0151c9bd4b2979626e17c592c76938c9e2b68b35
[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
20 static char fio_version_string[] = "fio 1.15.1";
21
22 #define FIO_RANDSEED            (0xb1899bedUL)
23
24 static char **ini_file;
25 static int max_jobs = MAX_JOBS;
26 static int dump_cmdline;
27
28 struct thread_data def_thread;
29 struct thread_data *threads = NULL;
30
31 int exitall_on_terminate = 0;
32 int terse_output = 0;
33 unsigned long long mlock_size = 0;
34 FILE *f_out = NULL;
35 FILE *f_err = NULL;
36
37 int write_bw_log = 0;
38
39 static int def_timeout = 0;
40 static int write_lat_log = 0;
41
42 static int prev_group_jobs;
43
44 /*
45  * Command line options. These will contain the above, plus a few
46  * extra that only pertain to fio itself and not jobs.
47  */
48 static struct option long_options[FIO_NR_OPTIONS] = {
49         {
50                 .name           = "output",
51                 .has_arg        = required_argument,
52                 .val            = 'o',
53         },
54         {
55                 .name           = "timeout",
56                 .has_arg        = required_argument,
57                 .val            = 't',
58         },
59         {
60                 .name           = "latency-log",
61                 .has_arg        = required_argument,
62                 .val            = 'l',
63         },
64         {
65                 .name           = "bandwidth-log",
66                 .has_arg        = required_argument,
67                 .val            = 'b',
68         },
69         {
70                 .name           = "minimal",
71                 .has_arg        = optional_argument,
72                 .val            = 'm',
73         },
74         {
75                 .name           = "version",
76                 .has_arg        = no_argument,
77                 .val            = 'v',
78         },
79         {
80                 .name           = "help",
81                 .has_arg        = no_argument,
82                 .val            = 'h',
83         },
84         {
85                 .name           = "cmdhelp",
86                 .has_arg        = optional_argument,
87                 .val            = 'c',
88         },
89         {
90                 .name           = "showcmd",
91                 .has_arg        = no_argument,
92                 .val            = 's'
93         },
94         {
95                 .name           = NULL,
96         },
97 };
98
99 FILE *get_f_out()
100 {
101         return f_out;
102 }
103
104 FILE *get_f_err()
105 {
106         return f_err;
107 }
108
109 /*
110  * Return a free job structure.
111  */
112 static struct thread_data *get_new_job(int global, struct thread_data *parent)
113 {
114         struct thread_data *td;
115
116         if (global)
117                 return &def_thread;
118         if (thread_number >= max_jobs)
119                 return NULL;
120
121         td = &threads[thread_number++];
122         *td = *parent;
123
124         dup_files(td, parent);
125         options_mem_dupe(td);
126
127         td->thread_number = thread_number;
128         return td;
129 }
130
131 static void put_job(struct thread_data *td)
132 {
133         if (td == &def_thread)
134                 return;
135
136         if (td->error)
137                 log_info("fio: %s\n", td->verror);
138
139         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
140         thread_number--;
141 }
142
143 static int setup_rate(struct thread_data *td)
144 {
145         unsigned long nr_reads_per_msec;
146         unsigned long long rate;
147         unsigned int bs;
148
149         if (!td->o.rate && !td->o.rate_iops)
150                 return 0;
151
152         if (td_rw(td))
153                 bs = td->o.rw_min_bs;
154         else if (td_read(td))
155                 bs = td->o.min_bs[DDIR_READ];
156         else
157                 bs = td->o.min_bs[DDIR_WRITE];
158
159         if (td->o.rate) {
160                 rate = td->o.rate;
161                 nr_reads_per_msec = (rate * 1024 * 1000LL) / bs;
162         } else
163                 nr_reads_per_msec = td->o.rate_iops * 1000UL;
164
165         if (!nr_reads_per_msec) {
166                 log_err("rate lower than supported\n");
167                 return -1;
168         }
169
170         td->rate_usec_cycle = 1000000000ULL / nr_reads_per_msec;
171         td->rate_pending_usleep = 0;
172         return 0;
173 }
174
175 /*
176  * Lazy way of fixing up options that depend on each other. We could also
177  * define option callback handlers, but this is easier.
178  */
179 static int fixup_options(struct thread_data *td)
180 {
181         struct thread_options *o = &td->o;
182
183         if (o->rwmix[DDIR_READ] + o->rwmix[DDIR_WRITE] > 100)
184                 o->rwmix[DDIR_WRITE] = 100 - o->rwmix[DDIR_READ];
185
186         if (o->write_iolog_file && o->read_iolog_file) {
187                 log_err("fio: read iolog overrides write_iolog\n");
188                 free(o->write_iolog_file);
189                 o->write_iolog_file = NULL;
190         }
191
192         if (td->io_ops->flags & FIO_SYNCIO)
193                 o->iodepth = 1;
194         else {
195                 if (!o->iodepth)
196                         o->iodepth = o->open_files;
197         }
198
199         /*
200          * only really works for sequential io for now, and with 1 file
201          */
202         if (o->zone_size && td_random(td) && o->open_files == 1)
203                 o->zone_size = 0;
204
205         /*
206          * Reads can do overwrites, we always need to pre-create the file
207          */
208         if (td_read(td) || td_rw(td))
209                 o->overwrite = 1;
210
211         if (!o->min_bs[DDIR_READ])
212                 o->min_bs[DDIR_READ]= o->bs[DDIR_READ];
213         if (!o->max_bs[DDIR_READ])
214                 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
215         if (!o->min_bs[DDIR_WRITE])
216                 o->min_bs[DDIR_WRITE]= o->bs[DDIR_WRITE];
217         if (!o->max_bs[DDIR_WRITE])
218                 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
219
220         o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
221
222         if (!o->file_size_high)
223                 o->file_size_high = o->file_size_low;
224
225         if (td_read(td) && !td_rw(td))
226                 o->verify = 0;
227
228         if (o->norandommap && o->verify != VERIFY_NONE) {
229                 log_err("fio: norandommap given, verify disabled\n");
230                 o->verify = VERIFY_NONE;
231         }
232         if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
233                 log_err("fio: bs_unaligned may not work with raw io\n");
234
235         /*
236          * thinktime_spin must be less than thinktime
237          */
238         if (o->thinktime_spin > o->thinktime)
239                 o->thinktime_spin = o->thinktime;
240
241         /*
242          * The low water mark cannot be bigger than the iodepth
243          */
244         if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
245                 /*
246                  * syslet work around - if the workload is sequential,
247                  * we want to let the queue drain all the way down to
248                  * avoid seeking between async threads
249                  */
250                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
251                         o->iodepth_low = 1;
252                 else
253                         o->iodepth_low = o->iodepth;
254         }
255
256         /*
257          * If batch number isn't set, default to the same as iodepth
258          */
259         if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
260                 o->iodepth_batch = o->iodepth;
261
262         if (o->nr_files > td->files_index)
263                 o->nr_files = td->files_index;
264
265         if (o->open_files > o->nr_files || !o->open_files)
266                 o->open_files = o->nr_files;
267
268         if ((o->rate && o->rate_iops) || (o->ratemin && o->rate_iops_min)) {
269                 log_err("fio: rate and rate_iops are mutually exclusive\n");
270                 return 1;
271         }
272         if ((o->rate < o->ratemin) || (o->rate_iops < o->rate_iops_min)) {
273                 log_err("fio: minimum rate exceeds rate\n");
274                 return 1;
275         }
276
277         return 0;
278 }
279
280 /*
281  * This function leaks the buffer
282  */
283 static char *to_kmg(unsigned int val)
284 {
285         char *buf = malloc(32);
286         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
287         char *p = post;
288
289         do {
290                 if (val & 1023)
291                         break;
292
293                 val >>= 10;
294                 p++;
295         } while (*p);
296
297         snprintf(buf, 31, "%u%c", val, *p);
298         return buf;
299 }
300
301 /* External engines are specified by "external:name.o") */
302 static const char *get_engine_name(const char *str)
303 {
304         char *p = strstr(str, ":");
305
306         if (!p)
307                 return str;
308
309         p++;
310         strip_blank_front(&p);
311         strip_blank_end(p);
312         return p;
313 }
314
315 static int exists_and_not_file(const char *filename)
316 {
317         struct stat sb;
318
319         if (lstat(filename, &sb) == -1)
320                 return 0;
321
322         if (S_ISREG(sb.st_mode))
323                 return 0;
324
325         return 1;
326 }
327
328 /*
329  * Initialize the various random states we need (random io, block size ranges,
330  * read/write mix, etc).
331  */
332 static int init_random_state(struct thread_data *td)
333 {
334         unsigned long seeds[6];
335         int fd;
336
337         fd = open("/dev/urandom", O_RDONLY);
338         if (fd == -1) {
339                 td_verror(td, errno, "open");
340                 return 1;
341         }
342
343         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
344                 td_verror(td, EIO, "read");
345                 close(fd);
346                 return 1;
347         }
348
349         close(fd);
350
351         os_random_seed(seeds[0], &td->bsrange_state);
352         os_random_seed(seeds[1], &td->verify_state);
353         os_random_seed(seeds[2], &td->rwmix_state);
354
355         if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
356                 os_random_seed(seeds[3], &td->next_file_state);
357
358         os_random_seed(seeds[5], &td->file_size_state);
359
360         if (!td_random(td))
361                 return 0;
362
363         if (td->o.rand_repeatable)
364                 seeds[4] = FIO_RANDSEED * td->thread_number;
365
366         os_random_seed(seeds[4], &td->random_state);
367         return 0;
368 }
369
370 /*
371  * Adds a job to the list of things todo. Sanitizes the various options
372  * to make sure we don't have conflicts, and initializes various
373  * members of td.
374  */
375 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
376 {
377         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
378                                    "randread", "randwrite", "randrw" };
379         unsigned int i;
380         const char *engine;
381         char fname[PATH_MAX];
382         int numjobs, file_alloced;
383
384         /*
385          * the def_thread is just for options, it's not a real job
386          */
387         if (td == &def_thread)
388                 return 0;
389
390         /*
391          * if we are just dumping the output command line, don't add the job
392          */
393         if (dump_cmdline) {
394                 put_job(td);
395                 return 0;
396         }
397
398         engine = get_engine_name(td->o.ioengine);
399         td->io_ops = load_ioengine(td, engine);
400         if (!td->io_ops) {
401                 log_err("fio: failed to load engine %s\n", engine);
402                 goto err;
403         }
404
405         if (td->o.use_thread)
406                 nr_thread++;
407         else
408                 nr_process++;
409
410         if (td->o.odirect)
411                 td->io_ops->flags |= FIO_RAWIO;
412
413         file_alloced = 0;
414         if (!td->o.filename && !td->files_index) {
415                 file_alloced = 1;
416
417                 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
418                         add_file(td, jobname);
419                 else {
420                         for (i = 0; i < td->o.nr_files; i++) {
421                                 sprintf(fname, "%s.%d.%d", jobname, td->thread_number, i);
422                                 add_file(td, fname);
423                         }
424                 }
425         }
426
427         if (fixup_options(td))
428                 goto err;
429
430         td->mutex = fio_sem_init(0);
431
432         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
433         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
434         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
435         td->ddir_nr = td->o.ddir_nr;
436
437         if ((td->o.stonewall || td->o.numjobs > 1 || td->o.new_group)
438              && prev_group_jobs) {
439                 prev_group_jobs = 0;
440                 groupid++;
441         }
442
443         td->groupid = groupid;
444         prev_group_jobs++;
445
446         if (init_random_state(td))
447                 goto err;
448
449         if (setup_rate(td))
450                 goto err;
451
452         if (td->o.write_lat_log) {
453                 setup_log(&td->ts.slat_log);
454                 setup_log(&td->ts.clat_log);
455         }
456         if (td->o.write_bw_log)
457                 setup_log(&td->ts.bw_log);
458
459         if (!td->o.name)
460                 td->o.name = strdup(jobname);
461
462         if (!terse_output) {
463                 if (!job_add_num) {
464                         if (!strcmp(td->io_ops->name, "cpuio"))
465                                 log_info("%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->o.name, td->o.cpuload, td->o.cpucycle);
466                         else {
467                                 char *c1, *c2, *c3, *c4;
468
469                                 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
470                                 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
471                                 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
472                                 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
473
474                                 log_info("%s: (g=%d): rw=%s, bs=%s-%s/%s-%s, ioengine=%s, iodepth=%u\n", td->o.name, td->groupid, ddir_str[td->o.td_ddir], c1, c2, c3, c4, td->io_ops->name, td->o.iodepth);
475
476                                 free(c1);
477                                 free(c2);
478                                 free(c3);
479                                 free(c4);
480                         }
481                 } else if (job_add_num == 1)
482                         log_info("...\n");
483         }
484
485         /*
486          * recurse add identical jobs, clear numjobs and stonewall options
487          * as they don't apply to sub-jobs
488          */
489         numjobs = td->o.numjobs;
490         while (--numjobs) {
491                 struct thread_data *td_new = get_new_job(0, td);
492
493                 if (!td_new)
494                         goto err;
495
496                 td_new->o.numjobs = 1;
497                 td_new->o.stonewall = 0;
498                 td_new->o.new_group = 0;
499
500                 if (file_alloced) {
501                         td_new->o.filename = NULL;
502                         td_new->files_index = 0;
503                         td_new->files = NULL;
504                 }
505
506                 job_add_num = numjobs - 1;
507
508                 if (add_job(td_new, jobname, job_add_num))
509                         goto err;
510         }
511
512         return 0;
513 err:
514         put_job(td);
515         return -1;
516 }
517
518 static int is_empty_or_comment(char *line)
519 {
520         unsigned int i;
521
522         for (i = 0; i < strlen(line); i++) {
523                 if (line[i] == ';')
524                         return 1;
525                 if (line[i] == '#')
526                         return 1;
527                 if (!isspace(line[i]) && !iscntrl(line[i]))
528                         return 0;
529         }
530
531         return 1;
532 }
533
534 /*
535  * This is our [ini] type file parser.
536  */
537 static int parse_jobs_ini(char *file, int stonewall_flag)
538 {
539         unsigned int global;
540         struct thread_data *td;
541         char *string, *name;
542         fpos_t off;
543         FILE *f;
544         char *p;
545         int ret = 0, stonewall;
546
547         f = fopen(file, "r");
548         if (!f) {
549                 perror("fopen job file");
550                 return 1;
551         }
552
553         string = malloc(4096);
554         name = malloc(256);
555         memset(name, 0, 256);
556
557         stonewall = stonewall_flag;
558         do {
559                 p = fgets(string, 4095, f);
560                 if (!p)
561                         break;
562
563                 strip_blank_front(&p);
564                 strip_blank_end(p);
565
566                 if (is_empty_or_comment(p))
567                         continue;
568                 if (sscanf(p, "[%255s]", name) != 1) {
569                         log_err("fio: option <%s> outside of job section\n", p);
570                         continue;
571                 }
572
573                 global = !strncmp(name, "global", 6);
574
575                 name[strlen(name) - 1] = '\0';
576
577                 if (dump_cmdline) {
578                         log_info("fio ");
579                         if (!global)
580                                 log_info("--name=%s ", name);
581                 }
582
583                 td = get_new_job(global, &def_thread);
584                 if (!td) {
585                         ret = 1;
586                         break;
587                 }
588
589                 /*
590                  * Seperate multiple job files by a stonewall
591                  */
592                 if (!global && stonewall) {
593                         td->o.stonewall = stonewall;
594                         stonewall = 0;
595                 }
596
597                 fgetpos(f, &off);
598                 while ((p = fgets(string, 4096, f)) != NULL) {
599                         if (is_empty_or_comment(p))
600                                 continue;
601
602                         strip_blank_front(&p);
603
604                         if (p[0] == '[')
605                                 break;
606
607                         strip_blank_end(p);
608
609                         fgetpos(f, &off);
610
611                         /*
612                          * Don't break here, continue parsing options so we
613                          * dump all the bad ones. Makes trial/error fixups
614                          * easier on the user.
615                          */
616                         ret |= fio_option_parse(td, p);
617                         if (!ret && dump_cmdline)
618                                 log_info("--%s ", p);
619                 }
620
621                 if (!ret) {
622                         fsetpos(f, &off);
623                         ret = add_job(td, name, 0);
624                 } else {
625                         log_err("fio: job %s dropped\n", name);
626                         put_job(td);
627                 }
628         } while (!ret);
629
630         if (dump_cmdline)
631                 log_info("\n");
632
633         free(string);
634         free(name);
635         fclose(f);
636         return ret;
637 }
638
639 static int fill_def_thread(void)
640 {
641         memset(&def_thread, 0, sizeof(def_thread));
642
643         if (fio_getaffinity(getpid(), &def_thread.o.cpumask) == -1) {
644                 perror("sched_getaffinity");
645                 return 1;
646         }
647
648         /*
649          * fill default options
650          */
651         fio_fill_default_options(&def_thread);
652
653         def_thread.o.timeout = def_timeout;
654         def_thread.o.write_bw_log = write_bw_log;
655         def_thread.o.write_lat_log = write_lat_log;
656
657 #ifdef FIO_HAVE_DISK_UTIL
658         def_thread.o.do_disk_util = 1;
659 #endif
660
661         return 0;
662 }
663
664 static void free_shm(void)
665 {
666         struct shmid_ds sbuf;
667
668         if (threads) {
669                 shmdt((void *) threads);
670                 threads = NULL;
671                 shmctl(shm_id, IPC_RMID, &sbuf);
672         }
673 }
674
675 /*
676  * The thread area is shared between the main process and the job
677  * threads/processes. So setup a shared memory segment that will hold
678  * all the job info.
679  */
680 static int setup_thread_area(void)
681 {
682         /*
683          * 1024 is too much on some machines, scale max_jobs if
684          * we get a failure that looks like too large a shm segment
685          */
686         do {
687                 size_t size = max_jobs * sizeof(struct thread_data);
688
689                 shm_id = shmget(0, size, IPC_CREAT | 0600);
690                 if (shm_id != -1)
691                         break;
692                 if (errno != EINVAL) {
693                         perror("shmget");
694                         break;
695                 }
696
697                 max_jobs >>= 1;
698         } while (max_jobs);
699
700         if (shm_id == -1)
701                 return 1;
702
703         threads = shmat(shm_id, NULL, 0);
704         if (threads == (void *) -1) {
705                 perror("shmat");
706                 return 1;
707         }
708
709         atexit(free_shm);
710         return 0;
711 }
712
713 static void usage(void)
714 {
715         printf("%s\n", fio_version_string);
716         printf("\t--output\tWrite output to file\n");
717         printf("\t--timeout\tRuntime in seconds\n");
718         printf("\t--latency-log\tGenerate per-job latency logs\n");
719         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
720         printf("\t--minimal\tMinimal (terse) output\n");
721         printf("\t--version\tPrint version info and exit\n");
722         printf("\t--help\t\tPrint this page\n");
723         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
724         printf("\t--showcmd\tTurn a job file into command line options\n");
725 }
726
727 static int parse_cmd_line(int argc, char *argv[])
728 {
729         struct thread_data *td = NULL;
730         int c, ini_idx = 0, lidx, ret, dont_add_job = 0;
731
732         while ((c = getopt_long_only(argc, argv, "", long_options, &lidx)) != -1) {
733                 switch (c) {
734                 case 't':
735                         def_timeout = atoi(optarg);
736                         break;
737                 case 'l':
738                         write_lat_log = 1;
739                         break;
740                 case 'w':
741                         write_bw_log = 1;
742                         break;
743                 case 'o':
744                         f_out = fopen(optarg, "w+");
745                         if (!f_out) {
746                                 perror("fopen output");
747                                 exit(1);
748                         }
749                         f_err = f_out;
750                         break;
751                 case 'm':
752                         terse_output = 1;
753                         break;
754                 case 'h':
755                         usage();
756                         exit(0);
757                 case 'c':
758                         exit(fio_show_option_help(optarg));
759                 case 's':
760                         dump_cmdline = 1;
761                         break;
762                 case 'v':
763                         printf("%s\n", fio_version_string);
764                         exit(0);
765                 case FIO_GETOPT_JOB: {
766                         const char *opt = long_options[lidx].name;
767                         char *val = optarg;
768
769                         if (!strncmp(opt, "name", 4) && td) {
770                                 ret = add_job(td, td->o.name ?: "fio", 0);
771                                 if (ret) {
772                                         put_job(td);
773                                         return 0;
774                                 }
775                                 td = NULL;
776                         }
777                         if (!td) {
778                                 int global = !strncmp(val, "global", 6);
779
780                                 td = get_new_job(global, &def_thread);
781                                 if (!td)
782                                         return 0;
783                         }
784
785                         ret = fio_cmd_option_parse(td, opt, val);
786                         if (ret)
787                                 dont_add_job = 1;
788                         break;
789                 }
790                 default:
791                         break;
792                 }
793         }
794
795         if (td) {
796                 if (dont_add_job)
797                         put_job(td);
798                 else {
799                         ret = add_job(td, td->o.name ?: "fio", 0);
800                         if (ret)
801                                 put_job(td);
802                 }
803         }
804
805         while (optind < argc) {
806                 ini_idx++;
807                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
808                 ini_file[ini_idx - 1] = strdup(argv[optind]);
809                 optind++;
810         }
811
812         return ini_idx;
813 }
814
815
816 int parse_options(int argc, char *argv[])
817 {
818         int job_files, i;
819
820         f_out = stdout;
821         f_err = stderr;
822
823         fio_options_dup_and_init(long_options);
824
825         if (setup_thread_area())
826                 return 1;
827         if (fill_def_thread())
828                 return 1;
829
830         job_files = parse_cmd_line(argc, argv);
831
832         for (i = 0; i < job_files; i++) {
833                 if (fill_def_thread())
834                         return 1;
835                 if (parse_jobs_ini(ini_file[i], i))
836                         return 1;
837                 free(ini_file[i]);
838         }
839
840         free(ini_file);
841         options_mem_free(&def_thread);
842
843         if (!thread_number) {
844                 if (dump_cmdline)
845                         return 0;
846
847                 log_err("No jobs defined(s)\n");
848                 return 1;
849         }
850
851         return 0;
852 }