Only flag sync_file_range() support for newer Linux versions
[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 <getopt.h>
12 #include <sys/ipc.h>
13 #include <sys/shm.h>
14 #include <sys/types.h>
15 #include <sys/stat.h>
16
17 #include "fio.h"
18 #include "parse.h"
19 #include "smalloc.h"
20 #include "filehash.h"
21 #include "verify.h"
22 #include "profile.h"
23
24 static char fio_version_string[] = "fio 1.41.5";
25
26 #define FIO_RANDSEED            (0xb1899bedUL)
27
28 static char **ini_file;
29 static int max_jobs = MAX_JOBS;
30 static int dump_cmdline;
31
32 static struct thread_data def_thread;
33 struct thread_data *threads = NULL;
34
35 int exitall_on_terminate = 0;
36 int terse_output = 0;
37 int eta_print;
38 unsigned long long mlock_size = 0;
39 FILE *f_out = NULL;
40 FILE *f_err = NULL;
41 char *job_section = NULL;
42 char *exec_profile = NULL;
43
44 int write_bw_log = 0;
45 int read_only = 0;
46
47 static int def_timeout;
48 static int write_lat_log;
49
50 static int prev_group_jobs;
51
52 unsigned long fio_debug = 0;
53 unsigned int fio_debug_jobno = -1;
54 unsigned int *fio_debug_jobp = NULL;
55
56 /*
57  * Command line options. These will contain the above, plus a few
58  * extra that only pertain to fio itself and not jobs.
59  */
60 static struct option l_opts[FIO_NR_OPTIONS] = {
61         {
62                 .name           = "output",
63                 .has_arg        = required_argument,
64                 .val            = 'o',
65         },
66         {
67                 .name           = "timeout",
68                 .has_arg        = required_argument,
69                 .val            = 't',
70         },
71         {
72                 .name           = "latency-log",
73                 .has_arg        = required_argument,
74                 .val            = 'l',
75         },
76         {
77                 .name           = "bandwidth-log",
78                 .has_arg        = required_argument,
79                 .val            = 'b',
80         },
81         {
82                 .name           = "minimal",
83                 .has_arg        = optional_argument,
84                 .val            = 'm',
85         },
86         {
87                 .name           = "version",
88                 .has_arg        = no_argument,
89                 .val            = 'v',
90         },
91         {
92                 .name           = "help",
93                 .has_arg        = no_argument,
94                 .val            = 'h',
95         },
96         {
97                 .name           = "cmdhelp",
98                 .has_arg        = optional_argument,
99                 .val            = 'c',
100         },
101         {
102                 .name           = "showcmd",
103                 .has_arg        = no_argument,
104                 .val            = 's',
105         },
106         {
107                 .name           = "readonly",
108                 .has_arg        = no_argument,
109                 .val            = 'r',
110         },
111         {
112                 .name           = "eta",
113                 .has_arg        = required_argument,
114                 .val            = 'e',
115         },
116         {
117                 .name           = "debug",
118                 .has_arg        = required_argument,
119                 .val            = 'd',
120         },
121         {
122                 .name           = "section",
123                 .has_arg        = required_argument,
124                 .val            = 'x',
125         },
126         {
127                 .name           = "alloc-size",
128                 .has_arg        = required_argument,
129                 .val            = 'a',
130         },
131         {
132                 .name           = "profile",
133                 .has_arg        = required_argument,
134                 .val            = 'p',
135         },
136         {
137                 .name           = NULL,
138         },
139 };
140
141 FILE *get_f_out()
142 {
143         return f_out;
144 }
145
146 FILE *get_f_err()
147 {
148         return f_err;
149 }
150
151 /*
152  * Return a free job structure.
153  */
154 static struct thread_data *get_new_job(int global, struct thread_data *parent)
155 {
156         struct thread_data *td;
157
158         if (global)
159                 return &def_thread;
160         if (thread_number >= max_jobs)
161                 return NULL;
162
163         td = &threads[thread_number++];
164         *td = *parent;
165
166         td->o.uid = td->o.gid = -1U;
167
168         dup_files(td, parent);
169         options_mem_dupe(td);
170
171         profile_add_hooks(td);
172
173         td->thread_number = thread_number;
174         return td;
175 }
176
177 static void put_job(struct thread_data *td)
178 {
179         if (td == &def_thread)
180                 return;
181         
182         profile_td_exit(td);
183
184         if (td->error)
185                 log_info("fio: %s\n", td->verror);
186
187         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
188         thread_number--;
189 }
190
191 static int __setup_rate(struct thread_data *td, enum fio_ddir ddir)
192 {
193         unsigned int bs = td->o.min_bs[ddir];
194         unsigned long long bytes_per_sec;
195
196         if (td->o.rate[ddir])
197                 bytes_per_sec = td->o.rate[ddir];
198         else
199                 bytes_per_sec = td->o.rate_iops[ddir] * bs;
200
201         if (!bytes_per_sec) {
202                 log_err("rate lower than supported\n");
203                 return -1;
204         }
205
206         td->rate_nsec_cycle[ddir] = 1000000000ULL / bytes_per_sec;
207         td->rate_pending_usleep[ddir] = 0;
208         return 0;
209 }
210
211 static int setup_rate(struct thread_data *td)
212 {
213         int ret = 0;
214
215         if (td->o.rate[DDIR_READ] || td->o.rate_iops[DDIR_READ])
216                 ret = __setup_rate(td, DDIR_READ);
217         if (td->o.rate[DDIR_WRITE] || td->o.rate_iops[DDIR_WRITE])
218                 ret |= __setup_rate(td, DDIR_WRITE);
219
220         return ret;
221 }
222
223 static int fixed_block_size(struct thread_options *o)
224 {
225         return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
226                 o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
227                 o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE];
228 }
229
230 /*
231  * Lazy way of fixing up options that depend on each other. We could also
232  * define option callback handlers, but this is easier.
233  */
234 static int fixup_options(struct thread_data *td)
235 {
236         struct thread_options *o = &td->o;
237
238 #ifndef FIO_HAVE_PSHARED_MUTEX
239         if (!o->use_thread) {
240                 log_info("fio: this platform does not support process shared"
241                          " mutexes, forcing use of threads. Use the 'thread'"
242                          " option to get rid of this warning.\n");
243                 o->use_thread = 1;
244         }
245 #endif
246
247         if (o->write_iolog_file && o->read_iolog_file) {
248                 log_err("fio: read iolog overrides write_iolog\n");
249                 free(o->write_iolog_file);
250                 o->write_iolog_file = NULL;
251         }
252
253         /*
254          * only really works for sequential io for now, and with 1 file
255          */
256         if (o->zone_size && td_random(td) && o->open_files == 1)
257                 o->zone_size = 0;
258
259         /*
260          * Reads can do overwrites, we always need to pre-create the file
261          */
262         if (td_read(td) || td_rw(td))
263                 o->overwrite = 1;
264
265         if (!o->min_bs[DDIR_READ])
266                 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
267         if (!o->max_bs[DDIR_READ])
268                 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
269         if (!o->min_bs[DDIR_WRITE])
270                 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
271         if (!o->max_bs[DDIR_WRITE])
272                 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
273
274         o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
275
276         /*
277          * For random IO, allow blockalign offset other than min_bs.
278          */
279         if (!o->ba[DDIR_READ] || !td_random(td))
280                 o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
281         if (!o->ba[DDIR_WRITE] || !td_random(td))
282                 o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
283
284         if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
285             o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
286             !o->norandommap) {
287                 log_err("fio: Any use of blockalign= turns off randommap\n");
288                 o->norandommap = 1;
289         }
290
291         if (!o->file_size_high)
292                 o->file_size_high = o->file_size_low;
293
294         if (o->norandommap && o->verify != VERIFY_NONE
295             && !fixed_block_size(o))  {
296                 log_err("fio: norandommap given for variable block sizes, "
297                         "verify disabled\n");
298                 o->verify = VERIFY_NONE;
299         }
300         if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
301                 log_err("fio: bs_unaligned may not work with raw io\n");
302
303         /*
304          * thinktime_spin must be less than thinktime
305          */
306         if (o->thinktime_spin > o->thinktime)
307                 o->thinktime_spin = o->thinktime;
308
309         /*
310          * The low water mark cannot be bigger than the iodepth
311          */
312         if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
313                 /*
314                  * syslet work around - if the workload is sequential,
315                  * we want to let the queue drain all the way down to
316                  * avoid seeking between async threads
317                  */
318                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
319                         o->iodepth_low = 1;
320                 else
321                         o->iodepth_low = o->iodepth;
322         }
323
324         /*
325          * If batch number isn't set, default to the same as iodepth
326          */
327         if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
328                 o->iodepth_batch = o->iodepth;
329
330         if (o->nr_files > td->files_index)
331                 o->nr_files = td->files_index;
332
333         if (o->open_files > o->nr_files || !o->open_files)
334                 o->open_files = o->nr_files;
335
336         if (((o->rate[0] + o->rate[1]) && (o->rate_iops[0] + o->rate_iops[1]))||
337             ((o->ratemin[0] + o->ratemin[1]) && (o->rate_iops_min[0] +
338                 o->rate_iops_min[1]))) {
339                 log_err("fio: rate and rate_iops are mutually exclusive\n");
340                 return 1;
341         }
342         if ((o->rate[0] < o->ratemin[0]) || (o->rate[1] < o->ratemin[1]) ||
343             (o->rate_iops[0] < o->rate_iops_min[0]) ||
344             (o->rate_iops[1] < o->rate_iops_min[1])) {
345                 log_err("fio: minimum rate exceeds rate\n");
346                 return 1;
347         }
348
349         if (!o->timeout && o->time_based) {
350                 log_err("fio: time_based requires a runtime/timeout setting\n");
351                 o->time_based = 0;
352         }
353
354         if (o->fill_device && !o->size)
355                 o->size = -1ULL;
356
357         if (td_rw(td) && o->verify != VERIFY_NONE)
358                 log_info("fio: mixed read/write workload with verify. May not "
359                  "work as expected, unless you pre-populated the file\n");
360
361         if (o->verify != VERIFY_NONE) {
362                 o->refill_buffers = 1;
363                 if (o->max_bs[DDIR_WRITE] != o->min_bs[DDIR_WRITE] &&
364                     !o->verify_interval)
365                         o->verify_interval = o->min_bs[DDIR_WRITE];
366         }
367
368         if (o->pre_read) {
369                 o->invalidate_cache = 0;
370                 if (td->io_ops->flags & FIO_PIPEIO)
371                         log_info("fio: cannot pre-read files with an IO engine"
372                                  " that isn't seekable. Pre-read disabled.\n");
373         }
374
375 #ifndef FIO_HAVE_FDATASYNC
376         if (o->fdatasync_blocks) {
377                 log_info("fio: this platform does not support fdatasync()"
378                          " falling back to using fsync().  Use the 'fsync'"
379                          " option instead of 'fdatasync' to get rid of"
380                          " this warning\n");
381                 o->fsync_blocks = o->fdatasync_blocks;
382                 o->fdatasync_blocks = 0;
383         }
384 #endif
385
386         return 0;
387 }
388
389 /*
390  * This function leaks the buffer
391  */
392 static char *to_kmg(unsigned int val)
393 {
394         char *buf = malloc(32);
395         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
396         char *p = post;
397
398         do {
399                 if (val & 1023)
400                         break;
401
402                 val >>= 10;
403                 p++;
404         } while (*p);
405
406         snprintf(buf, 31, "%u%c", val, *p);
407         return buf;
408 }
409
410 /* External engines are specified by "external:name.o") */
411 static const char *get_engine_name(const char *str)
412 {
413         char *p = strstr(str, ":");
414
415         if (!p)
416                 return str;
417
418         p++;
419         strip_blank_front(&p);
420         strip_blank_end(p);
421         return p;
422 }
423
424 static int exists_and_not_file(const char *filename)
425 {
426         struct stat sb;
427
428         if (lstat(filename, &sb) == -1)
429                 return 0;
430
431         if (S_ISREG(sb.st_mode))
432                 return 0;
433
434         return 1;
435 }
436
437 void td_fill_rand_seeds(struct thread_data *td)
438 {
439         os_random_seed(td->rand_seeds[0], &td->bsrange_state);
440         os_random_seed(td->rand_seeds[1], &td->verify_state);
441         os_random_seed(td->rand_seeds[2], &td->rwmix_state);
442
443         if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
444                 os_random_seed(td->rand_seeds[3], &td->next_file_state);
445
446         os_random_seed(td->rand_seeds[5], &td->file_size_state);
447
448         if (!td_random(td))
449                 return;
450
451         if (td->o.rand_repeatable)
452                 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
453
454         os_random_seed(td->rand_seeds[4], &td->random_state);
455 }
456
457 /*
458  * Initialize the various random states we need (random io, block size ranges,
459  * read/write mix, etc).
460  */
461 static int init_random_state(struct thread_data *td)
462 {
463         int fd;
464
465         fd = open("/dev/urandom", O_RDONLY);
466         if (fd == -1) {
467                 td_verror(td, errno, "open");
468                 return 1;
469         }
470
471         if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
472             (int) sizeof(td->rand_seeds)) {
473                 td_verror(td, EIO, "read");
474                 close(fd);
475                 return 1;
476         }
477
478         close(fd);
479         td_fill_rand_seeds(td);
480         return 0;
481 }
482
483 /*
484  * Adds a job to the list of things todo. Sanitizes the various options
485  * to make sure we don't have conflicts, and initializes various
486  * members of td.
487  */
488 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
489 {
490         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
491                                    "randread", "randwrite", "randrw" };
492         unsigned int i;
493         const char *engine;
494         char fname[PATH_MAX];
495         int numjobs, file_alloced;
496
497         /*
498          * the def_thread is just for options, it's not a real job
499          */
500         if (td == &def_thread)
501                 return 0;
502
503         /*
504          * if we are just dumping the output command line, don't add the job
505          */
506         if (dump_cmdline) {
507                 put_job(td);
508                 return 0;
509         }
510
511         if (profile_td_init(td))
512                 return 1;
513
514         engine = get_engine_name(td->o.ioengine);
515         td->io_ops = load_ioengine(td, engine);
516         if (!td->io_ops) {
517                 log_err("fio: failed to load engine %s\n", engine);
518                 goto err;
519         }
520
521         if (td->o.use_thread)
522                 nr_thread++;
523         else
524                 nr_process++;
525
526         if (td->o.odirect)
527                 td->io_ops->flags |= FIO_RAWIO;
528
529         file_alloced = 0;
530         if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
531                 file_alloced = 1;
532
533                 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
534                         add_file(td, jobname);
535                 else {
536                         for (i = 0; i < td->o.nr_files; i++) {
537                                 sprintf(fname, "%s.%d.%d", jobname,
538                                                         td->thread_number, i);
539                                 add_file(td, fname);
540                         }
541                 }
542         }
543
544         if (fixup_options(td))
545                 goto err;
546
547         if (td->io_ops->flags & FIO_DISKLESSIO) {
548                 struct fio_file *f;
549
550                 for_each_file(td, f, i)
551                         f->real_file_size = -1ULL;
552         }
553
554         td->mutex = fio_mutex_init(0);
555
556         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
557         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
558         td->ts.lat_stat[0].min_val = td->ts.lat_stat[1].min_val = ULONG_MAX;
559         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
560         td->ddir_nr = td->o.ddir_nr;
561
562         if ((td->o.stonewall || td->o.new_group) && prev_group_jobs) {
563                 prev_group_jobs = 0;
564                 groupid++;
565         }
566
567         td->groupid = groupid;
568         prev_group_jobs++;
569
570         if (init_random_state(td))
571                 goto err;
572
573         if (setup_rate(td))
574                 goto err;
575
576         if (td->o.write_lat_log) {
577                 setup_log(&td->ts.lat_log);
578                 setup_log(&td->ts.slat_log);
579                 setup_log(&td->ts.clat_log);
580         }
581         if (td->o.write_bw_log)
582                 setup_log(&td->ts.bw_log);
583
584         if (!td->o.name)
585                 td->o.name = strdup(jobname);
586
587         if (!terse_output) {
588                 if (!job_add_num) {
589                         if (!strcmp(td->io_ops->name, "cpuio")) {
590                                 log_info("%s: ioengine=cpu, cpuload=%u,"
591                                          " cpucycle=%u\n", td->o.name,
592                                                         td->o.cpuload,
593                                                         td->o.cpucycle);
594                         } else {
595                                 char *c1, *c2, *c3, *c4;
596
597                                 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
598                                 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
599                                 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
600                                 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
601
602                                 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
603                                          " ioengine=%s, iodepth=%u\n",
604                                                 td->o.name, td->groupid,
605                                                 ddir_str[td->o.td_ddir],
606                                                 c1, c2, c3, c4,
607                                                 td->io_ops->name,
608                                                 td->o.iodepth);
609
610                                 free(c1);
611                                 free(c2);
612                                 free(c3);
613                                 free(c4);
614                         }
615                 } else if (job_add_num == 1)
616                         log_info("...\n");
617         }
618
619         /*
620          * recurse add identical jobs, clear numjobs and stonewall options
621          * as they don't apply to sub-jobs
622          */
623         numjobs = td->o.numjobs;
624         while (--numjobs) {
625                 struct thread_data *td_new = get_new_job(0, td);
626
627                 if (!td_new)
628                         goto err;
629
630                 td_new->o.numjobs = 1;
631                 td_new->o.stonewall = 0;
632                 td_new->o.new_group = 0;
633
634                 if (file_alloced) {
635                         td_new->o.filename = NULL;
636                         td_new->files_index = 0;
637                         td_new->files_size = 0;
638                         td_new->files = NULL;
639                 }
640
641                 job_add_num = numjobs - 1;
642
643                 if (add_job(td_new, jobname, job_add_num))
644                         goto err;
645         }
646
647         return 0;
648 err:
649         put_job(td);
650         return -1;
651 }
652
653 /*
654  * Parse as if 'o' was a command line
655  */
656 void add_job_opts(const char **o)
657 {
658         struct thread_data *td, *td_parent;
659         int i, in_global = 1;
660         char jobname[32];
661
662         i = 0;
663         td_parent = td = NULL;
664         while (o[i]) {
665                 if (!strncmp(o[i], "name", 4)) {
666                         in_global = 0;
667                         if (td)
668                                 add_job(td, jobname, 0);
669                         td = NULL;
670                         sprintf(jobname, "%s", o[i] + 5);
671                 }
672                 if (in_global && !td_parent)
673                         td_parent = get_new_job(1, &def_thread);
674                 else if (!in_global && !td) {
675                         if (!td_parent)
676                                 td_parent = &def_thread;
677                         td = get_new_job(0, td_parent);
678                 }
679                 if (in_global)
680                         fio_options_parse(td_parent, (char **) &o[i], 1);
681                 else
682                         fio_options_parse(td, (char **) &o[i], 1);
683                 i++;
684         }
685
686         if (td)
687                 add_job(td, jobname, 0);
688 }
689
690 static int skip_this_section(const char *name)
691 {
692         if (!job_section)
693                 return 0;
694         if (!strncmp(name, "global", 6))
695                 return 0;
696
697         return strcmp(job_section, name);
698 }
699
700 static int is_empty_or_comment(char *line)
701 {
702         unsigned int i;
703
704         for (i = 0; i < strlen(line); i++) {
705                 if (line[i] == ';')
706                         return 1;
707                 if (line[i] == '#')
708                         return 1;
709                 if (!isspace(line[i]) && !iscntrl(line[i]))
710                         return 0;
711         }
712
713         return 1;
714 }
715
716 /*
717  * This is our [ini] type file parser.
718  */
719 static int parse_jobs_ini(char *file, int stonewall_flag)
720 {
721         unsigned int global;
722         struct thread_data *td;
723         char *string, *name;
724         FILE *f;
725         char *p;
726         int ret = 0, stonewall;
727         int first_sect = 1;
728         int skip_fgets = 0;
729         int inside_skip = 0;
730         char **opts;
731         int i, alloc_opts, num_opts;
732
733         if (!strcmp(file, "-"))
734                 f = stdin;
735         else
736                 f = fopen(file, "r");
737
738         if (!f) {
739                 perror("fopen job file");
740                 return 1;
741         }
742
743         string = malloc(4096);
744
745         /*
746          * it's really 256 + small bit, 280 should suffice
747          */
748         name = malloc(280);
749         memset(name, 0, 280);
750
751         alloc_opts = 8;
752         opts = malloc(sizeof(char *) * alloc_opts);
753         num_opts = 0;
754
755         stonewall = stonewall_flag;
756         do {
757                 /*
758                  * if skip_fgets is set, we already have loaded a line we
759                  * haven't handled.
760                  */
761                 if (!skip_fgets) {
762                         p = fgets(string, 4095, f);
763                         if (!p)
764                                 break;
765                 }
766
767                 skip_fgets = 0;
768                 strip_blank_front(&p);
769                 strip_blank_end(p);
770
771                 if (is_empty_or_comment(p))
772                         continue;
773                 if (sscanf(p, "[%255s]", name) != 1) {
774                         if (inside_skip)
775                                 continue;
776                         log_err("fio: option <%s> outside of [] job section\n",
777                                                                         p);
778                         break;
779                 }
780
781                 name[strlen(name) - 1] = '\0';
782
783                 if (skip_this_section(name)) {
784                         inside_skip = 1;
785                         continue;
786                 } else
787                         inside_skip = 0;
788
789                 global = !strncmp(name, "global", 6);
790
791                 if (dump_cmdline) {
792                         if (first_sect)
793                                 log_info("fio ");
794                         if (!global)
795                                 log_info("--name=%s ", name);
796                         first_sect = 0;
797                 }
798
799                 td = get_new_job(global, &def_thread);
800                 if (!td) {
801                         ret = 1;
802                         break;
803                 }
804
805                 /*
806                  * Seperate multiple job files by a stonewall
807                  */
808                 if (!global && stonewall) {
809                         td->o.stonewall = stonewall;
810                         stonewall = 0;
811                 }
812
813                 num_opts = 0;
814                 memset(opts, 0, alloc_opts * sizeof(char *));
815
816                 while ((p = fgets(string, 4096, f)) != NULL) {
817                         if (is_empty_or_comment(p))
818                                 continue;
819
820                         strip_blank_front(&p);
821
822                         /*
823                          * new section, break out and make sure we don't
824                          * fgets() a new line at the top.
825                          */
826                         if (p[0] == '[') {
827                                 skip_fgets = 1;
828                                 break;
829                         }
830
831                         strip_blank_end(p);
832
833                         if (num_opts == alloc_opts) {
834                                 alloc_opts <<= 1;
835                                 opts = realloc(opts,
836                                                 alloc_opts * sizeof(char *));
837                         }
838
839                         opts[num_opts] = strdup(p);
840                         num_opts++;
841                 }
842
843                 ret = fio_options_parse(td, opts, num_opts);
844                 if (!ret) {
845                         if (dump_cmdline)
846                                 for (i = 0; i < num_opts; i++)
847                                         log_info("--%s ", opts[i]);
848
849                         ret = add_job(td, name, 0);
850                 } else {
851                         log_err("fio: job %s dropped\n", name);
852                         put_job(td);
853                 }
854
855                 for (i = 0; i < num_opts; i++)
856                         free(opts[i]);
857                 num_opts = 0;
858         } while (!ret);
859
860         if (dump_cmdline)
861                 log_info("\n");
862
863         for (i = 0; i < num_opts; i++)
864                 free(opts[i]);
865
866         free(string);
867         free(name);
868         free(opts);
869         if (f != stdin)
870                 fclose(f);
871         return ret;
872 }
873
874 static int fill_def_thread(void)
875 {
876         memset(&def_thread, 0, sizeof(def_thread));
877
878         fio_getaffinity(getpid(), &def_thread.o.cpumask);
879
880         /*
881          * fill default options
882          */
883         fio_fill_default_options(&def_thread);
884
885         def_thread.o.timeout = def_timeout;
886         return 0;
887 }
888
889 static void free_shm(void)
890 {
891         struct shmid_ds sbuf;
892
893         if (threads) {
894                 void *tp = threads;
895
896                 threads = NULL;
897                 file_hash_exit();
898                 fio_debug_jobp = NULL;
899                 shmdt(tp);
900                 shmctl(shm_id, IPC_RMID, &sbuf);
901         }
902
903         scleanup();
904 }
905
906 /*
907  * The thread area is shared between the main process and the job
908  * threads/processes. So setup a shared memory segment that will hold
909  * all the job info. We use the end of the region for keeping track of
910  * open files across jobs, for file sharing.
911  */
912 static int setup_thread_area(void)
913 {
914         void *hash;
915
916         /*
917          * 1024 is too much on some machines, scale max_jobs if
918          * we get a failure that looks like too large a shm segment
919          */
920         do {
921                 size_t size = max_jobs * sizeof(struct thread_data);
922
923                 size += file_hash_size;
924                 size += sizeof(unsigned int);
925
926                 shm_id = shmget(0, size, IPC_CREAT | 0600);
927                 if (shm_id != -1)
928                         break;
929                 if (errno != EINVAL) {
930                         perror("shmget");
931                         break;
932                 }
933
934                 max_jobs >>= 1;
935         } while (max_jobs);
936
937         if (shm_id == -1)
938                 return 1;
939
940         threads = shmat(shm_id, NULL, 0);
941         if (threads == (void *) -1) {
942                 perror("shmat");
943                 return 1;
944         }
945
946         memset(threads, 0, max_jobs * sizeof(struct thread_data));
947         hash = (void *) threads + max_jobs * sizeof(struct thread_data);
948         fio_debug_jobp = (void *) hash + file_hash_size;
949         *fio_debug_jobp = -1;
950         file_hash_init(hash);
951         atexit(free_shm);
952         return 0;
953 }
954
955 static void usage(const char *name)
956 {
957         printf("%s\n", fio_version_string);
958         printf("%s [options] [job options] <job file(s)>\n", name);
959         printf("\t--debug=options\tEnable debug logging\n");
960         printf("\t--output\tWrite output to file\n");
961         printf("\t--timeout\tRuntime in seconds\n");
962         printf("\t--latency-log\tGenerate per-job latency logs\n");
963         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
964         printf("\t--minimal\tMinimal (terse) output\n");
965         printf("\t--version\tPrint version info and exit\n");
966         printf("\t--help\t\tPrint this page\n");
967         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
968                 " them\n");
969         printf("\t--showcmd\tTurn a job file into command line options\n");
970         printf("\t--eta=when\tWhen ETA estimate should be printed\n");
971         printf("\t          \tMay be \"always\", \"never\" or \"auto\"\n");
972         printf("\t--readonly\tTurn on safety read-only checks, preventing"
973                 " writes\n");
974         printf("\t--section=name\tOnly run specified section in job file\n");
975         printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
976                 " (def 1024)\n");
977         printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>");
978         printf("\n                   Jens Axboe <jaxboe@fusionio.com>\n");
979 }
980
981 #ifdef FIO_INC_DEBUG
982 struct debug_level debug_levels[] = {
983         { .name = "process",    .shift = FD_PROCESS, },
984         { .name = "file",       .shift = FD_FILE, },
985         { .name = "io",         .shift = FD_IO, },
986         { .name = "mem",        .shift = FD_MEM, },
987         { .name = "blktrace",   .shift = FD_BLKTRACE },
988         { .name = "verify",     .shift = FD_VERIFY },
989         { .name = "random",     .shift = FD_RANDOM },
990         { .name = "parse",      .shift = FD_PARSE },
991         { .name = "diskutil",   .shift = FD_DISKUTIL },
992         { .name = "job",        .shift = FD_JOB },
993         { .name = "mutex",      .shift = FD_MUTEX },
994         { .name = "profile",    .shift = FD_PROFILE },
995         { .name = "time",       .shift = FD_TIME },
996         { .name = NULL, },
997 };
998
999 static int set_debug(const char *string)
1000 {
1001         struct debug_level *dl;
1002         char *p = (char *) string;
1003         char *opt;
1004         int i;
1005
1006         if (!strcmp(string, "?") || !strcmp(string, "help")) {
1007                 int i;
1008
1009                 log_info("fio: dumping debug options:");
1010                 for (i = 0; debug_levels[i].name; i++) {
1011                         dl = &debug_levels[i];
1012                         log_info("%s,", dl->name);
1013                 }
1014                 log_info("all\n");
1015                 return 1;
1016         }
1017
1018         while ((opt = strsep(&p, ",")) != NULL) {
1019                 int found = 0;
1020
1021                 if (!strncmp(opt, "all", 3)) {
1022                         log_info("fio: set all debug options\n");
1023                         fio_debug = ~0UL;
1024                         continue;
1025                 }
1026
1027                 for (i = 0; debug_levels[i].name; i++) {
1028                         dl = &debug_levels[i];
1029                         found = !strncmp(opt, dl->name, strlen(dl->name));
1030                         if (!found)
1031                                 continue;
1032
1033                         if (dl->shift == FD_JOB) {
1034                                 opt = strchr(opt, ':');
1035                                 if (!opt) {
1036                                         log_err("fio: missing job number\n");
1037                                         break;
1038                                 }
1039                                 opt++;
1040                                 fio_debug_jobno = atoi(opt);
1041                                 log_info("fio: set debug jobno %d\n",
1042                                                         fio_debug_jobno);
1043                         } else {
1044                                 log_info("fio: set debug option %s\n", opt);
1045                                 fio_debug |= (1UL << dl->shift);
1046                         }
1047                         break;
1048                 }
1049
1050                 if (!found)
1051                         log_err("fio: debug mask %s not found\n", opt);
1052         }
1053         return 0;
1054 }
1055 #else
1056 static int set_debug(const char *string)
1057 {
1058         log_err("fio: debug tracing not included in build\n");
1059         return 1;
1060 }
1061 #endif
1062
1063 static int parse_cmd_line(int argc, char *argv[])
1064 {
1065         struct thread_data *td = NULL;
1066         int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1067
1068         while ((c = getopt_long_only(argc, argv, "", l_opts, &lidx)) != -1) {
1069                 switch (c) {
1070                 case 'a':
1071                         smalloc_pool_size = atoi(optarg);
1072                         break;
1073                 case 't':
1074                         def_timeout = atoi(optarg);
1075                         break;
1076                 case 'l':
1077                         write_lat_log = 1;
1078                         break;
1079                 case 'b':
1080                         write_bw_log = 1;
1081                         break;
1082                 case 'o':
1083                         f_out = fopen(optarg, "w+");
1084                         if (!f_out) {
1085                                 perror("fopen output");
1086                                 exit(1);
1087                         }
1088                         f_err = f_out;
1089                         break;
1090                 case 'm':
1091                         terse_output = 1;
1092                         break;
1093                 case 'h':
1094                         usage(argv[0]);
1095                         exit(0);
1096                 case 'c':
1097                         exit(fio_show_option_help(optarg));
1098                 case 's':
1099                         dump_cmdline = 1;
1100                         break;
1101                 case 'r':
1102                         read_only = 1;
1103                         break;
1104                 case 'v':
1105                         printf("%s\n", fio_version_string);
1106                         exit(0);
1107                 case 'e':
1108                         if (!strcmp("always", optarg))
1109                                 eta_print = FIO_ETA_ALWAYS;
1110                         else if (!strcmp("never", optarg))
1111                                 eta_print = FIO_ETA_NEVER;
1112                         break;
1113                 case 'd':
1114                         if (set_debug(optarg))
1115                                 do_exit++;
1116                         break;
1117                 case 'x':
1118                         if (!strcmp(optarg, "global")) {
1119                                 log_err("fio: can't use global as only "
1120                                         "section\n");
1121                                 do_exit++;
1122                                 exit_val = 1;
1123                                 break;
1124                         }
1125                         if (job_section)
1126                                 free(job_section);
1127                         job_section = strdup(optarg);
1128                         break;
1129                 case 'p':
1130                         exec_profile = strdup(optarg);
1131                         break;
1132                 case FIO_GETOPT_JOB: {
1133                         const char *opt = l_opts[lidx].name;
1134                         char *val = optarg;
1135
1136                         if (!strncmp(opt, "name", 4) && td) {
1137                                 ret = add_job(td, td->o.name ?: "fio", 0);
1138                                 if (ret) {
1139                                         put_job(td);
1140                                         return 0;
1141                                 }
1142                                 td = NULL;
1143                         }
1144                         if (!td) {
1145                                 int is_section = !strncmp(opt, "name", 4);
1146                                 int global = 0;
1147
1148                                 if (!is_section || !strncmp(val, "global", 6))
1149                                         global = 1;
1150
1151                                 if (is_section && skip_this_section(val))
1152                                         continue;
1153
1154                                 td = get_new_job(global, &def_thread);
1155                                 if (!td)
1156                                         return 0;
1157                         }
1158
1159                         ret = fio_cmd_option_parse(td, opt, val);
1160                         break;
1161                 }
1162                 default:
1163                         do_exit++;
1164                         exit_val = 1;
1165                         break;
1166                 }
1167         }
1168
1169         if (do_exit)
1170                 exit(exit_val);
1171
1172         if (td) {
1173                 if (!ret)
1174                         ret = add_job(td, td->o.name ?: "fio", 0);
1175                 if (ret)
1176                         put_job(td);
1177         }
1178
1179         while (optind < argc) {
1180                 ini_idx++;
1181                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1182                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1183                 optind++;
1184         }
1185
1186         return ini_idx;
1187 }
1188
1189 int parse_options(int argc, char *argv[])
1190 {
1191         int job_files, i;
1192
1193         f_out = stdout;
1194         f_err = stderr;
1195
1196         fio_options_dup_and_init(l_opts);
1197
1198         if (setup_thread_area())
1199                 return 1;
1200         if (fill_def_thread())
1201                 return 1;
1202
1203         job_files = parse_cmd_line(argc, argv);
1204
1205         for (i = 0; i < job_files; i++) {
1206                 if (fill_def_thread())
1207                         return 1;
1208                 if (parse_jobs_ini(ini_file[i], i))
1209                         return 1;
1210                 free(ini_file[i]);
1211         }
1212
1213         free(ini_file);
1214         options_mem_free(&def_thread);
1215
1216         if (!thread_number) {
1217                 if (dump_cmdline)
1218                         return 0;
1219                 if (exec_profile)
1220                         return 0;
1221
1222                 log_err("No jobs defined(s)\n\n");
1223                 usage(argv[0]);
1224                 return 1;
1225         }
1226
1227         if (def_thread.o.gtod_offload) {
1228                 fio_gtod_init();
1229                 fio_gtod_offload = 1;
1230                 fio_gtod_cpu = def_thread.o.gtod_cpu;
1231         }
1232
1233         return 0;
1234 }