blob: 0b0a5e8e540851e032e0004e23175b7cfcd798e3 [file] [log] [blame]
Jens Axboeebac4652005-12-08 15:25:21 +01001/*
2 * fio - the flexible io tester
3 *
4 * Copyright (C) 2005 Jens Axboe <axboe@suse.de>
Jens Axboeaae22ca2006-09-05 10:46:22 +02005 * Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
Jens Axboeebac4652005-12-08 15:25:21 +01006 *
Jens Axboe8e9fe632006-10-24 15:11:09 +02007 * The license below covers all files distributed with fio unless otherwise
8 * noted in the file itself.
9 *
Jens Axboeebac4652005-12-08 15:25:21 +010010 * This program is free software; you can redistribute it and/or modify
Jens Axboe8e9fe632006-10-24 15:11:09 +020011 * it under the terms of the GNU General Public License version 2 as
12 * published by the Free Software Foundation.
Jens Axboeebac4652005-12-08 15:25:21 +010013 *
14 * This program is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
18 *
19 * You should have received a copy of the GNU General Public License
20 * along with this program; if not, write to the Free Software
21 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
22 *
23 */
Jens Axboeebac4652005-12-08 15:25:21 +010024#include <unistd.h>
25#include <fcntl.h>
26#include <string.h>
Jens Axboeebac4652005-12-08 15:25:21 +010027#include <signal.h>
28#include <time.h>
Jens Axboeebac4652005-12-08 15:25:21 +010029#include <assert.h>
Jens Axboeebac4652005-12-08 15:25:21 +010030#include <sys/stat.h>
31#include <sys/wait.h>
32#include <sys/ipc.h>
33#include <sys/shm.h>
34#include <sys/ioctl.h>
35#include <sys/mman.h>
36
37#include "fio.h"
38#include "os.h"
39
Jens Axboe29d610e2007-02-06 13:25:33 +010040static unsigned long page_mask;
41#define ALIGN(buf) \
42 (char *) (((unsigned long) (buf) + page_mask) & ~page_mask)
Jens Axboeebac4652005-12-08 15:25:21 +010043
44int groupid = 0;
45int thread_number = 0;
Jens Axboeebac4652005-12-08 15:25:21 +010046int shm_id = 0;
Jens Axboe53cdc682006-10-18 11:50:58 +020047int temp_stall_ts;
Jens Axboeebac4652005-12-08 15:25:21 +010048
Jens Axboebbfd6b02006-06-07 19:42:54 +020049static volatile int startup_sem;
Jens Axboe6ce15a32007-01-12 09:04:41 +010050static volatile int fio_abort;
Jens Axboeebac4652005-12-08 15:25:21 +010051
Jens Axboebb3884d2007-01-17 17:23:11 +110052struct io_log *agg_io_log[2];
53
Jens Axboeebac4652005-12-08 15:25:21 +010054#define TERMINATE_ALL (-1)
Jens Axboe75154842006-06-01 13:56:09 +020055#define JOB_START_TIMEOUT (5 * 1000)
Jens Axboeebac4652005-12-08 15:25:21 +010056
Jens Axboe6ce15a32007-01-12 09:04:41 +010057static inline void td_set_runstate(struct thread_data *td, int runstate)
58{
59 td->runstate = runstate;
60}
61
62static void terminate_threads(int group_id, int forced_kill)
Jens Axboeebac4652005-12-08 15:25:21 +010063{
Jens Axboe34572e22006-10-20 09:33:18 +020064 struct thread_data *td;
Jens Axboeebac4652005-12-08 15:25:21 +010065 int i;
66
Jens Axboe34572e22006-10-20 09:33:18 +020067 for_each_td(td, i) {
Jens Axboeebac4652005-12-08 15:25:21 +010068 if (group_id == TERMINATE_ALL || groupid == td->groupid) {
69 td->terminate = 1;
70 td->start_delay = 0;
Jens Axboe6ce15a32007-01-12 09:04:41 +010071 if (forced_kill)
72 td_set_runstate(td, TD_EXITED);
Jens Axboeebac4652005-12-08 15:25:21 +010073 }
74 }
75}
76
77static void sig_handler(int sig)
78{
79 switch (sig) {
80 case SIGALRM:
81 update_io_ticks();
82 disk_util_timer_arm();
83 print_thread_status();
84 break;
85 default:
Jens Axboe6ce15a32007-01-12 09:04:41 +010086 printf("\nfio: terminating on signal %d\n", sig);
Jens Axboeebac4652005-12-08 15:25:21 +010087 fflush(stdout);
Jens Axboe6ce15a32007-01-12 09:04:41 +010088 terminate_threads(TERMINATE_ALL, 0);
Jens Axboeebac4652005-12-08 15:25:21 +010089 break;
90 }
91}
92
Jens Axboe906c8d72006-06-13 09:37:56 +020093/*
Jens Axboe906c8d72006-06-13 09:37:56 +020094 * Check if we are above the minimum rate given.
95 */
Jens Axboeebac4652005-12-08 15:25:21 +010096static int check_min_rate(struct thread_data *td, struct timeval *now)
97{
98 unsigned long spent;
99 unsigned long rate;
100 int ddir = td->ddir;
101
102 /*
103 * allow a 2 second settle period in the beginning
104 */
105 if (mtime_since(&td->start, now) < 2000)
106 return 0;
107
108 /*
109 * if rate blocks is set, sample is running
110 */
111 if (td->rate_bytes) {
112 spent = mtime_since(&td->lastrate, now);
113 if (spent < td->ratecycle)
114 return 0;
115
116 rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent;
117 if (rate < td->ratemin) {
Jens Axboe1e97cce2006-12-05 11:44:16 +0100118 fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
Jens Axboeebac4652005-12-08 15:25:21 +0100119 return 1;
120 }
121 }
122
123 td->rate_bytes = td->this_io_bytes[ddir];
124 memcpy(&td->lastrate, now, sizeof(*now));
125 return 0;
126}
127
128static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
129{
130 if (!td->timeout)
131 return 0;
132 if (mtime_since(&td->epoch, t) >= td->timeout * 1000)
133 return 1;
134
135 return 0;
136}
137
Jens Axboe53cdc682006-10-18 11:50:58 +0200138static struct fio_file *get_next_file(struct thread_data *td)
139{
Jens Axboe0ab8db82006-10-18 17:16:23 +0200140 unsigned int old_next_file = td->next_file;
Jens Axboeb2a15192006-10-18 14:10:42 +0200141 struct fio_file *f;
Jens Axboe53cdc682006-10-18 11:50:58 +0200142
Jens Axboeb2a15192006-10-18 14:10:42 +0200143 do {
144 f = &td->files[td->next_file];
145
146 td->next_file++;
147 if (td->next_file >= td->nr_files)
148 td->next_file = 0;
149
150 if (f->fd != -1)
151 break;
152
153 f = NULL;
154 } while (td->next_file != old_next_file);
Jens Axboe53cdc682006-10-18 11:50:58 +0200155
156 return f;
157}
158
Jens Axboe906c8d72006-06-13 09:37:56 +0200159/*
160 * When job exits, we can cancel the in-flight IO if we are using async
161 * io. Attempt to do so.
162 */
Jens Axboeebac4652005-12-08 15:25:21 +0100163static void cleanup_pending_aio(struct thread_data *td)
164{
165 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
166 struct list_head *entry, *n;
167 struct io_completion_data icd;
168 struct io_u *io_u;
169 int r;
170
171 /*
172 * get immediately available events, if any
173 */
Jens Axboe45bee282006-10-18 15:26:44 +0200174 r = td_io_getevents(td, 0, td->cur_depth, &ts);
Jens Axboeebac4652005-12-08 15:25:21 +0100175 if (r > 0) {
176 icd.nr = r;
177 ios_completed(td, &icd);
178 }
179
180 /*
181 * now cancel remaining active events
182 */
Jens Axboe2866c822006-10-09 15:57:48 +0200183 if (td->io_ops->cancel) {
Jens Axboeebac4652005-12-08 15:25:21 +0100184 list_for_each_safe(entry, n, &td->io_u_busylist) {
185 io_u = list_entry(entry, struct io_u, list);
186
Jens Axboe2866c822006-10-09 15:57:48 +0200187 r = td->io_ops->cancel(td, io_u);
Jens Axboeebac4652005-12-08 15:25:21 +0100188 if (!r)
189 put_io_u(td, io_u);
190 }
191 }
192
193 if (td->cur_depth) {
Jens Axboe45bee282006-10-18 15:26:44 +0200194 r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL);
Jens Axboeebac4652005-12-08 15:25:21 +0100195 if (r > 0) {
196 icd.nr = r;
197 ios_completed(td, &icd);
198 }
199 }
200}
201
Jens Axboe906c8d72006-06-13 09:37:56 +0200202/*
Jens Axboe858a3d42006-10-24 14:54:44 +0200203 * Helper to handle the final sync of a file. Works just like the normal
204 * io path, just does everything sync.
205 */
206static int fio_io_sync(struct thread_data *td, struct fio_file *f)
207{
208 struct io_u *io_u = __get_io_u(td);
209 struct io_completion_data icd;
210 int ret;
211
212 if (!io_u)
213 return 1;
214
215 io_u->ddir = DDIR_SYNC;
216 io_u->file = f;
217
218 if (td_io_prep(td, io_u)) {
219 put_io_u(td, io_u);
220 return 1;
221 }
222
223 ret = td_io_queue(td, io_u);
224 if (ret) {
Jens Axboe353a7e02006-10-25 09:16:07 +0200225 td_verror(td, io_u->error);
Jens Axboe858a3d42006-10-24 14:54:44 +0200226 put_io_u(td, io_u);
Jens Axboe858a3d42006-10-24 14:54:44 +0200227 return 1;
228 }
229
230 ret = td_io_getevents(td, 1, td->cur_depth, NULL);
231 if (ret < 0) {
Jens Axboe353a7e02006-10-25 09:16:07 +0200232 td_verror(td, ret);
Jens Axboe858a3d42006-10-24 14:54:44 +0200233 return 1;
234 }
235
236 icd.nr = ret;
237 ios_completed(td, &icd);
238 if (icd.error) {
239 td_verror(td, icd.error);
240 return 1;
241 }
242
243 return 0;
244}
245
246/*
Jens Axboe906c8d72006-06-13 09:37:56 +0200247 * The main verify engine. Runs over the writes we previusly submitted,
248 * reads the blocks back in, and checks the crc/md5 of the data.
249 */
Jens Axboe1e97cce2006-12-05 11:44:16 +0100250static void do_verify(struct thread_data *td)
Jens Axboeebac4652005-12-08 15:25:21 +0100251{
Jens Axboeebac4652005-12-08 15:25:21 +0100252 struct io_u *io_u, *v_io_u = NULL;
253 struct io_completion_data icd;
Jens Axboe53cdc682006-10-18 11:50:58 +0200254 struct fio_file *f;
Jens Axboee5b401d2006-10-18 16:03:40 +0200255 int ret, i;
256
257 /*
258 * sync io first and invalidate cache, to make sure we really
259 * read from disk.
260 */
261 for_each_file(td, f, i) {
Jens Axboe858a3d42006-10-24 14:54:44 +0200262 fio_io_sync(td, f);
Jens Axboee5b401d2006-10-18 16:03:40 +0200263 file_invalidate_cache(td, f);
264 }
Jens Axboeebac4652005-12-08 15:25:21 +0100265
266 td_set_runstate(td, TD_VERIFYING);
267
268 do {
269 if (td->terminate)
270 break;
271
Jens Axboeebac4652005-12-08 15:25:21 +0100272 io_u = __get_io_u(td);
273 if (!io_u)
274 break;
275
Jens Axboe02bcaa82006-11-24 10:42:00 +0100276 if (runtime_exceeded(td, &io_u->start_time)) {
277 put_io_u(td, io_u);
278 break;
279 }
280
Jens Axboeaea47d42006-05-26 19:27:29 +0200281 if (get_next_verify(td, io_u)) {
Jens Axboeebac4652005-12-08 15:25:21 +0100282 put_io_u(td, io_u);
283 break;
284 }
285
Jens Axboe53cdc682006-10-18 11:50:58 +0200286 f = get_next_file(td);
287 if (!f)
288 break;
289
290 io_u->file = f;
291
Jens Axboeaea47d42006-05-26 19:27:29 +0200292 if (td_io_prep(td, io_u)) {
Jens Axboeebac4652005-12-08 15:25:21 +0100293 put_io_u(td, io_u);
294 break;
295 }
296
Jens Axboe45bee282006-10-18 15:26:44 +0200297 ret = td_io_queue(td, io_u);
Jens Axboeebac4652005-12-08 15:25:21 +0100298 if (ret) {
Jens Axboe353a7e02006-10-25 09:16:07 +0200299 td_verror(td, io_u->error);
Jens Axboeebac4652005-12-08 15:25:21 +0100300 put_io_u(td, io_u);
Jens Axboeebac4652005-12-08 15:25:21 +0100301 break;
302 }
303
304 /*
305 * we have one pending to verify, do that while
306 * we are doing io on the next one
307 */
308 if (do_io_u_verify(td, &v_io_u))
309 break;
310
Jens Axboe45bee282006-10-18 15:26:44 +0200311 ret = td_io_getevents(td, 1, 1, NULL);
Jens Axboeebac4652005-12-08 15:25:21 +0100312 if (ret != 1) {
313 if (ret < 0)
314 td_verror(td, ret);
315 break;
316 }
317
Jens Axboe2866c822006-10-09 15:57:48 +0200318 v_io_u = td->io_ops->event(td, 0);
Jens Axboeebac4652005-12-08 15:25:21 +0100319 icd.nr = 1;
320 icd.error = 0;
Jens Axboe02bcaa82006-11-24 10:42:00 +0100321 fio_gettime(&icd.time, NULL);
Jens Axboeebac4652005-12-08 15:25:21 +0100322 io_completed(td, v_io_u, &icd);
323
324 if (icd.error) {
325 td_verror(td, icd.error);
326 put_io_u(td, v_io_u);
327 v_io_u = NULL;
328 break;
329 }
330
Jens Axboeebac4652005-12-08 15:25:21 +0100331 /*
332 * if we can't submit more io, we need to verify now
333 */
334 if (queue_full(td) && do_io_u_verify(td, &v_io_u))
335 break;
336
337 } while (1);
338
339 do_io_u_verify(td, &v_io_u);
340
341 if (td->cur_depth)
342 cleanup_pending_aio(td);
343
344 td_set_runstate(td, TD_RUNNING);
345}
346
Jens Axboe32cd46a2006-06-07 13:40:40 +0200347/*
Jens Axboeb990b5c2006-09-14 09:48:22 +0200348 * Not really an io thread, all it does is burn CPU cycles in the specified
349 * manner.
350 */
351static void do_cpuio(struct thread_data *td)
352{
353 struct timeval e;
354 int split = 100 / td->cpuload;
355 int i = 0;
356
357 while (!td->terminate) {
Jens Axboe02bcaa82006-11-24 10:42:00 +0100358 fio_gettime(&e, NULL);
Jens Axboeb990b5c2006-09-14 09:48:22 +0200359
360 if (runtime_exceeded(td, &e))
361 break;
362
363 if (!(i % split))
364 __usec_sleep(10000);
365 else
366 usec_sleep(td, 10000);
367
368 i++;
369 }
370}
371
372/*
Jens Axboe906c8d72006-06-13 09:37:56 +0200373 * Main IO worker function. It retrieves io_u's to process and queues
Jens Axboe32cd46a2006-06-07 13:40:40 +0200374 * and reaps them, checking for rate and errors along the way.
375 */
Jens Axboeebac4652005-12-08 15:25:21 +0100376static void do_io(struct thread_data *td)
377{
378 struct io_completion_data icd;
Jens Axboe02bcaa82006-11-24 10:42:00 +0100379 struct timeval s;
Jens Axboeebac4652005-12-08 15:25:21 +0100380 unsigned long usec;
Jens Axboe53cdc682006-10-18 11:50:58 +0200381 struct fio_file *f;
Jens Axboe84585002006-10-19 20:26:22 +0200382 int i, ret = 0;
Jens Axboeebac4652005-12-08 15:25:21 +0100383
Jens Axboe5853e5a2006-06-08 14:41:05 +0200384 td_set_runstate(td, TD_RUNNING);
385
Jens Axboeebac4652005-12-08 15:25:21 +0100386 while (td->this_io_bytes[td->ddir] < td->io_size) {
Jens Axboeebac4652005-12-08 15:25:21 +0100387 struct timespec *timeout;
Jens Axboe84585002006-10-19 20:26:22 +0200388 int min_evts = 0;
Jens Axboeebac4652005-12-08 15:25:21 +0100389 struct io_u *io_u;
390
391 if (td->terminate)
392 break;
393
Jens Axboe53cdc682006-10-18 11:50:58 +0200394 f = get_next_file(td);
395 if (!f)
396 break;
397
398 io_u = get_io_u(td, f);
Jens Axboeebac4652005-12-08 15:25:21 +0100399 if (!io_u)
400 break;
401
402 memcpy(&s, &io_u->start_time, sizeof(s));
403
Jens Axboecec6b552007-02-06 20:15:38 +0100404requeue:
Jens Axboe45bee282006-10-18 15:26:44 +0200405 ret = td_io_queue(td, io_u);
Jens Axboeebac4652005-12-08 15:25:21 +0100406 if (ret) {
Jens Axboecec6b552007-02-06 20:15:38 +0100407 if (ret > 0 && (io_u->xfer_buflen != io_u->resid) &&
408 io_u->resid) {
409 /*
410 * short read/write. requeue.
411 */
412 io_u->xfer_buflen = io_u->resid;
413 io_u->xfer_buf += ret;
414 goto requeue;
415 } else {
416 td_verror(td, io_u->error);
417 put_io_u(td, io_u);
418 break;
419 }
Jens Axboeebac4652005-12-08 15:25:21 +0100420 }
421
422 add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
423
424 if (td->cur_depth < td->iodepth) {
Jens Axboe8d7ad8d2007-01-09 21:22:18 +0100425 struct timespec ts = { .tv_sec = 0, .tv_nsec = 0};
426
Jens Axboeebac4652005-12-08 15:25:21 +0100427 timeout = &ts;
428 min_evts = 0;
429 } else {
430 timeout = NULL;
431 min_evts = 1;
432 }
433
Jens Axboe45bee282006-10-18 15:26:44 +0200434 ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
Jens Axboeebac4652005-12-08 15:25:21 +0100435 if (ret < 0) {
Jens Axboe87dc1ab2006-10-24 14:41:26 +0200436 td_verror(td, ret);
Jens Axboeebac4652005-12-08 15:25:21 +0100437 break;
438 } else if (!ret)
439 continue;
440
441 icd.nr = ret;
442 ios_completed(td, &icd);
443 if (icd.error) {
444 td_verror(td, icd.error);
445 break;
446 }
447
448 /*
449 * the rate is batched for now, it should work for batches
450 * of completions except the very first one which may look
451 * a little bursty
452 */
Jens Axboe02bcaa82006-11-24 10:42:00 +0100453 usec = utime_since(&s, &icd.time);
Jens Axboeebac4652005-12-08 15:25:21 +0100454
Jens Axboea00735e2006-11-03 08:58:08 +0100455 rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir);
Jens Axboeebac4652005-12-08 15:25:21 +0100456
Jens Axboe02bcaa82006-11-24 10:42:00 +0100457 if (check_min_rate(td, &icd.time)) {
Jens Axboe98aa62d2006-11-07 14:16:19 +0100458 if (exitall_on_terminate)
Jens Axboe6ce15a32007-01-12 09:04:41 +0100459 terminate_threads(td->groupid, 0);
Jens Axboefd841462007-01-13 12:20:30 +0100460 td_verror(td, ENODATA);
Jens Axboeebac4652005-12-08 15:25:21 +0100461 break;
462 }
463
Jens Axboe02bcaa82006-11-24 10:42:00 +0100464 if (runtime_exceeded(td, &icd.time))
Jens Axboeebac4652005-12-08 15:25:21 +0100465 break;
466
Jens Axboe9c1f7432007-01-03 20:43:19 +0100467 if (td->thinktime) {
468 unsigned long long b;
469
470 b = td->io_blocks[0] + td->io_blocks[1];
Jens Axboe25309a32007-01-10 11:33:14 +0100471 if (!(b % td->thinktime_blocks))
Jens Axboe9c1f7432007-01-03 20:43:19 +0100472 usec_sleep(td, td->thinktime);
473 }
Jens Axboeebac4652005-12-08 15:25:21 +0100474 }
475
Jens Axboe4d2413c2006-11-23 11:58:59 +0100476 if (!td->error) {
Jens Axboe84585002006-10-19 20:26:22 +0200477 if (td->cur_depth)
478 cleanup_pending_aio(td);
Jens Axboeebac4652005-12-08 15:25:21 +0100479
Jens Axboe84585002006-10-19 20:26:22 +0200480 if (should_fsync(td) && td->end_fsync) {
481 td_set_runstate(td, TD_FSYNCING);
482 for_each_file(td, f, i)
Jens Axboe858a3d42006-10-24 14:54:44 +0200483 fio_io_sync(td, f);
Jens Axboe84585002006-10-19 20:26:22 +0200484 }
Jens Axboe5853e5a2006-06-08 14:41:05 +0200485 }
Jens Axboeebac4652005-12-08 15:25:21 +0100486}
487
Jens Axboeebac4652005-12-08 15:25:21 +0100488static void cleanup_io_u(struct thread_data *td)
489{
490 struct list_head *entry, *n;
491 struct io_u *io_u;
492
493 list_for_each_safe(entry, n, &td->io_u_freelist) {
494 io_u = list_entry(entry, struct io_u, list);
495
496 list_del(&io_u->list);
497 free(io_u);
498 }
499
Jens Axboe2f9ade32006-10-20 11:25:52 +0200500 free_io_mem(td);
Jens Axboeebac4652005-12-08 15:25:21 +0100501}
502
Jens Axboe6b9cea22006-10-30 17:00:24 +0100503/*
504 * "randomly" fill the buffer contents
505 */
Jens Axboe66eeb292006-11-01 08:35:44 +0100506static void fill_rand_buf(struct io_u *io_u, int max_bs)
Jens Axboe6b9cea22006-10-30 17:00:24 +0100507{
Jens Axboe66eeb292006-11-01 08:35:44 +0100508 int *ptr = io_u->buf;
Jens Axboe6b9cea22006-10-30 17:00:24 +0100509
510 while ((void *) ptr - io_u->buf < max_bs) {
511 *ptr = rand() * 0x9e370001;
512 ptr++;
513 }
514}
515
Jens Axboeebac4652005-12-08 15:25:21 +0100516static int init_io_u(struct thread_data *td)
517{
518 struct io_u *io_u;
Jens Axboea00735e2006-11-03 08:58:08 +0100519 unsigned int max_bs;
Jens Axboeebac4652005-12-08 15:25:21 +0100520 int i, max_units;
521 char *p;
522
Jens Axboe2866c822006-10-09 15:57:48 +0200523 if (td->io_ops->flags & FIO_CPUIO)
Jens Axboeb990b5c2006-09-14 09:48:22 +0200524 return 0;
525
Jens Axboe2866c822006-10-09 15:57:48 +0200526 if (td->io_ops->flags & FIO_SYNCIO)
Jens Axboeebac4652005-12-08 15:25:21 +0100527 max_units = 1;
528 else
529 max_units = td->iodepth;
530
Jens Axboea00735e2006-11-03 08:58:08 +0100531 max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
Jens Axboe74b025b2006-12-19 15:18:14 +0100532 td->orig_buffer_size = max_bs * max_units;
533
Jens Axboed0bdaf42006-12-20 14:40:44 +0100534 if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE)
Jens Axboe56bb17f2006-12-20 20:27:36 +0100535 td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1);
Jens Axboe74b025b2006-12-19 15:18:14 +0100536 else
Jens Axboe29d610e2007-02-06 13:25:33 +0100537 td->orig_buffer_size += page_mask;
Jens Axboeebac4652005-12-08 15:25:21 +0100538
Jens Axboe2f9ade32006-10-20 11:25:52 +0200539 if (allocate_io_mem(td))
540 return 1;
Jens Axboeebac4652005-12-08 15:25:21 +0100541
Jens Axboeebac4652005-12-08 15:25:21 +0100542 p = ALIGN(td->orig_buffer);
543 for (i = 0; i < max_units; i++) {
544 io_u = malloc(sizeof(*io_u));
545 memset(io_u, 0, sizeof(*io_u));
546 INIT_LIST_HEAD(&io_u->list);
547
Jens Axboea00735e2006-11-03 08:58:08 +0100548 io_u->buf = p + max_bs * i;
Jens Axboe6b9cea22006-10-30 17:00:24 +0100549 if (td_write(td) || td_rw(td))
Jens Axboea00735e2006-11-03 08:58:08 +0100550 fill_rand_buf(io_u, max_bs);
Jens Axboe6b9cea22006-10-30 17:00:24 +0100551
Jens Axboeb1ff3402006-02-17 11:35:58 +0100552 io_u->index = i;
Jens Axboeebac4652005-12-08 15:25:21 +0100553 list_add(&io_u->list, &td->io_u_freelist);
554 }
555
556 return 0;
557}
558
Jens Axboeda867742006-06-06 10:39:10 -0700559static int switch_ioscheduler(struct thread_data *td)
560{
561 char tmp[256], tmp2[128];
562 FILE *f;
563 int ret;
564
Jens Axboef48b4672006-10-27 11:30:07 +0200565 if (td->io_ops->flags & FIO_CPUIO)
566 return 0;
567
Jens Axboeda867742006-06-06 10:39:10 -0700568 sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
569
570 f = fopen(tmp, "r+");
571 if (!f) {
572 td_verror(td, errno);
573 return 1;
574 }
575
576 /*
577 * Set io scheduler.
578 */
579 ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f);
580 if (ferror(f) || ret != 1) {
581 td_verror(td, errno);
582 fclose(f);
583 return 1;
584 }
585
586 rewind(f);
587
588 /*
589 * Read back and check that the selected scheduler is now the default.
590 */
591 ret = fread(tmp, 1, sizeof(tmp), f);
592 if (ferror(f) || ret < 0) {
593 td_verror(td, errno);
594 fclose(f);
595 return 1;
596 }
597
598 sprintf(tmp2, "[%s]", td->ioscheduler);
599 if (!strstr(tmp, tmp2)) {
Jens Axboe3b70d7e2006-06-08 21:48:46 +0200600 log_err("fio: io scheduler %s not found\n", td->ioscheduler);
Jens Axboeda867742006-06-06 10:39:10 -0700601 td_verror(td, EINVAL);
602 fclose(f);
603 return 1;
604 }
605
606 fclose(f);
607 return 0;
608}
609
Jens Axboeebac4652005-12-08 15:25:21 +0100610static void clear_io_state(struct thread_data *td)
611{
Jens Axboe53cdc682006-10-18 11:50:58 +0200612 struct fio_file *f;
613 int i;
Jens Axboeebac4652005-12-08 15:25:21 +0100614
Jens Axboeebac4652005-12-08 15:25:21 +0100615 td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0;
616 td->this_io_bytes[0] = td->this_io_bytes[1] = 0;
Jens Axboe20dc95c2005-12-09 10:29:35 +0100617 td->zone_bytes = 0;
Jens Axboeebac4652005-12-08 15:25:21 +0100618
Jens Axboe53cdc682006-10-18 11:50:58 +0200619 for_each_file(td, f, i) {
620 f->last_pos = 0;
621 if (td->io_ops->flags & FIO_SYNCIO)
622 lseek(f->fd, SEEK_SET, 0);
623
624 if (f->file_map)
625 memset(f->file_map, 0, f->num_maps * sizeof(long));
626 }
Jens Axboeebac4652005-12-08 15:25:21 +0100627}
628
Jens Axboe906c8d72006-06-13 09:37:56 +0200629/*
630 * Entry point for the thread based jobs. The process based jobs end up
631 * here as well, after a little setup.
632 */
Jens Axboeebac4652005-12-08 15:25:21 +0100633static void *thread_main(void *data)
634{
Jens Axboe69008992006-11-24 13:12:56 +0100635 unsigned long long runtime[2];
Jens Axboeebac4652005-12-08 15:25:21 +0100636 struct thread_data *td = data;
Jens Axboeebac4652005-12-08 15:25:21 +0100637
638 if (!td->use_thread)
639 setsid();
640
641 td->pid = getpid();
642
Jens Axboeaea47d42006-05-26 19:27:29 +0200643 INIT_LIST_HEAD(&td->io_u_freelist);
644 INIT_LIST_HEAD(&td->io_u_busylist);
645 INIT_LIST_HEAD(&td->io_hist_list);
646 INIT_LIST_HEAD(&td->io_log_list);
647
Jens Axboeebac4652005-12-08 15:25:21 +0100648 if (init_io_u(td))
649 goto err;
650
651 if (fio_setaffinity(td) == -1) {
652 td_verror(td, errno);
653 goto err;
654 }
655
Jens Axboeaea47d42006-05-26 19:27:29 +0200656 if (init_iolog(td))
657 goto err;
658
Jens Axboeebac4652005-12-08 15:25:21 +0100659 if (td->ioprio) {
660 if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) {
661 td_verror(td, errno);
662 goto err;
663 }
664 }
665
Jens Axboe1056eaa2006-07-13 10:33:45 -0700666 if (nice(td->nice) == -1) {
Jens Axboeb6f4d882006-06-02 10:32:51 +0200667 td_verror(td, errno);
668 goto err;
669 }
670
Jens Axboe75154842006-06-01 13:56:09 +0200671 if (init_random_state(td))
672 goto err;
673
Jens Axboe56f94982006-10-24 09:33:42 +0200674 if (td->ioscheduler && switch_ioscheduler(td))
675 goto err;
Jens Axboeda867742006-06-06 10:39:10 -0700676
Jens Axboe75154842006-06-01 13:56:09 +0200677 td_set_runstate(td, TD_INITIALIZED);
Jens Axboebbfd6b02006-06-07 19:42:54 +0200678 fio_sem_up(&startup_sem);
679 fio_sem_down(&td->mutex);
Jens Axboeebac4652005-12-08 15:25:21 +0100680
Jens Axboe53cdc682006-10-18 11:50:58 +0200681 if (!td->create_serialize && setup_files(td))
Jens Axboeebac4652005-12-08 15:25:21 +0100682 goto err;
Jens Axboe21972cd2006-11-23 12:39:16 +0100683 if (open_files(td))
684 goto err;
Jens Axboeebac4652005-12-08 15:25:21 +0100685
Jens Axboe7d6c5282007-02-05 10:52:10 +0100686 /*
687 * Do this late, as some IO engines would like to have the
688 * files setup prior to initializing structures.
689 */
690 if (td_io_init(td))
691 goto err;
692
Jens Axboe69cfd7e2007-02-07 13:58:53 +0100693 if (td->exec_prerun) {
694 if (system(td->exec_prerun) < 0)
695 goto err;
696 }
Jens Axboe4e0ba8a2006-06-06 09:36:28 +0200697
Jens Axboe69008992006-11-24 13:12:56 +0100698 fio_gettime(&td->epoch, NULL);
Jens Axboe36dff962006-11-24 13:29:20 +0100699 getrusage(RUSAGE_SELF, &td->ru_start);
Jens Axboe69008992006-11-24 13:12:56 +0100700
701 runtime[0] = runtime[1] = 0;
Jens Axboeebac4652005-12-08 15:25:21 +0100702 while (td->loops--) {
Jens Axboe02bcaa82006-11-24 10:42:00 +0100703 fio_gettime(&td->start, NULL);
Jens Axboeebac4652005-12-08 15:25:21 +0100704 memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
705
706 if (td->ratemin)
707 memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
708
709 clear_io_state(td);
710 prune_io_piece_log(td);
711
Jens Axboe2866c822006-10-09 15:57:48 +0200712 if (td->io_ops->flags & FIO_CPUIO)
Jens Axboeb990b5c2006-09-14 09:48:22 +0200713 do_cpuio(td);
714 else
715 do_io(td);
Jens Axboeebac4652005-12-08 15:25:21 +0100716
Jens Axboe69008992006-11-24 13:12:56 +0100717 runtime[td->ddir] += utime_since_now(&td->start);
Jens Axboeaea47d42006-05-26 19:27:29 +0200718 if (td_rw(td) && td->io_bytes[td->ddir ^ 1])
Jens Axboe69008992006-11-24 13:12:56 +0100719 runtime[td->ddir ^ 1] = runtime[td->ddir];
Jens Axboe3d60d1e2006-05-25 06:31:06 +0200720
Jens Axboeebac4652005-12-08 15:25:21 +0100721 if (td->error || td->terminate)
722 break;
723
724 if (td->verify == VERIFY_NONE)
725 continue;
726
727 clear_io_state(td);
Jens Axboe02bcaa82006-11-24 10:42:00 +0100728 fio_gettime(&td->start, NULL);
Jens Axboeebac4652005-12-08 15:25:21 +0100729
730 do_verify(td);
731
Jens Axboe69008992006-11-24 13:12:56 +0100732 runtime[DDIR_READ] += utime_since_now(&td->start);
Jens Axboeebac4652005-12-08 15:25:21 +0100733
734 if (td->error || td->terminate)
735 break;
736 }
737
Jens Axboe36dff962006-11-24 13:29:20 +0100738 update_rusage_stat(td);
Jens Axboe69008992006-11-24 13:12:56 +0100739 fio_gettime(&td->end_time, NULL);
740 td->runtime[0] = runtime[0] / 1000;
741 td->runtime[1] = runtime[1] / 1000;
742
Jens Axboeebac4652005-12-08 15:25:21 +0100743 if (td->bw_log)
744 finish_log(td, td->bw_log, "bw");
745 if (td->slat_log)
746 finish_log(td, td->slat_log, "slat");
747 if (td->clat_log)
748 finish_log(td, td->clat_log, "clat");
Jens Axboe076efc72006-10-27 11:24:25 +0200749 if (td->write_iolog_file)
Jens Axboe843a7412006-06-01 21:14:21 -0700750 write_iolog_close(td);
Jens Axboe69cfd7e2007-02-07 13:58:53 +0100751 if (td->exec_postrun) {
752 if (system(td->exec_postrun) < 0)
753 log_err("fio: postrun %s failed\n", td->exec_postrun);
754 }
Jens Axboeebac4652005-12-08 15:25:21 +0100755
756 if (exitall_on_terminate)
Jens Axboe6ce15a32007-01-12 09:04:41 +0100757 terminate_threads(td->groupid, 0);
Jens Axboeebac4652005-12-08 15:25:21 +0100758
759err:
Jens Axboe53cdc682006-10-18 11:50:58 +0200760 close_files(td);
Jens Axboe2866c822006-10-09 15:57:48 +0200761 close_ioengine(td);
Jens Axboeebac4652005-12-08 15:25:21 +0100762 cleanup_io_u(td);
Jens Axboeebac4652005-12-08 15:25:21 +0100763 td_set_runstate(td, TD_EXITED);
764 return NULL;
765
766}
767
Jens Axboe906c8d72006-06-13 09:37:56 +0200768/*
769 * We cannot pass the td data into a forked process, so attach the td and
770 * pass it to the thread worker.
771 */
Jens Axboeebac4652005-12-08 15:25:21 +0100772static void *fork_main(int shmid, int offset)
773{
774 struct thread_data *td;
775 void *data;
776
777 data = shmat(shmid, NULL, 0);
778 if (data == (void *) -1) {
779 perror("shmat");
780 return NULL;
781 }
782
783 td = data + offset * sizeof(struct thread_data);
784 thread_main(td);
785 shmdt(data);
786 return NULL;
787}
788
Jens Axboe906c8d72006-06-13 09:37:56 +0200789/*
Jens Axboe906c8d72006-06-13 09:37:56 +0200790 * Run over the job map and reap the threads that have exited, if any.
791 */
Jens Axboeebac4652005-12-08 15:25:21 +0100792static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
793{
Jens Axboe34572e22006-10-20 09:33:18 +0200794 struct thread_data *td;
Jens Axboe4d2413c2006-11-23 11:58:59 +0100795 int i, cputhreads, pending;
Jens Axboeebac4652005-12-08 15:25:21 +0100796
797 /*
798 * reap exited threads (TD_EXITED -> TD_REAPED)
799 */
Jens Axboe4d2413c2006-11-23 11:58:59 +0100800 pending = cputhreads = 0;
Jens Axboe34572e22006-10-20 09:33:18 +0200801 for_each_td(td, i) {
Jens Axboe84585002006-10-19 20:26:22 +0200802 /*
803 * ->io_ops is NULL for a thread that has closed its
804 * io engine
805 */
806 if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
Jens Axboeb990b5c2006-09-14 09:48:22 +0200807 cputhreads++;
808
Jens Axboe4d2413c2006-11-23 11:58:59 +0100809 if (td->runstate != TD_EXITED) {
810 if (td->runstate < TD_RUNNING)
811 pending++;
812
Jens Axboeebac4652005-12-08 15:25:21 +0100813 continue;
Jens Axboe4d2413c2006-11-23 11:58:59 +0100814 }
Jens Axboeebac4652005-12-08 15:25:21 +0100815
816 td_set_runstate(td, TD_REAPED);
817
818 if (td->use_thread) {
819 long ret;
820
821 if (pthread_join(td->thread, (void *) &ret))
822 perror("thread_join");
823 } else
824 waitpid(td->pid, NULL, 0);
825
826 (*nr_running)--;
827 (*m_rate) -= td->ratemin;
828 (*t_rate) -= td->rate;
829 }
Jens Axboeb990b5c2006-09-14 09:48:22 +0200830
Jens Axboe4d2413c2006-11-23 11:58:59 +0100831 if (*nr_running == cputhreads && !pending)
Jens Axboe6ce15a32007-01-12 09:04:41 +0100832 terminate_threads(TERMINATE_ALL, 0);
Jens Axboeebac4652005-12-08 15:25:21 +0100833}
834
Jens Axboe906c8d72006-06-13 09:37:56 +0200835/*
836 * Main function for kicking off and reaping jobs, as needed.
837 */
Jens Axboeebac4652005-12-08 15:25:21 +0100838static void run_threads(void)
839{
Jens Axboeebac4652005-12-08 15:25:21 +0100840 struct thread_data *td;
841 unsigned long spent;
842 int i, todo, nr_running, m_rate, t_rate, nr_started;
Jens Axboefcb6ade2006-05-31 12:14:35 +0200843
Jens Axboe2f9ade32006-10-20 11:25:52 +0200844 if (fio_pin_memory())
845 return;
Jens Axboeebac4652005-12-08 15:25:21 +0100846
Jens Axboec6ae0a52006-06-12 11:04:44 +0200847 if (!terse_output) {
848 printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
849 fflush(stdout);
850 }
Jens Axboec04f7ec2006-05-31 10:13:16 +0200851
Jens Axboe4efa9702006-05-31 11:56:49 +0200852 signal(SIGINT, sig_handler);
853 signal(SIGALRM, sig_handler);
854
Jens Axboeebac4652005-12-08 15:25:21 +0100855 todo = thread_number;
856 nr_running = 0;
857 nr_started = 0;
858 m_rate = t_rate = 0;
859
Jens Axboe34572e22006-10-20 09:33:18 +0200860 for_each_td(td, i) {
Jens Axboe263e5292006-10-18 15:37:01 +0200861 print_status_init(td->thread_number - 1);
Jens Axboeebac4652005-12-08 15:25:21 +0100862
Jens Axboe380cf262007-01-11 14:01:27 +0100863 if (!td->create_serialize) {
864 init_disk_util(td);
Jens Axboeebac4652005-12-08 15:25:21 +0100865 continue;
Jens Axboe380cf262007-01-11 14:01:27 +0100866 }
Jens Axboeebac4652005-12-08 15:25:21 +0100867
868 /*
869 * do file setup here so it happens sequentially,
870 * we don't want X number of threads getting their
871 * client data interspersed on disk
872 */
Jens Axboe53cdc682006-10-18 11:50:58 +0200873 if (setup_files(td)) {
Jens Axboeebac4652005-12-08 15:25:21 +0100874 td_set_runstate(td, TD_REAPED);
875 todo--;
876 }
Jens Axboe380cf262007-01-11 14:01:27 +0100877
878 init_disk_util(td);
Jens Axboeebac4652005-12-08 15:25:21 +0100879 }
880
Jens Axboeebac4652005-12-08 15:25:21 +0100881 while (todo) {
Jens Axboe75154842006-06-01 13:56:09 +0200882 struct thread_data *map[MAX_JOBS];
883 struct timeval this_start;
884 int this_jobs = 0, left;
885
Jens Axboeebac4652005-12-08 15:25:21 +0100886 /*
887 * create threads (TD_NOT_CREATED -> TD_CREATED)
888 */
Jens Axboe34572e22006-10-20 09:33:18 +0200889 for_each_td(td, i) {
Jens Axboeebac4652005-12-08 15:25:21 +0100890 if (td->runstate != TD_NOT_CREATED)
891 continue;
892
893 /*
894 * never got a chance to start, killed by other
895 * thread for some reason
896 */
897 if (td->terminate) {
898 todo--;
899 continue;
900 }
901
902 if (td->start_delay) {
Jens Axboe263e5292006-10-18 15:37:01 +0200903 spent = mtime_since_genesis();
Jens Axboeebac4652005-12-08 15:25:21 +0100904
905 if (td->start_delay * 1000 > spent)
906 continue;
907 }
908
909 if (td->stonewall && (nr_started || nr_running))
910 break;
911
Jens Axboe75154842006-06-01 13:56:09 +0200912 /*
913 * Set state to created. Thread will transition
914 * to TD_INITIALIZED when it's done setting up.
915 */
Jens Axboeebac4652005-12-08 15:25:21 +0100916 td_set_runstate(td, TD_CREATED);
Jens Axboe75154842006-06-01 13:56:09 +0200917 map[this_jobs++] = td;
Jens Axboebbfd6b02006-06-07 19:42:54 +0200918 fio_sem_init(&startup_sem, 1);
Jens Axboeebac4652005-12-08 15:25:21 +0100919 nr_started++;
920
921 if (td->use_thread) {
922 if (pthread_create(&td->thread, NULL, thread_main, td)) {
923 perror("thread_create");
924 nr_started--;
925 }
926 } else {
927 if (fork())
Jens Axboebbfd6b02006-06-07 19:42:54 +0200928 fio_sem_down(&startup_sem);
Jens Axboeebac4652005-12-08 15:25:21 +0100929 else {
930 fork_main(shm_id, i);
931 exit(0);
932 }
933 }
934 }
935
936 /*
Jens Axboe75154842006-06-01 13:56:09 +0200937 * Wait for the started threads to transition to
938 * TD_INITIALIZED.
Jens Axboeebac4652005-12-08 15:25:21 +0100939 */
Jens Axboe02bcaa82006-11-24 10:42:00 +0100940 fio_gettime(&this_start, NULL);
Jens Axboe75154842006-06-01 13:56:09 +0200941 left = this_jobs;
Jens Axboe6ce15a32007-01-12 09:04:41 +0100942 while (left && !fio_abort) {
Jens Axboe75154842006-06-01 13:56:09 +0200943 if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
944 break;
945
946 usleep(100000);
947
948 for (i = 0; i < this_jobs; i++) {
949 td = map[i];
950 if (!td)
951 continue;
Jens Axboeb6f4d882006-06-02 10:32:51 +0200952 if (td->runstate == TD_INITIALIZED) {
Jens Axboe75154842006-06-01 13:56:09 +0200953 map[i] = NULL;
954 left--;
Jens Axboeb6f4d882006-06-02 10:32:51 +0200955 } else if (td->runstate >= TD_EXITED) {
956 map[i] = NULL;
957 left--;
958 todo--;
959 nr_running++; /* work-around... */
Jens Axboe75154842006-06-01 13:56:09 +0200960 }
961 }
962 }
963
964 if (left) {
Jens Axboe3b70d7e2006-06-08 21:48:46 +0200965 log_err("fio: %d jobs failed to start\n", left);
Jens Axboe75154842006-06-01 13:56:09 +0200966 for (i = 0; i < this_jobs; i++) {
967 td = map[i];
968 if (!td)
969 continue;
970 kill(td->pid, SIGTERM);
971 }
972 break;
973 }
974
975 /*
Jens Axboeb6f4d882006-06-02 10:32:51 +0200976 * start created threads (TD_INITIALIZED -> TD_RUNNING).
Jens Axboe75154842006-06-01 13:56:09 +0200977 */
Jens Axboe34572e22006-10-20 09:33:18 +0200978 for_each_td(td, i) {
Jens Axboe75154842006-06-01 13:56:09 +0200979 if (td->runstate != TD_INITIALIZED)
Jens Axboeebac4652005-12-08 15:25:21 +0100980 continue;
981
982 td_set_runstate(td, TD_RUNNING);
983 nr_running++;
984 nr_started--;
985 m_rate += td->ratemin;
986 t_rate += td->rate;
Jens Axboe75154842006-06-01 13:56:09 +0200987 todo--;
Jens Axboebbfd6b02006-06-07 19:42:54 +0200988 fio_sem_up(&td->mutex);
Jens Axboeebac4652005-12-08 15:25:21 +0100989 }
990
991 reap_threads(&nr_running, &t_rate, &m_rate);
992
993 if (todo)
994 usleep(100000);
995 }
996
997 while (nr_running) {
998 reap_threads(&nr_running, &t_rate, &m_rate);
999 usleep(10000);
1000 }
1001
1002 update_io_ticks();
Jens Axboe2f9ade32006-10-20 11:25:52 +02001003 fio_unpin_memory();
Jens Axboeebac4652005-12-08 15:25:21 +01001004}
1005
Jens Axboeebac4652005-12-08 15:25:21 +01001006int main(int argc, char *argv[])
1007{
Jens Axboe29d610e2007-02-06 13:25:33 +01001008 long ps;
1009
Jens Axboeebac4652005-12-08 15:25:21 +01001010 if (parse_options(argc, argv))
1011 return 1;
1012
1013 if (!thread_number) {
Jens Axboe3b70d7e2006-06-08 21:48:46 +02001014 log_err("Nothing to do\n");
Jens Axboeebac4652005-12-08 15:25:21 +01001015 return 1;
1016 }
1017
Jens Axboe29d610e2007-02-06 13:25:33 +01001018 ps = sysconf(_SC_PAGESIZE);
1019 if (ps < 0) {
1020 log_err("Failed to get page size\n");
1021 return 1;
1022 }
1023
1024 page_mask = ps - 1;
1025
Jens Axboebb3884d2007-01-17 17:23:11 +11001026 if (write_bw_log) {
1027 setup_log(&agg_io_log[DDIR_READ]);
1028 setup_log(&agg_io_log[DDIR_WRITE]);
1029 }
1030
Jens Axboeebac4652005-12-08 15:25:21 +01001031 disk_util_timer_arm();
1032
1033 run_threads();
Jens Axboe6ce15a32007-01-12 09:04:41 +01001034
Jens Axboebb3884d2007-01-17 17:23:11 +11001035 if (!fio_abort) {
Jens Axboe6ce15a32007-01-12 09:04:41 +01001036 show_run_stats();
Jens Axboebb3884d2007-01-17 17:23:11 +11001037 if (write_bw_log) {
1038 __finish_log(agg_io_log[DDIR_READ],"agg-read_bw.log");
1039 __finish_log(agg_io_log[DDIR_WRITE],"agg-write_bw.log");
1040 }
1041 }
Jens Axboeebac4652005-12-08 15:25:21 +01001042
1043 return 0;
1044}