Solaris cpu affinity fix
[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 1.32";
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         return 0;
361 }
362
363 /*
364  * This function leaks the buffer
365  */
366 static char *to_kmg(unsigned int val)
367 {
368         char *buf = malloc(32);
369         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
370         char *p = post;
371
372         do {
373                 if (val & 1023)
374                         break;
375
376                 val >>= 10;
377                 p++;
378         } while (*p);
379
380         snprintf(buf, 31, "%u%c", val, *p);
381         return buf;
382 }
383
384 /* External engines are specified by "external:name.o") */
385 static const char *get_engine_name(const char *str)
386 {
387         char *p = strstr(str, ":");
388
389         if (!p)
390                 return str;
391
392         p++;
393         strip_blank_front(&p);
394         strip_blank_end(p);
395         return p;
396 }
397
398 static int exists_and_not_file(const char *filename)
399 {
400         struct stat sb;
401
402         if (lstat(filename, &sb) == -1)
403                 return 0;
404
405         if (S_ISREG(sb.st_mode))
406                 return 0;
407
408         return 1;
409 }
410
411 void td_fill_rand_seeds(struct thread_data *td)
412 {
413         os_random_seed(td->rand_seeds[0], &td->bsrange_state);
414         os_random_seed(td->rand_seeds[1], &td->verify_state);
415         os_random_seed(td->rand_seeds[2], &td->rwmix_state);
416
417         if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
418                 os_random_seed(td->rand_seeds[3], &td->next_file_state);
419
420         os_random_seed(td->rand_seeds[5], &td->file_size_state);
421
422         if (!td_random(td))
423                 return;
424
425         if (td->o.rand_repeatable)
426                 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
427
428         os_random_seed(td->rand_seeds[4], &td->random_state);
429 }
430
431 /*
432  * Initialize the various random states we need (random io, block size ranges,
433  * read/write mix, etc).
434  */
435 static int init_random_state(struct thread_data *td)
436 {
437         int fd;
438
439         fd = open("/dev/urandom", O_RDONLY);
440         if (fd == -1) {
441                 td_verror(td, errno, "open");
442                 return 1;
443         }
444
445         if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
446             (int) sizeof(td->rand_seeds)) {
447                 td_verror(td, EIO, "read");
448                 close(fd);
449                 return 1;
450         }
451
452         close(fd);
453         td_fill_rand_seeds(td);
454         return 0;
455 }
456
457 /*
458  * Adds a job to the list of things todo. Sanitizes the various options
459  * to make sure we don't have conflicts, and initializes various
460  * members of td.
461  */
462 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
463 {
464         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
465                                    "randread", "randwrite", "randrw" };
466         unsigned int i;
467         const char *engine;
468         char fname[PATH_MAX];
469         int numjobs, file_alloced;
470
471         /*
472          * the def_thread is just for options, it's not a real job
473          */
474         if (td == &def_thread)
475                 return 0;
476
477         /*
478          * if we are just dumping the output command line, don't add the job
479          */
480         if (dump_cmdline) {
481                 put_job(td);
482                 return 0;
483         }
484
485         engine = get_engine_name(td->o.ioengine);
486         td->io_ops = load_ioengine(td, engine);
487         if (!td->io_ops) {
488                 log_err("fio: failed to load engine %s\n", engine);
489                 goto err;
490         }
491
492         if (td->o.use_thread)
493                 nr_thread++;
494         else
495                 nr_process++;
496
497         if (td->o.odirect)
498                 td->io_ops->flags |= FIO_RAWIO;
499
500         file_alloced = 0;
501         if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
502                 file_alloced = 1;
503
504                 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
505                         add_file(td, jobname);
506                 else {
507                         for (i = 0; i < td->o.nr_files; i++) {
508                                 sprintf(fname, "%s.%d.%d", jobname,
509                                                         td->thread_number, i);
510                                 add_file(td, fname);
511                         }
512                 }
513         }
514
515         if (fixup_options(td))
516                 goto err;
517
518         if (td->io_ops->flags & FIO_DISKLESSIO) {
519                 struct fio_file *f;
520
521                 for_each_file(td, f, i)
522                         f->real_file_size = -1ULL;
523         }
524
525         td->mutex = fio_mutex_init(0);
526
527         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
528         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
529         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
530         td->ddir_nr = td->o.ddir_nr;
531
532         if ((td->o.stonewall || td->o.numjobs > 1 || td->o.new_group)
533              && prev_group_jobs) {
534                 prev_group_jobs = 0;
535                 groupid++;
536         }
537
538         td->groupid = groupid;
539         prev_group_jobs++;
540
541         if (init_random_state(td))
542                 goto err;
543
544         if (setup_rate(td))
545                 goto err;
546
547         if (td->o.write_lat_log) {
548                 setup_log(&td->ts.slat_log);
549                 setup_log(&td->ts.clat_log);
550         }
551         if (td->o.write_bw_log)
552                 setup_log(&td->ts.bw_log);
553
554         if (!td->o.name)
555                 td->o.name = strdup(jobname);
556
557         if (!terse_output) {
558                 if (!job_add_num) {
559                         if (!strcmp(td->io_ops->name, "cpuio")) {
560                                 log_info("%s: ioengine=cpu, cpuload=%u,"
561                                          " cpucycle=%u\n", td->o.name,
562                                                         td->o.cpuload,
563                                                         td->o.cpucycle);
564                         } else {
565                                 char *c1, *c2, *c3, *c4;
566
567                                 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
568                                 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
569                                 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
570                                 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
571
572                                 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
573                                          " ioengine=%s, iodepth=%u\n",
574                                                 td->o.name, td->groupid,
575                                                 ddir_str[td->o.td_ddir],
576                                                 c1, c2, c3, c4,
577                                                 td->io_ops->name,
578                                                 td->o.iodepth);
579
580                                 free(c1);
581                                 free(c2);
582                                 free(c3);
583                                 free(c4);
584                         }
585                 } else if (job_add_num == 1)
586                         log_info("...\n");
587         }
588
589         /*
590          * recurse add identical jobs, clear numjobs and stonewall options
591          * as they don't apply to sub-jobs
592          */
593         numjobs = td->o.numjobs;
594         while (--numjobs) {
595                 struct thread_data *td_new = get_new_job(0, td);
596
597                 if (!td_new)
598                         goto err;
599
600                 td_new->o.numjobs = 1;
601                 td_new->o.stonewall = 0;
602                 td_new->o.new_group = 0;
603
604                 if (file_alloced) {
605                         td_new->o.filename = NULL;
606                         td_new->files_index = 0;
607                         td_new->files_size = 0;
608                         td_new->files = NULL;
609                 }
610
611                 job_add_num = numjobs - 1;
612
613                 if (add_job(td_new, jobname, job_add_num))
614                         goto err;
615         }
616
617         return 0;
618 err:
619         put_job(td);
620         return -1;
621 }
622
623 static int skip_this_section(const char *name)
624 {
625         if (!job_section)
626                 return 0;
627         if (!strncmp(name, "global", 6))
628                 return 0;
629
630         return strcmp(job_section, name);
631 }
632
633 static int is_empty_or_comment(char *line)
634 {
635         unsigned int i;
636
637         for (i = 0; i < strlen(line); i++) {
638                 if (line[i] == ';')
639                         return 1;
640                 if (line[i] == '#')
641                         return 1;
642                 if (!isspace(line[i]) && !iscntrl(line[i]))
643                         return 0;
644         }
645
646         return 1;
647 }
648
649 /*
650  * This is our [ini] type file parser.
651  */
652 static int parse_jobs_ini(char *file, int stonewall_flag)
653 {
654         unsigned int global;
655         struct thread_data *td;
656         char *string, *name;
657         FILE *f;
658         char *p;
659         int ret = 0, stonewall;
660         int first_sect = 1;
661         int skip_fgets = 0;
662         int inside_skip = 0;
663         char **opts;
664         int i, alloc_opts, num_opts;
665
666         if (!strcmp(file, "-"))
667                 f = stdin;
668         else
669                 f = fopen(file, "r");
670
671         if (!f) {
672                 perror("fopen job file");
673                 return 1;
674         }
675
676         string = malloc(4096);
677
678         /*
679          * it's really 256 + small bit, 280 should suffice
680          */
681         name = malloc(280);
682         memset(name, 0, 280);
683
684         alloc_opts = 8;
685         opts = malloc(sizeof(char *) * alloc_opts);
686         num_opts = 0;
687
688         stonewall = stonewall_flag;
689         do {
690                 /*
691                  * if skip_fgets is set, we already have loaded a line we
692                  * haven't handled.
693                  */
694                 if (!skip_fgets) {
695                         p = fgets(string, 4095, f);
696                         if (!p)
697                                 break;
698                 }
699
700                 skip_fgets = 0;
701                 strip_blank_front(&p);
702                 strip_blank_end(p);
703
704                 if (is_empty_or_comment(p))
705                         continue;
706                 if (sscanf(p, "[%255s]", name) != 1) {
707                         if (inside_skip)
708                                 continue;
709                         log_err("fio: option <%s> outside of [] job section\n",
710                                                                         p);
711                         break;
712                 }
713
714                 if (skip_this_section(name)) {
715                         inside_skip = 1;
716                         continue;
717                 } else
718                         inside_skip = 0;
719
720                 global = !strncmp(name, "global", 6);
721
722                 name[strlen(name) - 1] = '\0';
723
724                 if (dump_cmdline) {
725                         if (first_sect)
726                                 log_info("fio ");
727                         if (!global)
728                                 log_info("--name=%s ", name);
729                         first_sect = 0;
730                 }
731
732                 td = get_new_job(global, &def_thread);
733                 if (!td) {
734                         ret = 1;
735                         break;
736                 }
737
738                 /*
739                  * Seperate multiple job files by a stonewall
740                  */
741                 if (!global && stonewall) {
742                         td->o.stonewall = stonewall;
743                         stonewall = 0;
744                 }
745
746                 num_opts = 0;
747                 memset(opts, 0, alloc_opts * sizeof(char *));
748
749                 while ((p = fgets(string, 4096, f)) != NULL) {
750                         if (is_empty_or_comment(p))
751                                 continue;
752
753                         strip_blank_front(&p);
754
755                         /*
756                          * new section, break out and make sure we don't
757                          * fgets() a new line at the top.
758                          */
759                         if (p[0] == '[') {
760                                 skip_fgets = 1;
761                                 break;
762                         }
763
764                         strip_blank_end(p);
765
766                         if (num_opts == alloc_opts) {
767                                 alloc_opts <<= 1;
768                                 opts = realloc(opts,
769                                                 alloc_opts * sizeof(char *));
770                         }
771
772                         opts[num_opts] = strdup(p);
773                         num_opts++;
774                 }
775
776                 ret = fio_options_parse(td, opts, num_opts);
777                 if (!ret) {
778                         if (dump_cmdline)
779                                 for (i = 0; i < num_opts; i++)
780                                         log_info("--%s ", opts[i]);
781
782                         ret = add_job(td, name, 0);
783                 } else {
784                         log_err("fio: job %s dropped\n", name);
785                         put_job(td);
786                 }
787
788                 for (i = 0; i < num_opts; i++)
789                         free(opts[i]);
790                 num_opts = 0;
791         } while (!ret);
792
793         if (dump_cmdline)
794                 log_info("\n");
795
796         for (i = 0; i < num_opts; i++)
797                 free(opts[i]);
798
799         free(string);
800         free(name);
801         free(opts);
802         if (f != stdin)
803                 fclose(f);
804         return ret;
805 }
806
807 static int fill_def_thread(void)
808 {
809         memset(&def_thread, 0, sizeof(def_thread));
810
811         fio_getaffinity(getpid(), &def_thread.o.cpumask);
812
813         /*
814          * fill default options
815          */
816         fio_fill_default_options(&def_thread);
817
818         def_thread.o.timeout = def_timeout;
819         def_thread.o.write_bw_log = write_bw_log;
820         def_thread.o.write_lat_log = write_lat_log;
821
822         return 0;
823 }
824
825 static void free_shm(void)
826 {
827         struct shmid_ds sbuf;
828
829         if (threads) {
830                 void *tp = threads;
831
832                 threads = NULL;
833                 file_hash_exit();
834                 fio_debug_jobp = NULL;
835                 shmdt(tp);
836                 shmctl(shm_id, IPC_RMID, &sbuf);
837         }
838
839         scleanup();
840 }
841
842 /*
843  * The thread area is shared between the main process and the job
844  * threads/processes. So setup a shared memory segment that will hold
845  * all the job info. We use the end of the region for keeping track of
846  * open files across jobs, for file sharing.
847  */
848 static int setup_thread_area(void)
849 {
850         void *hash;
851
852         /*
853          * 1024 is too much on some machines, scale max_jobs if
854          * we get a failure that looks like too large a shm segment
855          */
856         do {
857                 size_t size = max_jobs * sizeof(struct thread_data);
858
859                 size += file_hash_size;
860                 size += sizeof(unsigned int);
861
862                 shm_id = shmget(0, size, IPC_CREAT | 0600);
863                 if (shm_id != -1)
864                         break;
865                 if (errno != EINVAL) {
866                         perror("shmget");
867                         break;
868                 }
869
870                 max_jobs >>= 1;
871         } while (max_jobs);
872
873         if (shm_id == -1)
874                 return 1;
875
876         threads = shmat(shm_id, NULL, 0);
877         if (threads == (void *) -1) {
878                 perror("shmat");
879                 return 1;
880         }
881
882         memset(threads, 0, max_jobs * sizeof(struct thread_data));
883         hash = (void *) threads + max_jobs * sizeof(struct thread_data);
884         fio_debug_jobp = (void *) hash + file_hash_size;
885         *fio_debug_jobp = -1;
886         file_hash_init(hash);
887         atexit(free_shm);
888         return 0;
889 }
890
891 static void usage(const char *name)
892 {
893         printf("%s\n", fio_version_string);
894         printf("%s [options] [job options] <job file(s)>\n", name);
895         printf("\t--debug=options\tEnable debug logging\n");
896         printf("\t--output\tWrite output to file\n");
897         printf("\t--timeout\tRuntime in seconds\n");
898         printf("\t--latency-log\tGenerate per-job latency logs\n");
899         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
900         printf("\t--minimal\tMinimal (terse) output\n");
901         printf("\t--version\tPrint version info and exit\n");
902         printf("\t--help\t\tPrint this page\n");
903         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
904                 " them\n");
905         printf("\t--showcmd\tTurn a job file into command line options\n");
906         printf("\t--eta=when\tWhen ETA estimate should be printed\n");
907         printf("\t          \tMay be \"always\", \"never\" or \"auto\"\n");
908         printf("\t--readonly\tTurn on safety read-only checks, preventing"
909                 " writes\n");
910         printf("\t--section=name\tOnly run specified section in job file\n");
911         printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
912                 " (def 1024)\n");
913         printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>\n");
914 }
915
916 #ifdef FIO_INC_DEBUG
917 struct debug_level debug_levels[] = {
918         { .name = "process",    .shift = FD_PROCESS, },
919         { .name = "file",       .shift = FD_FILE, },
920         { .name = "io",         .shift = FD_IO, },
921         { .name = "mem",        .shift = FD_MEM, },
922         { .name = "blktrace",   .shift = FD_BLKTRACE },
923         { .name = "verify",     .shift = FD_VERIFY },
924         { .name = "random",     .shift = FD_RANDOM },
925         { .name = "parse",      .shift = FD_PARSE },
926         { .name = "diskutil",   .shift = FD_DISKUTIL },
927         { .name = "job",        .shift = FD_JOB },
928         { .name = "mutex",      .shift = FD_MUTEX },
929         { .name = NULL, },
930 };
931
932 static int set_debug(const char *string)
933 {
934         struct debug_level *dl;
935         char *p = (char *) string;
936         char *opt;
937         int i;
938
939         if (!strcmp(string, "?") || !strcmp(string, "help")) {
940                 int i;
941
942                 log_info("fio: dumping debug options:");
943                 for (i = 0; debug_levels[i].name; i++) {
944                         dl = &debug_levels[i];
945                         log_info("%s,", dl->name);
946                 }
947                 log_info("all\n");
948                 return 1;
949         }
950
951         while ((opt = strsep(&p, ",")) != NULL) {
952                 int found = 0;
953
954                 if (!strncmp(opt, "all", 3)) {
955                         log_info("fio: set all debug options\n");
956                         fio_debug = ~0UL;
957                         continue;
958                 }
959
960                 for (i = 0; debug_levels[i].name; i++) {
961                         dl = &debug_levels[i];
962                         found = !strncmp(opt, dl->name, strlen(dl->name));
963                         if (!found)
964                                 continue;
965
966                         if (dl->shift == FD_JOB) {
967                                 opt = strchr(opt, ':');
968                                 if (!opt) {
969                                         log_err("fio: missing job number\n");
970                                         break;
971                                 }
972                                 opt++;
973                                 fio_debug_jobno = atoi(opt);
974                                 log_info("fio: set debug jobno %d\n",
975                                                         fio_debug_jobno);
976                         } else {
977                                 log_info("fio: set debug option %s\n", opt);
978                                 fio_debug |= (1UL << dl->shift);
979                         }
980                         break;
981                 }
982
983                 if (!found)
984                         log_err("fio: debug mask %s not found\n", opt);
985         }
986         return 0;
987 }
988 #else
989 static int set_debug(const char *string)
990 {
991         log_err("fio: debug tracing not included in build\n");
992         return 1;
993 }
994 #endif
995
996 static int parse_cmd_line(int argc, char *argv[])
997 {
998         struct thread_data *td = NULL;
999         int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1000
1001         while ((c = getopt_long_only(argc, argv, "", l_opts, &lidx)) != -1) {
1002                 switch (c) {
1003                 case 'a':
1004                         smalloc_pool_size = atoi(optarg);
1005                         break;
1006                 case 't':
1007                         def_timeout = atoi(optarg);
1008                         break;
1009                 case 'l':
1010                         write_lat_log = 1;
1011                         break;
1012                 case 'w':
1013                         write_bw_log = 1;
1014                         break;
1015                 case 'o':
1016                         f_out = fopen(optarg, "w+");
1017                         if (!f_out) {
1018                                 perror("fopen output");
1019                                 exit(1);
1020                         }
1021                         f_err = f_out;
1022                         break;
1023                 case 'm':
1024                         terse_output = 1;
1025                         break;
1026                 case 'h':
1027                         usage(argv[0]);
1028                         exit(0);
1029                 case 'c':
1030                         exit(fio_show_option_help(optarg));
1031                 case 's':
1032                         dump_cmdline = 1;
1033                         break;
1034                 case 'r':
1035                         read_only = 1;
1036                         break;
1037                 case 'v':
1038                         printf("%s\n", fio_version_string);
1039                         exit(0);
1040                 case 'e':
1041                         if (!strcmp("always", optarg))
1042                                 eta_print = FIO_ETA_ALWAYS;
1043                         else if (!strcmp("never", optarg))
1044                                 eta_print = FIO_ETA_NEVER;
1045                         break;
1046                 case 'd':
1047                         if (set_debug(optarg))
1048                                 do_exit++;
1049                         break;
1050                 case 'x':
1051                         if (!strcmp(optarg, "global")) {
1052                                 log_err("fio: can't use global as only "
1053                                         "section\n");
1054                                 do_exit++;
1055                                 exit_val = 1;
1056                                 break;
1057                         }
1058                         if (job_section)
1059                                 free(job_section);
1060                         job_section = strdup(optarg);
1061                         break;
1062                 case FIO_GETOPT_JOB: {
1063                         const char *opt = l_opts[lidx].name;
1064                         char *val = optarg;
1065
1066                         if (!strncmp(opt, "name", 4) && td) {
1067                                 ret = add_job(td, td->o.name ?: "fio", 0);
1068                                 if (ret) {
1069                                         put_job(td);
1070                                         return 0;
1071                                 }
1072                                 td = NULL;
1073                         }
1074                         if (!td) {
1075                                 int is_section = !strncmp(opt, "name", 4);
1076                                 int global = 0;
1077
1078                                 if (!is_section || !strncmp(val, "global", 6))
1079                                         global = 1;
1080
1081                                 if (is_section && skip_this_section(val))
1082                                         continue;
1083
1084                                 td = get_new_job(global, &def_thread);
1085                                 if (!td)
1086                                         return 0;
1087                         }
1088
1089                         ret = fio_cmd_option_parse(td, opt, val);
1090                         break;
1091                 }
1092                 default:
1093                         do_exit++;
1094                         exit_val = 1;
1095                         break;
1096                 }
1097         }
1098
1099         if (do_exit)
1100                 exit(exit_val);
1101
1102         if (td) {
1103                 if (!ret)
1104                         ret = add_job(td, td->o.name ?: "fio", 0);
1105                 if (ret)
1106                         put_job(td);
1107         }
1108
1109         while (optind < argc) {
1110                 ini_idx++;
1111                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1112                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1113                 optind++;
1114         }
1115
1116         return ini_idx;
1117 }
1118
1119 int parse_options(int argc, char *argv[])
1120 {
1121         int job_files, i;
1122
1123         f_out = stdout;
1124         f_err = stderr;
1125
1126         fio_options_dup_and_init(l_opts);
1127
1128         if (setup_thread_area())
1129                 return 1;
1130         if (fill_def_thread())
1131                 return 1;
1132
1133         job_files = parse_cmd_line(argc, argv);
1134
1135         for (i = 0; i < job_files; i++) {
1136                 if (fill_def_thread())
1137                         return 1;
1138                 if (parse_jobs_ini(ini_file[i], i))
1139                         return 1;
1140                 free(ini_file[i]);
1141         }
1142
1143         free(ini_file);
1144         options_mem_free(&def_thread);
1145
1146         if (!thread_number) {
1147                 if (dump_cmdline)
1148                         return 0;
1149
1150                 log_err("No jobs defined(s)\n\n");
1151                 usage(argv[0]);
1152                 return 1;
1153         }
1154
1155         if (def_thread.o.gtod_offload) {
1156                 fio_gtod_init();
1157                 fio_gtod_offload = 1;
1158                 fio_gtod_cpu = def_thread.o.gtod_cpu;
1159         }
1160
1161         return 0;
1162 }