Don't add the file offset twice for trim
[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 <sys/ipc.h>
12 #include <sys/shm.h>
13 #include <sys/types.h>
14 #include <sys/stat.h>
15
16 #include "fio.h"
17 #include "parse.h"
18 #include "smalloc.h"
19 #include "filehash.h"
20 #include "verify.h"
21 #include "profile.h"
22
23 #include "lib/getopt.h"
24
25 static char fio_version_string[] = "fio 1.43.2";
26
27 #define FIO_RANDSEED            (0xb1899bedUL)
28
29 static char **ini_file;
30 static int max_jobs = MAX_JOBS;
31 static int dump_cmdline;
32
33 static struct thread_data def_thread;
34 struct thread_data *threads = NULL;
35
36 int exitall_on_terminate = 0;
37 int terse_output = 0;
38 int eta_print;
39 unsigned long long mlock_size = 0;
40 FILE *f_out = NULL;
41 FILE *f_err = NULL;
42 char *job_section = NULL;
43 char *exec_profile = NULL;
44
45 int write_bw_log = 0;
46 int read_only = 0;
47
48 static int def_timeout;
49 static int write_lat_log;
50
51 static int prev_group_jobs;
52
53 unsigned long fio_debug = 0;
54 unsigned int fio_debug_jobno = -1;
55 unsigned int *fio_debug_jobp = NULL;
56
57 /*
58  * Command line options. These will contain the above, plus a few
59  * extra that only pertain to fio itself and not jobs.
60  */
61 static struct option l_opts[FIO_NR_OPTIONS] = {
62         {
63                 .name           = "output",
64                 .has_arg        = required_argument,
65                 .val            = 'o',
66         },
67         {
68                 .name           = "timeout",
69                 .has_arg        = required_argument,
70                 .val            = 't',
71         },
72         {
73                 .name           = "latency-log",
74                 .has_arg        = required_argument,
75                 .val            = 'l',
76         },
77         {
78                 .name           = "bandwidth-log",
79                 .has_arg        = required_argument,
80                 .val            = 'b',
81         },
82         {
83                 .name           = "minimal",
84                 .has_arg        = optional_argument,
85                 .val            = 'm',
86         },
87         {
88                 .name           = "version",
89                 .has_arg        = no_argument,
90                 .val            = 'v',
91         },
92         {
93                 .name           = "help",
94                 .has_arg        = no_argument,
95                 .val            = 'h',
96         },
97         {
98                 .name           = "cmdhelp",
99                 .has_arg        = optional_argument,
100                 .val            = 'c',
101         },
102         {
103                 .name           = "showcmd",
104                 .has_arg        = no_argument,
105                 .val            = 's',
106         },
107         {
108                 .name           = "readonly",
109                 .has_arg        = no_argument,
110                 .val            = 'r',
111         },
112         {
113                 .name           = "eta",
114                 .has_arg        = required_argument,
115                 .val            = 'e',
116         },
117         {
118                 .name           = "debug",
119                 .has_arg        = required_argument,
120                 .val            = 'd',
121         },
122         {
123                 .name           = "section",
124                 .has_arg        = required_argument,
125                 .val            = 'x',
126         },
127         {
128                 .name           = "alloc-size",
129                 .has_arg        = required_argument,
130                 .val            = 'a',
131         },
132         {
133                 .name           = "profile",
134                 .has_arg        = required_argument,
135                 .val            = 'p',
136         },
137         {
138                 .name           = NULL,
139         },
140 };
141
142 FILE *get_f_out()
143 {
144         return f_out;
145 }
146
147 FILE *get_f_err()
148 {
149         return f_err;
150 }
151
152 /*
153  * Return a free job structure.
154  */
155 static struct thread_data *get_new_job(int global, struct thread_data *parent)
156 {
157         struct thread_data *td;
158
159         if (global)
160                 return &def_thread;
161         if (thread_number >= max_jobs)
162                 return NULL;
163
164         td = &threads[thread_number++];
165         *td = *parent;
166
167         td->o.uid = td->o.gid = -1U;
168
169         dup_files(td, parent);
170         options_mem_dupe(td);
171
172         profile_add_hooks(td);
173
174         td->thread_number = thread_number;
175         return td;
176 }
177
178 static void put_job(struct thread_data *td)
179 {
180         if (td == &def_thread)
181                 return;
182         
183         profile_td_exit(td);
184
185         if (td->error)
186                 log_info("fio: %s\n", td->verror);
187
188         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
189         thread_number--;
190 }
191
192 static int __setup_rate(struct thread_data *td, enum fio_ddir ddir)
193 {
194         unsigned int bs = td->o.min_bs[ddir];
195         unsigned long long bytes_per_sec;
196
197         assert(ddir_rw(ddir));
198
199         if (td->o.rate[ddir])
200                 bytes_per_sec = td->o.rate[ddir];
201         else
202                 bytes_per_sec = td->o.rate_iops[ddir] * bs;
203
204         if (!bytes_per_sec) {
205                 log_err("rate lower than supported\n");
206                 return -1;
207         }
208
209         td->rate_nsec_cycle[ddir] = 1000000000ULL / bytes_per_sec;
210         td->rate_pending_usleep[ddir] = 0;
211         return 0;
212 }
213
214 static int setup_rate(struct thread_data *td)
215 {
216         int ret = 0;
217
218         if (td->o.rate[DDIR_READ] || td->o.rate_iops[DDIR_READ])
219                 ret = __setup_rate(td, DDIR_READ);
220         if (td->o.rate[DDIR_WRITE] || td->o.rate_iops[DDIR_WRITE])
221                 ret |= __setup_rate(td, DDIR_WRITE);
222
223         return ret;
224 }
225
226 static int fixed_block_size(struct thread_options *o)
227 {
228         return o->min_bs[DDIR_READ] == o->max_bs[DDIR_READ] &&
229                 o->min_bs[DDIR_WRITE] == o->max_bs[DDIR_WRITE] &&
230                 o->min_bs[DDIR_READ] == o->min_bs[DDIR_WRITE];
231 }
232
233 /*
234  * Lazy way of fixing up options that depend on each other. We could also
235  * define option callback handlers, but this is easier.
236  */
237 static int fixup_options(struct thread_data *td)
238 {
239         struct thread_options *o = &td->o;
240
241 #ifndef FIO_HAVE_PSHARED_MUTEX
242         if (!o->use_thread) {
243                 log_info("fio: this platform does not support process shared"
244                          " mutexes, forcing use of threads. Use the 'thread'"
245                          " option to get rid of this warning.\n");
246                 o->use_thread = 1;
247         }
248 #endif
249
250         if (o->write_iolog_file && o->read_iolog_file) {
251                 log_err("fio: read iolog overrides write_iolog\n");
252                 free(o->write_iolog_file);
253                 o->write_iolog_file = NULL;
254         }
255
256         /*
257          * only really works for sequential io for now, and with 1 file
258          */
259         if (o->zone_size && td_random(td) && o->open_files == 1)
260                 o->zone_size = 0;
261
262         /*
263          * Reads can do overwrites, we always need to pre-create the file
264          */
265         if (td_read(td) || td_rw(td))
266                 o->overwrite = 1;
267
268         if (!o->min_bs[DDIR_READ])
269                 o->min_bs[DDIR_READ] = o->bs[DDIR_READ];
270         if (!o->max_bs[DDIR_READ])
271                 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
272         if (!o->min_bs[DDIR_WRITE])
273                 o->min_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
274         if (!o->max_bs[DDIR_WRITE])
275                 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
276
277         o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
278
279         /*
280          * For random IO, allow blockalign offset other than min_bs.
281          */
282         if (!o->ba[DDIR_READ] || !td_random(td))
283                 o->ba[DDIR_READ] = o->min_bs[DDIR_READ];
284         if (!o->ba[DDIR_WRITE] || !td_random(td))
285                 o->ba[DDIR_WRITE] = o->min_bs[DDIR_WRITE];
286
287         if ((o->ba[DDIR_READ] != o->min_bs[DDIR_READ] ||
288             o->ba[DDIR_WRITE] != o->min_bs[DDIR_WRITE]) &&
289             !o->norandommap) {
290                 log_err("fio: Any use of blockalign= turns off randommap\n");
291                 o->norandommap = 1;
292         }
293
294         if (!o->file_size_high)
295                 o->file_size_high = o->file_size_low;
296
297         if (o->norandommap && o->verify != VERIFY_NONE
298             && !fixed_block_size(o))  {
299                 log_err("fio: norandommap given for variable block sizes, "
300                         "verify disabled\n");
301                 o->verify = VERIFY_NONE;
302         }
303         if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
304                 log_err("fio: bs_unaligned may not work with raw io\n");
305
306         /*
307          * thinktime_spin must be less than thinktime
308          */
309         if (o->thinktime_spin > o->thinktime)
310                 o->thinktime_spin = o->thinktime;
311
312         /*
313          * The low water mark cannot be bigger than the iodepth
314          */
315         if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
316                 /*
317                  * syslet work around - if the workload is sequential,
318                  * we want to let the queue drain all the way down to
319                  * avoid seeking between async threads
320                  */
321                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
322                         o->iodepth_low = 1;
323                 else
324                         o->iodepth_low = o->iodepth;
325         }
326
327         /*
328          * If batch number isn't set, default to the same as iodepth
329          */
330         if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
331                 o->iodepth_batch = o->iodepth;
332
333         if (o->nr_files > td->files_index)
334                 o->nr_files = td->files_index;
335
336         if (o->open_files > o->nr_files || !o->open_files)
337                 o->open_files = o->nr_files;
338
339         if (((o->rate[0] + o->rate[1]) && (o->rate_iops[0] + o->rate_iops[1]))||
340             ((o->ratemin[0] + o->ratemin[1]) && (o->rate_iops_min[0] +
341                 o->rate_iops_min[1]))) {
342                 log_err("fio: rate and rate_iops are mutually exclusive\n");
343                 return 1;
344         }
345         if ((o->rate[0] < o->ratemin[0]) || (o->rate[1] < o->ratemin[1]) ||
346             (o->rate_iops[0] < o->rate_iops_min[0]) ||
347             (o->rate_iops[1] < o->rate_iops_min[1])) {
348                 log_err("fio: minimum rate exceeds rate\n");
349                 return 1;
350         }
351
352         if (!o->timeout && o->time_based) {
353                 log_err("fio: time_based requires a runtime/timeout setting\n");
354                 o->time_based = 0;
355         }
356
357         if (o->fill_device && !o->size)
358                 o->size = -1ULL;
359
360         if (td_rw(td) && o->verify != VERIFY_NONE)
361                 log_info("fio: mixed read/write workload with verify. May not "
362                  "work as expected, unless you pre-populated the file\n");
363
364         if (o->verify != VERIFY_NONE) {
365                 o->refill_buffers = 1;
366                 if (o->max_bs[DDIR_WRITE] != o->min_bs[DDIR_WRITE] &&
367                     !o->verify_interval)
368                         o->verify_interval = o->min_bs[DDIR_WRITE];
369         }
370
371         if (o->pre_read) {
372                 o->invalidate_cache = 0;
373                 if (td->io_ops->flags & FIO_PIPEIO)
374                         log_info("fio: cannot pre-read files with an IO engine"
375                                  " that isn't seekable. Pre-read disabled.\n");
376         }
377
378 #ifndef FIO_HAVE_FDATASYNC
379         if (o->fdatasync_blocks) {
380                 log_info("fio: this platform does not support fdatasync()"
381                          " falling back to using fsync().  Use the 'fsync'"
382                          " option instead of 'fdatasync' to get rid of"
383                          " this warning\n");
384                 o->fsync_blocks = o->fdatasync_blocks;
385                 o->fdatasync_blocks = 0;
386         }
387 #endif
388
389         return 0;
390 }
391
392 /*
393  * This function leaks the buffer
394  */
395 static char *to_kmg(unsigned int val)
396 {
397         char *buf = malloc(32);
398         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
399         char *p = post;
400
401         do {
402                 if (val & 1023)
403                         break;
404
405                 val >>= 10;
406                 p++;
407         } while (*p);
408
409         snprintf(buf, 31, "%u%c", val, *p);
410         return buf;
411 }
412
413 /* External engines are specified by "external:name.o") */
414 static const char *get_engine_name(const char *str)
415 {
416         char *p = strstr(str, ":");
417
418         if (!p)
419                 return str;
420
421         p++;
422         strip_blank_front(&p);
423         strip_blank_end(p);
424         return p;
425 }
426
427 static int exists_and_not_file(const char *filename)
428 {
429         struct stat sb;
430
431         if (lstat(filename, &sb) == -1)
432                 return 0;
433
434         if (S_ISREG(sb.st_mode))
435                 return 0;
436
437         return 1;
438 }
439
440 void td_fill_rand_seeds(struct thread_data *td)
441 {
442         os_random_seed(td->rand_seeds[0], &td->bsrange_state);
443         os_random_seed(td->rand_seeds[1], &td->verify_state);
444         os_random_seed(td->rand_seeds[2], &td->rwmix_state);
445
446         if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
447                 os_random_seed(td->rand_seeds[3], &td->next_file_state);
448
449         os_random_seed(td->rand_seeds[5], &td->file_size_state);
450         os_random_seed(td->rand_seeds[6], &td->trim_state);
451
452         if (!td_random(td))
453                 return;
454
455         if (td->o.rand_repeatable)
456                 td->rand_seeds[4] = FIO_RANDSEED * td->thread_number;
457
458         os_random_seed(td->rand_seeds[4], &td->random_state);
459 }
460
461 /*
462  * Initialize the various random states we need (random io, block size ranges,
463  * read/write mix, etc).
464  */
465 static int init_random_state(struct thread_data *td)
466 {
467         int fd;
468
469         fd = open("/dev/urandom", O_RDONLY);
470         if (fd == -1) {
471                 td_verror(td, errno, "open");
472                 return 1;
473         }
474
475         if (read(fd, td->rand_seeds, sizeof(td->rand_seeds)) <
476             (int) sizeof(td->rand_seeds)) {
477                 td_verror(td, EIO, "read");
478                 close(fd);
479                 return 1;
480         }
481
482         close(fd);
483         td_fill_rand_seeds(td);
484         return 0;
485 }
486
487 /*
488  * Adds a job to the list of things todo. Sanitizes the various options
489  * to make sure we don't have conflicts, and initializes various
490  * members of td.
491  */
492 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
493 {
494         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
495                                    "randread", "randwrite", "randrw" };
496         unsigned int i;
497         const char *engine;
498         char fname[PATH_MAX];
499         int numjobs, file_alloced;
500
501         /*
502          * the def_thread is just for options, it's not a real job
503          */
504         if (td == &def_thread)
505                 return 0;
506
507         /*
508          * if we are just dumping the output command line, don't add the job
509          */
510         if (dump_cmdline) {
511                 put_job(td);
512                 return 0;
513         }
514
515         if (profile_td_init(td))
516                 return 1;
517
518         engine = get_engine_name(td->o.ioengine);
519         td->io_ops = load_ioengine(td, engine);
520         if (!td->io_ops) {
521                 log_err("fio: failed to load engine %s\n", engine);
522                 goto err;
523         }
524
525         if (td->o.use_thread)
526                 nr_thread++;
527         else
528                 nr_process++;
529
530         if (td->o.odirect)
531                 td->io_ops->flags |= FIO_RAWIO;
532
533         file_alloced = 0;
534         if (!td->o.filename && !td->files_index && !td->o.read_iolog_file) {
535                 file_alloced = 1;
536
537                 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
538                         add_file(td, jobname);
539                 else {
540                         for (i = 0; i < td->o.nr_files; i++) {
541                                 sprintf(fname, "%s.%d.%d", jobname,
542                                                         td->thread_number, i);
543                                 add_file(td, fname);
544                         }
545                 }
546         }
547
548         if (fixup_options(td))
549                 goto err;
550
551         if (td->io_ops->flags & FIO_DISKLESSIO) {
552                 struct fio_file *f;
553
554                 for_each_file(td, f, i)
555                         f->real_file_size = -1ULL;
556         }
557
558         td->mutex = fio_mutex_init(0);
559
560         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
561         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
562         td->ts.lat_stat[0].min_val = td->ts.lat_stat[1].min_val = ULONG_MAX;
563         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
564         td->ddir_seq_nr = td->o.ddir_seq_nr + 1;
565
566         if ((td->o.stonewall || td->o.new_group) && prev_group_jobs) {
567                 prev_group_jobs = 0;
568                 groupid++;
569         }
570
571         td->groupid = groupid;
572         prev_group_jobs++;
573
574         if (init_random_state(td))
575                 goto err;
576
577         if (setup_rate(td))
578                 goto err;
579
580         if (td->o.write_lat_log) {
581                 setup_log(&td->ts.lat_log);
582                 setup_log(&td->ts.slat_log);
583                 setup_log(&td->ts.clat_log);
584         }
585         if (td->o.write_bw_log)
586                 setup_log(&td->ts.bw_log);
587
588         if (!td->o.name)
589                 td->o.name = strdup(jobname);
590
591         if (!terse_output) {
592                 if (!job_add_num) {
593                         if (!strcmp(td->io_ops->name, "cpuio")) {
594                                 log_info("%s: ioengine=cpu, cpuload=%u,"
595                                          " cpucycle=%u\n", td->o.name,
596                                                         td->o.cpuload,
597                                                         td->o.cpucycle);
598                         } else {
599                                 char *c1, *c2, *c3, *c4;
600
601                                 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
602                                 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
603                                 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
604                                 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
605
606                                 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s,"
607                                          " ioengine=%s, iodepth=%u\n",
608                                                 td->o.name, td->groupid,
609                                                 ddir_str[td->o.td_ddir],
610                                                 c1, c2, c3, c4,
611                                                 td->io_ops->name,
612                                                 td->o.iodepth);
613
614                                 free(c1);
615                                 free(c2);
616                                 free(c3);
617                                 free(c4);
618                         }
619                 } else if (job_add_num == 1)
620                         log_info("...\n");
621         }
622
623         /*
624          * recurse add identical jobs, clear numjobs and stonewall options
625          * as they don't apply to sub-jobs
626          */
627         numjobs = td->o.numjobs;
628         while (--numjobs) {
629                 struct thread_data *td_new = get_new_job(0, td);
630
631                 if (!td_new)
632                         goto err;
633
634                 td_new->o.numjobs = 1;
635                 td_new->o.stonewall = 0;
636                 td_new->o.new_group = 0;
637
638                 if (file_alloced) {
639                         td_new->o.filename = NULL;
640                         td_new->files_index = 0;
641                         td_new->files_size = 0;
642                         td_new->files = NULL;
643                 }
644
645                 job_add_num = numjobs - 1;
646
647                 if (add_job(td_new, jobname, job_add_num))
648                         goto err;
649         }
650
651         return 0;
652 err:
653         put_job(td);
654         return -1;
655 }
656
657 /*
658  * Parse as if 'o' was a command line
659  */
660 void add_job_opts(const char **o)
661 {
662         struct thread_data *td, *td_parent;
663         int i, in_global = 1;
664         char jobname[32];
665
666         i = 0;
667         td_parent = td = NULL;
668         while (o[i]) {
669                 if (!strncmp(o[i], "name", 4)) {
670                         in_global = 0;
671                         if (td)
672                                 add_job(td, jobname, 0);
673                         td = NULL;
674                         sprintf(jobname, "%s", o[i] + 5);
675                 }
676                 if (in_global && !td_parent)
677                         td_parent = get_new_job(1, &def_thread);
678                 else if (!in_global && !td) {
679                         if (!td_parent)
680                                 td_parent = &def_thread;
681                         td = get_new_job(0, td_parent);
682                 }
683                 if (in_global)
684                         fio_options_parse(td_parent, (char **) &o[i], 1);
685                 else
686                         fio_options_parse(td, (char **) &o[i], 1);
687                 i++;
688         }
689
690         if (td)
691                 add_job(td, jobname, 0);
692 }
693
694 static int skip_this_section(const char *name)
695 {
696         if (!job_section)
697                 return 0;
698         if (!strncmp(name, "global", 6))
699                 return 0;
700
701         return strcmp(job_section, name);
702 }
703
704 static int is_empty_or_comment(char *line)
705 {
706         unsigned int i;
707
708         for (i = 0; i < strlen(line); i++) {
709                 if (line[i] == ';')
710                         return 1;
711                 if (line[i] == '#')
712                         return 1;
713                 if (!isspace(line[i]) && !iscntrl(line[i]))
714                         return 0;
715         }
716
717         return 1;
718 }
719
720 /*
721  * This is our [ini] type file parser.
722  */
723 static int parse_jobs_ini(char *file, int stonewall_flag)
724 {
725         unsigned int global;
726         struct thread_data *td;
727         char *string, *name;
728         FILE *f;
729         char *p;
730         int ret = 0, stonewall;
731         int first_sect = 1;
732         int skip_fgets = 0;
733         int inside_skip = 0;
734         char **opts;
735         int i, alloc_opts, num_opts;
736
737         if (!strcmp(file, "-"))
738                 f = stdin;
739         else
740                 f = fopen(file, "r");
741
742         if (!f) {
743                 perror("fopen job file");
744                 return 1;
745         }
746
747         string = malloc(4096);
748
749         /*
750          * it's really 256 + small bit, 280 should suffice
751          */
752         name = malloc(280);
753         memset(name, 0, 280);
754
755         alloc_opts = 8;
756         opts = malloc(sizeof(char *) * alloc_opts);
757         num_opts = 0;
758
759         stonewall = stonewall_flag;
760         do {
761                 /*
762                  * if skip_fgets is set, we already have loaded a line we
763                  * haven't handled.
764                  */
765                 if (!skip_fgets) {
766                         p = fgets(string, 4095, f);
767                         if (!p)
768                                 break;
769                 }
770
771                 skip_fgets = 0;
772                 strip_blank_front(&p);
773                 strip_blank_end(p);
774
775                 if (is_empty_or_comment(p))
776                         continue;
777                 if (sscanf(p, "[%255s]", name) != 1) {
778                         if (inside_skip)
779                                 continue;
780                         log_err("fio: option <%s> outside of [] job section\n",
781                                                                         p);
782                         break;
783                 }
784
785                 name[strlen(name) - 1] = '\0';
786
787                 if (skip_this_section(name)) {
788                         inside_skip = 1;
789                         continue;
790                 } else
791                         inside_skip = 0;
792
793                 global = !strncmp(name, "global", 6);
794
795                 if (dump_cmdline) {
796                         if (first_sect)
797                                 log_info("fio ");
798                         if (!global)
799                                 log_info("--name=%s ", name);
800                         first_sect = 0;
801                 }
802
803                 td = get_new_job(global, &def_thread);
804                 if (!td) {
805                         ret = 1;
806                         break;
807                 }
808
809                 /*
810                  * Seperate multiple job files by a stonewall
811                  */
812                 if (!global && stonewall) {
813                         td->o.stonewall = stonewall;
814                         stonewall = 0;
815                 }
816
817                 num_opts = 0;
818                 memset(opts, 0, alloc_opts * sizeof(char *));
819
820                 while ((p = fgets(string, 4096, f)) != NULL) {
821                         if (is_empty_or_comment(p))
822                                 continue;
823
824                         strip_blank_front(&p);
825
826                         /*
827                          * new section, break out and make sure we don't
828                          * fgets() a new line at the top.
829                          */
830                         if (p[0] == '[') {
831                                 skip_fgets = 1;
832                                 break;
833                         }
834
835                         strip_blank_end(p);
836
837                         if (num_opts == alloc_opts) {
838                                 alloc_opts <<= 1;
839                                 opts = realloc(opts,
840                                                 alloc_opts * sizeof(char *));
841                         }
842
843                         opts[num_opts] = strdup(p);
844                         num_opts++;
845                 }
846
847                 ret = fio_options_parse(td, opts, num_opts);
848                 if (!ret) {
849                         if (dump_cmdline)
850                                 for (i = 0; i < num_opts; i++)
851                                         log_info("--%s ", opts[i]);
852
853                         ret = add_job(td, name, 0);
854                 } else {
855                         log_err("fio: job %s dropped\n", name);
856                         put_job(td);
857                 }
858
859                 for (i = 0; i < num_opts; i++)
860                         free(opts[i]);
861                 num_opts = 0;
862         } while (!ret);
863
864         if (dump_cmdline)
865                 log_info("\n");
866
867         for (i = 0; i < num_opts; i++)
868                 free(opts[i]);
869
870         free(string);
871         free(name);
872         free(opts);
873         if (f != stdin)
874                 fclose(f);
875         return ret;
876 }
877
878 static int fill_def_thread(void)
879 {
880         memset(&def_thread, 0, sizeof(def_thread));
881
882         fio_getaffinity(getpid(), &def_thread.o.cpumask);
883
884         /*
885          * fill default options
886          */
887         fio_fill_default_options(&def_thread);
888
889         def_thread.o.timeout = def_timeout;
890         return 0;
891 }
892
893 static void free_shm(void)
894 {
895         struct shmid_ds sbuf;
896
897         if (threads) {
898                 void *tp = threads;
899
900                 threads = NULL;
901                 file_hash_exit();
902                 fio_debug_jobp = NULL;
903                 shmdt(tp);
904                 shmctl(shm_id, IPC_RMID, &sbuf);
905         }
906
907         scleanup();
908 }
909
910 /*
911  * The thread area is shared between the main process and the job
912  * threads/processes. So setup a shared memory segment that will hold
913  * all the job info. We use the end of the region for keeping track of
914  * open files across jobs, for file sharing.
915  */
916 static int setup_thread_area(void)
917 {
918         void *hash;
919
920         /*
921          * 1024 is too much on some machines, scale max_jobs if
922          * we get a failure that looks like too large a shm segment
923          */
924         do {
925                 size_t size = max_jobs * sizeof(struct thread_data);
926
927                 size += file_hash_size;
928                 size += sizeof(unsigned int);
929
930                 shm_id = shmget(0, size, IPC_CREAT | 0600);
931                 if (shm_id != -1)
932                         break;
933                 if (errno != EINVAL) {
934                         perror("shmget");
935                         break;
936                 }
937
938                 max_jobs >>= 1;
939         } while (max_jobs);
940
941         if (shm_id == -1)
942                 return 1;
943
944         threads = shmat(shm_id, NULL, 0);
945         if (threads == (void *) -1) {
946                 perror("shmat");
947                 return 1;
948         }
949
950         memset(threads, 0, max_jobs * sizeof(struct thread_data));
951         hash = (void *) threads + max_jobs * sizeof(struct thread_data);
952         fio_debug_jobp = (void *) hash + file_hash_size;
953         *fio_debug_jobp = -1;
954         file_hash_init(hash);
955         atexit(free_shm);
956         return 0;
957 }
958
959 static void usage(const char *name)
960 {
961         printf("%s\n", fio_version_string);
962         printf("%s [options] [job options] <job file(s)>\n", name);
963         printf("\t--debug=options\tEnable debug logging\n");
964         printf("\t--output\tWrite output to file\n");
965         printf("\t--timeout\tRuntime in seconds\n");
966         printf("\t--latency-log\tGenerate per-job latency logs\n");
967         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
968         printf("\t--minimal\tMinimal (terse) output\n");
969         printf("\t--version\tPrint version info and exit\n");
970         printf("\t--help\t\tPrint this page\n");
971         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of"
972                 " them\n");
973         printf("\t--showcmd\tTurn a job file into command line options\n");
974         printf("\t--eta=when\tWhen ETA estimate should be printed\n");
975         printf("\t          \tMay be \"always\", \"never\" or \"auto\"\n");
976         printf("\t--readonly\tTurn on safety read-only checks, preventing"
977                 " writes\n");
978         printf("\t--section=name\tOnly run specified section in job file\n");
979         printf("\t--alloc-size=kb\tSet smalloc pool to this size in kb"
980                 " (def 1024)\n");
981         printf("\nFio was written by Jens Axboe <jens.axboe@oracle.com>");
982         printf("\n                   Jens Axboe <jaxboe@fusionio.com>\n");
983 }
984
985 #ifdef FIO_INC_DEBUG
986 struct debug_level debug_levels[] = {
987         { .name = "process",    .shift = FD_PROCESS, },
988         { .name = "file",       .shift = FD_FILE, },
989         { .name = "io",         .shift = FD_IO, },
990         { .name = "mem",        .shift = FD_MEM, },
991         { .name = "blktrace",   .shift = FD_BLKTRACE },
992         { .name = "verify",     .shift = FD_VERIFY },
993         { .name = "random",     .shift = FD_RANDOM },
994         { .name = "parse",      .shift = FD_PARSE },
995         { .name = "diskutil",   .shift = FD_DISKUTIL },
996         { .name = "job",        .shift = FD_JOB },
997         { .name = "mutex",      .shift = FD_MUTEX },
998         { .name = "profile",    .shift = FD_PROFILE },
999         { .name = "time",       .shift = FD_TIME },
1000         { .name = NULL, },
1001 };
1002
1003 static int set_debug(const char *string)
1004 {
1005         struct debug_level *dl;
1006         char *p = (char *) string;
1007         char *opt;
1008         int i;
1009
1010         if (!strcmp(string, "?") || !strcmp(string, "help")) {
1011                 int i;
1012
1013                 log_info("fio: dumping debug options:");
1014                 for (i = 0; debug_levels[i].name; i++) {
1015                         dl = &debug_levels[i];
1016                         log_info("%s,", dl->name);
1017                 }
1018                 log_info("all\n");
1019                 return 1;
1020         }
1021
1022         while ((opt = strsep(&p, ",")) != NULL) {
1023                 int found = 0;
1024
1025                 if (!strncmp(opt, "all", 3)) {
1026                         log_info("fio: set all debug options\n");
1027                         fio_debug = ~0UL;
1028                         continue;
1029                 }
1030
1031                 for (i = 0; debug_levels[i].name; i++) {
1032                         dl = &debug_levels[i];
1033                         found = !strncmp(opt, dl->name, strlen(dl->name));
1034                         if (!found)
1035                                 continue;
1036
1037                         if (dl->shift == FD_JOB) {
1038                                 opt = strchr(opt, ':');
1039                                 if (!opt) {
1040                                         log_err("fio: missing job number\n");
1041                                         break;
1042                                 }
1043                                 opt++;
1044                                 fio_debug_jobno = atoi(opt);
1045                                 log_info("fio: set debug jobno %d\n",
1046                                                         fio_debug_jobno);
1047                         } else {
1048                                 log_info("fio: set debug option %s\n", opt);
1049                                 fio_debug |= (1UL << dl->shift);
1050                         }
1051                         break;
1052                 }
1053
1054                 if (!found)
1055                         log_err("fio: debug mask %s not found\n", opt);
1056         }
1057         return 0;
1058 }
1059 #else
1060 static int set_debug(const char *string)
1061 {
1062         log_err("fio: debug tracing not included in build\n");
1063         return 1;
1064 }
1065 #endif
1066
1067 static int parse_cmd_line(int argc, char *argv[])
1068 {
1069         struct thread_data *td = NULL;
1070         int c, ini_idx = 0, lidx, ret = 0, do_exit = 0, exit_val = 0;
1071
1072         while ((c = getopt_long_only(argc, argv, "", l_opts, &lidx)) != -1) {
1073                 switch (c) {
1074                 case 'a':
1075                         smalloc_pool_size = atoi(optarg);
1076                         break;
1077                 case 't':
1078                         def_timeout = atoi(optarg);
1079                         break;
1080                 case 'l':
1081                         write_lat_log = 1;
1082                         break;
1083                 case 'b':
1084                         write_bw_log = 1;
1085                         break;
1086                 case 'o':
1087                         f_out = fopen(optarg, "w+");
1088                         if (!f_out) {
1089                                 perror("fopen output");
1090                                 exit(1);
1091                         }
1092                         f_err = f_out;
1093                         break;
1094                 case 'm':
1095                         terse_output = 1;
1096                         break;
1097                 case 'h':
1098                         usage(argv[0]);
1099                         exit(0);
1100                 case 'c':
1101                         exit(fio_show_option_help(optarg));
1102                 case 's':
1103                         dump_cmdline = 1;
1104                         break;
1105                 case 'r':
1106                         read_only = 1;
1107                         break;
1108                 case 'v':
1109                         printf("%s\n", fio_version_string);
1110                         exit(0);
1111                 case 'e':
1112                         if (!strcmp("always", optarg))
1113                                 eta_print = FIO_ETA_ALWAYS;
1114                         else if (!strcmp("never", optarg))
1115                                 eta_print = FIO_ETA_NEVER;
1116                         break;
1117                 case 'd':
1118                         if (set_debug(optarg))
1119                                 do_exit++;
1120                         break;
1121                 case 'x':
1122                         if (!strcmp(optarg, "global")) {
1123                                 log_err("fio: can't use global as only "
1124                                         "section\n");
1125                                 do_exit++;
1126                                 exit_val = 1;
1127                                 break;
1128                         }
1129                         if (job_section)
1130                                 free(job_section);
1131                         job_section = strdup(optarg);
1132                         break;
1133                 case 'p':
1134                         exec_profile = strdup(optarg);
1135                         break;
1136                 case FIO_GETOPT_JOB: {
1137                         const char *opt = l_opts[lidx].name;
1138                         char *val = optarg;
1139
1140                         if (!strncmp(opt, "name", 4) && td) {
1141                                 ret = add_job(td, td->o.name ?: "fio", 0);
1142                                 if (ret) {
1143                                         put_job(td);
1144                                         return 0;
1145                                 }
1146                                 td = NULL;
1147                         }
1148                         if (!td) {
1149                                 int is_section = !strncmp(opt, "name", 4);
1150                                 int global = 0;
1151
1152                                 if (!is_section || !strncmp(val, "global", 6))
1153                                         global = 1;
1154
1155                                 if (is_section && skip_this_section(val))
1156                                         continue;
1157
1158                                 td = get_new_job(global, &def_thread);
1159                                 if (!td)
1160                                         return 0;
1161                         }
1162
1163                         ret = fio_cmd_option_parse(td, opt, val);
1164                         break;
1165                 }
1166                 default:
1167                         do_exit++;
1168                         exit_val = 1;
1169                         break;
1170                 }
1171         }
1172
1173         if (do_exit)
1174                 exit(exit_val);
1175
1176         if (td) {
1177                 if (!ret)
1178                         ret = add_job(td, td->o.name ?: "fio", 0);
1179                 if (ret)
1180                         put_job(td);
1181         }
1182
1183         while (optind < argc) {
1184                 ini_idx++;
1185                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1186                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1187                 optind++;
1188         }
1189
1190         return ini_idx;
1191 }
1192
1193 int parse_options(int argc, char *argv[])
1194 {
1195         int job_files, i;
1196
1197         f_out = stdout;
1198         f_err = stderr;
1199
1200         fio_options_dup_and_init(l_opts);
1201
1202         if (setup_thread_area())
1203                 return 1;
1204         if (fill_def_thread())
1205                 return 1;
1206
1207         job_files = parse_cmd_line(argc, argv);
1208
1209         for (i = 0; i < job_files; i++) {
1210                 if (fill_def_thread())
1211                         return 1;
1212                 if (parse_jobs_ini(ini_file[i], i))
1213                         return 1;
1214                 free(ini_file[i]);
1215         }
1216
1217         free(ini_file);
1218         options_mem_free(&def_thread);
1219
1220         if (!thread_number) {
1221                 if (dump_cmdline)
1222                         return 0;
1223                 if (exec_profile)
1224                         return 0;
1225
1226                 log_err("No jobs defined(s)\n\n");
1227                 usage(argv[0]);
1228                 return 1;
1229         }
1230
1231         if (def_thread.o.gtod_offload) {
1232                 fio_gtod_init();
1233                 fio_gtod_offload = 1;
1234                 fio_gtod_cpu = def_thread.o.gtod_cpu;
1235         }
1236
1237         return 0;
1238 }