Commit | Line | Data |
---|---|---|
ebac4655 JA |
1 | /* |
2 | * fio - the flexible io tester | |
3 | * | |
4 | * Copyright (C) 2005 Jens Axboe <axboe@suse.de> | |
aae22ca7 | 5 | * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk> |
ebac4655 JA |
6 | * |
7 | * This program is free software; you can redistribute it and/or modify | |
8 | * it under the terms of the GNU General Public License as published by | |
9 | * the Free Software Foundation; either version 2 of the License, or | |
10 | * (at your option) any later version. | |
11 | * | |
12 | * This program is distributed in the hope that it will be useful, | |
13 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
14 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
15 | * GNU General Public License for more details. | |
16 | * | |
17 | * You should have received a copy of the GNU General Public License | |
18 | * along with this program; if not, write to the Free Software | |
19 | * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA | |
20 | * | |
21 | */ | |
ebac4655 JA |
22 | #include <unistd.h> |
23 | #include <fcntl.h> | |
24 | #include <string.h> | |
ebac4655 JA |
25 | #include <signal.h> |
26 | #include <time.h> | |
ebac4655 | 27 | #include <assert.h> |
ebac4655 JA |
28 | #include <sys/stat.h> |
29 | #include <sys/wait.h> | |
30 | #include <sys/ipc.h> | |
31 | #include <sys/shm.h> | |
32 | #include <sys/ioctl.h> | |
33 | #include <sys/mman.h> | |
34 | ||
35 | #include "fio.h" | |
36 | #include "os.h" | |
37 | ||
38 | #define MASK (4095) | |
39 | ||
40 | #define ALIGN(buf) (char *) (((unsigned long) (buf) + MASK) & ~(MASK)) | |
41 | ||
42 | int groupid = 0; | |
43 | int thread_number = 0; | |
44 | static char run_str[MAX_JOBS + 1]; | |
45 | int shm_id = 0; | |
5289b847 | 46 | static struct timeval genesis; |
53cdc686 | 47 | int temp_stall_ts; |
c1d5725e | 48 | char *fio_inst_prefix = _INST_PREFIX; |
ebac4655 | 49 | |
ebac4655 JA |
50 | static void print_thread_status(void); |
51 | ||
c04f7ec3 JA |
52 | extern unsigned long long mlock_size; |
53 | ||
ebac4655 | 54 | /* |
79809113 JA |
55 | * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it |
56 | * will never back again. It may cycle between running/verififying/fsyncing. | |
57 | * Once the thread reaches TD_EXITED, it is just waiting for the core to | |
58 | * reap it. | |
ebac4655 JA |
59 | */ |
60 | enum { | |
61 | TD_NOT_CREATED = 0, | |
62 | TD_CREATED, | |
75154845 | 63 | TD_INITIALIZED, |
ebac4655 JA |
64 | TD_RUNNING, |
65 | TD_VERIFYING, | |
5853e5a8 | 66 | TD_FSYNCING, |
ebac4655 JA |
67 | TD_EXITED, |
68 | TD_REAPED, | |
69 | }; | |
70 | ||
3d60d1ed | 71 | #define should_fsync(td) ((td_write(td) || td_rw(td)) && (!(td)->odirect || (td)->override_sync)) |
ebac4655 | 72 | |
bbfd6b00 | 73 | static volatile int startup_sem; |
ebac4655 JA |
74 | |
75 | #define TERMINATE_ALL (-1) | |
75154845 | 76 | #define JOB_START_TIMEOUT (5 * 1000) |
ebac4655 JA |
77 | |
78 | static void terminate_threads(int group_id) | |
79 | { | |
80 | int i; | |
81 | ||
82 | for (i = 0; i < thread_number; i++) { | |
83 | struct thread_data *td = &threads[i]; | |
84 | ||
85 | if (group_id == TERMINATE_ALL || groupid == td->groupid) { | |
86 | td->terminate = 1; | |
87 | td->start_delay = 0; | |
88 | } | |
89 | } | |
90 | } | |
91 | ||
92 | static void sig_handler(int sig) | |
93 | { | |
94 | switch (sig) { | |
95 | case SIGALRM: | |
96 | update_io_ticks(); | |
97 | disk_util_timer_arm(); | |
98 | print_thread_status(); | |
99 | break; | |
100 | default: | |
101 | printf("\nfio: terminating on signal\n"); | |
102 | fflush(stdout); | |
103 | terminate_threads(TERMINATE_ALL); | |
104 | break; | |
105 | } | |
106 | } | |
107 | ||
906c8d75 JA |
108 | /* |
109 | * The ->file_map[] contains a map of blocks we have or have not done io | |
110 | * to yet. Used to make sure we cover the entire range in a fair fashion. | |
111 | */ | |
53cdc686 JA |
112 | static int random_map_free(struct thread_data *td, struct fio_file *f, |
113 | unsigned long long block) | |
ebac4655 | 114 | { |
53cdc686 JA |
115 | unsigned int idx = RAND_MAP_IDX(td, f, block); |
116 | unsigned int bit = RAND_MAP_BIT(td, f, block); | |
ebac4655 | 117 | |
53cdc686 | 118 | return (f->file_map[idx] & (1UL << bit)) == 0; |
ebac4655 JA |
119 | } |
120 | ||
906c8d75 JA |
121 | /* |
122 | * Return the next free block in the map. | |
123 | */ | |
53cdc686 JA |
124 | static int get_next_free_block(struct thread_data *td, struct fio_file *f, |
125 | unsigned long long *b) | |
ebac4655 JA |
126 | { |
127 | int i; | |
128 | ||
129 | *b = 0; | |
130 | i = 0; | |
53cdc686 JA |
131 | while ((*b) * td->min_bs < f->file_size) { |
132 | if (f->file_map[i] != -1UL) { | |
133 | *b += ffz(f->file_map[i]); | |
ebac4655 JA |
134 | return 0; |
135 | } | |
136 | ||
137 | *b += BLOCKS_PER_MAP; | |
138 | i++; | |
139 | } | |
140 | ||
141 | return 1; | |
142 | } | |
143 | ||
906c8d75 JA |
144 | /* |
145 | * Mark a given offset as used in the map. | |
146 | */ | |
53cdc686 JA |
147 | static void mark_random_map(struct thread_data *td, struct fio_file *f, |
148 | struct io_u *io_u) | |
ebac4655 | 149 | { |
200bc855 | 150 | unsigned long long block = io_u->offset / (unsigned long long) td->min_bs; |
ebac4655 JA |
151 | unsigned int blocks = 0; |
152 | ||
153 | while (blocks < (io_u->buflen / td->min_bs)) { | |
154 | unsigned int idx, bit; | |
155 | ||
53cdc686 | 156 | if (!random_map_free(td, f, block)) |
ebac4655 JA |
157 | break; |
158 | ||
53cdc686 JA |
159 | idx = RAND_MAP_IDX(td, f, block); |
160 | bit = RAND_MAP_BIT(td, f, block); | |
ebac4655 | 161 | |
53cdc686 | 162 | assert(idx < f->num_maps); |
ebac4655 | 163 | |
53cdc686 | 164 | f->file_map[idx] |= (1UL << bit); |
ebac4655 JA |
165 | block++; |
166 | blocks++; | |
167 | } | |
168 | ||
169 | if ((blocks * td->min_bs) < io_u->buflen) | |
170 | io_u->buflen = blocks * td->min_bs; | |
171 | } | |
172 | ||
906c8d75 JA |
173 | /* |
174 | * For random io, generate a random new block and see if it's used. Repeat | |
175 | * until we find a free one. For sequential io, just return the end of | |
176 | * the last io issued. | |
177 | */ | |
53cdc686 JA |
178 | static int get_next_offset(struct thread_data *td, struct fio_file *f, |
179 | unsigned long long *offset) | |
20dc95c4 JA |
180 | { |
181 | unsigned long long b, rb; | |
182 | long r; | |
183 | ||
184 | if (!td->sequential) { | |
085227ad | 185 | unsigned long long max_blocks = td->io_size / td->min_bs; |
20dc95c4 JA |
186 | int loops = 50; |
187 | ||
188 | do { | |
6dfd46b9 | 189 | r = os_random_long(&td->random_state); |
085227ad | 190 | b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0)); |
53cdc686 | 191 | rb = b + (f->file_offset / td->min_bs); |
20dc95c4 | 192 | loops--; |
53cdc686 | 193 | } while (!random_map_free(td, f, rb) && loops); |
20dc95c4 JA |
194 | |
195 | if (!loops) { | |
53cdc686 | 196 | if (get_next_free_block(td, f, &b)) |
20dc95c4 JA |
197 | return 1; |
198 | } | |
199 | } else | |
53cdc686 | 200 | b = f->last_pos / td->min_bs; |
20dc95c4 | 201 | |
53cdc686 JA |
202 | *offset = (b * td->min_bs) + f->file_offset; |
203 | if (*offset > f->file_size) | |
20dc95c4 JA |
204 | return 1; |
205 | ||
206 | return 0; | |
207 | } | |
208 | ||
209 | static unsigned int get_next_buflen(struct thread_data *td) | |
210 | { | |
211 | unsigned int buflen; | |
212 | long r; | |
213 | ||
214 | if (td->min_bs == td->max_bs) | |
215 | buflen = td->min_bs; | |
216 | else { | |
6dfd46b9 | 217 | r = os_random_long(&td->bsrange_state); |
20dc95c4 JA |
218 | buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0)); |
219 | buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1); | |
220 | } | |
221 | ||
222 | if (buflen > td->io_size - td->this_io_bytes[td->ddir]) | |
223 | buflen = td->io_size - td->this_io_bytes[td->ddir]; | |
224 | ||
225 | return buflen; | |
226 | } | |
227 | ||
906c8d75 JA |
228 | /* |
229 | * Check if we are above the minimum rate given. | |
230 | */ | |
ebac4655 JA |
231 | static int check_min_rate(struct thread_data *td, struct timeval *now) |
232 | { | |
233 | unsigned long spent; | |
234 | unsigned long rate; | |
235 | int ddir = td->ddir; | |
236 | ||
237 | /* | |
238 | * allow a 2 second settle period in the beginning | |
239 | */ | |
240 | if (mtime_since(&td->start, now) < 2000) | |
241 | return 0; | |
242 | ||
243 | /* | |
244 | * if rate blocks is set, sample is running | |
245 | */ | |
246 | if (td->rate_bytes) { | |
247 | spent = mtime_since(&td->lastrate, now); | |
248 | if (spent < td->ratecycle) | |
249 | return 0; | |
250 | ||
251 | rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent; | |
252 | if (rate < td->ratemin) { | |
eb8bbf48 | 253 | fprintf(f_out, "%s: min rate %d not met, got %ldKiB/sec\n", td->name, td->ratemin, rate); |
ebac4655 JA |
254 | if (rate_quit) |
255 | terminate_threads(td->groupid); | |
256 | return 1; | |
257 | } | |
258 | } | |
259 | ||
260 | td->rate_bytes = td->this_io_bytes[ddir]; | |
261 | memcpy(&td->lastrate, now, sizeof(*now)); | |
262 | return 0; | |
263 | } | |
264 | ||
265 | static inline int runtime_exceeded(struct thread_data *td, struct timeval *t) | |
266 | { | |
267 | if (!td->timeout) | |
268 | return 0; | |
269 | if (mtime_since(&td->epoch, t) >= td->timeout * 1000) | |
270 | return 1; | |
271 | ||
272 | return 0; | |
273 | } | |
274 | ||
275 | static void fill_random_bytes(struct thread_data *td, | |
276 | unsigned char *p, unsigned int len) | |
277 | { | |
278 | unsigned int todo; | |
279 | double r; | |
280 | ||
281 | while (len) { | |
6dfd46b9 | 282 | r = os_random_double(&td->verify_state); |
ebac4655 JA |
283 | |
284 | /* | |
285 | * lrand48_r seems to be broken and only fill the bottom | |
286 | * 32-bits, even on 64-bit archs with 64-bit longs | |
287 | */ | |
288 | todo = sizeof(r); | |
289 | if (todo > len) | |
290 | todo = len; | |
291 | ||
292 | memcpy(p, &r, todo); | |
293 | ||
294 | len -= todo; | |
295 | p += todo; | |
296 | } | |
297 | } | |
298 | ||
299 | static void hexdump(void *buffer, int len) | |
300 | { | |
301 | unsigned char *p = buffer; | |
302 | int i; | |
303 | ||
304 | for (i = 0; i < len; i++) | |
eb8bbf48 JA |
305 | fprintf(f_out, "%02x", p[i]); |
306 | fprintf(f_out, "\n"); | |
ebac4655 JA |
307 | } |
308 | ||
309 | static int verify_io_u_crc32(struct verify_header *hdr, struct io_u *io_u) | |
310 | { | |
311 | unsigned char *p = (unsigned char *) io_u->buf; | |
312 | unsigned long c; | |
ebac4655 JA |
313 | |
314 | p += sizeof(*hdr); | |
315 | c = crc32(p, hdr->len - sizeof(*hdr)); | |
ebac4655 | 316 | |
22f78b32 | 317 | if (c != hdr->crc32) { |
3b70d7e5 JA |
318 | log_err("crc32: verify failed at %llu/%u\n", io_u->offset, io_u->buflen); |
319 | log_err("crc32: wanted %lx, got %lx\n", hdr->crc32, c); | |
22f78b32 | 320 | return 1; |
ebac4655 JA |
321 | } |
322 | ||
22f78b32 | 323 | return 0; |
ebac4655 JA |
324 | } |
325 | ||
326 | static int verify_io_u_md5(struct verify_header *hdr, struct io_u *io_u) | |
327 | { | |
328 | unsigned char *p = (unsigned char *) io_u->buf; | |
329 | struct md5_ctx md5_ctx; | |
ebac4655 JA |
330 | |
331 | memset(&md5_ctx, 0, sizeof(md5_ctx)); | |
332 | p += sizeof(*hdr); | |
333 | md5_update(&md5_ctx, p, hdr->len - sizeof(*hdr)); | |
334 | ||
22f78b32 | 335 | if (memcmp(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash))) { |
3b70d7e5 | 336 | log_err("md5: verify failed at %llu/%u\n", io_u->offset, io_u->buflen); |
ebac4655 JA |
337 | hexdump(hdr->md5_digest, sizeof(hdr->md5_digest)); |
338 | hexdump(md5_ctx.hash, sizeof(md5_ctx.hash)); | |
22f78b32 | 339 | return 1; |
ebac4655 JA |
340 | } |
341 | ||
22f78b32 | 342 | return 0; |
ebac4655 JA |
343 | } |
344 | ||
345 | static int verify_io_u(struct io_u *io_u) | |
346 | { | |
347 | struct verify_header *hdr = (struct verify_header *) io_u->buf; | |
348 | int ret; | |
349 | ||
350 | if (hdr->fio_magic != FIO_HDR_MAGIC) | |
351 | return 1; | |
352 | ||
353 | if (hdr->verify_type == VERIFY_MD5) | |
354 | ret = verify_io_u_md5(hdr, io_u); | |
355 | else if (hdr->verify_type == VERIFY_CRC32) | |
356 | ret = verify_io_u_crc32(hdr, io_u); | |
357 | else { | |
3b70d7e5 | 358 | log_err("Bad verify type %d\n", hdr->verify_type); |
ebac4655 JA |
359 | ret = 1; |
360 | } | |
361 | ||
362 | return ret; | |
363 | } | |
364 | ||
365 | static void fill_crc32(struct verify_header *hdr, void *p, unsigned int len) | |
366 | { | |
367 | hdr->crc32 = crc32(p, len); | |
368 | } | |
369 | ||
370 | static void fill_md5(struct verify_header *hdr, void *p, unsigned int len) | |
371 | { | |
372 | struct md5_ctx md5_ctx; | |
373 | ||
374 | memset(&md5_ctx, 0, sizeof(md5_ctx)); | |
375 | md5_update(&md5_ctx, p, len); | |
376 | memcpy(hdr->md5_digest, md5_ctx.hash, sizeof(md5_ctx.hash)); | |
377 | } | |
378 | ||
906c8d75 JA |
379 | /* |
380 | * Return the data direction for the next io_u. If the job is a | |
381 | * mixed read/write workload, check the rwmix cycle and switch if | |
382 | * necessary. | |
383 | */ | |
a6ccc7be | 384 | static int get_rw_ddir(struct thread_data *td) |
3d60d1ed | 385 | { |
a6ccc7be JA |
386 | if (td_rw(td)) { |
387 | struct timeval now; | |
388 | unsigned long elapsed; | |
3d60d1ed | 389 | |
a6ccc7be JA |
390 | gettimeofday(&now, NULL); |
391 | elapsed = mtime_since_now(&td->rwmix_switch); | |
3d60d1ed | 392 | |
a6ccc7be JA |
393 | /* |
394 | * Check if it's time to seed a new data direction. | |
395 | */ | |
396 | if (elapsed >= td->rwmixcycle) { | |
c1ee2ca4 | 397 | int v; |
a6ccc7be | 398 | long r; |
3d60d1ed | 399 | |
c1ee2ca4 JA |
400 | r = os_random_long(&td->rwmix_state); |
401 | v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0))); | |
a6ccc7be JA |
402 | if (v < td->rwmixread) |
403 | td->rwmix_ddir = DDIR_READ; | |
404 | else | |
405 | td->rwmix_ddir = DDIR_WRITE; | |
406 | memcpy(&td->rwmix_switch, &now, sizeof(now)); | |
407 | } | |
408 | return td->rwmix_ddir; | |
409 | } else if (td_read(td)) | |
3d60d1ed JA |
410 | return DDIR_READ; |
411 | else | |
412 | return DDIR_WRITE; | |
413 | } | |
414 | ||
ebac4655 JA |
415 | /* |
416 | * fill body of io_u->buf with random data and add a header with the | |
22f78b32 | 417 | * crc32 or md5 sum of that data. |
ebac4655 JA |
418 | */ |
419 | static void populate_io_u(struct thread_data *td, struct io_u *io_u) | |
420 | { | |
421 | unsigned char *p = (unsigned char *) io_u->buf; | |
422 | struct verify_header hdr; | |
423 | ||
424 | hdr.fio_magic = FIO_HDR_MAGIC; | |
425 | hdr.len = io_u->buflen; | |
426 | p += sizeof(hdr); | |
427 | fill_random_bytes(td, p, io_u->buflen - sizeof(hdr)); | |
428 | ||
429 | if (td->verify == VERIFY_MD5) { | |
430 | fill_md5(&hdr, p, io_u->buflen - sizeof(hdr)); | |
431 | hdr.verify_type = VERIFY_MD5; | |
432 | } else { | |
433 | fill_crc32(&hdr, p, io_u->buflen - sizeof(hdr)); | |
434 | hdr.verify_type = VERIFY_CRC32; | |
435 | } | |
436 | ||
437 | memcpy(io_u->buf, &hdr, sizeof(hdr)); | |
438 | } | |
439 | ||
aea47d44 | 440 | static int td_io_prep(struct thread_data *td, struct io_u *io_u) |
20dc95c4 | 441 | { |
2866c82d | 442 | if (td->io_ops->prep && td->io_ops->prep(td, io_u)) |
20dc95c4 JA |
443 | return 1; |
444 | ||
445 | return 0; | |
446 | } | |
447 | ||
b1ff3403 | 448 | void put_io_u(struct thread_data *td, struct io_u *io_u) |
ebac4655 | 449 | { |
53cdc686 | 450 | io_u->file = NULL; |
ebac4655 JA |
451 | list_del(&io_u->list); |
452 | list_add(&io_u->list, &td->io_u_freelist); | |
453 | td->cur_depth--; | |
454 | } | |
455 | ||
53cdc686 JA |
456 | static int fill_io_u(struct thread_data *td, struct fio_file *f, |
457 | struct io_u *io_u) | |
843a7413 JA |
458 | { |
459 | /* | |
460 | * If using an iolog, grab next piece if any available. | |
461 | */ | |
462 | if (td->read_iolog) | |
463 | return read_iolog_get(td, io_u); | |
464 | ||
aea47d44 JA |
465 | /* |
466 | * No log, let the seq/rand engine retrieve the next position. | |
467 | */ | |
53cdc686 | 468 | if (!get_next_offset(td, f, &io_u->offset)) { |
aea47d44 JA |
469 | io_u->buflen = get_next_buflen(td); |
470 | ||
471 | if (io_u->buflen) { | |
472 | io_u->ddir = get_rw_ddir(td); | |
843a7413 JA |
473 | |
474 | /* | |
475 | * If using a write iolog, store this entry. | |
476 | */ | |
477 | if (td->write_iolog) | |
478 | write_iolog_put(td, io_u); | |
479 | ||
53cdc686 | 480 | io_u->file = f; |
aea47d44 JA |
481 | return 0; |
482 | } | |
483 | } | |
484 | ||
485 | return 1; | |
486 | } | |
487 | ||
22f78b32 | 488 | #define queue_full(td) list_empty(&(td)->io_u_freelist) |
ebac4655 | 489 | |
b1ff3403 | 490 | struct io_u *__get_io_u(struct thread_data *td) |
ebac4655 | 491 | { |
22f78b32 | 492 | struct io_u *io_u = NULL; |
ebac4655 | 493 | |
22f78b32 JA |
494 | if (!queue_full(td)) { |
495 | io_u = list_entry(td->io_u_freelist.next, struct io_u, list); | |
496 | ||
497 | io_u->error = 0; | |
498 | io_u->resid = 0; | |
499 | list_del(&io_u->list); | |
500 | list_add(&io_u->list, &td->io_u_busylist); | |
501 | td->cur_depth++; | |
502 | } | |
ebac4655 | 503 | |
ebac4655 JA |
504 | return io_u; |
505 | } | |
506 | ||
906c8d75 JA |
507 | /* |
508 | * Return an io_u to be processed. Gets a buflen and offset, sets direction, | |
509 | * etc. The returned io_u is fully ready to be prepped and submitted. | |
510 | */ | |
53cdc686 | 511 | static struct io_u *get_io_u(struct thread_data *td, struct fio_file *f) |
ebac4655 JA |
512 | { |
513 | struct io_u *io_u; | |
514 | ||
515 | io_u = __get_io_u(td); | |
516 | if (!io_u) | |
517 | return NULL; | |
518 | ||
20dc95c4 JA |
519 | if (td->zone_bytes >= td->zone_size) { |
520 | td->zone_bytes = 0; | |
53cdc686 | 521 | f->last_pos += td->zone_skip; |
20dc95c4 JA |
522 | } |
523 | ||
53cdc686 | 524 | if (fill_io_u(td, f, io_u)) { |
ebac4655 JA |
525 | put_io_u(td, io_u); |
526 | return NULL; | |
527 | } | |
528 | ||
53cdc686 JA |
529 | if (io_u->buflen + io_u->offset > f->file_size) |
530 | io_u->buflen = f->file_size - io_u->offset; | |
ebac4655 JA |
531 | |
532 | if (!io_u->buflen) { | |
533 | put_io_u(td, io_u); | |
534 | return NULL; | |
535 | } | |
536 | ||
843a7413 | 537 | if (!td->read_iolog && !td->sequential) |
53cdc686 | 538 | mark_random_map(td, f, io_u); |
ebac4655 | 539 | |
53cdc686 | 540 | f->last_pos += io_u->buflen; |
ebac4655 JA |
541 | |
542 | if (td->verify != VERIFY_NONE) | |
543 | populate_io_u(td, io_u); | |
544 | ||
aea47d44 | 545 | if (td_io_prep(td, io_u)) { |
ebac4655 JA |
546 | put_io_u(td, io_u); |
547 | return NULL; | |
548 | } | |
549 | ||
550 | gettimeofday(&io_u->start_time, NULL); | |
551 | return io_u; | |
552 | } | |
553 | ||
554 | static inline void td_set_runstate(struct thread_data *td, int runstate) | |
555 | { | |
ebac4655 JA |
556 | td->runstate = runstate; |
557 | } | |
558 | ||
aea47d44 | 559 | static int get_next_verify(struct thread_data *td, struct io_u *io_u) |
ebac4655 JA |
560 | { |
561 | struct io_piece *ipo; | |
562 | ||
22f78b32 JA |
563 | if (!list_empty(&td->io_hist_list)) { |
564 | ipo = list_entry(td->io_hist_list.next, struct io_piece, list); | |
ebac4655 | 565 | |
22f78b32 | 566 | list_del(&ipo->list); |
ebac4655 | 567 | |
22f78b32 JA |
568 | io_u->offset = ipo->offset; |
569 | io_u->buflen = ipo->len; | |
570 | io_u->ddir = DDIR_READ; | |
571 | free(ipo); | |
572 | return 0; | |
573 | } | |
574 | ||
575 | return 1; | |
ebac4655 JA |
576 | } |
577 | ||
53cdc686 JA |
578 | static struct fio_file *get_next_file(struct thread_data *td) |
579 | { | |
580 | struct fio_file *f = &td->files[td->next_file]; | |
581 | ||
582 | td->next_file++; | |
583 | if (td->next_file >= td->nr_files) | |
584 | td->next_file = 0; | |
585 | ||
586 | return f; | |
587 | } | |
588 | ||
589 | static int td_io_sync(struct thread_data *td, struct fio_file *f) | |
ebac4655 | 590 | { |
2866c82d | 591 | if (td->io_ops->sync) |
53cdc686 | 592 | return td->io_ops->sync(td, f); |
ebac4655 JA |
593 | |
594 | return 0; | |
595 | } | |
596 | ||
597 | static int io_u_getevents(struct thread_data *td, int min, int max, | |
598 | struct timespec *t) | |
599 | { | |
2866c82d | 600 | return td->io_ops->getevents(td, min, max, t); |
ebac4655 JA |
601 | } |
602 | ||
603 | static int io_u_queue(struct thread_data *td, struct io_u *io_u) | |
604 | { | |
605 | gettimeofday(&io_u->issue_time, NULL); | |
606 | ||
2866c82d | 607 | return td->io_ops->queue(td, io_u); |
ebac4655 JA |
608 | } |
609 | ||
610 | #define iocb_time(iocb) ((unsigned long) (iocb)->data) | |
611 | ||
612 | static void io_completed(struct thread_data *td, struct io_u *io_u, | |
613 | struct io_completion_data *icd) | |
614 | { | |
615 | struct timeval e; | |
616 | unsigned long msec; | |
617 | ||
618 | gettimeofday(&e, NULL); | |
619 | ||
620 | if (!io_u->error) { | |
20dc95c4 | 621 | unsigned int bytes = io_u->buflen - io_u->resid; |
3d60d1ed | 622 | const int idx = io_u->ddir; |
ebac4655 JA |
623 | |
624 | td->io_blocks[idx]++; | |
20dc95c4 JA |
625 | td->io_bytes[idx] += bytes; |
626 | td->zone_bytes += bytes; | |
627 | td->this_io_bytes[idx] += bytes; | |
ebac4655 JA |
628 | |
629 | msec = mtime_since(&io_u->issue_time, &e); | |
630 | ||
3d60d1ed JA |
631 | add_clat_sample(td, idx, msec); |
632 | add_bw_sample(td, idx); | |
ebac4655 | 633 | |
3d60d1ed | 634 | if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE) |
ebac4655 JA |
635 | log_io_piece(td, io_u); |
636 | ||
20dc95c4 | 637 | icd->bytes_done[idx] += bytes; |
ebac4655 JA |
638 | } else |
639 | icd->error = io_u->error; | |
640 | } | |
641 | ||
642 | static void ios_completed(struct thread_data *td,struct io_completion_data *icd) | |
643 | { | |
644 | struct io_u *io_u; | |
645 | int i; | |
646 | ||
647 | icd->error = 0; | |
648 | icd->bytes_done[0] = icd->bytes_done[1] = 0; | |
649 | ||
650 | for (i = 0; i < icd->nr; i++) { | |
2866c82d | 651 | io_u = td->io_ops->event(td, i); |
ebac4655 JA |
652 | |
653 | io_completed(td, io_u, icd); | |
654 | put_io_u(td, io_u); | |
655 | } | |
656 | } | |
657 | ||
906c8d75 JA |
658 | /* |
659 | * When job exits, we can cancel the in-flight IO if we are using async | |
660 | * io. Attempt to do so. | |
661 | */ | |
ebac4655 JA |
662 | static void cleanup_pending_aio(struct thread_data *td) |
663 | { | |
664 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0}; | |
665 | struct list_head *entry, *n; | |
666 | struct io_completion_data icd; | |
667 | struct io_u *io_u; | |
668 | int r; | |
669 | ||
670 | /* | |
671 | * get immediately available events, if any | |
672 | */ | |
673 | r = io_u_getevents(td, 0, td->cur_depth, &ts); | |
674 | if (r > 0) { | |
675 | icd.nr = r; | |
676 | ios_completed(td, &icd); | |
677 | } | |
678 | ||
679 | /* | |
680 | * now cancel remaining active events | |
681 | */ | |
2866c82d | 682 | if (td->io_ops->cancel) { |
ebac4655 JA |
683 | list_for_each_safe(entry, n, &td->io_u_busylist) { |
684 | io_u = list_entry(entry, struct io_u, list); | |
685 | ||
2866c82d | 686 | r = td->io_ops->cancel(td, io_u); |
ebac4655 JA |
687 | if (!r) |
688 | put_io_u(td, io_u); | |
689 | } | |
690 | } | |
691 | ||
692 | if (td->cur_depth) { | |
693 | r = io_u_getevents(td, td->cur_depth, td->cur_depth, NULL); | |
694 | if (r > 0) { | |
695 | icd.nr = r; | |
696 | ios_completed(td, &icd); | |
697 | } | |
698 | } | |
699 | } | |
700 | ||
701 | static int do_io_u_verify(struct thread_data *td, struct io_u **io_u) | |
702 | { | |
703 | struct io_u *v_io_u = *io_u; | |
704 | int ret = 0; | |
705 | ||
706 | if (v_io_u) { | |
707 | ret = verify_io_u(v_io_u); | |
708 | put_io_u(td, v_io_u); | |
709 | *io_u = NULL; | |
710 | } | |
711 | ||
712 | return ret; | |
713 | } | |
714 | ||
906c8d75 JA |
715 | /* |
716 | * The main verify engine. Runs over the writes we previusly submitted, | |
717 | * reads the blocks back in, and checks the crc/md5 of the data. | |
718 | */ | |
ebac4655 JA |
719 | static void do_verify(struct thread_data *td) |
720 | { | |
721 | struct timeval t; | |
722 | struct io_u *io_u, *v_io_u = NULL; | |
723 | struct io_completion_data icd; | |
53cdc686 | 724 | struct fio_file *f; |
ebac4655 JA |
725 | int ret; |
726 | ||
727 | td_set_runstate(td, TD_VERIFYING); | |
728 | ||
729 | do { | |
730 | if (td->terminate) | |
731 | break; | |
732 | ||
733 | gettimeofday(&t, NULL); | |
734 | if (runtime_exceeded(td, &t)) | |
735 | break; | |
736 | ||
737 | io_u = __get_io_u(td); | |
738 | if (!io_u) | |
739 | break; | |
740 | ||
aea47d44 | 741 | if (get_next_verify(td, io_u)) { |
ebac4655 JA |
742 | put_io_u(td, io_u); |
743 | break; | |
744 | } | |
745 | ||
53cdc686 JA |
746 | f = get_next_file(td); |
747 | if (!f) | |
748 | break; | |
749 | ||
750 | io_u->file = f; | |
751 | ||
aea47d44 | 752 | if (td_io_prep(td, io_u)) { |
ebac4655 JA |
753 | put_io_u(td, io_u); |
754 | break; | |
755 | } | |
756 | ||
757 | ret = io_u_queue(td, io_u); | |
758 | if (ret) { | |
759 | put_io_u(td, io_u); | |
760 | td_verror(td, ret); | |
761 | break; | |
762 | } | |
763 | ||
764 | /* | |
765 | * we have one pending to verify, do that while | |
766 | * we are doing io on the next one | |
767 | */ | |
768 | if (do_io_u_verify(td, &v_io_u)) | |
769 | break; | |
770 | ||
771 | ret = io_u_getevents(td, 1, 1, NULL); | |
772 | if (ret != 1) { | |
773 | if (ret < 0) | |
774 | td_verror(td, ret); | |
775 | break; | |
776 | } | |
777 | ||
2866c82d | 778 | v_io_u = td->io_ops->event(td, 0); |
ebac4655 JA |
779 | icd.nr = 1; |
780 | icd.error = 0; | |
781 | io_completed(td, v_io_u, &icd); | |
782 | ||
783 | if (icd.error) { | |
784 | td_verror(td, icd.error); | |
785 | put_io_u(td, v_io_u); | |
786 | v_io_u = NULL; | |
787 | break; | |
788 | } | |
789 | ||
ebac4655 JA |
790 | /* |
791 | * if we can't submit more io, we need to verify now | |
792 | */ | |
793 | if (queue_full(td) && do_io_u_verify(td, &v_io_u)) | |
794 | break; | |
795 | ||
796 | } while (1); | |
797 | ||
798 | do_io_u_verify(td, &v_io_u); | |
799 | ||
800 | if (td->cur_depth) | |
801 | cleanup_pending_aio(td); | |
802 | ||
803 | td_set_runstate(td, TD_RUNNING); | |
804 | } | |
805 | ||
b990b5c0 JA |
806 | /* |
807 | * Not really an io thread, all it does is burn CPU cycles in the specified | |
808 | * manner. | |
809 | */ | |
810 | static void do_cpuio(struct thread_data *td) | |
811 | { | |
812 | struct timeval e; | |
813 | int split = 100 / td->cpuload; | |
814 | int i = 0; | |
815 | ||
816 | while (!td->terminate) { | |
817 | gettimeofday(&e, NULL); | |
818 | ||
819 | if (runtime_exceeded(td, &e)) | |
820 | break; | |
821 | ||
822 | if (!(i % split)) | |
823 | __usec_sleep(10000); | |
824 | else | |
825 | usec_sleep(td, 10000); | |
826 | ||
827 | i++; | |
828 | } | |
829 | } | |
830 | ||
32cd46a0 | 831 | /* |
906c8d75 | 832 | * Main IO worker function. It retrieves io_u's to process and queues |
32cd46a0 JA |
833 | * and reaps them, checking for rate and errors along the way. |
834 | */ | |
ebac4655 JA |
835 | static void do_io(struct thread_data *td) |
836 | { | |
837 | struct io_completion_data icd; | |
838 | struct timeval s, e; | |
839 | unsigned long usec; | |
53cdc686 JA |
840 | struct fio_file *f; |
841 | int i; | |
ebac4655 | 842 | |
5853e5a8 JA |
843 | td_set_runstate(td, TD_RUNNING); |
844 | ||
ebac4655 JA |
845 | while (td->this_io_bytes[td->ddir] < td->io_size) { |
846 | struct timespec ts = { .tv_sec = 0, .tv_nsec = 0}; | |
847 | struct timespec *timeout; | |
848 | int ret, min_evts = 0; | |
849 | struct io_u *io_u; | |
850 | ||
851 | if (td->terminate) | |
852 | break; | |
853 | ||
53cdc686 JA |
854 | f = get_next_file(td); |
855 | if (!f) | |
856 | break; | |
857 | ||
858 | io_u = get_io_u(td, f); | |
ebac4655 JA |
859 | if (!io_u) |
860 | break; | |
861 | ||
862 | memcpy(&s, &io_u->start_time, sizeof(s)); | |
863 | ||
864 | ret = io_u_queue(td, io_u); | |
865 | if (ret) { | |
866 | put_io_u(td, io_u); | |
867 | td_verror(td, ret); | |
868 | break; | |
869 | } | |
870 | ||
871 | add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time)); | |
872 | ||
873 | if (td->cur_depth < td->iodepth) { | |
874 | timeout = &ts; | |
875 | min_evts = 0; | |
876 | } else { | |
877 | timeout = NULL; | |
878 | min_evts = 1; | |
879 | } | |
880 | ||
881 | ret = io_u_getevents(td, min_evts, td->cur_depth, timeout); | |
882 | if (ret < 0) { | |
883 | td_verror(td, ret); | |
884 | break; | |
885 | } else if (!ret) | |
886 | continue; | |
887 | ||
888 | icd.nr = ret; | |
889 | ios_completed(td, &icd); | |
890 | if (icd.error) { | |
891 | td_verror(td, icd.error); | |
892 | break; | |
893 | } | |
894 | ||
895 | /* | |
896 | * the rate is batched for now, it should work for batches | |
897 | * of completions except the very first one which may look | |
898 | * a little bursty | |
899 | */ | |
900 | gettimeofday(&e, NULL); | |
901 | usec = utime_since(&s, &e); | |
902 | ||
903 | rate_throttle(td, usec, icd.bytes_done[td->ddir]); | |
904 | ||
905 | if (check_min_rate(td, &e)) { | |
906 | td_verror(td, ENOMEM); | |
907 | break; | |
908 | } | |
909 | ||
910 | if (runtime_exceeded(td, &e)) | |
911 | break; | |
912 | ||
913 | if (td->thinktime) | |
914 | usec_sleep(td, td->thinktime); | |
915 | ||
916 | if (should_fsync(td) && td->fsync_blocks && | |
917 | (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0) | |
53cdc686 | 918 | td_io_sync(td, f); |
ebac4655 JA |
919 | } |
920 | ||
921 | if (td->cur_depth) | |
922 | cleanup_pending_aio(td); | |
923 | ||
5853e5a8 JA |
924 | if (should_fsync(td) && td->end_fsync) { |
925 | td_set_runstate(td, TD_FSYNCING); | |
53cdc686 JA |
926 | for_each_file(td, f, i) |
927 | td_io_sync(td, f); | |
5853e5a8 | 928 | } |
ebac4655 JA |
929 | } |
930 | ||
ebac4655 JA |
931 | static int init_io(struct thread_data *td) |
932 | { | |
2866c82d JA |
933 | if (td->io_ops->init) |
934 | return td->io_ops->init(td); | |
935 | ||
936 | return 0; | |
ebac4655 JA |
937 | } |
938 | ||
939 | static void cleanup_io_u(struct thread_data *td) | |
940 | { | |
941 | struct list_head *entry, *n; | |
942 | struct io_u *io_u; | |
943 | ||
944 | list_for_each_safe(entry, n, &td->io_u_freelist) { | |
945 | io_u = list_entry(entry, struct io_u, list); | |
946 | ||
947 | list_del(&io_u->list); | |
948 | free(io_u); | |
949 | } | |
950 | ||
951 | if (td->mem_type == MEM_MALLOC) | |
952 | free(td->orig_buffer); | |
953 | else if (td->mem_type == MEM_SHM) { | |
954 | struct shmid_ds sbuf; | |
955 | ||
956 | shmdt(td->orig_buffer); | |
957 | shmctl(td->shm_id, IPC_RMID, &sbuf); | |
958 | } else if (td->mem_type == MEM_MMAP) | |
959 | munmap(td->orig_buffer, td->orig_buffer_size); | |
960 | else | |
3b70d7e5 | 961 | log_err("Bad memory type %d\n", td->mem_type); |
ebac4655 JA |
962 | |
963 | td->orig_buffer = NULL; | |
964 | } | |
965 | ||
966 | static int init_io_u(struct thread_data *td) | |
967 | { | |
968 | struct io_u *io_u; | |
969 | int i, max_units; | |
970 | char *p; | |
971 | ||
2866c82d | 972 | if (td->io_ops->flags & FIO_CPUIO) |
b990b5c0 JA |
973 | return 0; |
974 | ||
2866c82d | 975 | if (td->io_ops->flags & FIO_SYNCIO) |
ebac4655 JA |
976 | max_units = 1; |
977 | else | |
978 | max_units = td->iodepth; | |
979 | ||
980 | td->orig_buffer_size = td->max_bs * max_units + MASK; | |
981 | ||
982 | if (td->mem_type == MEM_MALLOC) | |
983 | td->orig_buffer = malloc(td->orig_buffer_size); | |
984 | else if (td->mem_type == MEM_SHM) { | |
985 | td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600); | |
986 | if (td->shm_id < 0) { | |
987 | td_verror(td, errno); | |
988 | perror("shmget"); | |
989 | return 1; | |
990 | } | |
991 | ||
992 | td->orig_buffer = shmat(td->shm_id, NULL, 0); | |
993 | if (td->orig_buffer == (void *) -1) { | |
994 | td_verror(td, errno); | |
995 | perror("shmat"); | |
996 | td->orig_buffer = NULL; | |
997 | return 1; | |
998 | } | |
999 | } else if (td->mem_type == MEM_MMAP) { | |
1000 | td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0); | |
1001 | if (td->orig_buffer == MAP_FAILED) { | |
1002 | td_verror(td, errno); | |
1003 | perror("mmap"); | |
1004 | td->orig_buffer = NULL; | |
1005 | return 1; | |
1006 | } | |
1007 | } | |
1008 | ||
ebac4655 JA |
1009 | p = ALIGN(td->orig_buffer); |
1010 | for (i = 0; i < max_units; i++) { | |
1011 | io_u = malloc(sizeof(*io_u)); | |
1012 | memset(io_u, 0, sizeof(*io_u)); | |
1013 | INIT_LIST_HEAD(&io_u->list); | |
1014 | ||
1015 | io_u->buf = p + td->max_bs * i; | |
b1ff3403 | 1016 | io_u->index = i; |
ebac4655 JA |
1017 | list_add(&io_u->list, &td->io_u_freelist); |
1018 | } | |
1019 | ||
1020 | return 0; | |
1021 | } | |
1022 | ||
da86774e JA |
1023 | static int switch_ioscheduler(struct thread_data *td) |
1024 | { | |
1025 | char tmp[256], tmp2[128]; | |
1026 | FILE *f; | |
1027 | int ret; | |
1028 | ||
1029 | sprintf(tmp, "%s/queue/scheduler", td->sysfs_root); | |
1030 | ||
1031 | f = fopen(tmp, "r+"); | |
1032 | if (!f) { | |
1033 | td_verror(td, errno); | |
1034 | return 1; | |
1035 | } | |
1036 | ||
1037 | /* | |
1038 | * Set io scheduler. | |
1039 | */ | |
1040 | ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f); | |
1041 | if (ferror(f) || ret != 1) { | |
1042 | td_verror(td, errno); | |
1043 | fclose(f); | |
1044 | return 1; | |
1045 | } | |
1046 | ||
1047 | rewind(f); | |
1048 | ||
1049 | /* | |
1050 | * Read back and check that the selected scheduler is now the default. | |
1051 | */ | |
1052 | ret = fread(tmp, 1, sizeof(tmp), f); | |
1053 | if (ferror(f) || ret < 0) { | |
1054 | td_verror(td, errno); | |
1055 | fclose(f); | |
1056 | return 1; | |
1057 | } | |
1058 | ||
1059 | sprintf(tmp2, "[%s]", td->ioscheduler); | |
1060 | if (!strstr(tmp, tmp2)) { | |
3b70d7e5 | 1061 | log_err("fio: io scheduler %s not found\n", td->ioscheduler); |
da86774e JA |
1062 | td_verror(td, EINVAL); |
1063 | fclose(f); | |
1064 | return 1; | |
1065 | } | |
1066 | ||
1067 | fclose(f); | |
1068 | return 0; | |
1069 | } | |
1070 | ||
ebac4655 JA |
1071 | static void clear_io_state(struct thread_data *td) |
1072 | { | |
53cdc686 JA |
1073 | struct fio_file *f; |
1074 | int i; | |
ebac4655 | 1075 | |
ebac4655 JA |
1076 | td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0; |
1077 | td->this_io_bytes[0] = td->this_io_bytes[1] = 0; | |
20dc95c4 | 1078 | td->zone_bytes = 0; |
ebac4655 | 1079 | |
53cdc686 JA |
1080 | for_each_file(td, f, i) { |
1081 | f->last_pos = 0; | |
1082 | if (td->io_ops->flags & FIO_SYNCIO) | |
1083 | lseek(f->fd, SEEK_SET, 0); | |
1084 | ||
1085 | if (f->file_map) | |
1086 | memset(f->file_map, 0, f->num_maps * sizeof(long)); | |
1087 | } | |
ebac4655 JA |
1088 | } |
1089 | ||
906c8d75 JA |
1090 | /* |
1091 | * Entry point for the thread based jobs. The process based jobs end up | |
1092 | * here as well, after a little setup. | |
1093 | */ | |
ebac4655 JA |
1094 | static void *thread_main(void *data) |
1095 | { | |
1096 | struct thread_data *td = data; | |
ebac4655 JA |
1097 | |
1098 | if (!td->use_thread) | |
1099 | setsid(); | |
1100 | ||
1101 | td->pid = getpid(); | |
1102 | ||
aea47d44 JA |
1103 | INIT_LIST_HEAD(&td->io_u_freelist); |
1104 | INIT_LIST_HEAD(&td->io_u_busylist); | |
1105 | INIT_LIST_HEAD(&td->io_hist_list); | |
1106 | INIT_LIST_HEAD(&td->io_log_list); | |
1107 | ||
ebac4655 JA |
1108 | if (init_io_u(td)) |
1109 | goto err; | |
1110 | ||
1111 | if (fio_setaffinity(td) == -1) { | |
1112 | td_verror(td, errno); | |
1113 | goto err; | |
1114 | } | |
1115 | ||
1116 | if (init_io(td)) | |
1117 | goto err; | |
1118 | ||
aea47d44 JA |
1119 | if (init_iolog(td)) |
1120 | goto err; | |
1121 | ||
ebac4655 JA |
1122 | if (td->ioprio) { |
1123 | if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) { | |
1124 | td_verror(td, errno); | |
1125 | goto err; | |
1126 | } | |
1127 | } | |
1128 | ||
1056eaad | 1129 | if (nice(td->nice) == -1) { |
b6f4d880 JA |
1130 | td_verror(td, errno); |
1131 | goto err; | |
1132 | } | |
1133 | ||
75154845 JA |
1134 | if (init_random_state(td)) |
1135 | goto err; | |
1136 | ||
da86774e JA |
1137 | if (td->ioscheduler && switch_ioscheduler(td)) |
1138 | goto err; | |
1139 | ||
75154845 | 1140 | td_set_runstate(td, TD_INITIALIZED); |
bbfd6b00 JA |
1141 | fio_sem_up(&startup_sem); |
1142 | fio_sem_down(&td->mutex); | |
ebac4655 | 1143 | |
53cdc686 | 1144 | if (!td->create_serialize && setup_files(td)) |
ebac4655 JA |
1145 | goto err; |
1146 | ||
ebac4655 JA |
1147 | gettimeofday(&td->epoch, NULL); |
1148 | ||
4e0ba8af JA |
1149 | if (td->exec_prerun) |
1150 | system(td->exec_prerun); | |
1151 | ||
ebac4655 JA |
1152 | while (td->loops--) { |
1153 | getrusage(RUSAGE_SELF, &td->ru_start); | |
1154 | gettimeofday(&td->start, NULL); | |
1155 | memcpy(&td->stat_sample_time, &td->start, sizeof(td->start)); | |
1156 | ||
1157 | if (td->ratemin) | |
1158 | memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate)); | |
1159 | ||
1160 | clear_io_state(td); | |
1161 | prune_io_piece_log(td); | |
1162 | ||
2866c82d | 1163 | if (td->io_ops->flags & FIO_CPUIO) |
b990b5c0 JA |
1164 | do_cpuio(td); |
1165 | else | |
1166 | do_io(td); | |
ebac4655 JA |
1167 | |
1168 | td->runtime[td->ddir] += mtime_since_now(&td->start); | |
aea47d44 | 1169 | if (td_rw(td) && td->io_bytes[td->ddir ^ 1]) |
3d60d1ed JA |
1170 | td->runtime[td->ddir ^ 1] = td->runtime[td->ddir]; |
1171 | ||
ebac4655 JA |
1172 | update_rusage_stat(td); |
1173 | ||
1174 | if (td->error || td->terminate) | |
1175 | break; | |
1176 | ||
1177 | if (td->verify == VERIFY_NONE) | |
1178 | continue; | |
1179 | ||
1180 | clear_io_state(td); | |
1181 | gettimeofday(&td->start, NULL); | |
1182 | ||
1183 | do_verify(td); | |
1184 | ||
1185 | td->runtime[DDIR_READ] += mtime_since_now(&td->start); | |
1186 | ||
1187 | if (td->error || td->terminate) | |
1188 | break; | |
1189 | } | |
1190 | ||
ebac4655 JA |
1191 | if (td->bw_log) |
1192 | finish_log(td, td->bw_log, "bw"); | |
1193 | if (td->slat_log) | |
1194 | finish_log(td, td->slat_log, "slat"); | |
1195 | if (td->clat_log) | |
1196 | finish_log(td, td->clat_log, "clat"); | |
843a7413 JA |
1197 | if (td->write_iolog) |
1198 | write_iolog_close(td); | |
4e0ba8af JA |
1199 | if (td->exec_postrun) |
1200 | system(td->exec_postrun); | |
ebac4655 JA |
1201 | |
1202 | if (exitall_on_terminate) | |
1203 | terminate_threads(td->groupid); | |
1204 | ||
1205 | err: | |
53cdc686 | 1206 | close_files(td); |
2866c82d | 1207 | close_ioengine(td); |
ebac4655 | 1208 | cleanup_io_u(td); |
ebac4655 JA |
1209 | td_set_runstate(td, TD_EXITED); |
1210 | return NULL; | |
1211 | ||
1212 | } | |
1213 | ||
906c8d75 JA |
1214 | /* |
1215 | * We cannot pass the td data into a forked process, so attach the td and | |
1216 | * pass it to the thread worker. | |
1217 | */ | |
ebac4655 JA |
1218 | static void *fork_main(int shmid, int offset) |
1219 | { | |
1220 | struct thread_data *td; | |
1221 | void *data; | |
1222 | ||
1223 | data = shmat(shmid, NULL, 0); | |
1224 | if (data == (void *) -1) { | |
1225 | perror("shmat"); | |
1226 | return NULL; | |
1227 | } | |
1228 | ||
1229 | td = data + offset * sizeof(struct thread_data); | |
1230 | thread_main(td); | |
1231 | shmdt(data); | |
1232 | return NULL; | |
1233 | } | |
1234 | ||
906c8d75 JA |
1235 | /* |
1236 | * Sets the status of the 'td' in the printed status map. | |
1237 | */ | |
ebac4655 JA |
1238 | static void check_str_update(struct thread_data *td) |
1239 | { | |
1240 | char c = run_str[td->thread_number - 1]; | |
1241 | ||
ebac4655 JA |
1242 | switch (td->runstate) { |
1243 | case TD_REAPED: | |
1244 | c = '_'; | |
1245 | break; | |
1246 | case TD_EXITED: | |
1247 | c = 'E'; | |
1248 | break; | |
1249 | case TD_RUNNING: | |
3d60d1ed JA |
1250 | if (td_rw(td)) { |
1251 | if (td->sequential) | |
1252 | c = 'M'; | |
1253 | else | |
1254 | c = 'm'; | |
1255 | } else if (td_read(td)) { | |
ebac4655 JA |
1256 | if (td->sequential) |
1257 | c = 'R'; | |
1258 | else | |
1259 | c = 'r'; | |
1260 | } else { | |
1261 | if (td->sequential) | |
1262 | c = 'W'; | |
1263 | else | |
1264 | c = 'w'; | |
1265 | } | |
1266 | break; | |
1267 | case TD_VERIFYING: | |
1268 | c = 'V'; | |
1269 | break; | |
5853e5a8 JA |
1270 | case TD_FSYNCING: |
1271 | c = 'F'; | |
1272 | break; | |
ebac4655 JA |
1273 | case TD_CREATED: |
1274 | c = 'C'; | |
1275 | break; | |
75154845 JA |
1276 | case TD_INITIALIZED: |
1277 | c = 'I'; | |
1278 | break; | |
ebac4655 JA |
1279 | case TD_NOT_CREATED: |
1280 | c = 'P'; | |
1281 | break; | |
1282 | default: | |
3b70d7e5 | 1283 | log_err("state %d\n", td->runstate); |
ebac4655 JA |
1284 | } |
1285 | ||
1286 | run_str[td->thread_number - 1] = c; | |
ebac4655 JA |
1287 | } |
1288 | ||
906c8d75 JA |
1289 | /* |
1290 | * Convert seconds to a printable string. | |
1291 | */ | |
5289b847 JA |
1292 | static void eta_to_str(char *str, int eta_sec) |
1293 | { | |
1294 | unsigned int d, h, m, s; | |
1295 | static int always_d, always_h; | |
1296 | ||
1297 | d = h = m = s = 0; | |
1298 | ||
1299 | s = eta_sec % 60; | |
1300 | eta_sec /= 60; | |
1301 | m = eta_sec % 60; | |
1302 | eta_sec /= 60; | |
1303 | h = eta_sec % 24; | |
1304 | eta_sec /= 24; | |
1305 | d = eta_sec; | |
1306 | ||
1307 | if (d || always_d) { | |
1308 | always_d = 1; | |
1309 | str += sprintf(str, "%02dd:", d); | |
1310 | } | |
1311 | if (h || always_h) { | |
1312 | always_h = 1; | |
1313 | str += sprintf(str, "%02dh:", h); | |
1314 | } | |
1315 | ||
1316 | str += sprintf(str, "%02dm:", m); | |
1317 | str += sprintf(str, "%02ds", s); | |
1318 | } | |
1319 | ||
906c8d75 JA |
1320 | /* |
1321 | * Best effort calculation of the estimated pending runtime of a job. | |
1322 | */ | |
6a0106a0 JA |
1323 | static int thread_eta(struct thread_data *td, unsigned long elapsed) |
1324 | { | |
1325 | unsigned long long bytes_total, bytes_done; | |
1326 | unsigned int eta_sec = 0; | |
1327 | ||
1328 | bytes_total = td->total_io_size; | |
3d60d1ed JA |
1329 | |
1330 | /* | |
1331 | * if writing, bytes_total will be twice the size. If mixing, | |
1332 | * assume a 50/50 split and thus bytes_total will be 50% larger. | |
1333 | */ | |
1334 | if (td->verify) { | |
1335 | if (td_rw(td)) | |
1336 | bytes_total = bytes_total * 3 / 2; | |
1337 | else | |
1338 | bytes_total <<= 1; | |
1339 | } | |
6a0106a0 JA |
1340 | if (td->zone_size && td->zone_skip) |
1341 | bytes_total /= (td->zone_skip / td->zone_size); | |
1342 | ||
1343 | if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING) { | |
1344 | double perc; | |
8b611c34 | 1345 | |
6a0106a0 JA |
1346 | bytes_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE]; |
1347 | perc = (double) bytes_done / (double) bytes_total; | |
1348 | if (perc > 1.0) | |
1349 | perc = 1.0; | |
1350 | ||
1351 | eta_sec = (elapsed * (1.0 / perc)) - elapsed; | |
1352 | ||
8b611c34 | 1353 | if (td->timeout && eta_sec > (td->timeout - elapsed)) |
6a0106a0 | 1354 | eta_sec = td->timeout - elapsed; |
75154845 JA |
1355 | } else if (td->runstate == TD_NOT_CREATED || td->runstate == TD_CREATED |
1356 | || td->runstate == TD_INITIALIZED) { | |
6a0106a0 JA |
1357 | int t_eta = 0, r_eta = 0; |
1358 | ||
1359 | /* | |
1360 | * We can only guess - assume it'll run the full timeout | |
1361 | * if given, otherwise assume it'll run at the specified rate. | |
1362 | */ | |
1363 | if (td->timeout) | |
1364 | t_eta = td->timeout + td->start_delay - elapsed; | |
1365 | if (td->rate) { | |
1366 | r_eta = (bytes_total / 1024) / td->rate; | |
1367 | r_eta += td->start_delay - elapsed; | |
1368 | } | |
1369 | ||
1370 | if (r_eta && t_eta) | |
1371 | eta_sec = min(r_eta, t_eta); | |
1372 | else if (r_eta) | |
1373 | eta_sec = r_eta; | |
1374 | else if (t_eta) | |
1375 | eta_sec = t_eta; | |
1376 | else | |
972cfd25 | 1377 | eta_sec = 0; |
6a0106a0 JA |
1378 | } else { |
1379 | /* | |
5853e5a8 | 1380 | * thread is already done or waiting for fsync |
6a0106a0 JA |
1381 | */ |
1382 | eta_sec = 0; | |
1383 | } | |
1384 | ||
1385 | return eta_sec; | |
1386 | } | |
1387 | ||
906c8d75 JA |
1388 | /* |
1389 | * Print status of the jobs we know about. This includes rate estimates, | |
1390 | * ETA, thread state, etc. | |
1391 | */ | |
ebac4655 JA |
1392 | static void print_thread_status(void) |
1393 | { | |
6a0106a0 | 1394 | unsigned long elapsed = time_since_now(&genesis); |
71a751ce | 1395 | int i, nr_running, nr_pending, t_rate, m_rate, *eta_secs, eta_sec; |
5289b847 | 1396 | char eta_str[32]; |
6a0106a0 JA |
1397 | double perc = 0.0; |
1398 | ||
c6ae0a5b | 1399 | if (temp_stall_ts || terse_output) |
88c6ed80 JA |
1400 | return; |
1401 | ||
6a0106a0 JA |
1402 | eta_secs = malloc(thread_number * sizeof(int)); |
1403 | memset(eta_secs, 0, thread_number * sizeof(int)); | |
ebac4655 | 1404 | |
71a751ce | 1405 | nr_pending = nr_running = t_rate = m_rate = 0; |
ebac4655 JA |
1406 | for (i = 0; i < thread_number; i++) { |
1407 | struct thread_data *td = &threads[i]; | |
1408 | ||
5853e5a8 JA |
1409 | if (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING|| |
1410 | td->runstate == TD_FSYNCING) { | |
ebac4655 JA |
1411 | nr_running++; |
1412 | t_rate += td->rate; | |
1413 | m_rate += td->ratemin; | |
71a751ce JA |
1414 | } else if (td->runstate < TD_RUNNING) |
1415 | nr_pending++; | |
ebac4655 | 1416 | |
6a0106a0 JA |
1417 | if (elapsed >= 3) |
1418 | eta_secs[i] = thread_eta(td, elapsed); | |
8b611c34 JA |
1419 | else |
1420 | eta_secs[i] = INT_MAX; | |
ebac4655 JA |
1421 | |
1422 | check_str_update(td); | |
1423 | } | |
1424 | ||
6a0106a0 JA |
1425 | if (exitall_on_terminate) |
1426 | eta_sec = INT_MAX; | |
1427 | else | |
1428 | eta_sec = 0; | |
5289b847 | 1429 | |
6a0106a0 JA |
1430 | for (i = 0; i < thread_number; i++) { |
1431 | if (exitall_on_terminate) { | |
1432 | if (eta_secs[i] < eta_sec) | |
1433 | eta_sec = eta_secs[i]; | |
1434 | } else { | |
1435 | if (eta_secs[i] > eta_sec) | |
1436 | eta_sec = eta_secs[i]; | |
5289b847 | 1437 | } |
6a0106a0 | 1438 | } |
5289b847 | 1439 | |
8b611c34 | 1440 | if (eta_sec != INT_MAX && elapsed) { |
6a0106a0 JA |
1441 | perc = (double) elapsed / (double) (elapsed + eta_sec); |
1442 | eta_to_str(eta_str, eta_sec); | |
ebac4655 JA |
1443 | } |
1444 | ||
71a751ce | 1445 | if (!nr_running && !nr_pending) |
f44c1d46 JA |
1446 | return; |
1447 | ||
eb8bbf48 | 1448 | printf("Threads running: %d", nr_running); |
ebac4655 JA |
1449 | if (m_rate || t_rate) |
1450 | printf(", commitrate %d/%dKiB/sec", t_rate, m_rate); | |
30521453 | 1451 | if (eta_sec != INT_MAX && nr_running) { |
6a0106a0 JA |
1452 | perc *= 100.0; |
1453 | printf(": [%s] [%3.2f%% done] [eta %s]", run_str, perc,eta_str); | |
1454 | } | |
5289b847 | 1455 | printf("\r"); |
ebac4655 | 1456 | fflush(stdout); |
6a0106a0 | 1457 | free(eta_secs); |
ebac4655 JA |
1458 | } |
1459 | ||
906c8d75 JA |
1460 | /* |
1461 | * Run over the job map and reap the threads that have exited, if any. | |
1462 | */ | |
ebac4655 JA |
1463 | static void reap_threads(int *nr_running, int *t_rate, int *m_rate) |
1464 | { | |
b990b5c0 | 1465 | int i, cputhreads; |
ebac4655 JA |
1466 | |
1467 | /* | |
1468 | * reap exited threads (TD_EXITED -> TD_REAPED) | |
1469 | */ | |
b990b5c0 | 1470 | for (i = 0, cputhreads = 0; i < thread_number; i++) { |
ebac4655 JA |
1471 | struct thread_data *td = &threads[i]; |
1472 | ||
2866c82d | 1473 | if (td->io_ops->flags & FIO_CPUIO) |
b990b5c0 JA |
1474 | cputhreads++; |
1475 | ||
ebac4655 JA |
1476 | if (td->runstate != TD_EXITED) |
1477 | continue; | |
1478 | ||
1479 | td_set_runstate(td, TD_REAPED); | |
1480 | ||
1481 | if (td->use_thread) { | |
1482 | long ret; | |
1483 | ||
1484 | if (pthread_join(td->thread, (void *) &ret)) | |
1485 | perror("thread_join"); | |
1486 | } else | |
1487 | waitpid(td->pid, NULL, 0); | |
1488 | ||
1489 | (*nr_running)--; | |
1490 | (*m_rate) -= td->ratemin; | |
1491 | (*t_rate) -= td->rate; | |
1492 | } | |
b990b5c0 JA |
1493 | |
1494 | if (*nr_running == cputhreads) | |
1495 | terminate_threads(TERMINATE_ALL); | |
ebac4655 JA |
1496 | } |
1497 | ||
fcb6ade2 JA |
1498 | static void fio_unpin_memory(void *pinned) |
1499 | { | |
1500 | if (pinned) { | |
1501 | if (munlock(pinned, mlock_size) < 0) | |
1502 | perror("munlock"); | |
1503 | munmap(pinned, mlock_size); | |
1504 | } | |
1505 | } | |
1506 | ||
1507 | static void *fio_pin_memory(void) | |
1508 | { | |
32cd46a0 | 1509 | unsigned long long phys_mem; |
fcb6ade2 JA |
1510 | void *ptr; |
1511 | ||
1512 | if (!mlock_size) | |
1513 | return NULL; | |
1514 | ||
1515 | /* | |
1516 | * Don't allow mlock of more than real_mem-128MB | |
1517 | */ | |
32cd46a0 JA |
1518 | phys_mem = os_phys_mem(); |
1519 | if (phys_mem) { | |
1520 | if ((mlock_size + 128 * 1024 * 1024) > phys_mem) { | |
1521 | mlock_size = phys_mem - 128 * 1024 * 1024; | |
eb8bbf48 | 1522 | fprintf(f_out, "fio: limiting mlocked memory to %lluMiB\n", mlock_size >> 20); |
fcb6ade2 JA |
1523 | } |
1524 | } | |
1525 | ||
1526 | ptr = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0); | |
1527 | if (!ptr) { | |
1528 | perror("malloc locked mem"); | |
1529 | return NULL; | |
1530 | } | |
1531 | if (mlock(ptr, mlock_size) < 0) { | |
1532 | munmap(ptr, mlock_size); | |
1533 | perror("mlock"); | |
1534 | return NULL; | |
1535 | } | |
1536 | ||
1537 | return ptr; | |
1538 | } | |
1539 | ||
906c8d75 JA |
1540 | /* |
1541 | * Main function for kicking off and reaping jobs, as needed. | |
1542 | */ | |
ebac4655 JA |
1543 | static void run_threads(void) |
1544 | { | |
ebac4655 JA |
1545 | struct thread_data *td; |
1546 | unsigned long spent; | |
1547 | int i, todo, nr_running, m_rate, t_rate, nr_started; | |
fcb6ade2 JA |
1548 | void *mlocked_mem; |
1549 | ||
1550 | mlocked_mem = fio_pin_memory(); | |
ebac4655 | 1551 | |
c6ae0a5b JA |
1552 | if (!terse_output) { |
1553 | printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : ""); | |
1554 | fflush(stdout); | |
1555 | } | |
c04f7ec3 | 1556 | |
4efa970e JA |
1557 | signal(SIGINT, sig_handler); |
1558 | signal(SIGALRM, sig_handler); | |
1559 | ||
ebac4655 JA |
1560 | todo = thread_number; |
1561 | nr_running = 0; | |
1562 | nr_started = 0; | |
1563 | m_rate = t_rate = 0; | |
1564 | ||
1565 | for (i = 0; i < thread_number; i++) { | |
1566 | td = &threads[i]; | |
1567 | ||
1568 | run_str[td->thread_number - 1] = 'P'; | |
1569 | ||
1570 | init_disk_util(td); | |
1571 | ||
1572 | if (!td->create_serialize) | |
1573 | continue; | |
1574 | ||
1575 | /* | |
1576 | * do file setup here so it happens sequentially, | |
1577 | * we don't want X number of threads getting their | |
1578 | * client data interspersed on disk | |
1579 | */ | |
53cdc686 | 1580 | if (setup_files(td)) { |
ebac4655 JA |
1581 | td_set_runstate(td, TD_REAPED); |
1582 | todo--; | |
1583 | } | |
1584 | } | |
1585 | ||
1586 | gettimeofday(&genesis, NULL); | |
1587 | ||
1588 | while (todo) { | |
75154845 JA |
1589 | struct thread_data *map[MAX_JOBS]; |
1590 | struct timeval this_start; | |
1591 | int this_jobs = 0, left; | |
1592 | ||
ebac4655 JA |
1593 | /* |
1594 | * create threads (TD_NOT_CREATED -> TD_CREATED) | |
1595 | */ | |
1596 | for (i = 0; i < thread_number; i++) { | |
1597 | td = &threads[i]; | |
1598 | ||
1599 | if (td->runstate != TD_NOT_CREATED) | |
1600 | continue; | |
1601 | ||
1602 | /* | |
1603 | * never got a chance to start, killed by other | |
1604 | * thread for some reason | |
1605 | */ | |
1606 | if (td->terminate) { | |
1607 | todo--; | |
1608 | continue; | |
1609 | } | |
1610 | ||
1611 | if (td->start_delay) { | |
1612 | spent = mtime_since_now(&genesis); | |
1613 | ||
1614 | if (td->start_delay * 1000 > spent) | |
1615 | continue; | |
1616 | } | |
1617 | ||
1618 | if (td->stonewall && (nr_started || nr_running)) | |
1619 | break; | |
1620 | ||
75154845 JA |
1621 | /* |
1622 | * Set state to created. Thread will transition | |
1623 | * to TD_INITIALIZED when it's done setting up. | |
1624 | */ | |
ebac4655 | 1625 | td_set_runstate(td, TD_CREATED); |
75154845 | 1626 | map[this_jobs++] = td; |
bbfd6b00 | 1627 | fio_sem_init(&startup_sem, 1); |
ebac4655 JA |
1628 | nr_started++; |
1629 | ||
1630 | if (td->use_thread) { | |
1631 | if (pthread_create(&td->thread, NULL, thread_main, td)) { | |
1632 | perror("thread_create"); | |
1633 | nr_started--; | |
1634 | } | |
1635 | } else { | |
1636 | if (fork()) | |
bbfd6b00 | 1637 | fio_sem_down(&startup_sem); |
ebac4655 JA |
1638 | else { |
1639 | fork_main(shm_id, i); | |
1640 | exit(0); | |
1641 | } | |
1642 | } | |
1643 | } | |
1644 | ||
1645 | /* | |
75154845 JA |
1646 | * Wait for the started threads to transition to |
1647 | * TD_INITIALIZED. | |
ebac4655 | 1648 | */ |
75154845 JA |
1649 | gettimeofday(&this_start, NULL); |
1650 | left = this_jobs; | |
1651 | while (left) { | |
1652 | if (mtime_since_now(&this_start) > JOB_START_TIMEOUT) | |
1653 | break; | |
1654 | ||
1655 | usleep(100000); | |
1656 | ||
1657 | for (i = 0; i < this_jobs; i++) { | |
1658 | td = map[i]; | |
1659 | if (!td) | |
1660 | continue; | |
b6f4d880 | 1661 | if (td->runstate == TD_INITIALIZED) { |
75154845 JA |
1662 | map[i] = NULL; |
1663 | left--; | |
b6f4d880 JA |
1664 | } else if (td->runstate >= TD_EXITED) { |
1665 | map[i] = NULL; | |
1666 | left--; | |
1667 | todo--; | |
1668 | nr_running++; /* work-around... */ | |
75154845 JA |
1669 | } |
1670 | } | |
1671 | } | |
1672 | ||
1673 | if (left) { | |
3b70d7e5 | 1674 | log_err("fio: %d jobs failed to start\n", left); |
75154845 JA |
1675 | for (i = 0; i < this_jobs; i++) { |
1676 | td = map[i]; | |
1677 | if (!td) | |
1678 | continue; | |
1679 | kill(td->pid, SIGTERM); | |
1680 | } | |
1681 | break; | |
1682 | } | |
1683 | ||
1684 | /* | |
b6f4d880 | 1685 | * start created threads (TD_INITIALIZED -> TD_RUNNING). |
75154845 | 1686 | */ |
ebac4655 JA |
1687 | for (i = 0; i < thread_number; i++) { |
1688 | td = &threads[i]; | |
1689 | ||
75154845 | 1690 | if (td->runstate != TD_INITIALIZED) |
ebac4655 JA |
1691 | continue; |
1692 | ||
1693 | td_set_runstate(td, TD_RUNNING); | |
1694 | nr_running++; | |
1695 | nr_started--; | |
1696 | m_rate += td->ratemin; | |
1697 | t_rate += td->rate; | |
75154845 | 1698 | todo--; |
bbfd6b00 | 1699 | fio_sem_up(&td->mutex); |
ebac4655 JA |
1700 | } |
1701 | ||
1702 | reap_threads(&nr_running, &t_rate, &m_rate); | |
1703 | ||
1704 | if (todo) | |
1705 | usleep(100000); | |
1706 | } | |
1707 | ||
1708 | while (nr_running) { | |
1709 | reap_threads(&nr_running, &t_rate, &m_rate); | |
1710 | usleep(10000); | |
1711 | } | |
1712 | ||
1713 | update_io_ticks(); | |
fcb6ade2 | 1714 | fio_unpin_memory(mlocked_mem); |
ebac4655 JA |
1715 | } |
1716 | ||
ebac4655 JA |
1717 | int main(int argc, char *argv[]) |
1718 | { | |
1719 | if (parse_options(argc, argv)) | |
1720 | return 1; | |
1721 | ||
1722 | if (!thread_number) { | |
3b70d7e5 | 1723 | log_err("Nothing to do\n"); |
ebac4655 JA |
1724 | return 1; |
1725 | } | |
1726 | ||
1727 | disk_util_timer_arm(); | |
1728 | ||
1729 | run_threads(); | |
1730 | show_run_stats(); | |
1731 | ||
1732 | return 0; | |
1733 | } |