cd32725c4b9893ed2983419e44835f707b4a1935
[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 1.28";
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 #ifndef FIO_HAVE_CPU_AFFINITY
237         if (td->o.gtod_cpu) {
238                 log_err("fio: platform must support CPU affinity for"
239                         "gettimeofday() offloading\n");
240                 return 1;
241         }
242 #endif
243
244 #if 0
245         if (read_only && td_write(td)) {
246                 log_err("fio: job <%s> has write bit set, but fio is in"
247                         " read-only mode\n", td->o.name);
248                 return 1;
249         }
250 #endif
251
252         if (o->write_iolog_file && o->read_iolog_file) {
253                 log_err("fio: read iolog overrides write_iolog\n");
254                 free(o->write_iolog_file);
255                 o->write_iolog_file = NULL;
256         }
257
258         /*
259          * only really works for sequential io for now, and with 1 file
260          */
261         if (o->zone_size && td_random(td) && o->open_files == 1)
262                 o->zone_size = 0;
263
264         /*
265          * Reads can do overwrites, we always need to pre-create the file
266          */
267         if (td_read(td) || td_rw(td))
268                 o->overwrite = 1;
269
270         if (!o->min_bs[DDIR_READ])
271                 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
272         if (!o->max_bs[DDIR_READ])
273                 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
274         if (!o->min_bs[DDIR_WRITE])
275                 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
276         if (!o->max_bs[DDIR_WRITE])
277                 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
278
279         o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
280
281         /*
282          * For random IO, allow blockalign offset other than min_bs.
283          */
284         if (!o->ba[DDIR_READ] || !td_random(td))
285                 o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
286         if (!o->ba[DDIR_WRITE] || !td_random(td))
287                 o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
288
289         if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
290             o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
291             !td->o.norandommap) {
292                 log_err("fio: Any use of blockalign= turns off randommap\n");
293                 td->o.norandommap = 1;
294         }
295
296         if (!o->file_size_high)
297                 o->file_size_high = o->file_size_low;
298
299         if (o->norandommap && o->verify != VERIFY_NONE
300             && !fixed_block_size(o))  {
301                 log_err("fio: norandommap given for variable block sizes, "
302                         "verify disabled\n");
303                 o->verify = VERIFY_NONE;
304         }
305         if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
306                 log_err("fio: bs_unaligned may not work with raw io\n");
307
308         /*
309          * thinktime_spin must be less than thinktime
310          */
311         if (o->thinktime_spin > o->thinktime)
312                 o->thinktime_spin = o->thinktime;
313
314         /*
315          * The low water mark cannot be bigger than the iodepth
316          */
317         if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
318                 /*
319                  * syslet work around - if the workload is sequential,
320                  * we want to let the queue drain all the way down to
321                  * avoid seeking between async threads
322                  */
323                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
324                         o->iodepth_low = 1;
325                 else
326                         o->iodepth_low = o->iodepth;
327         }
328
329         /*
330          * If batch number isn't set, default to the same as iodepth
331          */
332         if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
333                 o->iodepth_batch = o->iodepth;
334
335         if (o->nr_files > td->files_index)
336                 o->nr_files = td->files_index;
337
338         if (o->open_files > o->nr_files || !o->open_files)
339                 o->open_files = o->nr_files;
340
341         if (((o->rate[0] + o->rate[1]) && (o->rate_iops[0] + o->rate_iops[1]))||
342             ((o->ratemin[0] + o->ratemin[1]) && (o->rate_iops_min[0] +
343                 o->rate_iops_min[1]))) {
344                 log_err("fio: rate and rate_iops are mutually exclusive\n");
345                 return 1;
346         }
347         if ((o->rate[0] < o->ratemin[0]) || (o->rate[1] < o->ratemin[1]) ||
348             (o->rate_iops[0] < o->rate_iops_min[0]) ||
349             (o->rate_iops[1] < o->rate_iops_min[1])) {
350                 log_err("fio: minimum rate exceeds rate\n");
351                 return 1;
352         }
353
354         if (!o->timeout && o->time_based) {
355                 log_err("fio: time_based requires a runtime/timeout setting\n");
356                 o->time_based = 0;
357         }
358
359         if (o->fill_device && !o->size)
360                 o->size = -1ULL;
361
362         if (td_rw(td) && td->o.verify != VERIFY_NONE)
363                 log_info("fio: mixed read/write workload with verify. May not "
364                  "work as expected, unless you pre-populated the file\n");
365
366         if (td->o.verify != VERIFY_NONE)
367                 td->o.refill_buffers = 1;
368
369         if (td->o.pre_read)
370                 td->o.invalidate_cache = 0;
371
372         if (td->o.mem_align) {
373                 if (td->o.odirect && !is_power_of_2(td->o.mem_align)) {
374                         log_err("fio: given IO mem alignment conflicts with"
375                                         " direct=1. Resetting.\n");
376                         td->o.mem_align = page_mask;
377                 }
378         }
379                         
380         return 0;
381 }
382
383 /*
384  * This function leaks the buffer
385  */
386 static char *to_kmg(unsigned int val)
387 {
388         char *buf = malloc(32);
389         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
390         char *p = post;
391
392         do {
393                 if (val & 1023)
394                         break;
395
396                 val >>= 10;
397                 p++;
398         } while (*p);
399
400         snprintf(buf, 31, "%u%c", val, *p);
401         return buf;
402 }
403
404 /* External engines are specified by "external:name.o") */
405 static const char *get_engine_name(const char *str)
406 {
407         char *p = strstr(str, ":");
408
409         if (!p)
410                 return str;
411
412         p++;
413         strip_blank_front(&p);
414         strip_blank_end(p);
415         return p;
416 }
417
418 static int exists_and_not_file(const char *filename)
419 {
420         struct stat sb;
421
422         if (lstat(filename, &sb) == -1)
423                 return 0;
424
425         if (S_ISREG(sb.st_mode))
426                 return 0;
427
428         return 1;
429 }
430
431 void td_fill_rand_seeds(struct thread_data *td)
432 {
433         os_random_seed(td->rand_seeds[0], &td->bsrange_state);
434         os_random_seed(td->rand_seeds[1], &td->verify_state);
435         os_random_seed(td->rand_seeds[2], &td->rwmix_state);
436
437         if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
438                 os_random_seed(td->rand_seeds[3], &td->next_file_state);
439
440         os_random_seed(td->rand_seeds[5], &td->file_size_state);
441
442         if (!td_random(td))
443                 return;
444
445         if (td->o.rand_repeatable)
446                 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
447
448         os_random_seed(td->rand_seeds[4], &td->random_state);
449 }
450
451 /*
452  * Initialize the various random states we need (random io, block size ranges,
453  * read/write mix, etc).
454  */
455 static int init_random_state(struct thread_data *td)
456 {
457         int fd;
458
459         fd = open("/dev/urandom", O_RDONLY);
460         if (fd == -1) {
461                 td_verror(td, errno, "open");
462                 return 1;
463         }
464
465         if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
466             (int) sizeof(td->rand_seeds)) {
467                 td_verror(td, EIO, "read");
468                 close(fd);
469                 return 1;
470         }
471
472         close(fd);
473         td_fill_rand_seeds(td);
474         return 0;
475 }
476
477 /*
478  * Adds a job to the list of things todo. Sanitizes the various options
479  * to make sure we don't have conflicts, and initializes various
480  * members of td.
481  */
482 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
483 {
484         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
485                                    "randread", "randwrite", "randrw" };
486         unsigned int i;
487         const char *engine;
488         char fname[PATH_MAX];
489         int numjobs, file_alloced;
490
491         /*
492          * the def_thread is just for options, it's not a real job
493          */
494         if (td == &def_thread)
495                 return 0;
496
497         /*
498          * if we are just dumping the output command line, don't add the job
499          */
500         if (dump_cmdline) {
501                 put_job(td);
502                 return 0;
503         }
504
505         engine = get_engine_name(td->o.ioengine);
506         td->io_ops = load_ioengine(td, engine);
507         if (!td->io_ops) {
508                 log_err("fio: failed to load engine %s\n", engine);
509                 goto err;
510         }
511
512         if (td->o.use_thread)
513                 nr_thread++;
514         else
515                 nr_process++;
516
517         if (td->o.odirect)
518                 td->io_ops->flags |= FIO_RAWIO;
519
520         file_alloced = 0;
521         if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
522                 file_alloced = 1;
523
524                 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
525                         add_file(td, jobname);
526                 else {
527                         for (i = 0; i < td->o.nr_files; i++) {
528                                 sprintf(fname, "%s.%d.%d", jobname,
529                                                         td->thread_number, i);
530                                 add_file(td, fname);
531                         }
532                 }
533         }
534
535         if (fixup_options(td))
536                 goto err;
537
538         if (td->io_ops->flags & FIO_DISKLESSIO) {
539                 struct fio_file *f;
540
541                 for_each_file(td, f, i)
542                         f->real_file_size = -1ULL;
543         }
544
545         td->mutex = fio_mutex_init(0);
546
547         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
548         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
549         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
550         td->ddir_nr = td->o.ddir_nr;
551
552         if ((td->o.stonewall || td->o.numjobs > 1 || td->o.new_group)
553              && prev_group_jobs) {
554                 prev_group_jobs = 0;
555                 groupid++;
556         }
557
558         td->groupid = groupid;
559         prev_group_jobs++;
560
561         if (init_random_state(td))
562                 goto err;
563
564         if (setup_rate(td))
565                 goto err;
566
567         if (td->o.write_lat_log) {
568                 setup_log(&td->ts.slat_log);
569                 setup_log(&td->ts.clat_log);
570         }
571         if (td->o.write_bw_log)
572                 setup_log(&td->ts.bw_log);
573
574         if (!td->o.name)
575                 td->o.name = strdup(jobname);
576
577         if (!terse_output) {
578                 if (!job_add_num) {
579                         if (!strcmp(td->io_ops->name, "cpuio")) {
580                                 log_info("%s: ioengine=cpu, cpuload=%u,"
581                                          " cpucycle=%u\n", td->o.name,
582                                                         td->o.cpuload,
583                                                         td->o.cpucycle);
584                         } else {
585                                 char *c1, *c2, *c3, *c4;
586
587                                 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
588                                 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
589                                 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
590                                 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
591
592                                 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
593                                          " ioengine=%s, iodepth=%u\n",
594                                                 td->o.name, td->groupid,
595                                                 ddir_str[td->o.td_ddir],
596                                                 c1, c2, c3, c4,
597                                                 td->io_ops->name,
598                                                 td->o.iodepth);
599
600                                 free(c1);
601                                 free(c2);
602                                 free(c3);
603                                 free(c4);
604                         }
605                 } else if (job_add_num == 1)
606                         log_info("...\n");
607         }
608
609         /*
610          * recurse add identical jobs, clear numjobs and stonewall options
611          * as they don't apply to sub-jobs
612          */
613         numjobs = td->o.numjobs;
614         while (--numjobs) {
615                 struct thread_data *td_new = get_new_job(0, td);
616
617                 if (!td_new)
618                         goto err;
619
620                 td_new->o.numjobs = 1;
621                 td_new->o.stonewall = 0;
622                 td_new->o.new_group = 0;
623
624                 if (file_alloced) {
625                         td_new->o.filename = NULL;
626                         td_new->files_index = 0;
627                         td_new->files_size = 0;
628                         td_new->files = NULL;
629                 }
630
631                 job_add_num = numjobs - 1;
632
633                 if (add_job(td_new, jobname, job_add_num))
634                         goto err;
635         }
636
637         return 0;
638 err:
639         put_job(td);
640         return -1;
641 }
642
643 static int skip_this_section(const char *name)
644 {
645         if (!job_section)
646                 return 0;
647         if (!strncmp(name, "global", 6))
648                 return 0;
649
650         return strcmp(job_section, name);
651 }
652
653 static int is_empty_or_comment(char *line)
654 {
655         unsigned int i;
656
657         for (i = 0; i < strlen(line); i++) {
658                 if (line[i] == ';')
659                         return 1;
660                 if (line[i] == '#')
661                         return 1;
662                 if (!isspace(line[i]) && !iscntrl(line[i]))
663                         return 0;
664         }
665
666         return 1;
667 }
668
669 /*
670  * This is our [ini] type file parser.
671  */
672 static int parse_jobs_ini(char *file, int stonewall_flag)
673 {
674         unsigned int global;
675         struct thread_data *td;
676         char *string, *name;
677         FILE *f;
678         char *p;
679         int ret = 0, stonewall;
680         int first_sect = 1;
681         int skip_fgets = 0;
682         int inside_skip = 0;
683         char **opts;
684         int i, alloc_opts, num_opts;
685
686         if (!strcmp(file, "-"))
687                 f = stdin;
688         else
689                 f = fopen(file, "r");
690
691         if (!f) {
692                 perror("fopen job file");
693                 return 1;
694         }
695
696         string = malloc(4096);
697
698         /*
699          * it's really 256 + small bit, 280 should suffice
700          */
701         name = malloc(280);
702         memset(name, 0, 280);
703
704         alloc_opts = 8;
705         opts = malloc(sizeof(char *) * alloc_opts);
706         num_opts = 0;
707
708         stonewall = stonewall_flag;
709         do {
710                 /*
711                  * if skip_fgets is set, we already have loaded a line we
712                  * haven't handled.
713                  */
714                 if (!skip_fgets) {
715                         p = fgets(string, 4095, f);
716                         if (!p)
717                                 break;
718                 }
719
720                 skip_fgets = 0;
721                 strip_blank_front(&p);
722                 strip_blank_end(p);
723
724                 if (is_empty_or_comment(p))
725                         continue;
726                 if (sscanf(p, "[%255s]", name) != 1) {
727                         if (inside_skip)
728                                 continue;
729                         log_err("fio: option <%s> outside of [] job section\n",
730                                                                         p);
731                         break;
732                 }
733
734                 if (skip_this_section(name)) {
735                         inside_skip = 1;
736                         continue;
737                 } else
738                         inside_skip = 0;
739
740                 global = !strncmp(name, "global", 6);
741
742                 name[strlen(name) - 1] = '\0';
743
744                 if (dump_cmdline) {
745                         if (first_sect)
746                                 log_info("fio ");
747                         if (!global)
748                                 log_info("--name=%s ", name);
749                         first_sect = 0;
750                 }
751
752                 td = get_new_job(global, &def_thread);
753                 if (!td) {
754                         ret = 1;
755                         break;
756                 }
757
758                 /*
759                  * Seperate multiple job files by a stonewall
760                  */
761                 if (!global && stonewall) {
762                         td->o.stonewall = stonewall;
763                         stonewall = 0;
764                 }
765
766                 num_opts = 0;
767                 memset(opts, 0, alloc_opts * sizeof(char *));
768
769                 while ((p = fgets(string, 4096, f)) != NULL) {
770                         if (is_empty_or_comment(p))
771                                 continue;
772
773                         strip_blank_front(&p);
774
775                         /*
776                          * new section, break out and make sure we don't
777                          * fgets() a new line at the top.
778                          */
779                         if (p[0] == '[') {
780                                 skip_fgets = 1;
781                                 break;
782                         }
783
784                         strip_blank_end(p);
785
786                         if (num_opts == alloc_opts) {
787                                 alloc_opts <<= 1;
788                                 opts = realloc(opts,
789                                                 alloc_opts * sizeof(char *));
790                         }
791
792                         opts[num_opts] = strdup(p);
793                         num_opts++;
794                 }
795
796                 ret = fio_options_parse(td, opts, num_opts);
797                 if (!ret) {
798                         if (dump_cmdline)
799                                 for (i = 0; i < num_opts; i++)
800                                         log_info("--%s ", opts[i]);
801
802                         ret = add_job(td, name, 0);
803                 } else {
804                         log_err("fio: job %s dropped\n", name);
805                         put_job(td);
806                 }
807
808                 for (i = 0; i < num_opts; i++)
809                         free(opts[i]);
810                 num_opts = 0;
811         } while (!ret);
812
813         if (dump_cmdline)
814                 log_info("\n");
815
816         for (i = 0; i < num_opts; i++)
817                 free(opts[i]);
818
819         free(string);
820         free(name);
821         free(opts);
822         if (f != stdin)
823                 fclose(f);
824         return ret;
825 }
826
827 static int fill_def_thread(void)
828 {
829         memset(&def_thread, 0, sizeof(def_thread));
830
831         fio_getaffinity(getpid(), &def_thread.o.cpumask);
832
833         /*
834          * fill default options
835          */
836         fio_fill_default_options(&def_thread);
837
838         def_thread.o.timeout = def_timeout;
839         def_thread.o.write_bw_log = write_bw_log;
840         def_thread.o.write_lat_log = write_lat_log;
841
842         return 0;
843 }
844
845 static void free_shm(void)
846 {
847         struct shmid_ds sbuf;
848
849         if (threads) {
850                 void *tp = threads;
851
852                 threads = NULL;
853                 file_hash_exit();
854                 fio_debug_jobp = NULL;
855                 shmdt(tp);
856                 shmctl(shm_id, IPC_RMID, &sbuf);
857         }
858
859         scleanup();
860 }
861
862 /*
863  * The thread area is shared between the main process and the job
864  * threads/processes. So setup a shared memory segment that will hold
865  * all the job info. We use the end of the region for keeping track of
866  * open files across jobs, for file sharing.
867  */
868 static int setup_thread_area(void)
869 {
870         void *hash;
871
872         /*
873          * 1024 is too much on some machines, scale max_jobs if
874          * we get a failure that looks like too large a shm segment
875          */
876         do {
877                 size_t size = max_jobs * sizeof(struct thread_data);
878
879                 size += file_hash_size;
880                 size += sizeof(unsigned int);
881
882                 shm_id = shmget(0, size, IPC_CREAT | 0600);
883                 if (shm_id != -1)
884                         break;
885                 if (errno != EINVAL) {
886                         perror("shmget");
887                         break;
888                 }
889
890                 max_jobs >>= 1;
891         } while (max_jobs);
892
893         if (shm_id == -1)
894                 return 1;
895
896         threads = shmat(shm_id, NULL, 0);
897         if (threads == (void *) -1) {
898                 perror("shmat");
899                 return 1;
900         }
901
902         memset(threads, 0, max_jobs * sizeof(struct thread_data));
903         hash = (void *) threads + max_jobs * sizeof(struct thread_data);
904         fio_debug_jobp = (void *) hash + file_hash_size;
905         *fio_debug_jobp = -1;
906         file_hash_init(hash);
907         atexit(free_shm);
908         return 0;
909 }
910
911 static void usage(const char *name)
912 {
913         printf("%s\n", fio_version_string);
914         printf("%s [options] [job options] <job file(s)>\n", name);
915         printf("\t--debug=options\tEnable debug logging\n");
916         printf("\t--output\tWrite output to file\n");
917         printf("\t--timeout\tRuntime in seconds\n");
918         printf("\t--latency-log\tGenerate per-job latency logs\n");
919         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
920         printf("\t--minimal\tMinimal (terse) output\n");
921         printf("\t--version\tPrint version info and exit\n");
922         printf("\t--help\t\tPrint this page\n");
923         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
924                 " them\n");
925         printf("\t--showcmd\tTurn a job file into command line options\n");
926         printf("\t--eta=when\tWhen ETA estimate should be printed\n");
927         printf("\t          \tMay be \"always\", \"never\" or \"auto\"\n");
928         printf("\t--readonly\tTurn on safety read-only checks, preventing"
929                 " writes\n");
930         printf("\t--section=name\tOnly run specified section in job file\n");
931         printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
932                 " (def 1024)\n");
933         printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>\n");
934 }
935
936 #ifdef FIO_INC_DEBUG
937 struct debug_level debug_levels[] = {
938         { .name = "process",    .shift = FD_PROCESS, },
939         { .name = "file",       .shift = FD_FILE, },
940         { .name = "io",         .shift = FD_IO, },
941         { .name = "mem",        .shift = FD_MEM, },
942         { .name = "blktrace",   .shift = FD_BLKTRACE },
943         { .name = "verify",     .shift = FD_VERIFY },
944         { .name = "random",     .shift = FD_RANDOM },
945         { .name = "parse",      .shift = FD_PARSE },
946         { .name = "diskutil",   .shift = FD_DISKUTIL },
947         { .name = "job",        .shift = FD_JOB },
948         { .name = "mutex",      .shift = FD_MUTEX },
949         { .name = NULL, },
950 };
951
952 static int set_debug(const char *string)
953 {
954         struct debug_level *dl;
955         char *p = (char *) string;
956         char *opt;
957         int i;
958
959         if (!strcmp(string, "?") || !strcmp(string, "help")) {
960                 int i;
961
962                 log_info("fio: dumping debug options:");
963                 for (i = 0; debug_levels[i].name; i++) {
964                         dl = &debug_levels[i];
965                         log_info("%s,", dl->name);
966                 }
967                 log_info("all\n");
968                 return 1;
969         }
970
971         while ((opt = strsep(&p, ",")) != NULL) {
972                 int found = 0;
973
974                 if (!strncmp(opt, "all", 3)) {
975                         log_info("fio: set all debug options\n");
976                         fio_debug = ~0UL;
977                         continue;
978                 }
979
980                 for (i = 0; debug_levels[i].name; i++) {
981                         dl = &debug_levels[i];
982                         found = !strncmp(opt, dl->name, strlen(dl->name));
983                         if (!found)
984                                 continue;
985
986                         if (dl->shift == FD_JOB) {
987                                 opt = strchr(opt, ':');
988                                 if (!opt) {
989                                         log_err("fio: missing job number\n");
990                                         break;
991                                 }
992                                 opt++;
993                                 fio_debug_jobno = atoi(opt);
994                                 log_info("fio: set debug jobno %d\n",
995                                                         fio_debug_jobno);
996                         } else {
997                                 log_info("fio: set debug option %s\n", opt);
998                                 fio_debug |= (1UL << dl->shift);
999                         }
1000                         break;
1001                 }
1002
1003                 if (!found)
1004                         log_err("fio: debug mask %s not found\n", opt);
1005         }
1006         return 0;
1007 }
1008 #else
1009 static int set_debug(const char *string)
1010 {
1011         log_err("fio: debug tracing not included in build\n");
1012         return 1;
1013 }
1014 #endif
1015
1016 static int parse_cmd_line(int argc, char *argv[])
1017 {
1018         struct thread_data *td = NULL;
1019         int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1020
1021         while ((c = getopt_long_only(argc, argv, "", l_opts, &lidx)) != -1) {
1022                 switch (c) {
1023                 case 'a':
1024                         smalloc_pool_size = atoi(optarg);
1025                         break;
1026                 case 't':
1027                         def_timeout = atoi(optarg);
1028                         break;
1029                 case 'l':
1030                         write_lat_log = 1;
1031                         break;
1032                 case 'w':
1033                         write_bw_log = 1;
1034                         break;
1035                 case 'o':
1036                         f_out = fopen(optarg, "w+");
1037                         if (!f_out) {
1038                                 perror("fopen output");
1039                                 exit(1);
1040                         }
1041                         f_err = f_out;
1042                         break;
1043                 case 'm':
1044                         terse_output = 1;
1045                         break;
1046                 case 'h':
1047                         usage(argv[0]);
1048                         exit(0);
1049                 case 'c':
1050                         exit(fio_show_option_help(optarg));
1051                 case 's':
1052                         dump_cmdline = 1;
1053                         break;
1054                 case 'r':
1055                         read_only = 1;
1056                         break;
1057                 case 'v':
1058                         printf("%s\n", fio_version_string);
1059                         exit(0);
1060                 case 'e':
1061                         if (!strcmp("always", optarg))
1062                                 eta_print = FIO_ETA_ALWAYS;
1063                         else if (!strcmp("never", optarg))
1064                                 eta_print = FIO_ETA_NEVER;
1065                         break;
1066                 case 'd':
1067                         if (set_debug(optarg))
1068                                 do_exit++;
1069                         break;
1070                 case 'x':
1071                         if (!strcmp(optarg, "global")) {
1072                                 log_err("fio: can't use global as only "
1073                                         "section\n");
1074                                 do_exit++;
1075                                 exit_val = 1;
1076                                 break;
1077                         }
1078                         if (job_section)
1079                                 free(job_section);
1080                         job_section = strdup(optarg);
1081                         break;
1082                 case FIO_GETOPT_JOB: {
1083                         const char *opt = l_opts[lidx].name;
1084                         char *val = optarg;
1085
1086                         if (!strncmp(opt, "name", 4) && td) {
1087                                 ret = add_job(td, td->o.name ?: "fio", 0);
1088                                 if (ret) {
1089                                         put_job(td);
1090                                         return 0;
1091                                 }
1092                                 td = NULL;
1093                         }
1094                         if (!td) {
1095                                 int is_section = !strncmp(opt, "name", 4);
1096                                 int global = 0;
1097
1098                                 if (!is_section || !strncmp(val, "global", 6))
1099                                         global = 1;
1100
1101                                 if (is_section && skip_this_section(val))
1102                                         continue;
1103
1104                                 td = get_new_job(global, &def_thread);
1105                                 if (!td)
1106                                         return 0;
1107                         }
1108
1109                         ret = fio_cmd_option_parse(td, opt, val);
1110                         break;
1111                 }
1112                 default:
1113                         do_exit++;
1114                         exit_val = 1;
1115                         break;
1116                 }
1117         }
1118
1119         if (do_exit)
1120                 exit(exit_val);
1121
1122         if (td) {
1123                 if (!ret)
1124                         ret = add_job(td, td->o.name ?: "fio", 0);
1125                 if (ret)
1126                         put_job(td);
1127         }
1128
1129         while (optind < argc) {
1130                 ini_idx++;
1131                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1132                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1133                 optind++;
1134         }
1135
1136         return ini_idx;
1137 }
1138
1139 int parse_options(int argc, char *argv[])
1140 {
1141         int job_files, i;
1142
1143         f_out = stdout;
1144         f_err = stderr;
1145
1146         fio_options_dup_and_init(l_opts);
1147
1148         if (setup_thread_area())
1149                 return 1;
1150         if (fill_def_thread())
1151                 return 1;
1152
1153         job_files = parse_cmd_line(argc, argv);
1154
1155         for (i = 0; i < job_files; i++) {
1156                 if (fill_def_thread())
1157                         return 1;
1158                 if (parse_jobs_ini(ini_file[i], i))
1159                         return 1;
1160                 free(ini_file[i]);
1161         }
1162
1163         free(ini_file);
1164         options_mem_free(&def_thread);
1165
1166         if (!thread_number) {
1167                 if (dump_cmdline)
1168                         return 0;
1169
1170                 log_err("No jobs defined(s)\n\n");
1171                 usage(argv[0]);
1172                 return 1;
1173         }
1174
1175         if (def_thread.o.gtod_offload) {
1176                 fio_gtod_init();
1177                 fio_gtod_offload = 1;
1178                 fio_gtod_cpu = def_thread.o.gtod_cpu;
1179         }
1180
1181         return 0;
1182 }