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