Switch to using our internal Tausworthe based random generator for offsets
[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
23 #include "lib/getopt.h"
24
25 static char fio_version_string[] = "fio 1.51";
26
27 #define FIO_RANDSEED            (0xb1899bedUL)
28
29 static char **ini_file;
30 static int max_jobs = MAX_JOBS;
31 static int dump_cmdline;
32
33 static struct thread_data def_thread;
34 struct thread_data *threads = NULL;
35
36 int exitall_on_terminate = 0;
37 int terse_output = 0;
38 int eta_print;
39 unsigned long long mlock_size = 0;
40 FILE *f_out = NULL;
41 FILE *f_err = NULL;
42 char **job_sections = NULL;
43 int nr_job_sections = 0;
44 char *exec_profile = NULL;
45 int warnings_fatal = 0;
46
47 int write_bw_log = 0;
48 int read_only = 0;
49
50 static int def_timeout;
51 static int write_lat_log;
52
53 static int prev_group_jobs;
54
55 unsigned long fio_debug = 0;
56 unsigned int fio_debug_jobno = -1;
57 unsigned int *fio_debug_jobp = NULL;
58
59 static char cmd_optstr[256];
60
61 /*
62  * Command line options. These will contain the above, plus a few
63  * extra that only pertain to fio itself and not jobs.
64  */
65 static struct option l_opts[FIO_NR_OPTIONS] = {
66         {
67                 .name           = "output",
68                 .has_arg        = required_argument,
69                 .val            = 'o',
70         },
71         {
72                 .name           = "timeout",
73                 .has_arg        = required_argument,
74                 .val            = 't',
75         },
76         {
77                 .name           = "latency-log",
78                 .has_arg        = required_argument,
79                 .val            = 'l',
80         },
81         {
82                 .name           = "bandwidth-log",
83                 .has_arg        = required_argument,
84                 .val            = 'b',
85         },
86         {
87                 .name           = "minimal",
88                 .has_arg        = optional_argument,
89                 .val            = 'm',
90         },
91         {
92                 .name           = "version",
93                 .has_arg        = no_argument,
94                 .val            = 'v',
95         },
96         {
97                 .name           = "help",
98                 .has_arg        = no_argument,
99                 .val            = 'h',
100         },
101         {
102                 .name           = "cmdhelp",
103                 .has_arg        = optional_argument,
104                 .val            = 'c',
105         },
106         {
107                 .name           = "showcmd",
108                 .has_arg        = no_argument,
109                 .val            = 's',
110         },
111         {
112                 .name           = "readonly",
113                 .has_arg        = no_argument,
114                 .val            = 'r',
115         },
116         {
117                 .name           = "eta",
118                 .has_arg        = required_argument,
119                 .val            = 'e',
120         },
121         {
122                 .name           = "debug",
123                 .has_arg        = required_argument,
124                 .val            = 'd',
125         },
126         {
127                 .name           = "section",
128                 .has_arg        = required_argument,
129                 .val            = 'x',
130         },
131         {
132                 .name           = "alloc-size",
133                 .has_arg        = required_argument,
134                 .val            = 'a',
135         },
136         {
137                 .name           = "profile",
138                 .has_arg        = required_argument,
139                 .val            = 'p',
140         },
141         {
142                 .name           = "warnings-fatal",
143                 .has_arg        = no_argument,
144                 .val            = 'w',
145         },
146         {
147                 .name           = NULL,
148         },
149 };
150
151 FILE *get_f_out()
152 {
153         return f_out;
154 }
155
156 FILE *get_f_err()
157 {
158         return f_err;
159 }
160
161 /*
162  * Return a free job structure.
163  */
164 static struct thread_data *get_new_job(int global, struct thread_data *parent)
165 {
166         struct thread_data *td;
167
168         if (global)
169                 return &def_thread;
170         if (thread_number >= max_jobs) {
171                 log_err("error: maximum number of jobs (%d) reached.\n",
172                                 max_jobs);
173                 return NULL;
174         }
175
176         td = &threads[thread_number++];
177         *td = *parent;
178
179         td->o.uid = td->o.gid = -1U;
180
181         dup_files(td, parent);
182         options_mem_dupe(td);
183
184         profile_add_hooks(td);
185
186         td->thread_number = thread_number;
187         return td;
188 }
189
190 static void put_job(struct thread_data *td)
191 {
192         if (td == &def_thread)
193                 return;
194         
195         profile_td_exit(td);
196
197         if (td->error)
198                 log_info("fio: %s\n", td->verror);
199
200         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
201         thread_number--;
202 }
203
204 static int __setup_rate(struct thread_data *td, enum fio_ddir ddir)
205 {
206         unsigned int bs = td->o.min_bs[ddir];
207         unsigned long long bytes_per_sec;
208
209         assert(ddir_rw(ddir));
210
211         if (td->o.rate[ddir])
212                 bytes_per_sec = td->o.rate[ddir];
213         else
214                 bytes_per_sec = td->o.rate_iops[ddir] * bs;
215
216         if (!bytes_per_sec) {
217                 log_err("rate lower than supported\n");
218                 return -1;
219         }
220
221         td->rate_nsec_cycle[ddir] = 1000000000ULL / bytes_per_sec;
222         td->rate_pending_usleep[ddir] = 0;
223         return 0;
224 }
225
226 static int setup_rate(struct thread_data *td)
227 {
228         int ret = 0;
229
230         if (td->o.rate[DDIR_READ] || td->o.rate_iops[DDIR_READ])
231                 ret = __setup_rate(td, DDIR_READ);
232         if (td->o.rate[DDIR_WRITE] || td->o.rate_iops[DDIR_WRITE])
233                 ret |= __setup_rate(td, DDIR_WRITE);
234
235         return ret;
236 }
237
238 static int fixed_block_size(struct thread_options *o)
239 {
240         return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
241                 o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
242                 o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE];
243 }
244
245 /*
246  * Lazy way of fixing up options that depend on each other. We could also
247  * define option callback handlers, but this is easier.
248  */
249 static int fixup_options(struct thread_data *td)
250 {
251         struct thread_options *o = &td->o;
252         int ret = 0;
253
254 #ifndef FIO_HAVE_PSHARED_MUTEX
255         if (!o->use_thread) {
256                 log_info("fio: this platform does not support process shared"
257                          " mutexes, forcing use of threads. Use the 'thread'"
258                          " option to get rid of this warning.\n");
259                 o->use_thread = 1;
260                 ret = warnings_fatal;
261         }
262 #endif
263
264         if (o->write_iolog_file && o->read_iolog_file) {
265                 log_err("fio: read iolog overrides write_iolog\n");
266                 free(o->write_iolog_file);
267                 o->write_iolog_file = NULL;
268                 ret = warnings_fatal;
269         }
270
271         /*
272          * only really works for sequential io for now, and with 1 file
273          */
274         if (o->zone_size && td_random(td) && o->open_files == 1)
275                 o->zone_size = 0;
276
277         /*
278          * Reads can do overwrites, we always need to pre-create the file
279          */
280         if (td_read(td) || td_rw(td))
281                 o->overwrite = 1;
282
283         if (!o->min_bs[DDIR_READ])
284                 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
285         if (!o->max_bs[DDIR_READ])
286                 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
287         if (!o->min_bs[DDIR_WRITE])
288                 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
289         if (!o->max_bs[DDIR_WRITE])
290                 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
291
292         o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
293
294         /*
295          * For random IO, allow blockalign offset other than min_bs.
296          */
297         if (!o->ba[DDIR_READ] || !td_random(td))
298                 o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
299         if (!o->ba[DDIR_WRITE] || !td_random(td))
300                 o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
301
302         if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
303             o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
304             !o->norandommap) {
305                 log_err("fio: Any use of blockalign= turns off randommap\n");
306                 o->norandommap = 1;
307                 ret = warnings_fatal;
308         }
309
310         if (!o->file_size_high)
311                 o->file_size_high = o->file_size_low;
312
313         if (o->norandommap && o->verify != VERIFY_NONE
314             && !fixed_block_size(o))  {
315                 log_err("fio: norandommap given for variable block sizes, "
316                         "verify disabled\n");
317                 o->verify = VERIFY_NONE;
318                 ret = warnings_fatal;
319         }
320         if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
321                 log_err("fio: bs_unaligned may not work with raw io\n");
322
323         /*
324          * thinktime_spin must be less than thinktime
325          */
326         if (o->thinktime_spin > o->thinktime)
327                 o->thinktime_spin = o->thinktime;
328
329         /*
330          * The low water mark cannot be bigger than the iodepth
331          */
332         if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
333                 /*
334                  * syslet work around - if the workload is sequential,
335                  * we want to let the queue drain all the way down to
336                  * avoid seeking between async threads
337                  */
338                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
339                         o->iodepth_low = 1;
340                 else
341                         o->iodepth_low = o->iodepth;
342         }
343
344         /*
345          * If batch number isn't set, default to the same as iodepth
346          */
347         if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
348                 o->iodepth_batch = o->iodepth;
349
350         if (o->nr_files > td->files_index)
351                 o->nr_files = td->files_index;
352
353         if (o->open_files > o->nr_files || !o->open_files)
354                 o->open_files = o->nr_files;
355
356         if (((o->rate[0] + o->rate[1]) && (o->rate_iops[0] + o->rate_iops[1]))||
357             ((o->ratemin[0] + o->ratemin[1]) && (o->rate_iops_min[0] +
358                 o->rate_iops_min[1]))) {
359                 log_err("fio: rate and rate_iops are mutually exclusive\n");
360                 ret = 1;
361         }
362         if ((o->rate[0] < o->ratemin[0]) || (o->rate[1] < o->ratemin[1]) ||
363             (o->rate_iops[0] < o->rate_iops_min[0]) ||
364             (o->rate_iops[1] < o->rate_iops_min[1])) {
365                 log_err("fio: minimum rate exceeds rate\n");
366                 ret = 1;
367         }
368
369         if (!o->timeout && o->time_based) {
370                 log_err("fio: time_based requires a runtime/timeout setting\n");
371                 o->time_based = 0;
372                 ret = warnings_fatal;
373         }
374
375         if (o->fill_device && !o->size)
376                 o->size = -1ULL;
377
378         if (o->verify != VERIFY_NONE) {
379                 if (td_write(td) && o->do_verify && o->numjobs > 1) {
380                         log_info("Multiple writers may overwrite blocks that "
381                                 "belong to other jobs. This can cause "
382                                 "verification failures.\n");
383                         ret = warnings_fatal;
384                 }
385
386                 o->refill_buffers = 1;
387                 if (o->max_bs[DDIR_WRITE] != o->min_bs[DDIR_WRITE] &&
388                     !o->verify_interval)
389                         o->verify_interval = o->min_bs[DDIR_WRITE];
390         }
391
392         if (o->pre_read) {
393                 o->invalidate_cache = 0;
394                 if (td->io_ops->flags & FIO_PIPEIO) {
395                         log_info("fio: cannot pre-read files with an IO engine"
396                                  " that isn't seekable. Pre-read disabled.\n");
397                         ret = warnings_fatal;
398                 }
399         }
400
401 #ifndef FIO_HAVE_FDATASYNC
402         if (o->fdatasync_blocks) {
403                 log_info("fio: this platform does not support fdatasync()"
404                          " falling back to using fsync().  Use the 'fsync'"
405                          " option instead of 'fdatasync' to get rid of"
406                          " this warning\n");
407                 o->fsync_blocks = o->fdatasync_blocks;
408                 o->fdatasync_blocks = 0;
409                 ret = warnings_fatal;
410         }
411 #endif
412
413         return ret;
414 }
415
416 /*
417  * This function leaks the buffer
418  */
419 static char *to_kmg(unsigned int val)
420 {
421         char *buf = malloc(32);
422         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
423         char *p = post;
424
425         do {
426                 if (val & 1023)
427                         break;
428
429                 val >>= 10;
430                 p++;
431         } while (*p);
432
433         snprintf(buf, 31, "%u%c", val, *p);
434         return buf;
435 }
436
437 /* External engines are specified by "external:name.o") */
438 static const char *get_engine_name(const char *str)
439 {
440         char *p = strstr(str, ":");
441
442         if (!p)
443                 return str;
444
445         p++;
446         strip_blank_front(&p);
447         strip_blank_end(p);
448         return p;
449 }
450
451 static int exists_and_not_file(const char *filename)
452 {
453         struct stat sb;
454
455         if (lstat(filename, &sb) == -1)
456                 return 0;
457
458         /* \\.\ is the device namespace in Windows, where every file
459          * is a device node */
460         if (S_ISREG(sb.st_mode) && strncmp(filename, "\\\\.\\", 4) != 0)
461                 return 0;
462
463         return 1;
464 }
465
466 void td_fill_rand_seeds(struct thread_data *td)
467 {
468         os_random_seed(td->rand_seeds[0], &td->bsrange_state);
469         os_random_seed(td->rand_seeds[1], &td->verify_state);
470         os_random_seed(td->rand_seeds[2], &td->rwmix_state);
471
472         if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
473                 os_random_seed(td->rand_seeds[3], &td->next_file_state);
474
475         os_random_seed(td->rand_seeds[5], &td->file_size_state);
476         os_random_seed(td->rand_seeds[6], &td->trim_state);
477
478         if (!td_random(td))
479                 return;
480
481         if (td->o.rand_repeatable)
482                 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
483
484         os_random_seed(td->rand_seeds[4], &td->random_state);
485         init_rand_seed(&td->__random_state, td->rand_seeds[4]);
486 }
487
488 /*
489  * Initialize the various random states we need (random io, block size ranges,
490  * read/write mix, etc).
491  */
492 static int init_random_state(struct thread_data *td)
493 {
494         int fd;
495
496         fd = open("/dev/urandom", O_RDONLY);
497         if (fd == -1) {
498                 td_verror(td, errno, "open");
499                 return 1;
500         }
501
502         if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
503             (int) sizeof(td->rand_seeds)) {
504                 td_verror(td, EIO, "read");
505                 close(fd);
506                 return 1;
507         }
508
509         close(fd);
510         td_fill_rand_seeds(td);
511         return 0;
512 }
513
514 /*
515  * Adds a job to the list of things todo. Sanitizes the various options
516  * to make sure we don't have conflicts, and initializes various
517  * members of td.
518  */
519 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
520 {
521         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
522                                    "randread", "randwrite", "randrw" };
523         unsigned int i;
524         const char *engine;
525         char fname[PATH_MAX];
526         int numjobs, file_alloced;
527
528         /*
529          * the def_thread is just for options, it's not a real job
530          */
531         if (td == &def_thread)
532                 return 0;
533
534         /*
535          * if we are just dumping the output command line, don't add the job
536          */
537         if (dump_cmdline) {
538                 put_job(td);
539                 return 0;
540         }
541
542         if (profile_td_init(td))
543                 return 1;
544
545         engine = get_engine_name(td->o.ioengine);
546         td->io_ops = load_ioengine(td, engine);
547         if (!td->io_ops) {
548                 log_err("fio: failed to load engine %s\n", engine);
549                 goto err;
550         }
551
552         if (td->o.use_thread)
553                 nr_thread++;
554         else
555                 nr_process++;
556
557         if (td->o.odirect)
558                 td->io_ops->flags |= FIO_RAWIO;
559
560         file_alloced = 0;
561         if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
562                 file_alloced = 1;
563
564                 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
565                         add_file(td, jobname);
566                 else {
567                         for (i = 0; i < td->o.nr_files; i++) {
568                                 sprintf(fname, "%s.%d.%d", jobname,
569                                                         td->thread_number, i);
570                                 add_file(td, fname);
571                         }
572                 }
573         }
574
575         if (fixup_options(td))
576                 goto err;
577
578         if (td->io_ops->flags & FIO_DISKLESSIO) {
579                 struct fio_file *f;
580
581                 for_each_file(td, f, i)
582                         f->real_file_size = -1ULL;
583         }
584
585         td->mutex = fio_mutex_init(0);
586
587         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
588         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
589         td->ts.lat_stat[0].min_val = td->ts.lat_stat[1].min_val = ULONG_MAX;
590         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
591         td->ddir_seq_nr = td->o.ddir_seq_nr;
592
593         if ((td->o.stonewall || td->o.new_group) && prev_group_jobs) {
594                 prev_group_jobs = 0;
595                 groupid++;
596         }
597
598         td->groupid = groupid;
599         prev_group_jobs++;
600
601         if (init_random_state(td))
602                 goto err;
603
604         if (setup_rate(td))
605                 goto err;
606
607         if (td->o.write_lat_log) {
608                 setup_log(&td->ts.lat_log);
609                 setup_log(&td->ts.slat_log);
610                 setup_log(&td->ts.clat_log);
611         }
612         if (td->o.write_bw_log)
613                 setup_log(&td->ts.bw_log);
614
615         if (!td->o.name)
616                 td->o.name = strdup(jobname);
617
618         if (!terse_output) {
619                 if (!job_add_num) {
620                         if (!strcmp(td->io_ops->name, "cpuio")) {
621                                 log_info("%s: ioengine=cpu, cpuload=%u,"
622                                          " cpucycle=%u\n", td->o.name,
623                                                         td->o.cpuload,
624                                                         td->o.cpucycle);
625                         } else {
626                                 char *c1, *c2, *c3, *c4;
627
628                                 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
629                                 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
630                                 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
631                                 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
632
633                                 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
634                                          " ioengine=%s, iodepth=%u\n",
635                                                 td->o.name, td->groupid,
636                                                 ddir_str[td->o.td_ddir],
637                                                 c1, c2, c3, c4,
638                                                 td->io_ops->name,
639                                                 td->o.iodepth);
640
641                                 free(c1);
642                                 free(c2);
643                                 free(c3);
644                                 free(c4);
645                         }
646                 } else if (job_add_num == 1)
647                         log_info("...\n");
648         }
649
650         /*
651          * recurse add identical jobs, clear numjobs and stonewall options
652          * as they don't apply to sub-jobs
653          */
654         numjobs = td->o.numjobs;
655         while (--numjobs) {
656                 struct thread_data *td_new = get_new_job(0, td);
657
658                 if (!td_new)
659                         goto err;
660
661                 td_new->o.numjobs = 1;
662                 td_new->o.stonewall = 0;
663                 td_new->o.new_group = 0;
664
665                 if (file_alloced) {
666                         td_new->o.filename = NULL;
667                         td_new->files_index = 0;
668                         td_new->files_size = 0;
669                         td_new->files = NULL;
670                 }
671
672                 job_add_num = numjobs - 1;
673
674                 if (add_job(td_new, jobname, job_add_num))
675                         goto err;
676         }
677
678         return 0;
679 err:
680         put_job(td);
681         return -1;
682 }
683
684 /*
685  * Parse as if 'o' was a command line
686  */
687 void add_job_opts(const char **o)
688 {
689         struct thread_data *td, *td_parent;
690         int i, in_global = 1;
691         char jobname[32];
692
693         i = 0;
694         td_parent = td = NULL;
695         while (o[i]) {
696                 if (!strncmp(o[i], "name", 4)) {
697                         in_global = 0;
698                         if (td)
699                                 add_job(td, jobname, 0);
700                         td = NULL;
701                         sprintf(jobname, "%s", o[i] + 5);
702                 }
703                 if (in_global && !td_parent)
704                         td_parent = get_new_job(1, &def_thread);
705                 else if (!in_global && !td) {
706                         if (!td_parent)
707                                 td_parent = &def_thread;
708                         td = get_new_job(0, td_parent);
709                 }
710                 if (in_global)
711                         fio_options_parse(td_parent, (char **) &o[i], 1);
712                 else
713                         fio_options_parse(td, (char **) &o[i], 1);
714                 i++;
715         }
716
717         if (td)
718                 add_job(td, jobname, 0);
719 }
720
721 static int skip_this_section(const char *name)
722 {
723         int i;
724
725         if (!nr_job_sections)
726                 return 0;
727         if (!strncmp(name, "global", 6))
728                 return 0;
729
730         for (i = 0; i < nr_job_sections; i++)
731                 if (!strcmp(job_sections[i], name))
732                         return 0;
733
734         return 1;
735 }
736
737 static int is_empty_or_comment(char *line)
738 {
739         unsigned int i;
740
741         for (i = 0; i < strlen(line); i++) {
742                 if (line[i] == ';')
743                         return 1;
744                 if (line[i] == '#')
745                         return 1;
746                 if (!isspace(line[i]) && !iscntrl(line[i]))
747                         return 0;
748         }
749
750         return 1;
751 }
752
753 /*
754  * This is our [ini] type file parser.
755  */
756 static int parse_jobs_ini(char *file, int stonewall_flag)
757 {
758         unsigned int global;
759         struct thread_data *td;
760         char *string, *name;
761         FILE *f;
762         char *p;
763         int ret = 0, stonewall;
764         int first_sect = 1;
765         int skip_fgets = 0;
766         int inside_skip = 0;
767         char **opts;
768         int i, alloc_opts, num_opts;
769
770         if (!strcmp(file, "-"))
771                 f = stdin;
772         else
773                 f = fopen(file, "r");
774
775         if (!f) {
776                 perror("fopen job file");
777                 return 1;
778         }
779
780         string = malloc(4096);
781
782         /*
783          * it's really 256 + small bit, 280 should suffice
784          */
785         name = malloc(280);
786         memset(name, 0, 280);
787
788         alloc_opts = 8;
789         opts = malloc(sizeof(char *) * alloc_opts);
790         num_opts = 0;
791
792         stonewall = stonewall_flag;
793         do {
794                 /*
795                  * if skip_fgets is set, we already have loaded a line we
796                  * haven't handled.
797                  */
798                 if (!skip_fgets) {
799                         p = fgets(string, 4095, f);
800                         if (!p)
801                                 break;
802                 }
803
804                 skip_fgets = 0;
805                 strip_blank_front(&p);
806                 strip_blank_end(p);
807
808                 if (is_empty_or_comment(p))
809                         continue;
810                 if (sscanf(p, "[%255s]", name) != 1) {
811                         if (inside_skip)
812                                 continue;
813                         log_err("fio: option <%s> outside of [] job section\n",
814                                                                         p);
815                         break;
816                 }
817
818                 name[strlen(name) - 1] = '\0';
819
820                 if (skip_this_section(name)) {
821                         inside_skip = 1;
822                         continue;
823                 } else
824                         inside_skip = 0;
825
826                 global = !strncmp(name, "global", 6);
827
828                 if (dump_cmdline) {
829                         if (first_sect)
830                                 log_info("fio ");
831                         if (!global)
832                                 log_info("--name=%s ", name);
833                         first_sect = 0;
834                 }
835
836                 td = get_new_job(global, &def_thread);
837                 if (!td) {
838                         ret = 1;
839                         break;
840                 }
841
842                 /*
843                  * Seperate multiple job files by a stonewall
844                  */
845                 if (!global && stonewall) {
846                         td->o.stonewall = stonewall;
847                         stonewall = 0;
848                 }
849
850                 num_opts = 0;
851                 memset(opts, 0, alloc_opts * sizeof(char *));
852
853                 while ((p = fgets(string, 4096, f)) != NULL) {
854                         if (is_empty_or_comment(p))
855                                 continue;
856
857                         strip_blank_front(&p);
858
859                         /*
860                          * new section, break out and make sure we don't
861                          * fgets() a new line at the top.
862                          */
863                         if (p[0] == '[') {
864                                 skip_fgets = 1;
865                                 break;
866                         }
867
868                         strip_blank_end(p);
869
870                         if (num_opts == alloc_opts) {
871                                 alloc_opts <<= 1;
872                                 opts = realloc(opts,
873                                                 alloc_opts * sizeof(char *));
874                         }
875
876                         opts[num_opts] = strdup(p);
877                         num_opts++;
878                 }
879
880                 ret = fio_options_parse(td, opts, num_opts);
881                 if (!ret) {
882                         if (dump_cmdline)
883                                 for (i = 0; i < num_opts; i++)
884                                         log_info("--%s ", opts[i]);
885
886                         ret = add_job(td, name, 0);
887                 } else {
888                         log_err("fio: job %s dropped\n", name);
889                         put_job(td);
890                 }
891
892                 for (i = 0; i < num_opts; i++)
893                         free(opts[i]);
894                 num_opts = 0;
895         } while (!ret);
896
897         if (dump_cmdline)
898                 log_info("\n");
899
900         for (i = 0; i < num_opts; i++)
901                 free(opts[i]);
902
903         free(string);
904         free(name);
905         free(opts);
906         if (f != stdin)
907                 fclose(f);
908         return ret;
909 }
910
911 static int fill_def_thread(void)
912 {
913         memset(&def_thread, 0, sizeof(def_thread));
914
915         fio_getaffinity(getpid(), &def_thread.o.cpumask);
916
917         /*
918          * fill default options
919          */
920         fio_fill_default_options(&def_thread);
921
922         def_thread.o.timeout = def_timeout;
923         return 0;
924 }
925
926 static void free_shm(void)
927 {
928         struct shmid_ds sbuf;
929
930         if (threads) {
931                 void *tp = threads;
932
933                 threads = NULL;
934                 file_hash_exit();
935                 fio_debug_jobp = NULL;
936                 shmdt(tp);
937                 shmctl(shm_id, IPC_RMID, &sbuf);
938         }
939
940         scleanup();
941 }
942
943 /*
944  * The thread area is shared between the main process and the job
945  * threads/processes. So setup a shared memory segment that will hold
946  * all the job info. We use the end of the region for keeping track of
947  * open files across jobs, for file sharing.
948  */
949 static int setup_thread_area(void)
950 {
951         void *hash;
952
953         /*
954          * 1024 is too much on some machines, scale max_jobs if
955          * we get a failure that looks like too large a shm segment
956          */
957         do {
958                 size_t size = max_jobs * sizeof(struct thread_data);
959
960                 size += file_hash_size;
961                 size += sizeof(unsigned int);
962
963                 shm_id = shmget(0, size, IPC_CREAT | 0600);
964                 if (shm_id != -1)
965                         break;
966                 if (errno != EINVAL) {
967                         perror("shmget");
968                         break;
969                 }
970
971                 max_jobs >>= 1;
972         } while (max_jobs);
973
974         if (shm_id == -1)
975                 return 1;
976
977         threads = shmat(shm_id, NULL, 0);
978         if (threads == (void *) -1) {
979                 perror("shmat");
980                 return 1;
981         }
982
983         memset(threads, 0, max_jobs * sizeof(struct thread_data));
984         hash = (void *) threads + max_jobs * sizeof(struct thread_data);
985         fio_debug_jobp = (void *) hash + file_hash_size;
986         *fio_debug_jobp = -1;
987         file_hash_init(hash);
988         atexit(free_shm);
989         return 0;
990 }
991
992 static void usage(const char *name)
993 {
994         printf("%s [options] [job options] <job file(s)>\n", name);
995         printf("\t--debug=options\tEnable debug logging\n");
996         printf("\t--output\tWrite output to file\n");
997         printf("\t--timeout\tRuntime in seconds\n");
998         printf("\t--latency-log\tGenerate per-job latency logs\n");
999         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1000         printf("\t--minimal\tMinimal (terse) output\n");
1001         printf("\t--version\tPrint version info and exit\n");
1002         printf("\t--help\t\tPrint this page\n");
1003         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
1004                 " them\n");
1005         printf("\t--showcmd\tTurn a job file into command line options\n");
1006         printf("\t--eta=when\tWhen ETA estimate should be printed\n");
1007         printf("\t          \tMay be \"always\", \"never\" or \"auto\"\n");
1008         printf("\t--readonly\tTurn on safety read-only checks, preventing"
1009                 " writes\n");
1010         printf("\t--section=name\tOnly run specified section in job file\n");
1011         printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
1012                 " (def 1024)\n");
1013         printf("\t--warnings-fatal Fio parser warnings are fatal\n");
1014         printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>");
1015         printf("\n                   Jens Axboe <jaxboe@fusionio.com>\n");
1016 }
1017
1018 #ifdef FIO_INC_DEBUG
1019 struct debug_level debug_levels[] = {
1020         { .name = "process",    .shift = FD_PROCESS, },
1021         { .name = "file",       .shift = FD_FILE, },
1022         { .name = "io",         .shift = FD_IO, },
1023         { .name = "mem",        .shift = FD_MEM, },
1024         { .name = "blktrace",   .shift = FD_BLKTRACE },
1025         { .name = "verify",     .shift = FD_VERIFY },
1026         { .name = "random",     .shift = FD_RANDOM },
1027         { .name = "parse",      .shift = FD_PARSE },
1028         { .name = "diskutil",   .shift = FD_DISKUTIL },
1029         { .name = "job",        .shift = FD_JOB },
1030         { .name = "mutex",      .shift = FD_MUTEX },
1031         { .name = "profile",    .shift = FD_PROFILE },
1032         { .name = "time",       .shift = FD_TIME },
1033         { .name = NULL, },
1034 };
1035
1036 static int set_debug(const char *string)
1037 {
1038         struct debug_level *dl;
1039         char *p = (char *) string;
1040         char *opt;
1041         int i;
1042
1043         if (!strcmp(string, "?") || !strcmp(string, "help")) {
1044                 log_info("fio: dumping debug options:");
1045                 for (i = 0; debug_levels[i].name; i++) {
1046                         dl = &debug_levels[i];
1047                         log_info("%s,", dl->name);
1048                 }
1049                 log_info("all\n");
1050                 return 1;
1051         }
1052
1053         while ((opt = strsep(&p, ",")) != NULL) {
1054                 int found = 0;
1055
1056                 if (!strncmp(opt, "all", 3)) {
1057                         log_info("fio: set all debug options\n");
1058                         fio_debug = ~0UL;
1059                         continue;
1060                 }
1061
1062                 for (i = 0; debug_levels[i].name; i++) {
1063                         dl = &debug_levels[i];
1064                         found = !strncmp(opt, dl->name, strlen(dl->name));
1065                         if (!found)
1066                                 continue;
1067
1068                         if (dl->shift == FD_JOB) {
1069                                 opt = strchr(opt, ':');
1070                                 if (!opt) {
1071                                         log_err("fio: missing job number\n");
1072                                         break;
1073                                 }
1074                                 opt++;
1075                                 fio_debug_jobno = atoi(opt);
1076                                 log_info("fio: set debug jobno %d\n",
1077                                                         fio_debug_jobno);
1078                         } else {
1079                                 log_info("fio: set debug option %s\n", opt);
1080                                 fio_debug |= (1UL << dl->shift);
1081                         }
1082                         break;
1083                 }
1084
1085                 if (!found)
1086                         log_err("fio: debug mask %s not found\n", opt);
1087         }
1088         return 0;
1089 }
1090 #else
1091 static int set_debug(const char *string)
1092 {
1093         log_err("fio: debug tracing not included in build\n");
1094         return 1;
1095 }
1096 #endif
1097
1098 static void fio_options_fill_optstring(void)
1099 {
1100         char *ostr = cmd_optstr;
1101         int i, c;
1102
1103         c = i = 0;
1104         while (l_opts[i].name) {
1105                 ostr[c++] = l_opts[i].val;
1106                 if (l_opts[i].has_arg == required_argument)
1107                         ostr[c++] = ':';
1108                 else if (l_opts[i].has_arg == optional_argument) {
1109                         ostr[c++] = ':';
1110                         ostr[c++] = ':';
1111                 }
1112                 i++;
1113         }
1114         ostr[c] = '\0';
1115 }
1116
1117 static int parse_cmd_line(int argc, char *argv[])
1118 {
1119         struct thread_data *td = NULL;
1120         int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1121         char *ostr = cmd_optstr;
1122
1123         while ((c = getopt_long_only(argc, argv, ostr, l_opts, &lidx)) != -1) {
1124                 switch (c) {
1125                 case 'a':
1126                         smalloc_pool_size = atoi(optarg);
1127                         break;
1128                 case 't':
1129                         def_timeout = atoi(optarg);
1130                         break;
1131                 case 'l':
1132                         write_lat_log = 1;
1133                         break;
1134                 case 'b':
1135                         write_bw_log = 1;
1136                         break;
1137                 case 'o':
1138                         f_out = fopen(optarg, "w+");
1139                         if (!f_out) {
1140                                 perror("fopen output");
1141                                 exit(1);
1142                         }
1143                         f_err = f_out;
1144                         break;
1145                 case 'm':
1146                         terse_output = 1;
1147                         break;
1148                 case 'h':
1149                         usage(argv[0]);
1150                         exit(0);
1151                 case 'c':
1152                         exit(fio_show_option_help(optarg));
1153                 case 's':
1154                         dump_cmdline = 1;
1155                         break;
1156                 case 'r':
1157                         read_only = 1;
1158                         break;
1159                 case 'v':
1160                         log_info("%s\n", fio_version_string);
1161                         exit(0);
1162                 case 'e':
1163                         if (!strcmp("always", optarg))
1164                                 eta_print = FIO_ETA_ALWAYS;
1165                         else if (!strcmp("never", optarg))
1166                                 eta_print = FIO_ETA_NEVER;
1167                         break;
1168                 case 'd':
1169                         if (set_debug(optarg))
1170                                 do_exit++;
1171                         break;
1172                 case 'x': {
1173                         size_t new_size;
1174
1175                         if (!strcmp(optarg, "global")) {
1176                                 log_err("fio: can't use global as only "
1177                                         "section\n");
1178                                 do_exit++;
1179                                 exit_val = 1;
1180                                 break;
1181                         }
1182                         new_size = (nr_job_sections + 1) * sizeof(char *);
1183                         job_sections = realloc(job_sections, new_size);
1184                         job_sections[nr_job_sections] = strdup(optarg);
1185                         nr_job_sections++;
1186                         break;
1187                         }
1188                 case 'p':
1189                         exec_profile = strdup(optarg);
1190                         break;
1191                 case FIO_GETOPT_JOB: {
1192                         const char *opt = l_opts[lidx].name;
1193                         char *val = optarg;
1194
1195                         if (!strncmp(opt, "name", 4) && td) {
1196                                 ret = add_job(td, td->o.name ?: "fio", 0);
1197                                 if (ret) {
1198                                         put_job(td);
1199                                         return 0;
1200                                 }
1201                                 td = NULL;
1202                         }
1203                         if (!td) {
1204                                 int is_section = !strncmp(opt, "name", 4);
1205                                 int global = 0;
1206
1207                                 if (!is_section || !strncmp(val, "global", 6))
1208                                         global = 1;
1209
1210                                 if (is_section && skip_this_section(val))
1211                                         continue;
1212
1213                                 td = get_new_job(global, &def_thread);
1214                                 if (!td)
1215                                         return 0;
1216                         }
1217
1218                         ret = fio_cmd_option_parse(td, opt, val);
1219                         break;
1220                 }
1221                 case 'w':
1222                         warnings_fatal = 1;
1223                         break;
1224                 default:
1225                         do_exit++;
1226                         exit_val = 1;
1227                         break;
1228                 }
1229         }
1230
1231         if (do_exit)
1232                 exit(exit_val);
1233
1234         if (td) {
1235                 if (!ret)
1236                         ret = add_job(td, td->o.name ?: "fio", 0);
1237                 if (ret)
1238                         put_job(td);
1239         }
1240
1241         while (optind < argc) {
1242                 ini_idx++;
1243                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1244                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1245                 optind++;
1246         }
1247
1248         return ini_idx;
1249 }
1250
1251 int parse_options(int argc, char *argv[])
1252 {
1253         int job_files, i;
1254
1255         f_out = stdout;
1256         f_err = stderr;
1257
1258         fio_options_fill_optstring();
1259         fio_options_dup_and_init(l_opts);
1260
1261         if (setup_thread_area())
1262                 return 1;
1263         if (fill_def_thread())
1264                 return 1;
1265
1266         job_files = parse_cmd_line(argc, argv);
1267
1268         for (i = 0; i < job_files; i++) {
1269                 if (fill_def_thread())
1270                         return 1;
1271                 if (parse_jobs_ini(ini_file[i], i))
1272                         return 1;
1273                 free(ini_file[i]);
1274         }
1275
1276         free(ini_file);
1277         options_mem_free(&def_thread);
1278
1279         if (!thread_number) {
1280                 if (dump_cmdline)
1281                         return 0;
1282                 if (exec_profile)
1283                         return 0;
1284
1285                 log_err("No jobs(s) defined\n\n");
1286                 usage(argv[0]);
1287                 return 1;
1288         }
1289
1290         if (def_thread.o.gtod_offload) {
1291                 fio_gtod_init();
1292                 fio_gtod_offload = 1;
1293                 fio_gtod_cpu = def_thread.o.gtod_cpu;
1294         }
1295
1296         log_info("%s\n", fio_version_string);
1297         return 0;
1298 }