAGS_PRCHK_APPTAG; | |
1507 | } | |
1508 | } else { | |
1509 | fprintf(stderr, "Unknown key '%s'\n", key); | |
11fdf7f2 TL |
1510 | } |
1511 | } | |
1512 | ||
1513 | return 0; | |
1514 | } | |
1515 | ||
7c673cae FG |
1516 | static int |
1517 | parse_args(int argc, char **argv) | |
1518 | { | |
1519 | const char *workload_type; | |
1520 | int op; | |
1521 | bool mix_specified = false; | |
9f95a23c | 1522 | long int val; |
7c673cae | 1523 | |
11fdf7f2 | 1524 | /* default value */ |
7c673cae FG |
1525 | g_queue_depth = 0; |
1526 | g_io_size_bytes = 0; | |
1527 | workload_type = NULL; | |
1528 | g_time_in_sec = 0; | |
1529 | g_rw_percentage = -1; | |
1530 | g_core_mask = NULL; | |
1531 | g_max_completions = 0; | |
1532 | ||
9f95a23c | 1533 | while ((op = getopt(argc, argv, "c:e:i:lm:n:o:q:r:k:s:t:w:DGHILM:U:")) != -1) { |
7c673cae | 1534 | switch (op) { |
9f95a23c TL |
1535 | case 'i': |
1536 | case 'm': | |
1537 | case 'n': | |
1538 | case 'o': | |
1539 | case 'q': | |
1540 | case 'k': | |
1541 | case 's': | |
1542 | case 't': | |
1543 | case 'M': | |
1544 | case 'U': | |
1545 | val = spdk_strtol(optarg, 10); | |
1546 | if (val < 0) { | |
1547 | fprintf(stderr, "Converting a string to integer failed\n"); | |
1548 | return val; | |
1549 | } | |
1550 | switch (op) { | |
1551 | case 'i': | |
1552 | g_shm_id = val; | |
1553 | break; | |
1554 | case 'm': | |
1555 | g_max_completions = val; | |
1556 | break; | |
1557 | case 'n': | |
1558 | g_nr_io_queues_per_ns = val; | |
1559 | break; | |
1560 | case 'o': | |
1561 | g_io_size_bytes = val; | |
1562 | break; | |
1563 | case 'q': | |
1564 | g_queue_depth = val; | |
1565 | break; | |
1566 | case 'k': | |
1567 | g_keep_alive_timeout_in_ms = val; | |
1568 | break; | |
1569 | case 's': | |
1570 | g_dpdk_mem = val; | |
1571 | break; | |
1572 | case 't': | |
1573 | g_time_in_sec = val; | |
1574 | break; | |
1575 | case 'M': | |
1576 | g_rw_percentage = val; | |
1577 | mix_specified = true; | |
1578 | break; | |
1579 | case 'U': | |
1580 | g_nr_unused_io_queues = val; | |
1581 | break; | |
1582 | } | |
1583 | break; | |
7c673cae FG |
1584 | case 'c': |
1585 | g_core_mask = optarg; | |
1586 | break; | |
11fdf7f2 TL |
1587 | case 'e': |
1588 | if (parse_metadata(optarg)) { | |
1589 | usage(argv[0]); | |
1590 | return 1; | |
1591 | } | |
7c673cae | 1592 | break; |
7c673cae | 1593 | case 'l': |
11fdf7f2 | 1594 | g_latency_ssd_tracking_enable = true; |
7c673cae | 1595 | break; |
7c673cae FG |
1596 | case 'r': |
1597 | if (add_trid(optarg)) { | |
1598 | usage(argv[0]); | |
1599 | return 1; | |
1600 | } | |
1601 | break; | |
7c673cae FG |
1602 | case 'w': |
1603 | workload_type = optarg; | |
1604 | break; | |
11fdf7f2 TL |
1605 | case 'D': |
1606 | g_disable_sq_cmb = 1; | |
1607 | break; | |
9f95a23c TL |
1608 | case 'G': |
1609 | #ifndef DEBUG | |
1610 | fprintf(stderr, "%s must be configured with --enable-debug for -G flag\n", | |
1611 | argv[0]); | |
1612 | usage(argv[0]); | |
1613 | return 1; | |
1614 | #else | |
1615 | spdk_log_set_flag("nvme"); | |
1616 | spdk_log_set_print_level(SPDK_LOG_DEBUG); | |
1617 | break; | |
1618 | #endif | |
1619 | case 'H': | |
1620 | g_header_digest = 1; | |
1621 | break; | |
1622 | case 'I': | |
1623 | g_data_digest = 1; | |
1624 | break; | |
11fdf7f2 TL |
1625 | case 'L': |
1626 | g_latency_sw_tracking_level++; | |
1627 | break; | |
7c673cae FG |
1628 | default: |
1629 | usage(argv[0]); | |
1630 | return 1; | |
1631 | } | |
1632 | } | |
1633 | ||
9f95a23c TL |
1634 | if (!g_nr_io_queues_per_ns) { |
1635 | usage(argv[0]); | |
1636 | return 1; | |
1637 | } | |
1638 | ||
7c673cae FG |
1639 | if (!g_queue_depth) { |
1640 | usage(argv[0]); | |
1641 | return 1; | |
1642 | } | |
1643 | if (!g_io_size_bytes) { | |
1644 | usage(argv[0]); | |
1645 | return 1; | |
1646 | } | |
1647 | if (!workload_type) { | |
1648 | usage(argv[0]); | |
1649 | return 1; | |
1650 | } | |
1651 | if (!g_time_in_sec) { | |
1652 | usage(argv[0]); | |
1653 | return 1; | |
1654 | } | |
1655 | ||
1656 | if (strcmp(workload_type, "read") && | |
1657 | strcmp(workload_type, "write") && | |
1658 | strcmp(workload_type, "randread") && | |
1659 | strcmp(workload_type, "randwrite") && | |
1660 | strcmp(workload_type, "rw") && | |
1661 | strcmp(workload_type, "randrw")) { | |
1662 | fprintf(stderr, | |
1663 | "io pattern type must be one of\n" | |
1664 | "(read, write, randread, randwrite, rw, randrw)\n"); | |
1665 | return 1; | |
1666 | } | |
1667 | ||
1668 | if (!strcmp(workload_type, "read") || | |
1669 | !strcmp(workload_type, "randread")) { | |
1670 | g_rw_percentage = 100; | |
1671 | } | |
1672 | ||
1673 | if (!strcmp(workload_type, "write") || | |
1674 | !strcmp(workload_type, "randwrite")) { | |
1675 | g_rw_percentage = 0; | |
1676 | } | |
1677 | ||
1678 | if (!strcmp(workload_type, "read") || | |
1679 | !strcmp(workload_type, "randread") || | |
1680 | !strcmp(workload_type, "write") || | |
1681 | !strcmp(workload_type, "randwrite")) { | |
1682 | if (mix_specified) { | |
1683 | fprintf(stderr, "Ignoring -M option... Please use -M option" | |
1684 | " only when using rw or randrw.\n"); | |
1685 | } | |
1686 | } | |
1687 | ||
1688 | if (!strcmp(workload_type, "rw") || | |
1689 | !strcmp(workload_type, "randrw")) { | |
1690 | if (g_rw_percentage < 0 || g_rw_percentage > 100) { | |
1691 | fprintf(stderr, | |
1692 | "-M must be specified to value from 0 to 100 " | |
1693 | "for rw or randrw.\n"); | |
1694 | return 1; | |
1695 | } | |
1696 | } | |
1697 | ||
1698 | if (!strcmp(workload_type, "read") || | |
1699 | !strcmp(workload_type, "write") || | |
1700 | !strcmp(workload_type, "rw")) { | |
1701 | g_is_random = 0; | |
1702 | } else { | |
1703 | g_is_random = 1; | |
1704 | } | |
1705 | ||
1706 | if (TAILQ_EMPTY(&g_trid_list)) { | |
1707 | /* If no transport IDs specified, default to enumerating all local PCIe devices */ | |
1708 | add_trid("trtype:PCIe"); | |
11fdf7f2 TL |
1709 | } else { |
1710 | struct trid_entry *trid_entry, *trid_entry_tmp; | |
1711 | ||
1712 | g_no_pci = true; | |
1713 | /* check whether there is local PCIe type */ | |
1714 | TAILQ_FOREACH_SAFE(trid_entry, &g_trid_list, tailq, trid_entry_tmp) { | |
1715 | if (trid_entry->trid.trtype == SPDK_NVME_TRANSPORT_PCIE) { | |
1716 | g_no_pci = false; | |
1717 | break; | |
1718 | } | |
1719 | } | |
7c673cae FG |
1720 | } |
1721 | ||
1722 | g_aio_optind = optind; | |
11fdf7f2 | 1723 | |
7c673cae FG |
1724 | return 0; |
1725 | } | |
1726 | ||
1727 | static int | |
1728 | register_workers(void) | |
1729 | { | |
1730 | uint32_t i; | |
1731 | struct worker_thread *worker; | |
1732 | ||
1733 | g_workers = NULL; | |
1734 | g_num_workers = 0; | |
1735 | ||
1736 | SPDK_ENV_FOREACH_CORE(i) { | |
1737 | worker = calloc(1, sizeof(*worker)); | |
1738 | if (worker == NULL) { | |
1739 | fprintf(stderr, "Unable to allocate worker\n"); | |
1740 | return -1; | |
1741 | } | |
1742 | ||
1743 | worker->lcore = i; | |
1744 | worker->next = g_workers; | |
1745 | g_workers = worker; | |
1746 | g_num_workers++; | |
1747 | } | |
1748 | ||
1749 | return 0; | |
1750 | } | |
1751 | ||
1752 | static void | |
1753 | unregister_workers(void) | |
1754 | { | |
1755 | struct worker_thread *worker = g_workers; | |
1756 | ||
1757 | /* Free namespace context and worker thread */ | |
1758 | while (worker) { | |
1759 | struct worker_thread *next_worker = worker->next; | |
1760 | struct ns_worker_ctx *ns_ctx = worker->ns_ctx; | |
1761 | ||
1762 | while (ns_ctx) { | |
1763 | struct ns_worker_ctx *next_ns_ctx = ns_ctx->next; | |
11fdf7f2 | 1764 | spdk_histogram_data_free(ns_ctx->histogram); |
7c673cae FG |
1765 | free(ns_ctx); |
1766 | ns_ctx = next_ns_ctx; | |
1767 | } | |
1768 | ||
1769 | free(worker); | |
1770 | worker = next_worker; | |
1771 | } | |
1772 | } | |
1773 | ||
1774 | static bool | |
1775 | probe_cb(void *cb_ctx, const struct spdk_nvme_transport_id *trid, | |
1776 | struct spdk_nvme_ctrlr_opts *opts) | |
1777 | { | |
7c673cae FG |
1778 | if (trid->trtype != SPDK_NVME_TRANSPORT_PCIE) { |
1779 | printf("Attaching to NVMe over Fabrics controller at %s:%s: %s\n", | |
1780 | trid->traddr, trid->trsvcid, | |
1781 | trid->subnqn); | |
1782 | } else { | |
11fdf7f2 TL |
1783 | if (g_disable_sq_cmb) { |
1784 | opts->use_cmb_sqs = false; | |
7c673cae FG |
1785 | } |
1786 | ||
11fdf7f2 TL |
1787 | printf("Attaching to NVMe Controller at %s\n", |
1788 | trid->traddr); | |
7c673cae FG |
1789 | } |
1790 | ||
11fdf7f2 TL |
1791 | /* Set io_queue_size to UINT16_MAX, NVMe driver |
1792 | * will then reduce this to MQES to maximize | |
1793 | * the io_queue_size as much as possible. | |
1794 | */ | |
1795 | opts->io_queue_size = UINT16_MAX; | |
7c673cae | 1796 | |
9f95a23c TL |
1797 | /* Set the header and data_digest */ |
1798 | opts->header_digest = g_header_digest; | |
1799 | opts->data_digest = g_data_digest; | |
1800 | opts->keep_alive_timeout_ms = spdk_max(opts->keep_alive_timeout_ms, | |
1801 | g_keep_alive_timeout_in_ms); | |
1802 | ||
7c673cae FG |
1803 | return true; |
1804 | } | |
1805 | ||
1806 | static void | |
1807 | attach_cb(void *cb_ctx, const struct spdk_nvme_transport_id *trid, | |
1808 | struct spdk_nvme_ctrlr *ctrlr, const struct spdk_nvme_ctrlr_opts *opts) | |
1809 | { | |
11fdf7f2 | 1810 | struct trid_entry *trid_entry = cb_ctx; |
7c673cae FG |
1811 | struct spdk_pci_addr pci_addr; |
1812 | struct spdk_pci_device *pci_dev; | |
1813 | struct spdk_pci_id pci_id; | |
1814 | ||
11fdf7f2 | 1815 | g_controllers_found++; |
7c673cae FG |
1816 | if (trid->trtype != SPDK_NVME_TRANSPORT_PCIE) { |
1817 | printf("Attached to NVMe over Fabrics controller at %s:%s: %s\n", | |
1818 | trid->traddr, trid->trsvcid, | |
1819 | trid->subnqn); | |
1820 | } else { | |
1821 | if (spdk_pci_addr_parse(&pci_addr, trid->traddr)) { | |
1822 | return; | |
1823 | } | |
1824 | ||
11fdf7f2 | 1825 | pci_dev = spdk_nvme_ctrlr_get_pci_device(ctrlr); |
7c673cae FG |
1826 | if (!pci_dev) { |
1827 | return; | |
1828 | } | |
1829 | ||
1830 | pci_id = spdk_pci_device_get_id(pci_dev); | |
1831 | ||
1832 | printf("Attached to NVMe Controller at %s [%04x:%04x]\n", | |
1833 | trid->traddr, | |
1834 | pci_id.vendor_id, pci_id.device_id); | |
1835 | } | |
1836 | ||
11fdf7f2 | 1837 | register_ctrlr(ctrlr, trid_entry); |
7c673cae FG |
1838 | } |
1839 | ||
1840 | static int | |
1841 | register_controllers(void) | |
1842 | { | |
1843 | struct trid_entry *trid_entry; | |
1844 | ||
1845 | printf("Initializing NVMe Controllers\n"); | |
1846 | ||
1847 | TAILQ_FOREACH(trid_entry, &g_trid_list, tailq) { | |
11fdf7f2 | 1848 | if (spdk_nvme_probe(&trid_entry->trid, trid_entry, probe_cb, attach_cb, NULL) != 0) { |
7c673cae FG |
1849 | fprintf(stderr, "spdk_nvme_probe() failed for transport address '%s'\n", |
1850 | trid_entry->trid.traddr); | |
1851 | return -1; | |
1852 | } | |
1853 | } | |
1854 | ||
1855 | return 0; | |
1856 | } | |
1857 | ||
1858 | static void | |
1859 | unregister_controllers(void) | |
1860 | { | |
1861 | struct ctrlr_entry *entry = g_controllers; | |
1862 | ||
1863 | while (entry) { | |
1864 | struct ctrlr_entry *next = entry->next; | |
11fdf7f2 TL |
1865 | spdk_dma_free(entry->latency_page); |
1866 | if (g_latency_ssd_tracking_enable && | |
1867 | spdk_nvme_ctrlr_is_feature_supported(entry->ctrlr, SPDK_NVME_INTEL_FEAT_LATENCY_TRACKING)) { | |
7c673cae | 1868 | set_latency_tracking_feature(entry->ctrlr, false); |
11fdf7f2 | 1869 | } |
7c673cae | 1870 | |
9f95a23c TL |
1871 | if (g_nr_unused_io_queues) { |
1872 | int i; | |
7c673cae | 1873 | |
9f95a23c TL |
1874 | for (i = 0; i < g_nr_unused_io_queues; i++) { |
1875 | spdk_nvme_ctrlr_free_io_qpair(entry->unused_qpairs[i]); | |
1876 | } | |
1877 | ||
1878 | free(entry->unused_qpairs); | |
7c673cae | 1879 | } |
7c673cae | 1880 | |
9f95a23c TL |
1881 | spdk_nvme_detach(entry->ctrlr); |
1882 | free(entry); | |
1883 | entry = next; | |
1884 | } | |
7c673cae FG |
1885 | } |
1886 | ||
1887 | static int | |
1888 | associate_workers_with_ns(void) | |
1889 | { | |
1890 | struct ns_entry *entry = g_namespaces; | |
1891 | struct worker_thread *worker = g_workers; | |
1892 | struct ns_worker_ctx *ns_ctx; | |
1893 | int i, count; | |
1894 | ||
1895 | count = g_num_namespaces > g_num_workers ? g_num_namespaces : g_num_workers; | |
1896 | ||
1897 | for (i = 0; i < count; i++) { | |
1898 | if (entry == NULL) { | |
1899 | break; | |
1900 | } | |
1901 | ||
9f95a23c | 1902 | ns_ctx = calloc(1, sizeof(struct ns_worker_ctx)); |
7c673cae FG |
1903 | if (!ns_ctx) { |
1904 | return -1; | |
1905 | } | |
7c673cae FG |
1906 | |
1907 | printf("Associating %s with lcore %d\n", entry->name, worker->lcore); | |
1908 | ns_ctx->min_tsc = UINT64_MAX; | |
1909 | ns_ctx->entry = entry; | |
1910 | ns_ctx->next = worker->ns_ctx; | |
11fdf7f2 | 1911 | ns_ctx->histogram = spdk_histogram_data_alloc(); |
7c673cae FG |
1912 | worker->ns_ctx = ns_ctx; |
1913 | ||
1914 | worker = worker->next; | |
1915 | if (worker == NULL) { | |
1916 | worker = g_workers; | |
1917 | } | |
1918 | ||
1919 | entry = entry->next; | |
1920 | if (entry == NULL) { | |
1921 | entry = g_namespaces; | |
1922 | } | |
1923 | ||
1924 | } | |
1925 | ||
1926 | return 0; | |
1927 | } | |
1928 | ||
9f95a23c TL |
1929 | static void * |
1930 | nvme_poll_ctrlrs(void *arg) | |
1931 | { | |
1932 | struct ctrlr_entry *entry; | |
1933 | int oldstate; | |
1934 | ||
1935 | spdk_unaffinitize_thread(); | |
1936 | ||
1937 | while (true) { | |
1938 | pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate); | |
1939 | ||
1940 | entry = g_controllers; | |
1941 | while (entry) { | |
1942 | if (entry->trtype != SPDK_NVME_TRANSPORT_PCIE) { | |
1943 | spdk_nvme_ctrlr_process_admin_completions(entry->ctrlr); | |
1944 | } | |
1945 | entry = entry->next; | |
1946 | } | |
1947 | ||
1948 | pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &oldstate); | |
1949 | ||
1950 | /* This is a pthread cancellation point and cannot be removed. */ | |
1951 | sleep(1); | |
1952 | } | |
1953 | ||
1954 | return NULL; | |
1955 | } | |
1956 | ||
7c673cae FG |
1957 | int main(int argc, char **argv) |
1958 | { | |
1959 | int rc; | |
1960 | struct worker_thread *worker, *master_worker; | |
1961 | unsigned master_core; | |
7c673cae | 1962 | struct spdk_env_opts opts; |
9f95a23c | 1963 | pthread_t thread_id = 0; |
7c673cae FG |
1964 | |
1965 | rc = parse_args(argc, argv); | |
1966 | if (rc != 0) { | |
1967 | return rc; | |
1968 | } | |
1969 | ||
1970 | spdk_env_opts_init(&opts); | |
1971 | opts.name = "perf"; | |
1972 | opts.shm_id = g_shm_id; | |
1973 | if (g_core_mask) { | |
1974 | opts.core_mask = g_core_mask; | |
1975 | } | |
1976 | ||
1977 | if (g_dpdk_mem) { | |
11fdf7f2 TL |
1978 | opts.mem_size = g_dpdk_mem; |
1979 | } | |
1980 | if (g_no_pci) { | |
1981 | opts.no_pci = g_no_pci; | |
1982 | } | |
1983 | if (spdk_env_init(&opts) < 0) { | |
1984 | fprintf(stderr, "Unable to initialize SPDK env\n"); | |
1985 | rc = -1; | |
1986 | goto cleanup; | |
7c673cae | 1987 | } |
7c673cae FG |
1988 | |
1989 | g_tsc_rate = spdk_get_ticks_hz(); | |
1990 | ||
1991 | if (register_workers() != 0) { | |
1992 | rc = -1; | |
1993 | goto cleanup; | |
1994 | } | |
1995 | ||
9f95a23c | 1996 | #if HAVE_LIBAIO |
7c673cae FG |
1997 | if (register_aio_files(argc, argv) != 0) { |
1998 | rc = -1; | |
1999 | goto cleanup; | |
2000 | } | |
9f95a23c | 2001 | #endif |
7c673cae FG |
2002 | |
2003 | if (register_controllers() != 0) { | |
2004 | rc = -1; | |
2005 | goto cleanup; | |
2006 | } | |
2007 | ||
11fdf7f2 TL |
2008 | if (g_warn) { |
2009 | printf("WARNING: Some requested NVMe devices were skipped\n"); | |
7c673cae FG |
2010 | } |
2011 | ||
11fdf7f2 TL |
2012 | if (g_num_namespaces == 0) { |
2013 | fprintf(stderr, "No valid NVMe controllers or AIO devices found\n"); | |
9f95a23c TL |
2014 | goto cleanup; |
2015 | } | |
2016 | ||
2017 | rc = pthread_create(&thread_id, NULL, &nvme_poll_ctrlrs, NULL); | |
2018 | if (rc != 0) { | |
2019 | fprintf(stderr, "Unable to spawn a thread to poll admin queues.\n"); | |
2020 | goto cleanup; | |
11fdf7f2 | 2021 | } |
7c673cae | 2022 | |
11fdf7f2 | 2023 | if (associate_workers_with_ns() != 0) { |
7c673cae FG |
2024 | rc = -1; |
2025 | goto cleanup; | |
2026 | } | |
2027 | ||
2028 | printf("Initialization complete. Launching workers.\n"); | |
2029 | ||
2030 | /* Launch all of the slave workers */ | |
11fdf7f2 | 2031 | master_core = spdk_env_get_current_core(); |
7c673cae FG |
2032 | master_worker = NULL; |
2033 | worker = g_workers; | |
2034 | while (worker != NULL) { | |
2035 | if (worker->lcore != master_core) { | |
11fdf7f2 | 2036 | spdk_env_thread_launch_pinned(worker->lcore, work_fn, worker); |
7c673cae FG |
2037 | } else { |
2038 | assert(master_worker == NULL); | |
2039 | master_worker = worker; | |
2040 | } | |
2041 | worker = worker->next; | |
2042 | } | |
2043 | ||
2044 | assert(master_worker != NULL); | |
2045 | rc = work_fn(master_worker); | |
2046 | ||
11fdf7f2 | 2047 | spdk_env_thread_wait_all(); |
7c673cae FG |
2048 | |
2049 | print_stats(); | |
2050 | ||
2051 | cleanup: | |
9f95a23c TL |
2052 | if (thread_id && pthread_cancel(thread_id) == 0) { |
2053 | pthread_join(thread_id, NULL); | |
2054 | } | |
7c673cae FG |
2055 | unregister_trids(); |
2056 | unregister_namespaces(); | |
2057 | unregister_controllers(); | |
2058 | unregister_workers(); | |
2059 | ||
2060 | if (rc != 0) { | |
2061 | fprintf(stderr, "%s: errors occured\n", argv[0]); | |
2062 | } | |
2063 | ||
2064 | return rc; | |
2065 | } |