22b97950fb8ffaa5e08149ab483a802ccbb0c4cf
[fio.git] / init.c
1 /*
2  * This file contains job initialization and setup functions.
3  */
4 #include <stdio.h>
5 #include <stdlib.h>
6 #include <unistd.h>
7 #include <fcntl.h>
8 #include <ctype.h>
9 #include <string.h>
10 #include <errno.h>
11 #include <sys/ipc.h>
12 #include <sys/shm.h>
13 #include <sys/types.h>
14 #include <sys/stat.h>
15
16 #include "fio.h"
17 #include "parse.h"
18 #include "smalloc.h"
19 #include "filehash.h"
20 #include "verify.h"
21 #include "profile.h"
22 #include "server.h"
23
24 #include "lib/getopt.h"
25
26 static char fio_version_string[] = "fio 1.58";
27
28 #define FIO_RANDSEED            (0xb1899bedUL)
29
30 static char **ini_file;
31 static int max_jobs = FIO_MAX_JOBS;
32 static int dump_cmdline;
33
34 static struct thread_data def_thread;
35 struct thread_data *threads = NULL;
36
37 int exitall_on_terminate = 0;
38 int terse_output = 0;
39 int eta_print;
40 unsigned long long mlock_size = 0;
41 FILE *f_out = NULL;
42 FILE *f_err = NULL;
43 char **job_sections = NULL;
44 int nr_job_sections = 0;
45 char *exec_profile = NULL;
46 int warnings_fatal = 0;
47 int terse_version = 2;
48 int is_backend = 0;
49 int nr_clients = 0;
50 int log_syslog = 0;
51
52 int write_bw_log = 0;
53 int read_only = 0;
54
55 static int write_lat_log;
56
57 static int prev_group_jobs;
58
59 unsigned long fio_debug = 0;
60 unsigned int fio_debug_jobno = -1;
61 unsigned int *fio_debug_jobp = NULL;
62
63 static char cmd_optstr[256];
64
65 /*
66  * Command line options. These will contain the above, plus a few
67  * extra that only pertain to fio itself and not jobs.
68  */
69 static struct option l_opts[FIO_NR_OPTIONS] = {
70         {
71                 .name           = (char *) "output",
72                 .has_arg        = required_argument,
73                 .val            = 'o',
74         },
75         {
76                 .name           = (char *) "timeout",
77                 .has_arg        = required_argument,
78                 .val            = 't',
79         },
80         {
81                 .name           = (char *) "latency-log",
82                 .has_arg        = required_argument,
83                 .val            = 'l',
84         },
85         {
86                 .name           = (char *) "bandwidth-log",
87                 .has_arg        = required_argument,
88                 .val            = 'b',
89         },
90         {
91                 .name           = (char *) "minimal",
92                 .has_arg        = optional_argument,
93                 .val            = 'm',
94         },
95         {
96                 .name           = (char *) "version",
97                 .has_arg        = no_argument,
98                 .val            = 'v',
99         },
100         {
101                 .name           = (char *) "help",
102                 .has_arg        = no_argument,
103                 .val            = 'h',
104         },
105         {
106                 .name           = (char *) "cmdhelp",
107                 .has_arg        = optional_argument,
108                 .val            = 'c',
109         },
110         {
111                 .name           = (char *) "showcmd",
112                 .has_arg        = no_argument,
113                 .val            = 's',
114         },
115         {
116                 .name           = (char *) "readonly",
117                 .has_arg        = no_argument,
118                 .val            = 'r',
119         },
120         {
121                 .name           = (char *) "eta",
122                 .has_arg        = required_argument,
123                 .val            = 'e',
124         },
125         {
126                 .name           = (char *) "debug",
127                 .has_arg        = required_argument,
128                 .val            = 'd',
129         },
130         {
131                 .name           = (char *) "section",
132                 .has_arg        = required_argument,
133                 .val            = 'x',
134         },
135         {
136                 .name           = (char *) "alloc-size",
137                 .has_arg        = required_argument,
138                 .val            = 'a',
139         },
140         {
141                 .name           = (char *) "profile",
142                 .has_arg        = required_argument,
143                 .val            = 'p',
144         },
145         {
146                 .name           = (char *) "warnings-fatal",
147                 .has_arg        = no_argument,
148                 .val            = 'w',
149         },
150         {
151                 .name           = (char *) "max-jobs",
152                 .has_arg        = required_argument,
153                 .val            = 'j',
154         },
155         {
156                 .name           = (char *) "terse-version",
157                 .has_arg        = required_argument,
158                 .val            = 'V',
159         },
160         {
161                 .name           = (char *) "server",
162                 .has_arg        = no_argument,
163                 .val            = 'S',
164         },
165         {       .name           = (char *) "daemonize",
166                 .has_arg        = no_argument,
167                 .val            = 'D',
168         },
169         {
170                 .name           = (char *) "net-port",
171                 .has_arg        = required_argument,
172                 .val            = 'P',
173         },
174         {
175                 .name           = (char *) "client",
176                 .has_arg        = required_argument,
177                 .val            = 'C',
178         },
179         {
180                 .name           = NULL,
181         },
182 };
183
184 /*
185  * Return a free job structure.
186  */
187 static struct thread_data *get_new_job(int global, struct thread_data *parent)
188 {
189         struct thread_data *td;
190
191         if (global)
192                 return &def_thread;
193         if (thread_number >= max_jobs) {
194                 log_err("error: maximum number of jobs (%d) reached.\n",
195                                 max_jobs);
196                 return NULL;
197         }
198
199         td = &threads[thread_number++];
200         *td = *parent;
201
202         td->o.uid = td->o.gid = -1U;
203
204         dup_files(td, parent);
205         options_mem_dupe(td);
206
207         profile_add_hooks(td);
208
209         td->thread_number = thread_number;
210         return td;
211 }
212
213 static void put_job(struct thread_data *td)
214 {
215         if (td == &def_thread)
216                 return;
217
218         profile_td_exit(td);
219
220         if (td->error)
221                 log_info("fio: %s\n", td->verror);
222
223         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
224         thread_number--;
225 }
226
227 static int __setup_rate(struct thread_data *td, enum fio_ddir ddir)
228 {
229         unsigned int bs = td->o.min_bs[ddir];
230         unsigned long long bytes_per_sec;
231
232         assert(ddir_rw(ddir));
233
234         if (td->o.rate[ddir])
235                 bytes_per_sec = td->o.rate[ddir];
236         else
237                 bytes_per_sec = td->o.rate_iops[ddir] * bs;
238
239         if (!bytes_per_sec) {
240                 log_err("rate lower than supported\n");
241                 return -1;
242         }
243
244         td->rate_nsec_cycle[ddir] = 1000000000ULL / bytes_per_sec;
245         td->rate_pending_usleep[ddir] = 0;
246         return 0;
247 }
248
249 static int setup_rate(struct thread_data *td)
250 {
251         int ret = 0;
252
253         if (td->o.rate[DDIR_READ] || td->o.rate_iops[DDIR_READ])
254                 ret = __setup_rate(td, DDIR_READ);
255         if (td->o.rate[DDIR_WRITE] || td->o.rate_iops[DDIR_WRITE])
256                 ret |= __setup_rate(td, DDIR_WRITE);
257
258         return ret;
259 }
260
261 static int fixed_block_size(struct thread_options *o)
262 {
263         return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
264                 o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
265                 o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE];
266 }
267
268 /*
269  * Lazy way of fixing up options that depend on each other. We could also
270  * define option callback handlers, but this is easier.
271  */
272 static int fixup_options(struct thread_data *td)
273 {
274         struct thread_options *o = &td->o;
275         int ret = 0;
276
277 #ifndef FIO_HAVE_PSHARED_MUTEX
278         if (!o->use_thread) {
279                 log_info("fio: this platform does not support process shared"
280                          " mutexes, forcing use of threads. Use the 'thread'"
281                          " option to get rid of this warning.\n");
282                 o->use_thread = 1;
283                 ret = warnings_fatal;
284         }
285 #endif
286
287         if (o->write_iolog_file && o->read_iolog_file) {
288                 log_err("fio: read iolog overrides write_iolog\n");
289                 free(o->write_iolog_file);
290                 o->write_iolog_file = NULL;
291                 ret = warnings_fatal;
292         }
293
294         /*
295          * only really works for sequential io for now, and with 1 file
296          */
297         if (o->zone_size && td_random(td) && o->open_files == 1)
298                 o->zone_size = 0;
299
300         /*
301          * Reads can do overwrites, we always need to pre-create the file
302          */
303         if (td_read(td) || td_rw(td))
304                 o->overwrite = 1;
305
306         if (!o->min_bs[DDIR_READ])
307                 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
308         if (!o->max_bs[DDIR_READ])
309                 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
310         if (!o->min_bs[DDIR_WRITE])
311                 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
312         if (!o->max_bs[DDIR_WRITE])
313                 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
314
315         o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
316
317         /*
318          * For random IO, allow blockalign offset other than min_bs.
319          */
320         if (!o->ba[DDIR_READ] || !td_random(td))
321                 o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
322         if (!o->ba[DDIR_WRITE] || !td_random(td))
323                 o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
324
325         if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
326             o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
327             !o->norandommap) {
328                 log_err("fio: Any use of blockalign= turns off randommap\n");
329                 o->norandommap = 1;
330                 ret = warnings_fatal;
331         }
332
333         if (!o->file_size_high)
334                 o->file_size_high = o->file_size_low;
335
336         if (o->norandommap && o->verify != VERIFY_NONE
337             && !fixed_block_size(o))  {
338                 log_err("fio: norandommap given for variable block sizes, "
339                         "verify disabled\n");
340                 o->verify = VERIFY_NONE;
341                 ret = warnings_fatal;
342         }
343         if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
344                 log_err("fio: bs_unaligned may not work with raw io\n");
345
346         /*
347          * thinktime_spin must be less than thinktime
348          */
349         if (o->thinktime_spin > o->thinktime)
350                 o->thinktime_spin = o->thinktime;
351
352         /*
353          * The low water mark cannot be bigger than the iodepth
354          */
355         if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
356                 /*
357                  * syslet work around - if the workload is sequential,
358                  * we want to let the queue drain all the way down to
359                  * avoid seeking between async threads
360                  */
361                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
362                         o->iodepth_low = 1;
363                 else
364                         o->iodepth_low = o->iodepth;
365         }
366
367         /*
368          * If batch number isn't set, default to the same as iodepth
369          */
370         if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
371                 o->iodepth_batch = o->iodepth;
372
373         if (o->nr_files > td->files_index)
374                 o->nr_files = td->files_index;
375
376         if (o->open_files > o->nr_files || !o->open_files)
377                 o->open_files = o->nr_files;
378
379         if (((o->rate[0] + o->rate[1]) && (o->rate_iops[0] + o->rate_iops[1]))||
380             ((o->ratemin[0] + o->ratemin[1]) && (o->rate_iops_min[0] +
381                 o->rate_iops_min[1]))) {
382                 log_err("fio: rate and rate_iops are mutually exclusive\n");
383                 ret = 1;
384         }
385         if ((o->rate[0] < o->ratemin[0]) || (o->rate[1] < o->ratemin[1]) ||
386             (o->rate_iops[0] < o->rate_iops_min[0]) ||
387             (o->rate_iops[1] < o->rate_iops_min[1])) {
388                 log_err("fio: minimum rate exceeds rate\n");
389                 ret = 1;
390         }
391
392         if (!o->timeout && o->time_based) {
393                 log_err("fio: time_based requires a runtime/timeout setting\n");
394                 o->time_based = 0;
395                 ret = warnings_fatal;
396         }
397
398         if (o->fill_device && !o->size)
399                 o->size = -1ULL;
400
401         if (o->verify != VERIFY_NONE) {
402                 if (td_write(td) && o->do_verify && o->numjobs > 1) {
403                         log_info("Multiple writers may overwrite blocks that "
404                                 "belong to other jobs. This can cause "
405                                 "verification failures.\n");
406                         ret = warnings_fatal;
407                 }
408
409                 o->refill_buffers = 1;
410                 if (o->max_bs[DDIR_WRITE] != o->min_bs[DDIR_WRITE] &&
411                     !o->verify_interval)
412                         o->verify_interval = o->min_bs[DDIR_WRITE];
413         }
414
415         if (o->pre_read) {
416                 o->invalidate_cache = 0;
417                 if (td->io_ops->flags & FIO_PIPEIO) {
418                         log_info("fio: cannot pre-read files with an IO engine"
419                                  " that isn't seekable. Pre-read disabled.\n");
420                         ret = warnings_fatal;
421                 }
422         }
423
424 #ifndef FIO_HAVE_FDATASYNC
425         if (o->fdatasync_blocks) {
426                 log_info("fio: this platform does not support fdatasync()"
427                          " falling back to using fsync().  Use the 'fsync'"
428                          " option instead of 'fdatasync' to get rid of"
429                          " this warning\n");
430                 o->fsync_blocks = o->fdatasync_blocks;
431                 o->fdatasync_blocks = 0;
432                 ret = warnings_fatal;
433         }
434 #endif
435
436         return ret;
437 }
438
439 /*
440  * This function leaks the buffer
441  */
442 static char *to_kmg(unsigned int val)
443 {
444         char *buf = malloc(32);
445         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
446         char *p = post;
447
448         do {
449                 if (val & 1023)
450                         break;
451
452                 val >>= 10;
453                 p++;
454         } while (*p);
455
456         snprintf(buf, 31, "%u%c", val, *p);
457         return buf;
458 }
459
460 /* External engines are specified by "external:name.o") */
461 static const char *get_engine_name(const char *str)
462 {
463         char *p = strstr(str, ":");
464
465         if (!p)
466                 return str;
467
468         p++;
469         strip_blank_front(&p);
470         strip_blank_end(p);
471         return p;
472 }
473
474 static int exists_and_not_file(const char *filename)
475 {
476         struct stat sb;
477
478         if (lstat(filename, &sb) == -1)
479                 return 0;
480
481         /* \\.\ is the device namespace in Windows, where every file
482          * is a device node */
483         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
484                 return 0;
485
486         return 1;
487 }
488
489 static void td_fill_rand_seeds_os(struct thread_data *td)
490 {
491         os_random_seed(td->rand_seeds[0], &td->bsrange_state);
492         os_random_seed(td->rand_seeds[1], &td->verify_state);
493         os_random_seed(td->rand_seeds[2], &td->rwmix_state);
494
495         if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
496                 os_random_seed(td->rand_seeds[3], &td->next_file_state);
497
498         os_random_seed(td->rand_seeds[5], &td->file_size_state);
499         os_random_seed(td->rand_seeds[6], &td->trim_state);
500
501         if (!td_random(td))
502                 return;
503
504         if (td->o.rand_repeatable)
505                 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
506
507         os_random_seed(td->rand_seeds[4], &td->random_state);
508 }
509
510 static void td_fill_rand_seeds_internal(struct thread_data *td)
511 {
512         init_rand_seed(&td->__bsrange_state, td->rand_seeds[0]);
513         init_rand_seed(&td->__verify_state, td->rand_seeds[1]);
514         init_rand_seed(&td->__rwmix_state, td->rand_seeds[2]);
515
516         if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
517                 init_rand_seed(&td->__next_file_state, td->rand_seeds[3]);
518
519         init_rand_seed(&td->__file_size_state, td->rand_seeds[5]);
520         init_rand_seed(&td->__trim_state, td->rand_seeds[6]);
521
522         if (!td_random(td))
523                 return;
524
525         if (td->o.rand_repeatable)
526                 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
527
528         init_rand_seed(&td->__random_state, td->rand_seeds[4]);
529 }
530
531 void td_fill_rand_seeds(struct thread_data *td)
532 {
533         if (td->o.use_os_rand)
534                 td_fill_rand_seeds_os(td);
535         else
536                 td_fill_rand_seeds_internal(td);
537
538         init_rand_seed(&td->buf_state, td->rand_seeds[7]);
539 }
540
541 /*
542  * Initialize the various random states we need (random io, block size ranges,
543  * read/write mix, etc).
544  */
545 static int init_random_state(struct thread_data *td)
546 {
547         int fd;
548
549         fd = open("/dev/urandom", O_RDONLY);
550         if (fd == -1) {
551                 td_verror(td, errno, "open");
552                 return 1;
553         }
554
555         if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
556             (int) sizeof(td->rand_seeds)) {
557                 td_verror(td, EIO, "read");
558                 close(fd);
559                 return 1;
560         }
561
562         close(fd);
563         td_fill_rand_seeds(td);
564         return 0;
565 }
566
567 /*
568  * Adds a job to the list of things todo. Sanitizes the various options
569  * to make sure we don't have conflicts, and initializes various
570  * members of td.
571  */
572 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
573 {
574         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
575                                    "randread", "randwrite", "randrw" };
576         unsigned int i;
577         const char *engine;
578         char fname[PATH_MAX];
579         int numjobs, file_alloced;
580
581         /*
582          * the def_thread is just for options, it's not a real job
583          */
584         if (td == &def_thread)
585                 return 0;
586
587         /*
588          * if we are just dumping the output command line, don't add the job
589          */
590         if (dump_cmdline) {
591                 put_job(td);
592                 return 0;
593         }
594
595         if (profile_td_init(td))
596                 goto err;
597
598         engine = get_engine_name(td->o.ioengine);
599         td->io_ops = load_ioengine(td, engine);
600         if (!td->io_ops) {
601                 log_err("fio: failed to load engine %s\n", engine);
602                 goto err;
603         }
604
605         if (td->o.use_thread)
606                 nr_thread++;
607         else
608                 nr_process++;
609
610         if (td->o.odirect)
611                 td->io_ops->flags |= FIO_RAWIO;
612
613         file_alloced = 0;
614         if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
615                 file_alloced = 1;
616
617                 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
618                         add_file(td, jobname);
619                 else {
620                         for (i = 0; i < td->o.nr_files; i++) {
621                                 sprintf(fname, "%s.%d.%d", jobname,
622                                                         td->thread_number, i);
623                                 add_file(td, fname);
624                         }
625                 }
626         }
627
628         if (fixup_options(td))
629                 goto err;
630
631         if (td->io_ops->flags & FIO_DISKLESSIO) {
632                 struct fio_file *f;
633
634                 for_each_file(td, f, i)
635                         f->real_file_size = -1ULL;
636         }
637
638         td->mutex = fio_mutex_init(0);
639
640         td->ts.clat_percentiles = td->o.clat_percentiles;
641         if (td->o.overwrite_plist)
642                 td->ts.percentile_list = td->o.percentile_list;
643         else
644                 td->ts.percentile_list = NULL;
645
646         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
647         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
648         td->ts.lat_stat[0].min_val = td->ts.lat_stat[1].min_val = ULONG_MAX;
649         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
650         td->ddir_seq_nr = td->o.ddir_seq_nr;
651
652         if ((td->o.stonewall || td->o.new_group) && prev_group_jobs) {
653                 prev_group_jobs = 0;
654                 groupid++;
655         }
656
657         td->groupid = groupid;
658         prev_group_jobs++;
659
660         if (init_random_state(td))
661                 goto err;
662
663         if (setup_rate(td))
664                 goto err;
665
666         if (td->o.write_lat_log) {
667                 setup_log(&td->lat_log);
668                 setup_log(&td->slat_log);
669                 setup_log(&td->clat_log);
670         }
671         if (td->o.write_bw_log)
672                 setup_log(&td->bw_log);
673
674         if (!td->o.name)
675                 td->o.name = strdup(jobname);
676
677         if (!terse_output) {
678                 if (!job_add_num) {
679                         if (!strcmp(td->io_ops->name, "cpuio")) {
680                                 log_info("%s: ioengine=cpu, cpuload=%u,"
681                                          " cpucycle=%u\n", td->o.name,
682                                                         td->o.cpuload,
683                                                         td->o.cpucycle);
684                         } else {
685                                 char *c1, *c2, *c3, *c4;
686
687                                 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
688                                 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
689                                 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
690                                 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
691
692                                 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
693                                          " ioengine=%s, iodepth=%u\n",
694                                                 td->o.name, td->groupid,
695                                                 ddir_str[td->o.td_ddir],
696                                                 c1, c2, c3, c4,
697                                                 td->io_ops->name,
698                                                 td->o.iodepth);
699
700                                 free(c1);
701                                 free(c2);
702                                 free(c3);
703                                 free(c4);
704                         }
705                 } else if (job_add_num == 1)
706                         log_info("...\n");
707         }
708
709         /*
710          * recurse add identical jobs, clear numjobs and stonewall options
711          * as they don't apply to sub-jobs
712          */
713         numjobs = td->o.numjobs;
714         while (--numjobs) {
715                 struct thread_data *td_new = get_new_job(0, td);
716
717                 if (!td_new)
718                         goto err;
719
720                 td_new->o.numjobs = 1;
721                 td_new->o.stonewall = 0;
722                 td_new->o.new_group = 0;
723
724                 if (file_alloced) {
725                         td_new->o.filename = NULL;
726                         td_new->files_index = 0;
727                         td_new->files_size = 0;
728                         td_new->files = NULL;
729                 }
730
731                 job_add_num = numjobs - 1;
732
733                 if (add_job(td_new, jobname, job_add_num))
734                         goto err;
735         }
736
737         return 0;
738 err:
739         put_job(td);
740         return -1;
741 }
742
743 /*
744  * Parse as if 'o' was a command line
745  */
746 void add_job_opts(const char **o)
747 {
748         struct thread_data *td, *td_parent;
749         int i, in_global = 1;
750         char jobname[32];
751
752         i = 0;
753         td_parent = td = NULL;
754         while (o[i]) {
755                 if (!strncmp(o[i], "name", 4)) {
756                         in_global = 0;
757                         if (td)
758                                 add_job(td, jobname, 0);
759                         td = NULL;
760                         sprintf(jobname, "%s", o[i] + 5);
761                 }
762                 if (in_global && !td_parent)
763                         td_parent = get_new_job(1, &def_thread);
764                 else if (!in_global && !td) {
765                         if (!td_parent)
766                                 td_parent = &def_thread;
767                         td = get_new_job(0, td_parent);
768                 }
769                 if (in_global)
770                         fio_options_parse(td_parent, (char **) &o[i], 1);
771                 else
772                         fio_options_parse(td, (char **) &o[i], 1);
773                 i++;
774         }
775
776         if (td)
777                 add_job(td, jobname, 0);
778 }
779
780 static int skip_this_section(const char *name)
781 {
782         int i;
783
784         if (!nr_job_sections)
785                 return 0;
786         if (!strncmp(name, "global", 6))
787                 return 0;
788
789         for (i = 0; i < nr_job_sections; i++)
790                 if (!strcmp(job_sections[i], name))
791                         return 0;
792
793         return 1;
794 }
795
796 static int is_empty_or_comment(char *line)
797 {
798         unsigned int i;
799
800         for (i = 0; i < strlen(line); i++) {
801                 if (line[i] == ';')
802                         return 1;
803                 if (line[i] == '#')
804                         return 1;
805                 if (!isspace((int) line[i]) && !iscntrl((int) line[i]))
806                         return 0;
807         }
808
809         return 1;
810 }
811
812 /*
813  * This is our [ini] type file parser.
814  */
815 int parse_jobs_ini(char *file, int is_buf, int stonewall_flag)
816 {
817         unsigned int global;
818         struct thread_data *td;
819         char *string, *name;
820         FILE *f;
821         char *p;
822         int ret = 0, stonewall;
823         int first_sect = 1;
824         int skip_fgets = 0;
825         int inside_skip = 0;
826         char **opts;
827         int i, alloc_opts, num_opts;
828
829         if (is_buf)
830                 f = NULL;
831         else {
832                 if (!strcmp(file, "-"))
833                         f = stdin;
834                 else
835                         f = fopen(file, "r");
836
837                 if (!f) {
838                         perror("fopen job file");
839                         return 1;
840                 }
841         }
842
843         string = malloc(4096);
844
845         /*
846          * it's really 256 + small bit, 280 should suffice
847          */
848         name = malloc(280);
849         memset(name, 0, 280);
850
851         alloc_opts = 8;
852         opts = malloc(sizeof(char *) * alloc_opts);
853         num_opts = 0;
854
855         stonewall = stonewall_flag;
856         do {
857                 /*
858                  * if skip_fgets is set, we already have loaded a line we
859                  * haven't handled.
860                  */
861                 if (!skip_fgets) {
862                         if (is_buf)
863                                 p = strsep(&file, "\n");
864                         else
865                                 p = fgets(string, 4095, f);
866                         if (!p)
867                                 break;
868                 }
869
870                 skip_fgets = 0;
871                 strip_blank_front(&p);
872                 strip_blank_end(p);
873
874                 if (is_empty_or_comment(p))
875                         continue;
876                 if (sscanf(p, "[%255[^\n]]", name) != 1) {
877                         if (inside_skip)
878                                 continue;
879                         log_err("fio: option <%s> outside of [] job section\n",
880                                                                         p);
881                         break;
882                 }
883
884                 name[strlen(name) - 1] = '\0';
885
886                 if (skip_this_section(name)) {
887                         inside_skip = 1;
888                         continue;
889                 } else
890                         inside_skip = 0;
891
892                 global = !strncmp(name, "global", 6);
893
894                 if (dump_cmdline) {
895                         if (first_sect)
896                                 log_info("fio ");
897                         if (!global)
898                                 log_info("--name=%s ", name);
899                         first_sect = 0;
900                 }
901
902                 td = get_new_job(global, &def_thread);
903                 if (!td) {
904                         ret = 1;
905                         break;
906                 }
907
908                 /*
909                  * Seperate multiple job files by a stonewall
910                  */
911                 if (!global && stonewall) {
912                         td->o.stonewall = stonewall;
913                         stonewall = 0;
914                 }
915
916                 num_opts = 0;
917                 memset(opts, 0, alloc_opts * sizeof(char *));
918
919                 while (1) {
920                         if (is_buf)
921                                 p = strsep(&file, "\n");
922                         else
923                                 p = fgets(string, 4096, f);
924                         if (!p)
925                                 break;
926
927                         if (is_empty_or_comment(p))
928                                 continue;
929
930                         strip_blank_front(&p);
931
932                         /*
933                          * new section, break out and make sure we don't
934                          * fgets() a new line at the top.
935                          */
936                         if (p[0] == '[') {
937                                 skip_fgets = 1;
938                                 break;
939                         }
940
941                         strip_blank_end(p);
942
943                         if (num_opts == alloc_opts) {
944                                 alloc_opts <<= 1;
945                                 opts = realloc(opts,
946                                                 alloc_opts * sizeof(char *));
947                         }
948
949                         opts[num_opts] = strdup(p);
950                         num_opts++;
951                 }
952
953                 ret = fio_options_parse(td, opts, num_opts);
954                 if (!ret) {
955                         if (dump_cmdline)
956                                 for (i = 0; i < num_opts; i++)
957                                         log_info("--%s ", opts[i]);
958
959                         ret = add_job(td, name, 0);
960                 } else {
961                         log_err("fio: job %s dropped\n", name);
962                         put_job(td);
963                 }
964
965                 for (i = 0; i < num_opts; i++)
966                         free(opts[i]);
967                 num_opts = 0;
968         } while (!ret);
969
970         if (dump_cmdline)
971                 log_info("\n");
972
973         for (i = 0; i < num_opts; i++)
974                 free(opts[i]);
975
976         free(string);
977         free(name);
978         free(opts);
979         if (!is_buf && f != stdin)
980                 fclose(f);
981         return ret;
982 }
983
984 static int fill_def_thread(void)
985 {
986         memset(&def_thread, 0, sizeof(def_thread));
987
988         fio_getaffinity(getpid(), &def_thread.o.cpumask);
989
990         /*
991          * fill default options
992          */
993         fio_fill_default_options(&def_thread);
994         return 0;
995 }
996
997 static void free_shm(void)
998 {
999         struct shmid_ds sbuf;
1000
1001         if (threads) {
1002                 void *tp = threads;
1003
1004                 threads = NULL;
1005                 file_hash_exit();
1006                 fio_debug_jobp = NULL;
1007                 shmdt(tp);
1008                 shmctl(shm_id, IPC_RMID, &sbuf);
1009         }
1010
1011         scleanup();
1012 }
1013
1014 /*
1015  * The thread area is shared between the main process and the job
1016  * threads/processes. So setup a shared memory segment that will hold
1017  * all the job info. We use the end of the region for keeping track of
1018  * open files across jobs, for file sharing.
1019  */
1020 static int setup_thread_area(void)
1021 {
1022         void *hash;
1023
1024         /*
1025          * 1024 is too much on some machines, scale max_jobs if
1026          * we get a failure that looks like too large a shm segment
1027          */
1028         do {
1029                 size_t size = max_jobs * sizeof(struct thread_data);
1030
1031                 size += file_hash_size;
1032                 size += sizeof(unsigned int);
1033
1034                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1035                 if (shm_id != -1)
1036                         break;
1037                 if (errno != EINVAL) {
1038                         perror("shmget");
1039                         break;
1040                 }
1041
1042                 max_jobs >>= 1;
1043         } while (max_jobs);
1044
1045         if (shm_id == -1)
1046                 return 1;
1047
1048         threads = shmat(shm_id, NULL, 0);
1049         if (threads == (void *) -1) {
1050                 perror("shmat");
1051                 return 1;
1052         }
1053
1054         memset(threads, 0, max_jobs * sizeof(struct thread_data));
1055         hash = (void *) threads + max_jobs * sizeof(struct thread_data);
1056         fio_debug_jobp = (void *) hash + file_hash_size;
1057         *fio_debug_jobp = -1;
1058         file_hash_init(hash);
1059         atexit(free_shm);
1060         return 0;
1061 }
1062
1063 static void usage(const char *name)
1064 {
1065         printf("%s\n", fio_version_string);
1066         printf("%s [options] [job options] <job file(s)>\n", name);
1067         printf("\t--debug=options\tEnable debug logging\n");
1068         printf("\t--output\tWrite output to file\n");
1069         printf("\t--timeout\tRuntime in seconds\n");
1070         printf("\t--latency-log\tGenerate per-job latency logs\n");
1071         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1072         printf("\t--minimal\tMinimal (terse) output\n");
1073         printf("\t--version\tPrint version info and exit\n");
1074         printf("\t--terse-version=x Terse version output format\n");
1075         printf("\t--help\t\tPrint this page\n");
1076         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
1077                 " them\n");
1078         printf("\t--showcmd\tTurn a job file into command line options\n");
1079         printf("\t--eta=when\tWhen ETA estimate should be printed\n");
1080         printf("\t          \tMay be \"always\", \"never\" or \"auto\"\n");
1081         printf("\t--readonly\tTurn on safety read-only checks, preventing"
1082                 " writes\n");
1083         printf("\t--section=name\tOnly run specified section in job file\n");
1084         printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
1085                 " (def 1024)\n");
1086         printf("\t--warnings-fatal Fio parser warnings are fatal\n");
1087         printf("\t--max-jobs\tMaximum number of threads/processes to support\n");
1088         printf("\t--server\tStart a backend fio server\n");
1089         printf("\t--client=hostname Talk to remove backend fio server at hostname\n");
1090         printf("\t--net-port=port\tUse specified port for client/server connection\n");
1091         printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>");
1092         printf("\n                   Jens Axboe <jaxboe@fusionio.com>\n");
1093 }
1094
1095 #ifdef FIO_INC_DEBUG
1096 struct debug_level debug_levels[] = {
1097         { .name = "process",    .shift = FD_PROCESS, },
1098         { .name = "file",       .shift = FD_FILE, },
1099         { .name = "io",         .shift = FD_IO, },
1100         { .name = "mem",        .shift = FD_MEM, },
1101         { .name = "blktrace",   .shift = FD_BLKTRACE },
1102         { .name = "verify",     .shift = FD_VERIFY },
1103         { .name = "random",     .shift = FD_RANDOM },
1104         { .name = "parse",      .shift = FD_PARSE },
1105         { .name = "diskutil",   .shift = FD_DISKUTIL },
1106         { .name = "job",        .shift = FD_JOB },
1107         { .name = "mutex",      .shift = FD_MUTEX },
1108         { .name = "profile",    .shift = FD_PROFILE },
1109         { .name = "time",       .shift = FD_TIME },
1110         { .name = "net",        .shift = FD_NET },
1111         { .name = NULL, },
1112 };
1113
1114 static int set_debug(const char *string)
1115 {
1116         struct debug_level *dl;
1117         char *p = (char *) string;
1118         char *opt;
1119         int i;
1120
1121         if (!strcmp(string, "?") || !strcmp(string, "help")) {
1122                 log_info("fio: dumping debug options:");
1123                 for (i = 0; debug_levels[i].name; i++) {
1124                         dl = &debug_levels[i];
1125                         log_info("%s,", dl->name);
1126                 }
1127                 log_info("all\n");
1128                 return 1;
1129         }
1130
1131         while ((opt = strsep(&p, ",")) != NULL) {
1132                 int found = 0;
1133
1134                 if (!strncmp(opt, "all", 3)) {
1135                         log_info("fio: set all debug options\n");
1136                         fio_debug = ~0UL;
1137                         continue;
1138                 }
1139
1140                 for (i = 0; debug_levels[i].name; i++) {
1141                         dl = &debug_levels[i];
1142                         found = !strncmp(opt, dl->name, strlen(dl->name));
1143                         if (!found)
1144                                 continue;
1145
1146                         if (dl->shift == FD_JOB) {
1147                                 opt = strchr(opt, ':');
1148                                 if (!opt) {
1149                                         log_err("fio: missing job number\n");
1150                                         break;
1151                                 }
1152                                 opt++;
1153                                 fio_debug_jobno = atoi(opt);
1154                                 log_info("fio: set debug jobno %d\n",
1155                                                         fio_debug_jobno);
1156                         } else {
1157                                 log_info("fio: set debug option %s\n", opt);
1158                                 fio_debug |= (1UL << dl->shift);
1159                         }
1160                         break;
1161                 }
1162
1163                 if (!found)
1164                         log_err("fio: debug mask %s not found\n", opt);
1165         }
1166         return 0;
1167 }
1168 #else
1169 static int set_debug(const char *string)
1170 {
1171         log_err("fio: debug tracing not included in build\n");
1172         return 1;
1173 }
1174 #endif
1175
1176 static void fio_options_fill_optstring(void)
1177 {
1178         char *ostr = cmd_optstr;
1179         int i, c;
1180
1181         c = i = 0;
1182         while (l_opts[i].name) {
1183                 ostr[c++] = l_opts[i].val;
1184                 if (l_opts[i].has_arg == required_argument)
1185                         ostr[c++] = ':';
1186                 else if (l_opts[i].has_arg == optional_argument) {
1187                         ostr[c++] = ':';
1188                         ostr[c++] = ':';
1189                 }
1190                 i++;
1191         }
1192         ostr[c] = '\0';
1193 }
1194
1195 static int parse_cmd_line(int argc, char *argv[])
1196 {
1197         struct thread_data *td = NULL;
1198         int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1199         char *ostr = cmd_optstr;
1200         int daemonize_server = 0;
1201
1202         while ((c = getopt_long_only(argc, argv, ostr, l_opts, &lidx)) != -1) {
1203                 switch (c) {
1204                 case 'a':
1205                         smalloc_pool_size = atoi(optarg);
1206                         break;
1207                 case 't':
1208                         def_thread.o.timeout = atoi(optarg);
1209                         break;
1210                 case 'l':
1211                         write_lat_log = 1;
1212                         break;
1213                 case 'b':
1214                         write_bw_log = 1;
1215                         break;
1216                 case 'o':
1217                         f_out = fopen(optarg, "w+");
1218                         if (!f_out) {
1219                                 perror("fopen output");
1220                                 exit(1);
1221                         }
1222                         f_err = f_out;
1223                         break;
1224                 case 'm':
1225                         terse_output = 1;
1226                         break;
1227                 case 'h':
1228                         usage(argv[0]);
1229                         exit(0);
1230                 case 'c':
1231                         exit(fio_show_option_help(optarg));
1232                 case 's':
1233                         dump_cmdline = 1;
1234                         break;
1235                 case 'r':
1236                         read_only = 1;
1237                         break;
1238                 case 'v':
1239                         log_info("%s\n", fio_version_string);
1240                         exit(0);
1241                 case 'V':
1242                         terse_version = atoi(optarg);
1243                         if (terse_version != 2) {
1244                                 log_err("fio: bad terse version format\n");
1245                                 exit_val = 1;
1246                                 do_exit++;
1247                         }
1248                         break;
1249                 case 'e':
1250                         if (!strcmp("always", optarg))
1251                                 eta_print = FIO_ETA_ALWAYS;
1252                         else if (!strcmp("never", optarg))
1253                                 eta_print = FIO_ETA_NEVER;
1254                         break;
1255                 case 'd':
1256                         if (set_debug(optarg))
1257                                 do_exit++;
1258                         break;
1259                 case 'x': {
1260                         size_t new_size;
1261
1262                         if (!strcmp(optarg, "global")) {
1263                                 log_err("fio: can't use global as only "
1264                                         "section\n");
1265                                 do_exit++;
1266                                 exit_val = 1;
1267                                 break;
1268                         }
1269                         new_size = (nr_job_sections + 1) * sizeof(char *);
1270                         job_sections = realloc(job_sections, new_size);
1271                         job_sections[nr_job_sections] = strdup(optarg);
1272                         nr_job_sections++;
1273                         break;
1274                         }
1275                 case 'p':
1276                         exec_profile = strdup(optarg);
1277                         break;
1278                 case FIO_GETOPT_JOB: {
1279                         const char *opt = l_opts[lidx].name;
1280                         char *val = optarg;
1281
1282                         if (!strncmp(opt, "name", 4) && td) {
1283                                 ret = add_job(td, td->o.name ?: "fio", 0);
1284                                 if (ret)
1285                                         return 0;
1286                                 td = NULL;
1287                         }
1288                         if (!td) {
1289                                 int is_section = !strncmp(opt, "name", 4);
1290                                 int global = 0;
1291
1292                                 if (!is_section || !strncmp(val, "global", 6))
1293                                         global = 1;
1294
1295                                 if (is_section && skip_this_section(val))
1296                                         continue;
1297
1298                                 td = get_new_job(global, &def_thread);
1299                                 if (!td)
1300                                         return 0;
1301                         }
1302
1303                         ret = fio_cmd_option_parse(td, opt, val);
1304                         break;
1305                 }
1306                 case 'w':
1307                         warnings_fatal = 1;
1308                         break;
1309                 case 'j':
1310                         max_jobs = atoi(optarg);
1311                         if (!max_jobs || max_jobs > REAL_MAX_JOBS) {
1312                                 log_err("fio: invalid max jobs: %d\n", max_jobs);
1313                                 do_exit++;
1314                                 exit_val = 1;
1315                         }
1316                         break;
1317                 case 'S':
1318                         if (nr_clients) {
1319                                 log_err("fio: can't be both client and server\n");
1320                                 do_exit++;
1321                                 exit_val = 1;
1322                                 break;
1323                         }
1324                         is_backend = 1;
1325                         break;
1326                 case 'D':
1327                         daemonize_server = 1;
1328                         break;
1329                 case 'P':
1330                         fio_net_port = atoi(optarg);
1331                         break;
1332                 case 'C':
1333                         if (is_backend) {
1334                                 log_err("fio: can't be both client and server\n");
1335                                 do_exit++;
1336                                 exit_val = 1;
1337                                 break;
1338                         }
1339                         fio_client_add(optarg);
1340                         break;
1341                 default:
1342                         do_exit++;
1343                         exit_val = 1;
1344                         break;
1345                 }
1346         }
1347
1348         if (do_exit)
1349                 exit(exit_val);
1350
1351         if (nr_clients && fio_clients_connect()) {
1352                 do_exit++;
1353                 exit_val = 1;
1354                 return 1;
1355         }
1356
1357         if (is_backend)
1358                 return fio_start_server(daemonize_server);
1359
1360         if (td) {
1361                 if (!ret)
1362                         ret = add_job(td, td->o.name ?: "fio", 0);
1363         }
1364
1365         while (optind < argc) {
1366                 ini_idx++;
1367                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1368                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1369                 optind++;
1370         }
1371
1372         return ini_idx;
1373 }
1374
1375 int parse_options(int argc, char *argv[])
1376 {
1377         int job_files, i;
1378
1379         f_out = stdout;
1380         f_err = stderr;
1381
1382         fio_options_fill_optstring();
1383         fio_options_dup_and_init(l_opts);
1384
1385         if (setup_thread_area())
1386                 return 1;
1387         if (fill_def_thread())
1388                 return 1;
1389
1390         job_files = parse_cmd_line(argc, argv);
1391
1392         for (i = 0; i < job_files; i++) {
1393                 if (fill_def_thread())
1394                         return 1;
1395                 if (nr_clients) {
1396                         if (fio_clients_send_ini(ini_file[i]))
1397                                 return 1;
1398                         free(ini_file[i]);
1399                 } else if (!is_backend) {
1400                         if (parse_jobs_ini(ini_file[i], 0, i))
1401                                 return 1;
1402                         free(ini_file[i]);
1403                 }
1404         }
1405
1406         free(ini_file);
1407         options_mem_free(&def_thread);
1408
1409         if (!thread_number) {
1410                 if (dump_cmdline)
1411                         return 0;
1412                 if (exec_profile)
1413                         return 0;
1414                 if (is_backend || nr_clients)
1415                         return 0;
1416
1417                 log_err("No jobs(s) defined\n\n");
1418                 usage(argv[0]);
1419                 return 1;
1420         }
1421
1422         if (def_thread.o.gtod_offload) {
1423                 fio_gtod_init();
1424                 fio_gtod_offload = 1;
1425                 fio_gtod_cpu = def_thread.o.gtod_cpu;
1426         }
1427
1428         log_info("%s\n", fio_version_string);
1429         return 0;
1430 }