Fio 1.29-rc1
[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
23 static char fio_version_string[] = "fio 1.29-rc1";
24
25 #define FIO_RANDSEED            (0xb1899bedUL)
26
27 static char **ini_file;
28 static int max_jobs = MAX_JOBS;
29 static int dump_cmdline;
30
31 static struct thread_data def_thread;
32 struct thread_data *threads = NULL;
33
34 int exitall_on_terminate = 0;
35 int terse_output = 0;
36 int eta_print;
37 unsigned long long mlock_size = 0;
38 FILE *f_out = NULL;
39 FILE *f_err = NULL;
40 char *job_section = NULL;
41
42 int write_bw_log = 0;
43 int read_only = 0;
44
45 static int def_timeout;
46 static int write_lat_log;
47
48 static int prev_group_jobs;
49
50 unsigned long fio_debug = 0;
51 unsigned int fio_debug_jobno = -1;
52 unsigned int *fio_debug_jobp = NULL;
53
54 /*
55  * Command line options. These will contain the above, plus a few
56  * extra that only pertain to fio itself and not jobs.
57  */
58 static struct option l_opts[FIO_NR_OPTIONS] = {
59         {
60                 .name           = "output",
61                 .has_arg        = required_argument,
62                 .val            = 'o',
63         },
64         {
65                 .name           = "timeout",
66                 .has_arg        = required_argument,
67                 .val            = 't',
68         },
69         {
70                 .name           = "latency-log",
71                 .has_arg        = required_argument,
72                 .val            = 'l',
73         },
74         {
75                 .name           = "bandwidth-log",
76                 .has_arg        = required_argument,
77                 .val            = 'b',
78         },
79         {
80                 .name           = "minimal",
81                 .has_arg        = optional_argument,
82                 .val            = 'm',
83         },
84         {
85                 .name           = "version",
86                 .has_arg        = no_argument,
87                 .val            = 'v',
88         },
89         {
90                 .name           = "help",
91                 .has_arg        = no_argument,
92                 .val            = 'h',
93         },
94         {
95                 .name           = "cmdhelp",
96                 .has_arg        = optional_argument,
97                 .val            = 'c',
98         },
99         {
100                 .name           = "showcmd",
101                 .has_arg        = no_argument,
102                 .val            = 's',
103         },
104         {
105                 .name           = "readonly",
106                 .has_arg        = no_argument,
107                 .val            = 'r',
108         },
109         {
110                 .name           = "eta",
111                 .has_arg        = required_argument,
112                 .val            = 'e',
113         },
114         {
115                 .name           = "debug",
116                 .has_arg        = required_argument,
117                 .val            = 'd',
118         },
119         {
120                 .name           = "section",
121                 .has_arg        = required_argument,
122                 .val            = 'x',
123         },
124         {
125                 .name           = "alloc-size",
126                 .has_arg        = required_argument,
127                 .val            = 'a',
128         },
129         {
130                 .name           = NULL,
131         },
132 };
133
134 FILE *get_f_out()
135 {
136         return f_out;
137 }
138
139 FILE *get_f_err()
140 {
141         return f_err;
142 }
143
144 /*
145  * Return a free job structure.
146  */
147 static struct thread_data *get_new_job(int global, struct thread_data *parent)
148 {
149         struct thread_data *td;
150
151         if (global)
152                 return &def_thread;
153         if (thread_number >= max_jobs)
154                 return NULL;
155
156         td = &threads[thread_number++];
157         *td = *parent;
158
159         dup_files(td, parent);
160         options_mem_dupe(td);
161
162         td->thread_number = thread_number;
163         return td;
164 }
165
166 static void put_job(struct thread_data *td)
167 {
168         if (td == &def_thread)
169                 return;
170
171         if (td->error)
172                 log_info("fio: %s\n", td->verror);
173
174         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
175         thread_number--;
176 }
177
178 static int __setup_rate(struct thread_data *td, enum fio_ddir ddir)
179 {
180         unsigned int bs = td->o.min_bs[ddir];
181         unsigned long long rate;
182         unsigned long ios_per_msec;
183
184         if (td->o.rate[ddir]) {
185                 rate = td->o.rate[ddir];
186                 ios_per_msec = (rate * 1000LL) / bs;
187         } else
188                 ios_per_msec = td->o.rate_iops[ddir] * 1000UL;
189
190         if (!ios_per_msec) {
191                 log_err("rate lower than supported\n");
192                 return -1;
193         }
194
195         td->rate_usec_cycle[ddir] = 1000000000ULL / ios_per_msec;
196         td->rate_pending_usleep[ddir] = 0;
197         return 0;
198 }
199
200 static int setup_rate(struct thread_data *td)
201 {
202         int ret = 0;
203
204         if (td->o.rate[DDIR_READ] || td->o.rate_iops[DDIR_READ])
205                 ret = __setup_rate(td, DDIR_READ);
206         if (td->o.rate[DDIR_WRITE] || td->o.rate_iops[DDIR_WRITE])
207                 ret |= __setup_rate(td, DDIR_WRITE);
208
209         return ret;
210 }
211
212 static int fixed_block_size(struct thread_options *o)
213 {
214         return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
215                 o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
216                 o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE];
217 }
218
219 /*
220  * Lazy way of fixing up options that depend on each other. We could also
221  * define option callback handlers, but this is easier.
222  */
223 static int fixup_options(struct thread_data *td)
224 {
225         struct thread_options *o = &td->o;
226
227 #ifndef FIO_HAVE_PSHARED_MUTEX
228         if (!td->o.use_thread) {
229                 log_info("fio: this platform does not support process shared"
230                          " mutexes, forcing use of threads. Use the 'thread'"
231                          " option to get rid of this warning.\n");
232                 td->o.use_thread = 1;
233         }
234 #endif
235
236         if (o->write_iolog_file && o->read_iolog_file) {
237                 log_err("fio: read iolog overrides write_iolog\n");
238                 free(o->write_iolog_file);
239                 o->write_iolog_file = NULL;
240         }
241
242         /*
243          * only really works for sequential io for now, and with 1 file
244          */
245         if (o->zone_size && td_random(td) && o->open_files == 1)
246                 o->zone_size = 0;
247
248         /*
249          * Reads can do overwrites, we always need to pre-create the file
250          */
251         if (td_read(td) || td_rw(td))
252                 o->overwrite = 1;
253
254         if (!o->min_bs[DDIR_READ])
255                 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
256         if (!o->max_bs[DDIR_READ])
257                 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
258         if (!o->min_bs[DDIR_WRITE])
259                 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
260         if (!o->max_bs[DDIR_WRITE])
261                 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
262
263         o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
264
265         /*
266          * For random IO, allow blockalign offset other than min_bs.
267          */
268         if (!o->ba[DDIR_READ] || !td_random(td))
269                 o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
270         if (!o->ba[DDIR_WRITE] || !td_random(td))
271                 o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
272
273         if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
274             o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
275             !td->o.norandommap) {
276                 log_err("fio: Any use of blockalign= turns off randommap\n");
277                 td->o.norandommap = 1;
278         }
279
280         if (!o->file_size_high)
281                 o->file_size_high = o->file_size_low;
282
283         if (o->norandommap && o->verify != VERIFY_NONE
284             && !fixed_block_size(o))  {
285                 log_err("fio: norandommap given for variable block sizes, "
286                         "verify disabled\n");
287                 o->verify = VERIFY_NONE;
288         }
289         if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
290                 log_err("fio: bs_unaligned may not work with raw io\n");
291
292         /*
293          * thinktime_spin must be less than thinktime
294          */
295         if (o->thinktime_spin > o->thinktime)
296                 o->thinktime_spin = o->thinktime;
297
298         /*
299          * The low water mark cannot be bigger than the iodepth
300          */
301         if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
302                 /*
303                  * syslet work around - if the workload is sequential,
304                  * we want to let the queue drain all the way down to
305                  * avoid seeking between async threads
306                  */
307                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
308                         o->iodepth_low = 1;
309                 else
310                         o->iodepth_low = o->iodepth;
311         }
312
313         /*
314          * If batch number isn't set, default to the same as iodepth
315          */
316         if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
317                 o->iodepth_batch = o->iodepth;
318
319         if (o->nr_files > td->files_index)
320                 o->nr_files = td->files_index;
321
322         if (o->open_files > o->nr_files || !o->open_files)
323                 o->open_files = o->nr_files;
324
325         if (((o->rate[0] + o->rate[1]) && (o->rate_iops[0] + o->rate_iops[1]))||
326             ((o->ratemin[0] + o->ratemin[1]) && (o->rate_iops_min[0] +
327                 o->rate_iops_min[1]))) {
328                 log_err("fio: rate and rate_iops are mutually exclusive\n");
329                 return 1;
330         }
331         if ((o->rate[0] < o->ratemin[0]) || (o->rate[1] < o->ratemin[1]) ||
332             (o->rate_iops[0] < o->rate_iops_min[0]) ||
333             (o->rate_iops[1] < o->rate_iops_min[1])) {
334                 log_err("fio: minimum rate exceeds rate\n");
335                 return 1;
336         }
337
338         if (!o->timeout && o->time_based) {
339                 log_err("fio: time_based requires a runtime/timeout setting\n");
340                 o->time_based = 0;
341         }
342
343         if (o->fill_device && !o->size)
344                 o->size = -1ULL;
345
346         if (td_rw(td) && td->o.verify != VERIFY_NONE)
347                 log_info("fio: mixed read/write workload with verify. May not "
348                  "work as expected, unless you pre-populated the file\n");
349
350         if (td->o.verify != VERIFY_NONE)
351                 td->o.refill_buffers = 1;
352
353         if (td->o.pre_read)
354                 td->o.invalidate_cache = 0;
355
356         if (td->o.mem_align) {
357                 if (td->o.odirect && !is_power_of_2(td->o.mem_align)) {
358                         log_err("fio: given IO mem alignment conflicts with"
359                                         " direct=1. Resetting.\n");
360                         td->o.mem_align = page_mask;
361                 }
362         }
363                         
364         return 0;
365 }
366
367 /*
368  * This function leaks the buffer
369  */
370 static char *to_kmg(unsigned int val)
371 {
372         char *buf = malloc(32);
373         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
374         char *p = post;
375
376         do {
377                 if (val & 1023)
378                         break;
379
380                 val >>= 10;
381                 p++;
382         } while (*p);
383
384         snprintf(buf, 31, "%u%c", val, *p);
385         return buf;
386 }
387
388 /* External engines are specified by "external:name.o") */
389 static const char *get_engine_name(const char *str)
390 {
391         char *p = strstr(str, ":");
392
393         if (!p)
394                 return str;
395
396         p++;
397         strip_blank_front(&p);
398         strip_blank_end(p);
399         return p;
400 }
401
402 static int exists_and_not_file(const char *filename)
403 {
404         struct stat sb;
405
406         if (lstat(filename, &sb) == -1)
407                 return 0;
408
409         if (S_ISREG(sb.st_mode))
410                 return 0;
411
412         return 1;
413 }
414
415 void td_fill_rand_seeds(struct thread_data *td)
416 {
417         os_random_seed(td->rand_seeds[0], &td->bsrange_state);
418         os_random_seed(td->rand_seeds[1], &td->verify_state);
419         os_random_seed(td->rand_seeds[2], &td->rwmix_state);
420
421         if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
422                 os_random_seed(td->rand_seeds[3], &td->next_file_state);
423
424         os_random_seed(td->rand_seeds[5], &td->file_size_state);
425
426         if (!td_random(td))
427                 return;
428
429         if (td->o.rand_repeatable)
430                 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
431
432         os_random_seed(td->rand_seeds[4], &td->random_state);
433 }
434
435 /*
436  * Initialize the various random states we need (random io, block size ranges,
437  * read/write mix, etc).
438  */
439 static int init_random_state(struct thread_data *td)
440 {
441         int fd;
442
443         fd = open("/dev/urandom", O_RDONLY);
444         if (fd == -1) {
445                 td_verror(td, errno, "open");
446                 return 1;
447         }
448
449         if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
450             (int) sizeof(td->rand_seeds)) {
451                 td_verror(td, EIO, "read");
452                 close(fd);
453                 return 1;
454         }
455
456         close(fd);
457         td_fill_rand_seeds(td);
458         return 0;
459 }
460
461 /*
462  * Adds a job to the list of things todo. Sanitizes the various options
463  * to make sure we don't have conflicts, and initializes various
464  * members of td.
465  */
466 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
467 {
468         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
469                                    "randread", "randwrite", "randrw" };
470         unsigned int i;
471         const char *engine;
472         char fname[PATH_MAX];
473         int numjobs, file_alloced;
474
475         /*
476          * the def_thread is just for options, it's not a real job
477          */
478         if (td == &def_thread)
479                 return 0;
480
481         /*
482          * if we are just dumping the output command line, don't add the job
483          */
484         if (dump_cmdline) {
485                 put_job(td);
486                 return 0;
487         }
488
489         engine = get_engine_name(td->o.ioengine);
490         td->io_ops = load_ioengine(td, engine);
491         if (!td->io_ops) {
492                 log_err("fio: failed to load engine %s\n", engine);
493                 goto err;
494         }
495
496         if (td->o.use_thread)
497                 nr_thread++;
498         else
499                 nr_process++;
500
501         if (td->o.odirect)
502                 td->io_ops->flags |= FIO_RAWIO;
503
504         file_alloced = 0;
505         if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
506                 file_alloced = 1;
507
508                 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
509                         add_file(td, jobname);
510                 else {
511                         for (i = 0; i < td->o.nr_files; i++) {
512                                 sprintf(fname, "%s.%d.%d", jobname,
513                                                         td->thread_number, i);
514                                 add_file(td, fname);
515                         }
516                 }
517         }
518
519         if (fixup_options(td))
520                 goto err;
521
522         if (td->io_ops->flags & FIO_DISKLESSIO) {
523                 struct fio_file *f;
524
525                 for_each_file(td, f, i)
526                         f->real_file_size = -1ULL;
527         }
528
529         td->mutex = fio_mutex_init(0);
530
531         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
532         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
533         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
534         td->ddir_nr = td->o.ddir_nr;
535
536         if ((td->o.stonewall || td->o.numjobs > 1 || td->o.new_group)
537              && prev_group_jobs) {
538                 prev_group_jobs = 0;
539                 groupid++;
540         }
541
542         td->groupid = groupid;
543         prev_group_jobs++;
544
545         if (init_random_state(td))
546                 goto err;
547
548         if (setup_rate(td))
549                 goto err;
550
551         if (td->o.write_lat_log) {
552                 setup_log(&td->ts.slat_log);
553                 setup_log(&td->ts.clat_log);
554         }
555         if (td->o.write_bw_log)
556                 setup_log(&td->ts.bw_log);
557
558         if (!td->o.name)
559                 td->o.name = strdup(jobname);
560
561         if (!terse_output) {
562                 if (!job_add_num) {
563                         if (!strcmp(td->io_ops->name, "cpuio")) {
564                                 log_info("%s: ioengine=cpu, cpuload=%u,"
565                                          " cpucycle=%u\n", td->o.name,
566                                                         td->o.cpuload,
567                                                         td->o.cpucycle);
568                         } else {
569                                 char *c1, *c2, *c3, *c4;
570
571                                 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
572                                 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
573                                 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
574                                 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
575
576                                 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
577                                          " ioengine=%s, iodepth=%u\n",
578                                                 td->o.name, td->groupid,
579                                                 ddir_str[td->o.td_ddir],
580                                                 c1, c2, c3, c4,
581                                                 td->io_ops->name,
582                                                 td->o.iodepth);
583
584                                 free(c1);
585                                 free(c2);
586                                 free(c3);
587                                 free(c4);
588                         }
589                 } else if (job_add_num == 1)
590                         log_info("...\n");
591         }
592
593         /*
594          * recurse add identical jobs, clear numjobs and stonewall options
595          * as they don't apply to sub-jobs
596          */
597         numjobs = td->o.numjobs;
598         while (--numjobs) {
599                 struct thread_data *td_new = get_new_job(0, td);
600
601                 if (!td_new)
602                         goto err;
603
604                 td_new->o.numjobs = 1;
605                 td_new->o.stonewall = 0;
606                 td_new->o.new_group = 0;
607
608                 if (file_alloced) {
609                         td_new->o.filename = NULL;
610                         td_new->files_index = 0;
611                         td_new->files_size = 0;
612                         td_new->files = NULL;
613                 }
614
615                 job_add_num = numjobs - 1;
616
617                 if (add_job(td_new, jobname, job_add_num))
618                         goto err;
619         }
620
621         return 0;
622 err:
623         put_job(td);
624         return -1;
625 }
626
627 static int skip_this_section(const char *name)
628 {
629         if (!job_section)
630                 return 0;
631         if (!strncmp(name, "global", 6))
632                 return 0;
633
634         return strcmp(job_section, name);
635 }
636
637 static int is_empty_or_comment(char *line)
638 {
639         unsigned int i;
640
641         for (i = 0; i < strlen(line); i++) {
642                 if (line[i] == ';')
643                         return 1;
644                 if (line[i] == '#')
645                         return 1;
646                 if (!isspace(line[i]) && !iscntrl(line[i]))
647                         return 0;
648         }
649
650         return 1;
651 }
652
653 /*
654  * This is our [ini] type file parser.
655  */
656 static int parse_jobs_ini(char *file, int stonewall_flag)
657 {
658         unsigned int global;
659         struct thread_data *td;
660         char *string, *name;
661         FILE *f;
662         char *p;
663         int ret = 0, stonewall;
664         int first_sect = 1;
665         int skip_fgets = 0;
666         int inside_skip = 0;
667         char **opts;
668         int i, alloc_opts, num_opts;
669
670         if (!strcmp(file, "-"))
671                 f = stdin;
672         else
673                 f = fopen(file, "r");
674
675         if (!f) {
676                 perror("fopen job file");
677                 return 1;
678         }
679
680         string = malloc(4096);
681
682         /*
683          * it's really 256 + small bit, 280 should suffice
684          */
685         name = malloc(280);
686         memset(name, 0, 280);
687
688         alloc_opts = 8;
689         opts = malloc(sizeof(char *) * alloc_opts);
690         num_opts = 0;
691
692         stonewall = stonewall_flag;
693         do {
694                 /*
695                  * if skip_fgets is set, we already have loaded a line we
696                  * haven't handled.
697                  */
698                 if (!skip_fgets) {
699                         p = fgets(string, 4095, f);
700                         if (!p)
701                                 break;
702                 }
703
704                 skip_fgets = 0;
705                 strip_blank_front(&p);
706                 strip_blank_end(p);
707
708                 if (is_empty_or_comment(p))
709                         continue;
710                 if (sscanf(p, "[%255s]", name) != 1) {
711                         if (inside_skip)
712                                 continue;
713                         log_err("fio: option <%s> outside of [] job section\n",
714                                                                         p);
715                         break;
716                 }
717
718                 if (skip_this_section(name)) {
719                         inside_skip = 1;
720                         continue;
721                 } else
722                         inside_skip = 0;
723
724                 global = !strncmp(name, "global", 6);
725
726                 name[strlen(name) - 1] = '\0';
727
728                 if (dump_cmdline) {
729                         if (first_sect)
730                                 log_info("fio ");
731                         if (!global)
732                                 log_info("--name=%s ", name);
733                         first_sect = 0;
734                 }
735
736                 td = get_new_job(global, &def_thread);
737                 if (!td) {
738                         ret = 1;
739                         break;
740                 }
741
742                 /*
743                  * Seperate multiple job files by a stonewall
744                  */
745                 if (!global && stonewall) {
746                         td->o.stonewall = stonewall;
747                         stonewall = 0;
748                 }
749
750                 num_opts = 0;
751                 memset(opts, 0, alloc_opts * sizeof(char *));
752
753                 while ((p = fgets(string, 4096, f)) != NULL) {
754                         if (is_empty_or_comment(p))
755                                 continue;
756
757                         strip_blank_front(&p);
758
759                         /*
760                          * new section, break out and make sure we don't
761                          * fgets() a new line at the top.
762                          */
763                         if (p[0] == '[') {
764                                 skip_fgets = 1;
765                                 break;
766                         }
767
768                         strip_blank_end(p);
769
770                         if (num_opts == alloc_opts) {
771                                 alloc_opts <<= 1;
772                                 opts = realloc(opts,
773                                                 alloc_opts * sizeof(char *));
774                         }
775
776                         opts[num_opts] = strdup(p);
777                         num_opts++;
778                 }
779
780                 ret = fio_options_parse(td, opts, num_opts);
781                 if (!ret) {
782                         if (dump_cmdline)
783                                 for (i = 0; i < num_opts; i++)
784                                         log_info("--%s ", opts[i]);
785
786                         ret = add_job(td, name, 0);
787                 } else {
788                         log_err("fio: job %s dropped\n", name);
789                         put_job(td);
790                 }
791
792                 for (i = 0; i < num_opts; i++)
793                         free(opts[i]);
794                 num_opts = 0;
795         } while (!ret);
796
797         if (dump_cmdline)
798                 log_info("\n");
799
800         for (i = 0; i < num_opts; i++)
801                 free(opts[i]);
802
803         free(string);
804         free(name);
805         free(opts);
806         if (f != stdin)
807                 fclose(f);
808         return ret;
809 }
810
811 static int fill_def_thread(void)
812 {
813         memset(&def_thread, 0, sizeof(def_thread));
814
815         fio_getaffinity(getpid(), &def_thread.o.cpumask);
816
817         /*
818          * fill default options
819          */
820         fio_fill_default_options(&def_thread);
821
822         def_thread.o.timeout = def_timeout;
823         def_thread.o.write_bw_log = write_bw_log;
824         def_thread.o.write_lat_log = write_lat_log;
825
826         return 0;
827 }
828
829 static void free_shm(void)
830 {
831         struct shmid_ds sbuf;
832
833         if (threads) {
834                 void *tp = threads;
835
836                 threads = NULL;
837                 file_hash_exit();
838                 fio_debug_jobp = NULL;
839                 shmdt(tp);
840                 shmctl(shm_id, IPC_RMID, &sbuf);
841         }
842
843         scleanup();
844 }
845
846 /*
847  * The thread area is shared between the main process and the job
848  * threads/processes. So setup a shared memory segment that will hold
849  * all the job info. We use the end of the region for keeping track of
850  * open files across jobs, for file sharing.
851  */
852 static int setup_thread_area(void)
853 {
854         void *hash;
855
856         /*
857          * 1024 is too much on some machines, scale max_jobs if
858          * we get a failure that looks like too large a shm segment
859          */
860         do {
861                 size_t size = max_jobs * sizeof(struct thread_data);
862
863                 size += file_hash_size;
864                 size += sizeof(unsigned int);
865
866                 shm_id = shmget(0, size, IPC_CREAT | 0600);
867                 if (shm_id != -1)
868                         break;
869                 if (errno != EINVAL) {
870                         perror("shmget");
871                         break;
872                 }
873
874                 max_jobs >>= 1;
875         } while (max_jobs);
876
877         if (shm_id == -1)
878                 return 1;
879
880         threads = shmat(shm_id, NULL, 0);
881         if (threads == (void *) -1) {
882                 perror("shmat");
883                 return 1;
884         }
885
886         memset(threads, 0, max_jobs * sizeof(struct thread_data));
887         hash = (void *) threads + max_jobs * sizeof(struct thread_data);
888         fio_debug_jobp = (void *) hash + file_hash_size;
889         *fio_debug_jobp = -1;
890         file_hash_init(hash);
891         atexit(free_shm);
892         return 0;
893 }
894
895 static void usage(const char *name)
896 {
897         printf("%s\n", fio_version_string);
898         printf("%s [options] [job options] <job file(s)>\n", name);
899         printf("\t--debug=options\tEnable debug logging\n");
900         printf("\t--output\tWrite output to file\n");
901         printf("\t--timeout\tRuntime in seconds\n");
902         printf("\t--latency-log\tGenerate per-job latency logs\n");
903         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
904         printf("\t--minimal\tMinimal (terse) output\n");
905         printf("\t--version\tPrint version info and exit\n");
906         printf("\t--help\t\tPrint this page\n");
907         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
908                 " them\n");
909         printf("\t--showcmd\tTurn a job file into command line options\n");
910         printf("\t--eta=when\tWhen ETA estimate should be printed\n");
911         printf("\t          \tMay be \"always\", \"never\" or \"auto\"\n");
912         printf("\t--readonly\tTurn on safety read-only checks, preventing"
913                 " writes\n");
914         printf("\t--section=name\tOnly run specified section in job file\n");
915         printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
916                 " (def 1024)\n");
917         printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>\n");
918 }
919
920 #ifdef FIO_INC_DEBUG
921 struct debug_level debug_levels[] = {
922         { .name = "process",    .shift = FD_PROCESS, },
923         { .name = "file",       .shift = FD_FILE, },
924         { .name = "io",         .shift = FD_IO, },
925         { .name = "mem",        .shift = FD_MEM, },
926         { .name = "blktrace",   .shift = FD_BLKTRACE },
927         { .name = "verify",     .shift = FD_VERIFY },
928         { .name = "random",     .shift = FD_RANDOM },
929         { .name = "parse",      .shift = FD_PARSE },
930         { .name = "diskutil",   .shift = FD_DISKUTIL },
931         { .name = "job",        .shift = FD_JOB },
932         { .name = "mutex",      .shift = FD_MUTEX },
933         { .name = NULL, },
934 };
935
936 static int set_debug(const char *string)
937 {
938         struct debug_level *dl;
939         char *p = (char *) string;
940         char *opt;
941         int i;
942
943         if (!strcmp(string, "?") || !strcmp(string, "help")) {
944                 int i;
945
946                 log_info("fio: dumping debug options:");
947                 for (i = 0; debug_levels[i].name; i++) {
948                         dl = &debug_levels[i];
949                         log_info("%s,", dl->name);
950                 }
951                 log_info("all\n");
952                 return 1;
953         }
954
955         while ((opt = strsep(&p, ",")) != NULL) {
956                 int found = 0;
957
958                 if (!strncmp(opt, "all", 3)) {
959                         log_info("fio: set all debug options\n");
960                         fio_debug = ~0UL;
961                         continue;
962                 }
963
964                 for (i = 0; debug_levels[i].name; i++) {
965                         dl = &debug_levels[i];
966                         found = !strncmp(opt, dl->name, strlen(dl->name));
967                         if (!found)
968                                 continue;
969
970                         if (dl->shift == FD_JOB) {
971                                 opt = strchr(opt, ':');
972                                 if (!opt) {
973                                         log_err("fio: missing job number\n");
974                                         break;
975                                 }
976                                 opt++;
977                                 fio_debug_jobno = atoi(opt);
978                                 log_info("fio: set debug jobno %d\n",
979                                                         fio_debug_jobno);
980                         } else {
981                                 log_info("fio: set debug option %s\n", opt);
982                                 fio_debug |= (1UL << dl->shift);
983                         }
984                         break;
985                 }
986
987                 if (!found)
988                         log_err("fio: debug mask %s not found\n", opt);
989         }
990         return 0;
991 }
992 #else
993 static int set_debug(const char *string)
994 {
995         log_err("fio: debug tracing not included in build\n");
996         return 1;
997 }
998 #endif
999
1000 static int parse_cmd_line(int argc, char *argv[])
1001 {
1002         struct thread_data *td = NULL;
1003         int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1004
1005         while ((c = getopt_long_only(argc, argv, "", l_opts, &lidx)) != -1) {
1006                 switch (c) {
1007                 case 'a':
1008                         smalloc_pool_size = atoi(optarg);
1009                         break;
1010                 case 't':
1011                         def_timeout = atoi(optarg);
1012                         break;
1013                 case 'l':
1014                         write_lat_log = 1;
1015                         break;
1016                 case 'w':
1017                         write_bw_log = 1;
1018                         break;
1019                 case 'o':
1020                         f_out = fopen(optarg, "w+");
1021                         if (!f_out) {
1022                                 perror("fopen output");
1023                                 exit(1);
1024                         }
1025                         f_err = f_out;
1026                         break;
1027                 case 'm':
1028                         terse_output = 1;
1029                         break;
1030                 case 'h':
1031                         usage(argv[0]);
1032                         exit(0);
1033                 case 'c':
1034                         exit(fio_show_option_help(optarg));
1035                 case 's':
1036                         dump_cmdline = 1;
1037                         break;
1038                 case 'r':
1039                         read_only = 1;
1040                         break;
1041                 case 'v':
1042                         printf("%s\n", fio_version_string);
1043                         exit(0);
1044                 case 'e':
1045                         if (!strcmp("always", optarg))
1046                                 eta_print = FIO_ETA_ALWAYS;
1047                         else if (!strcmp("never", optarg))
1048                                 eta_print = FIO_ETA_NEVER;
1049                         break;
1050                 case 'd':
1051                         if (set_debug(optarg))
1052                                 do_exit++;
1053                         break;
1054                 case 'x':
1055                         if (!strcmp(optarg, "global")) {
1056                                 log_err("fio: can't use global as only "
1057                                         "section\n");
1058                                 do_exit++;
1059                                 exit_val = 1;
1060                                 break;
1061                         }
1062                         if (job_section)
1063                                 free(job_section);
1064                         job_section = strdup(optarg);
1065                         break;
1066                 case FIO_GETOPT_JOB: {
1067                         const char *opt = l_opts[lidx].name;
1068                         char *val = optarg;
1069
1070                         if (!strncmp(opt, "name", 4) && td) {
1071                                 ret = add_job(td, td->o.name ?: "fio", 0);
1072                                 if (ret) {
1073                                         put_job(td);
1074                                         return 0;
1075                                 }
1076                                 td = NULL;
1077                         }
1078                         if (!td) {
1079                                 int is_section = !strncmp(opt, "name", 4);
1080                                 int global = 0;
1081
1082                                 if (!is_section || !strncmp(val, "global", 6))
1083                                         global = 1;
1084
1085                                 if (is_section && skip_this_section(val))
1086                                         continue;
1087
1088                                 td = get_new_job(global, &def_thread);
1089                                 if (!td)
1090                                         return 0;
1091                         }
1092
1093                         ret = fio_cmd_option_parse(td, opt, val);
1094                         break;
1095                 }
1096                 default:
1097                         do_exit++;
1098                         exit_val = 1;
1099                         break;
1100                 }
1101         }
1102
1103         if (do_exit)
1104                 exit(exit_val);
1105
1106         if (td) {
1107                 if (!ret)
1108                         ret = add_job(td, td->o.name ?: "fio", 0);
1109                 if (ret)
1110                         put_job(td);
1111         }
1112
1113         while (optind < argc) {
1114                 ini_idx++;
1115                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1116                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1117                 optind++;
1118         }
1119
1120         return ini_idx;
1121 }
1122
1123 int parse_options(int argc, char *argv[])
1124 {
1125         int job_files, i;
1126
1127         f_out = stdout;
1128         f_err = stderr;
1129
1130         fio_options_dup_and_init(l_opts);
1131
1132         if (setup_thread_area())
1133                 return 1;
1134         if (fill_def_thread())
1135                 return 1;
1136
1137         job_files = parse_cmd_line(argc, argv);
1138
1139         for (i = 0; i < job_files; i++) {
1140                 if (fill_def_thread())
1141                         return 1;
1142                 if (parse_jobs_ini(ini_file[i], i))
1143                         return 1;
1144                 free(ini_file[i]);
1145         }
1146
1147         free(ini_file);
1148         options_mem_free(&def_thread);
1149
1150         if (!thread_number) {
1151                 if (dump_cmdline)
1152                         return 0;
1153
1154                 log_err("No jobs defined(s)\n\n");
1155                 usage(argv[0]);
1156                 return 1;
1157         }
1158
1159         if (def_thread.o.gtod_offload) {
1160                 fio_gtod_init();
1161                 fio_gtod_offload = 1;
1162                 fio_gtod_cpu = def_thread.o.gtod_cpu;
1163         }
1164
1165         return 0;
1166 }