Add code to detect a task that exited prior to up'ing the startup mutex
[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                 if (td->io_ops->flags & FIO_PIPEIO)
356                         log_info("fio: cannot pre-read files with an IO engine"
357                                  " that isn't seekable. Pre-read disabled.\n");
358         }
359
360         if (td->o.mem_align) {
361                 if (td->o.odirect && !is_power_of_2(td->o.mem_align)) {
362                         log_err("fio: given IO mem alignment conflicts with"
363                                         " direct=1. Resetting.\n");
364                         td->o.mem_align = page_mask;
365                 }
366         }
367                         
368         return 0;
369 }
370
371 /*
372  * This function leaks the buffer
373  */
374 static char *to_kmg(unsigned int val)
375 {
376         char *buf = malloc(32);
377         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
378         char *p = post;
379
380         do {
381                 if (val & 1023)
382                         break;
383
384                 val >>= 10;
385                 p++;
386         } while (*p);
387
388         snprintf(buf, 31, "%u%c", val, *p);
389         return buf;
390 }
391
392 /* External engines are specified by "external:name.o") */
393 static const char *get_engine_name(const char *str)
394 {
395         char *p = strstr(str, ":");
396
397         if (!p)
398                 return str;
399
400         p++;
401         strip_blank_front(&p);
402         strip_blank_end(p);
403         return p;
404 }
405
406 static int exists_and_not_file(const char *filename)
407 {
408         struct stat sb;
409
410         if (lstat(filename, &sb) == -1)
411                 return 0;
412
413         if (S_ISREG(sb.st_mode))
414                 return 0;
415
416         return 1;
417 }
418
419 void td_fill_rand_seeds(struct thread_data *td)
420 {
421         os_random_seed(td->rand_seeds[0], &td->bsrange_state);
422         os_random_seed(td->rand_seeds[1], &td->verify_state);
423         os_random_seed(td->rand_seeds[2], &td->rwmix_state);
424
425         if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
426                 os_random_seed(td->rand_seeds[3], &td->next_file_state);
427
428         os_random_seed(td->rand_seeds[5], &td->file_size_state);
429
430         if (!td_random(td))
431                 return;
432
433         if (td->o.rand_repeatable)
434                 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
435
436         os_random_seed(td->rand_seeds[4], &td->random_state);
437 }
438
439 /*
440  * Initialize the various random states we need (random io, block size ranges,
441  * read/write mix, etc).
442  */
443 static int init_random_state(struct thread_data *td)
444 {
445         int fd;
446
447         fd = open("/dev/urandom", O_RDONLY);
448         if (fd == -1) {
449                 td_verror(td, errno, "open");
450                 return 1;
451         }
452
453         if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
454             (int) sizeof(td->rand_seeds)) {
455                 td_verror(td, EIO, "read");
456                 close(fd);
457                 return 1;
458         }
459
460         close(fd);
461         td_fill_rand_seeds(td);
462         return 0;
463 }
464
465 /*
466  * Adds a job to the list of things todo. Sanitizes the various options
467  * to make sure we don't have conflicts, and initializes various
468  * members of td.
469  */
470 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
471 {
472         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
473                                    "randread", "randwrite", "randrw" };
474         unsigned int i;
475         const char *engine;
476         char fname[PATH_MAX];
477         int numjobs, file_alloced;
478
479         /*
480          * the def_thread is just for options, it's not a real job
481          */
482         if (td == &def_thread)
483                 return 0;
484
485         /*
486          * if we are just dumping the output command line, don't add the job
487          */
488         if (dump_cmdline) {
489                 put_job(td);
490                 return 0;
491         }
492
493         engine = get_engine_name(td->o.ioengine);
494         td->io_ops = load_ioengine(td, engine);
495         if (!td->io_ops) {
496                 log_err("fio: failed to load engine %s\n", engine);
497                 goto err;
498         }
499
500         if (td->o.use_thread)
501                 nr_thread++;
502         else
503                 nr_process++;
504
505         if (td->o.odirect)
506                 td->io_ops->flags |= FIO_RAWIO;
507
508         file_alloced = 0;
509         if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
510                 file_alloced = 1;
511
512                 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
513                         add_file(td, jobname);
514                 else {
515                         for (i = 0; i < td->o.nr_files; i++) {
516                                 sprintf(fname, "%s.%d.%d", jobname,
517                                                         td->thread_number, i);
518                                 add_file(td, fname);
519                         }
520                 }
521         }
522
523         if (fixup_options(td))
524                 goto err;
525
526         if (td->io_ops->flags & FIO_DISKLESSIO) {
527                 struct fio_file *f;
528
529                 for_each_file(td, f, i)
530                         f->real_file_size = -1ULL;
531         }
532
533         td->mutex = fio_mutex_init(0);
534
535         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
536         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
537         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
538         td->ddir_nr = td->o.ddir_nr;
539
540         if ((td->o.stonewall || td->o.numjobs > 1 || td->o.new_group)
541              && prev_group_jobs) {
542                 prev_group_jobs = 0;
543                 groupid++;
544         }
545
546         td->groupid = groupid;
547         prev_group_jobs++;
548
549         if (init_random_state(td))
550                 goto err;
551
552         if (setup_rate(td))
553                 goto err;
554
555         if (td->o.write_lat_log) {
556                 setup_log(&td->ts.slat_log);
557                 setup_log(&td->ts.clat_log);
558         }
559         if (td->o.write_bw_log)
560                 setup_log(&td->ts.bw_log);
561
562         if (!td->o.name)
563                 td->o.name = strdup(jobname);
564
565         if (!terse_output) {
566                 if (!job_add_num) {
567                         if (!strcmp(td->io_ops->name, "cpuio")) {
568                                 log_info("%s: ioengine=cpu, cpuload=%u,"
569                                          " cpucycle=%u\n", td->o.name,
570                                                         td->o.cpuload,
571                                                         td->o.cpucycle);
572                         } else {
573                                 char *c1, *c2, *c3, *c4;
574
575                                 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
576                                 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
577                                 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
578                                 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
579
580                                 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
581                                          " ioengine=%s, iodepth=%u\n",
582                                                 td->o.name, td->groupid,
583                                                 ddir_str[td->o.td_ddir],
584                                                 c1, c2, c3, c4,
585                                                 td->io_ops->name,
586                                                 td->o.iodepth);
587
588                                 free(c1);
589                                 free(c2);
590                                 free(c3);
591                                 free(c4);
592                         }
593                 } else if (job_add_num == 1)
594                         log_info("...\n");
595         }
596
597         /*
598          * recurse add identical jobs, clear numjobs and stonewall options
599          * as they don't apply to sub-jobs
600          */
601         numjobs = td->o.numjobs;
602         while (--numjobs) {
603                 struct thread_data *td_new = get_new_job(0, td);
604
605                 if (!td_new)
606                         goto err;
607
608                 td_new->o.numjobs = 1;
609                 td_new->o.stonewall = 0;
610                 td_new->o.new_group = 0;
611
612                 if (file_alloced) {
613                         td_new->o.filename = NULL;
614                         td_new->files_index = 0;
615                         td_new->files_size = 0;
616                         td_new->files = NULL;
617                 }
618
619                 job_add_num = numjobs - 1;
620
621                 if (add_job(td_new, jobname, job_add_num))
622                         goto err;
623         }
624
625         return 0;
626 err:
627         put_job(td);
628         return -1;
629 }
630
631 static int skip_this_section(const char *name)
632 {
633         if (!job_section)
634                 return 0;
635         if (!strncmp(name, "global", 6))
636                 return 0;
637
638         return strcmp(job_section, name);
639 }
640
641 static int is_empty_or_comment(char *line)
642 {
643         unsigned int i;
644
645         for (i = 0; i < strlen(line); i++) {
646                 if (line[i] == ';')
647                         return 1;
648                 if (line[i] == '#')
649                         return 1;
650                 if (!isspace(line[i]) && !iscntrl(line[i]))
651                         return 0;
652         }
653
654         return 1;
655 }
656
657 /*
658  * This is our [ini] type file parser.
659  */
660 static int parse_jobs_ini(char *file, int stonewall_flag)
661 {
662         unsigned int global;
663         struct thread_data *td;
664         char *string, *name;
665         FILE *f;
666         char *p;
667         int ret = 0, stonewall;
668         int first_sect = 1;
669         int skip_fgets = 0;
670         int inside_skip = 0;
671         char **opts;
672         int i, alloc_opts, num_opts;
673
674         if (!strcmp(file, "-"))
675                 f = stdin;
676         else
677                 f = fopen(file, "r");
678
679         if (!f) {
680                 perror("fopen job file");
681                 return 1;
682         }
683
684         string = malloc(4096);
685
686         /*
687          * it's really 256 + small bit, 280 should suffice
688          */
689         name = malloc(280);
690         memset(name, 0, 280);
691
692         alloc_opts = 8;
693         opts = malloc(sizeof(char *) * alloc_opts);
694         num_opts = 0;
695
696         stonewall = stonewall_flag;
697         do {
698                 /*
699                  * if skip_fgets is set, we already have loaded a line we
700                  * haven't handled.
701                  */
702                 if (!skip_fgets) {
703                         p = fgets(string, 4095, f);
704                         if (!p)
705                                 break;
706                 }
707
708                 skip_fgets = 0;
709                 strip_blank_front(&p);
710                 strip_blank_end(p);
711
712                 if (is_empty_or_comment(p))
713                         continue;
714                 if (sscanf(p, "[%255s]", name) != 1) {
715                         if (inside_skip)
716                                 continue;
717                         log_err("fio: option <%s> outside of [] job section\n",
718                                                                         p);
719                         break;
720                 }
721
722                 if (skip_this_section(name)) {
723                         inside_skip = 1;
724                         continue;
725                 } else
726                         inside_skip = 0;
727
728                 global = !strncmp(name, "global", 6);
729
730                 name[strlen(name) - 1] = '\0';
731
732                 if (dump_cmdline) {
733                         if (first_sect)
734                                 log_info("fio ");
735                         if (!global)
736                                 log_info("--name=%s ", name);
737                         first_sect = 0;
738                 }
739
740                 td = get_new_job(global, &def_thread);
741                 if (!td) {
742                         ret = 1;
743                         break;
744                 }
745
746                 /*
747                  * Seperate multiple job files by a stonewall
748                  */
749                 if (!global && stonewall) {
750                         td->o.stonewall = stonewall;
751                         stonewall = 0;
752                 }
753
754                 num_opts = 0;
755                 memset(opts, 0, alloc_opts * sizeof(char *));
756
757                 while ((p = fgets(string, 4096, f)) != NULL) {
758                         if (is_empty_or_comment(p))
759                                 continue;
760
761                         strip_blank_front(&p);
762
763                         /*
764                          * new section, break out and make sure we don't
765                          * fgets() a new line at the top.
766                          */
767                         if (p[0] == '[') {
768                                 skip_fgets = 1;
769                                 break;
770                         }
771
772                         strip_blank_end(p);
773
774                         if (num_opts == alloc_opts) {
775                                 alloc_opts <<= 1;
776                                 opts = realloc(opts,
777                                                 alloc_opts * sizeof(char *));
778                         }
779
780                         opts[num_opts] = strdup(p);
781                         num_opts++;
782                 }
783
784                 ret = fio_options_parse(td, opts, num_opts);
785                 if (!ret) {
786                         if (dump_cmdline)
787                                 for (i = 0; i < num_opts; i++)
788                                         log_info("--%s ", opts[i]);
789
790                         ret = add_job(td, name, 0);
791                 } else {
792                         log_err("fio: job %s dropped\n", name);
793                         put_job(td);
794                 }
795
796                 for (i = 0; i < num_opts; i++)
797                         free(opts[i]);
798                 num_opts = 0;
799         } while (!ret);
800
801         if (dump_cmdline)
802                 log_info("\n");
803
804         for (i = 0; i < num_opts; i++)
805                 free(opts[i]);
806
807         free(string);
808         free(name);
809         free(opts);
810         if (f != stdin)
811                 fclose(f);
812         return ret;
813 }
814
815 static int fill_def_thread(void)
816 {
817         memset(&def_thread, 0, sizeof(def_thread));
818
819         fio_getaffinity(getpid(), &def_thread.o.cpumask);
820
821         /*
822          * fill default options
823          */
824         fio_fill_default_options(&def_thread);
825
826         def_thread.o.timeout = def_timeout;
827         def_thread.o.write_bw_log = write_bw_log;
828         def_thread.o.write_lat_log = write_lat_log;
829
830         return 0;
831 }
832
833 static void free_shm(void)
834 {
835         struct shmid_ds sbuf;
836
837         if (threads) {
838                 void *tp = threads;
839
840                 threads = NULL;
841                 file_hash_exit();
842                 fio_debug_jobp = NULL;
843                 shmdt(tp);
844                 shmctl(shm_id, IPC_RMID, &sbuf);
845         }
846
847         scleanup();
848 }
849
850 /*
851  * The thread area is shared between the main process and the job
852  * threads/processes. So setup a shared memory segment that will hold
853  * all the job info. We use the end of the region for keeping track of
854  * open files across jobs, for file sharing.
855  */
856 static int setup_thread_area(void)
857 {
858         void *hash;
859
860         /*
861          * 1024 is too much on some machines, scale max_jobs if
862          * we get a failure that looks like too large a shm segment
863          */
864         do {
865                 size_t size = max_jobs * sizeof(struct thread_data);
866
867                 size += file_hash_size;
868                 size += sizeof(unsigned int);
869
870                 shm_id = shmget(0, size, IPC_CREAT | 0600);
871                 if (shm_id != -1)
872                         break;
873                 if (errno != EINVAL) {
874                         perror("shmget");
875                         break;
876                 }
877
878                 max_jobs >>= 1;
879         } while (max_jobs);
880
881         if (shm_id == -1)
882                 return 1;
883
884         threads = shmat(shm_id, NULL, 0);
885         if (threads == (void *) -1) {
886                 perror("shmat");
887                 return 1;
888         }
889
890         memset(threads, 0, max_jobs * sizeof(struct thread_data));
891         hash = (void *) threads + max_jobs * sizeof(struct thread_data);
892         fio_debug_jobp = (void *) hash + file_hash_size;
893         *fio_debug_jobp = -1;
894         file_hash_init(hash);
895         atexit(free_shm);
896         return 0;
897 }
898
899 static void usage(const char *name)
900 {
901         printf("%s\n", fio_version_string);
902         printf("%s [options] [job options] <job file(s)>\n", name);
903         printf("\t--debug=options\tEnable debug logging\n");
904         printf("\t--output\tWrite output to file\n");
905         printf("\t--timeout\tRuntime in seconds\n");
906         printf("\t--latency-log\tGenerate per-job latency logs\n");
907         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
908         printf("\t--minimal\tMinimal (terse) output\n");
909         printf("\t--version\tPrint version info and exit\n");
910         printf("\t--help\t\tPrint this page\n");
911         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
912                 " them\n");
913         printf("\t--showcmd\tTurn a job file into command line options\n");
914         printf("\t--eta=when\tWhen ETA estimate should be printed\n");
915         printf("\t          \tMay be \"always\", \"never\" or \"auto\"\n");
916         printf("\t--readonly\tTurn on safety read-only checks, preventing"
917                 " writes\n");
918         printf("\t--section=name\tOnly run specified section in job file\n");
919         printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
920                 " (def 1024)\n");
921         printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>\n");
922 }
923
924 #ifdef FIO_INC_DEBUG
925 struct debug_level debug_levels[] = {
926         { .name = "process",    .shift = FD_PROCESS, },
927         { .name = "file",       .shift = FD_FILE, },
928         { .name = "io",         .shift = FD_IO, },
929         { .name = "mem",        .shift = FD_MEM, },
930         { .name = "blktrace",   .shift = FD_BLKTRACE },
931         { .name = "verify",     .shift = FD_VERIFY },
932         { .name = "random",     .shift = FD_RANDOM },
933         { .name = "parse",      .shift = FD_PARSE },
934         { .name = "diskutil",   .shift = FD_DISKUTIL },
935         { .name = "job",        .shift = FD_JOB },
936         { .name = "mutex",      .shift = FD_MUTEX },
937         { .name = NULL, },
938 };
939
940 static int set_debug(const char *string)
941 {
942         struct debug_level *dl;
943         char *p = (char *) string;
944         char *opt;
945         int i;
946
947         if (!strcmp(string, "?") || !strcmp(string, "help")) {
948                 int i;
949
950                 log_info("fio: dumping debug options:");
951                 for (i = 0; debug_levels[i].name; i++) {
952                         dl = &debug_levels[i];
953                         log_info("%s,", dl->name);
954                 }
955                 log_info("all\n");
956                 return 1;
957         }
958
959         while ((opt = strsep(&p, ",")) != NULL) {
960                 int found = 0;
961
962                 if (!strncmp(opt, "all", 3)) {
963                         log_info("fio: set all debug options\n");
964                         fio_debug = ~0UL;
965                         continue;
966                 }
967
968                 for (i = 0; debug_levels[i].name; i++) {
969                         dl = &debug_levels[i];
970                         found = !strncmp(opt, dl->name, strlen(dl->name));
971                         if (!found)
972                                 continue;
973
974                         if (dl->shift == FD_JOB) {
975                                 opt = strchr(opt, ':');
976                                 if (!opt) {
977                                         log_err("fio: missing job number\n");
978                                         break;
979                                 }
980                                 opt++;
981                                 fio_debug_jobno = atoi(opt);
982                                 log_info("fio: set debug jobno %d\n",
983                                                         fio_debug_jobno);
984                         } else {
985                                 log_info("fio: set debug option %s\n", opt);
986                                 fio_debug |= (1UL << dl->shift);
987                         }
988                         break;
989                 }
990
991                 if (!found)
992                         log_err("fio: debug mask %s not found\n", opt);
993         }
994         return 0;
995 }
996 #else
997 static int set_debug(const char *string)
998 {
999         log_err("fio: debug tracing not included in build\n");
1000         return 1;
1001 }
1002 #endif
1003
1004 static int parse_cmd_line(int argc, char *argv[])
1005 {
1006         struct thread_data *td = NULL;
1007         int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1008
1009         while ((c = getopt_long_only(argc, argv, "", l_opts, &lidx)) != -1) {
1010                 switch (c) {
1011                 case 'a':
1012                         smalloc_pool_size = atoi(optarg);
1013                         break;
1014                 case 't':
1015                         def_timeout = atoi(optarg);
1016                         break;
1017                 case 'l':
1018                         write_lat_log = 1;
1019                         break;
1020                 case 'w':
1021                         write_bw_log = 1;
1022                         break;
1023                 case 'o':
1024                         f_out = fopen(optarg, "w+");
1025                         if (!f_out) {
1026                                 perror("fopen output");
1027                                 exit(1);
1028                         }
1029                         f_err = f_out;
1030                         break;
1031                 case 'm':
1032                         terse_output = 1;
1033                         break;
1034                 case 'h':
1035                         usage(argv[0]);
1036                         exit(0);
1037                 case 'c':
1038                         exit(fio_show_option_help(optarg));
1039                 case 's':
1040                         dump_cmdline = 1;
1041                         break;
1042                 case 'r':
1043                         read_only = 1;
1044                         break;
1045                 case 'v':
1046                         printf("%s\n", fio_version_string);
1047                         exit(0);
1048                 case 'e':
1049                         if (!strcmp("always", optarg))
1050                                 eta_print = FIO_ETA_ALWAYS;
1051                         else if (!strcmp("never", optarg))
1052                                 eta_print = FIO_ETA_NEVER;
1053                         break;
1054                 case 'd':
1055                         if (set_debug(optarg))
1056                                 do_exit++;
1057                         break;
1058                 case 'x':
1059                         if (!strcmp(optarg, "global")) {
1060                                 log_err("fio: can't use global as only "
1061                                         "section\n");
1062                                 do_exit++;
1063                                 exit_val = 1;
1064                                 break;
1065                         }
1066                         if (job_section)
1067                                 free(job_section);
1068                         job_section = strdup(optarg);
1069                         break;
1070                 case FIO_GETOPT_JOB: {
1071                         const char *opt = l_opts[lidx].name;
1072                         char *val = optarg;
1073
1074                         if (!strncmp(opt, "name", 4) && td) {
1075                                 ret = add_job(td, td->o.name ?: "fio", 0);
1076                                 if (ret) {
1077                                         put_job(td);
1078                                         return 0;
1079                                 }
1080                                 td = NULL;
1081                         }
1082                         if (!td) {
1083                                 int is_section = !strncmp(opt, "name", 4);
1084                                 int global = 0;
1085
1086                                 if (!is_section || !strncmp(val, "global", 6))
1087                                         global = 1;
1088
1089                                 if (is_section && skip_this_section(val))
1090                                         continue;
1091
1092                                 td = get_new_job(global, &def_thread);
1093                                 if (!td)
1094                                         return 0;
1095                         }
1096
1097                         ret = fio_cmd_option_parse(td, opt, val);
1098                         break;
1099                 }
1100                 default:
1101                         do_exit++;
1102                         exit_val = 1;
1103                         break;
1104                 }
1105         }
1106
1107         if (do_exit)
1108                 exit(exit_val);
1109
1110         if (td) {
1111                 if (!ret)
1112                         ret = add_job(td, td->o.name ?: "fio", 0);
1113                 if (ret)
1114                         put_job(td);
1115         }
1116
1117         while (optind < argc) {
1118                 ini_idx++;
1119                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1120                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1121                 optind++;
1122         }
1123
1124         return ini_idx;
1125 }
1126
1127 int parse_options(int argc, char *argv[])
1128 {
1129         int job_files, i;
1130
1131         f_out = stdout;
1132         f_err = stderr;
1133
1134         fio_options_dup_and_init(l_opts);
1135
1136         if (setup_thread_area())
1137                 return 1;
1138         if (fill_def_thread())
1139                 return 1;
1140
1141         job_files = parse_cmd_line(argc, argv);
1142
1143         for (i = 0; i < job_files; i++) {
1144                 if (fill_def_thread())
1145                         return 1;
1146                 if (parse_jobs_ini(ini_file[i], i))
1147                         return 1;
1148                 free(ini_file[i]);
1149         }
1150
1151         free(ini_file);
1152         options_mem_free(&def_thread);
1153
1154         if (!thread_number) {
1155                 if (dump_cmdline)
1156                         return 0;
1157
1158                 log_err("No jobs defined(s)\n\n");
1159                 usage(argv[0]);
1160                 return 1;
1161         }
1162
1163         if (def_thread.o.gtod_offload) {
1164                 fio_gtod_init();
1165                 fio_gtod_offload = 1;
1166                 fio_gtod_cpu = def_thread.o.gtod_cpu;
1167         }
1168
1169         return 0;
1170 }