e5d776c22f6a13481508a565c2d76671d2475faf
[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 1.14a";
21
22 #define FIO_RANDSEED            (0xb1899bedUL)
23
24 static char **ini_file;
25 static int max_jobs = MAX_JOBS;
26
27 struct thread_data def_thread;
28 struct thread_data *threads = NULL;
29
30 int exitall_on_terminate = 0;
31 int terse_output = 0;
32 unsigned long long mlock_size = 0;
33 FILE *f_out = NULL;
34 FILE *f_err = NULL;
35
36 int write_bw_log = 0;
37
38 static int def_timeout = 0;
39 static int write_lat_log = 0;
40
41 static int prev_group_jobs;
42
43 /*
44  * Command line options. These will contain the above, plus a few
45  * extra that only pertain to fio itself and not jobs.
46  */
47 static struct option long_options[FIO_NR_OPTIONS] = {
48         {
49                 .name           = "output",
50                 .has_arg        = required_argument,
51                 .val            = 'o',
52         },
53         {
54                 .name           = "timeout",
55                 .has_arg        = required_argument,
56                 .val            = 't',
57         },
58         {
59                 .name           = "latency-log",
60                 .has_arg        = required_argument,
61                 .val            = 'l',
62         },
63         {
64                 .name           = "bandwidth-log",
65                 .has_arg        = required_argument,
66                 .val            = 'b',
67         },
68         {
69                 .name           = "minimal",
70                 .has_arg        = optional_argument,
71                 .val            = 'm',
72         },
73         {
74                 .name           = "version",
75                 .has_arg        = no_argument,
76                 .val            = 'v',
77         },
78         {
79                 .name           = "help",
80                 .has_arg        = no_argument,
81                 .val            = 'h',
82         },
83         {
84                 .name           = "cmdhelp",
85                 .has_arg        = optional_argument,
86                 .val            = 'c',
87         },
88         {
89                 .name           = NULL,
90         },
91 };
92
93 FILE *get_f_out()
94 {
95         return f_out;
96 }
97
98 FILE *get_f_err()
99 {
100         return f_err;
101 }
102
103 /*
104  * Return a free job structure.
105  */
106 static struct thread_data *get_new_job(int global, struct thread_data *parent)
107 {
108         struct thread_data *td;
109
110         if (global)
111                 return &def_thread;
112         if (thread_number >= max_jobs)
113                 return NULL;
114
115         td = &threads[thread_number++];
116         *td = *parent;
117
118         dup_files(td, parent);
119
120         td->thread_number = thread_number;
121         return td;
122 }
123
124 static void put_job(struct thread_data *td)
125 {
126         if (td == &def_thread)
127                 return;
128
129         if (td->error)
130                 log_info("fio: %s\n", td->verror);
131
132         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
133         thread_number--;
134 }
135
136 static int setup_rate(struct thread_data *td)
137 {
138         unsigned long nr_reads_per_msec;
139         unsigned long long rate;
140         unsigned int bs;
141
142         if (!td->o.rate && !td->o.rate_iops)
143                 return 0;
144
145         if (td_rw(td))
146                 bs = td->o.rw_min_bs;
147         else if (td_read(td))
148                 bs = td->o.min_bs[DDIR_READ];
149         else
150                 bs = td->o.min_bs[DDIR_WRITE];
151
152         if (td->o.rate) {
153                 rate = td->o.rate;
154                 nr_reads_per_msec = (rate * 1024 * 1000LL) / bs;
155         } else
156                 nr_reads_per_msec = td->o.rate_iops * 1000UL;
157
158         if (!nr_reads_per_msec) {
159                 log_err("rate lower than supported\n");
160                 return -1;
161         }
162
163         td->rate_usec_cycle = 1000000000ULL / nr_reads_per_msec;
164         td->rate_pending_usleep = 0;
165         return 0;
166 }
167
168 /*
169  * Lazy way of fixing up options that depend on each other. We could also
170  * define option callback handlers, but this is easier.
171  */
172 static int fixup_options(struct thread_data *td)
173 {
174         struct thread_options *o = &td->o;
175
176         if (o->rwmix[DDIR_READ] + o->rwmix[DDIR_WRITE] > 100)
177                 o->rwmix[DDIR_WRITE] = 100 - o->rwmix[DDIR_READ];
178
179         if (o->write_iolog_file && o->read_iolog_file) {
180                 log_err("fio: read iolog overrides write_iolog\n");
181                 free(o->write_iolog_file);
182                 o->write_iolog_file = NULL;
183         }
184
185         if (td->io_ops->flags & FIO_SYNCIO)
186                 o->iodepth = 1;
187         else {
188                 if (!o->iodepth)
189                         o->iodepth = o->open_files;
190         }
191
192         /*
193          * only really works for sequential io for now, and with 1 file
194          */
195         if (o->zone_size && td_random(td) && o->open_files == 1)
196                 o->zone_size = 0;
197
198         /*
199          * Reads can do overwrites, we always need to pre-create the file
200          */
201         if (td_read(td) || td_rw(td))
202                 o->overwrite = 1;
203
204         if (!o->min_bs[DDIR_READ])
205                 o->min_bs[DDIR_READ]= o->bs[DDIR_READ];
206         if (!o->max_bs[DDIR_READ])
207                 o->max_bs[DDIR_READ] = o->bs[DDIR_READ];
208         if (!o->min_bs[DDIR_WRITE])
209                 o->min_bs[DDIR_WRITE]= o->bs[DDIR_WRITE];
210         if (!o->max_bs[DDIR_WRITE])
211                 o->max_bs[DDIR_WRITE] = o->bs[DDIR_WRITE];
212
213         o->rw_min_bs = min(o->min_bs[DDIR_READ], o->min_bs[DDIR_WRITE]);
214
215         if (!o->file_size_high)
216                 o->file_size_high = o->file_size_low;
217
218         if (td_read(td) && !td_rw(td))
219                 o->verify = 0;
220
221         if (o->norandommap && o->verify != VERIFY_NONE) {
222                 log_err("fio: norandommap given, verify disabled\n");
223                 o->verify = VERIFY_NONE;
224         }
225         if (o->bs_unaligned && (o->odirect || td->io_ops->flags & FIO_RAWIO))
226                 log_err("fio: bs_unaligned may not work with raw io\n");
227
228         /*
229          * thinktime_spin must be less than thinktime
230          */
231         if (o->thinktime_spin > o->thinktime)
232                 o->thinktime_spin = o->thinktime;
233
234         /*
235          * The low water mark cannot be bigger than the iodepth
236          */
237         if (o->iodepth_low > o->iodepth || !o->iodepth_low) {
238                 /*
239                  * syslet work around - if the workload is sequential,
240                  * we want to let the queue drain all the way down to
241                  * avoid seeking between async threads
242                  */
243                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
244                         o->iodepth_low = 1;
245                 else
246                         o->iodepth_low = o->iodepth;
247         }
248
249         /*
250          * If batch number isn't set, default to the same as iodepth
251          */
252         if (o->iodepth_batch > o->iodepth || !o->iodepth_batch)
253                 o->iodepth_batch = o->iodepth;
254
255         if (o->nr_files > td->files_index)
256                 o->nr_files = td->files_index;
257
258         if (o->open_files > o->nr_files || !o->open_files)
259                 o->open_files = o->nr_files;
260
261         if ((o->rate && o->rate_iops) || (o->ratemin && o->rate_iops_min)) {
262                 log_err("fio: rate and rate_iops are mutually exclusive\n");
263                 return 1;
264         }
265         if ((o->rate < o->ratemin) || (o->rate_iops < o->rate_iops_min)) {
266                 log_err("fio: minimum rate exceeds rate\n");
267                 return 1;
268         }
269
270         return 0;
271 }
272
273 /*
274  * This function leaks the buffer
275  */
276 static char *to_kmg(unsigned int val)
277 {
278         char *buf = malloc(32);
279         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
280         char *p = post;
281
282         do {
283                 if (val & 1023)
284                         break;
285
286                 val >>= 10;
287                 p++;
288         } while (*p);
289
290         snprintf(buf, 31, "%u%c", val, *p);
291         return buf;
292 }
293
294 /* External engines are specified by "external:name.o") */
295 static const char *get_engine_name(const char *str)
296 {
297         char *p = strstr(str, ":");
298
299         if (!p)
300                 return str;
301
302         p++;
303         strip_blank_front(&p);
304         strip_blank_end(p);
305         return p;
306 }
307
308 static int exists_and_not_file(const char *filename)
309 {
310         struct stat sb;
311
312         if (lstat(filename, &sb) == -1)
313                 return 0;
314
315         if (S_ISREG(sb.st_mode))
316                 return 0;
317
318         return 1;
319 }
320
321 /*
322  * Initialize the various random states we need (random io, block size ranges,
323  * read/write mix, etc).
324  */
325 static int init_random_state(struct thread_data *td)
326 {
327         unsigned long seeds[6];
328         int fd;
329
330         fd = open("/dev/urandom", O_RDONLY);
331         if (fd == -1) {
332                 td_verror(td, errno, "open");
333                 return 1;
334         }
335
336         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
337                 td_verror(td, EIO, "read");
338                 close(fd);
339                 return 1;
340         }
341
342         close(fd);
343
344         os_random_seed(seeds[0], &td->bsrange_state);
345         os_random_seed(seeds[1], &td->verify_state);
346         os_random_seed(seeds[2], &td->rwmix_state);
347
348         if (td->o.file_service_type == FIO_FSERVICE_RANDOM)
349                 os_random_seed(seeds[3], &td->next_file_state);
350
351         os_random_seed(seeds[5], &td->file_size_state);
352
353         if (!td_random(td))
354                 return 0;
355
356         if (td->o.rand_repeatable)
357                 seeds[4] = FIO_RANDSEED * td->thread_number;
358
359         os_random_seed(seeds[4], &td->random_state);
360         return 0;
361 }
362
363 /*
364  * Adds a job to the list of things todo. Sanitizes the various options
365  * to make sure we don't have conflicts, and initializes various
366  * members of td.
367  */
368 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
369 {
370         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
371                                    "randread", "randwrite", "randrw" };
372         unsigned int i;
373         const char *engine;
374         char fname[PATH_MAX];
375         int numjobs, file_alloced;
376
377         /*
378          * the def_thread is just for options, it's not a real job
379          */
380         if (td == &def_thread)
381                 return 0;
382
383         engine = get_engine_name(td->o.ioengine);
384         td->io_ops = load_ioengine(td, engine);
385         if (!td->io_ops) {
386                 log_err("fio: failed to load engine %s\n", engine);
387                 goto err;
388         }
389
390         if (td->o.use_thread)
391                 nr_thread++;
392         else
393                 nr_process++;
394
395         if (td->o.odirect)
396                 td->io_ops->flags |= FIO_RAWIO;
397
398         file_alloced = 0;
399         if (!td->o.filename && !td->files_index) {
400                 file_alloced = 1;
401
402                 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
403                         add_file(td, jobname);
404                 else {
405                         for (i = 0; i < td->o.nr_files; i++) {
406                                 sprintf(fname, "%s.%d.%d", jobname, td->thread_number, i);
407                                 add_file(td, fname);
408                         }
409                 }
410         }
411
412         if (fixup_options(td))
413                 goto err;
414
415         td->mutex = fio_sem_init(0);
416
417         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
418         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
419         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
420         td->ddir_nr = td->o.ddir_nr;
421
422         if ((td->o.stonewall || td->o.numjobs > 1 || td->o.new_group)
423              && prev_group_jobs) {
424                 prev_group_jobs = 0;
425                 groupid++;
426         }
427
428         td->groupid = groupid;
429         prev_group_jobs++;
430
431         if (init_random_state(td))
432                 goto err;
433
434         if (setup_rate(td))
435                 goto err;
436
437         if (td->o.write_lat_log) {
438                 setup_log(&td->ts.slat_log);
439                 setup_log(&td->ts.clat_log);
440         }
441         if (td->o.write_bw_log)
442                 setup_log(&td->ts.bw_log);
443
444         if (!td->o.name)
445                 td->o.name = strdup(jobname);
446
447         if (!terse_output) {
448                 if (!job_add_num) {
449                         if (!strcmp(td->io_ops->name, "cpuio"))
450                                 log_info("%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->o.name, td->o.cpuload, td->o.cpucycle);
451                         else {
452                                 char *c1, *c2, *c3, *c4;
453
454                                 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
455                                 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
456                                 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
457                                 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
458
459                                 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);
460
461                                 free(c1);
462                                 free(c2);
463                                 free(c3);
464                                 free(c4);
465                         }
466                 } else if (job_add_num == 1)
467                         log_info("...\n");
468         }
469
470         /*
471          * recurse add identical jobs, clear numjobs and stonewall options
472          * as they don't apply to sub-jobs
473          */
474         numjobs = td->o.numjobs;
475         while (--numjobs) {
476                 struct thread_data *td_new = get_new_job(0, td);
477
478                 if (!td_new)
479                         goto err;
480
481                 td_new->o.numjobs = 1;
482                 td_new->o.stonewall = 0;
483
484                 if (file_alloced) {
485                         td_new->o.filename = NULL;
486                         td_new->files_index = 0;
487                         td_new->files = NULL;
488                 }
489
490                 job_add_num = numjobs - 1;
491
492                 if (add_job(td_new, jobname, job_add_num))
493                         goto err;
494         }
495
496         return 0;
497 err:
498         put_job(td);
499         return -1;
500 }
501
502 static int is_empty_or_comment(char *line)
503 {
504         unsigned int i;
505
506         for (i = 0; i < strlen(line); i++) {
507                 if (line[i] == ';')
508                         return 1;
509                 if (line[i] == '#')
510                         return 1;
511                 if (!isspace(line[i]) && !iscntrl(line[i]))
512                         return 0;
513         }
514
515         return 1;
516 }
517
518 /*
519  * This is our [ini] type file parser.
520  */
521 static int parse_jobs_ini(char *file, int stonewall_flag)
522 {
523         unsigned int global;
524         struct thread_data *td;
525         char *string, *name;
526         fpos_t off;
527         FILE *f;
528         char *p;
529         int ret = 0, stonewall;
530
531         f = fopen(file, "r");
532         if (!f) {
533                 perror("fopen job file");
534                 return 1;
535         }
536
537         string = malloc(4096);
538         name = malloc(256);
539         memset(name, 0, 256);
540
541         stonewall = stonewall_flag;
542         do {
543                 p = fgets(string, 4095, f);
544                 if (!p)
545                         break;
546                 if (is_empty_or_comment(p))
547                         continue;
548                 if (sscanf(p, "[%255s]", name) != 1)
549                         continue;
550
551                 global = !strncmp(name, "global", 6);
552
553                 name[strlen(name) - 1] = '\0';
554
555                 td = get_new_job(global, &def_thread);
556                 if (!td) {
557                         ret = 1;
558                         break;
559                 }
560
561                 /*
562                  * Seperate multiple job files by a stonewall
563                  */
564                 if (!global && stonewall) {
565                         td->o.stonewall = stonewall;
566                         stonewall = 0;
567                 }
568
569                 fgetpos(f, &off);
570                 while ((p = fgets(string, 4096, f)) != NULL) {
571                         if (is_empty_or_comment(p))
572                                 continue;
573
574                         strip_blank_front(&p);
575
576                         if (p[0] == '[')
577                                 break;
578
579                         strip_blank_end(p);
580
581                         fgetpos(f, &off);
582
583                         /*
584                          * Don't break here, continue parsing options so we
585                          * dump all the bad ones. Makes trial/error fixups
586                          * easier on the user.
587                          */
588                         ret |= fio_option_parse(td, p);
589                 }
590
591                 if (!ret) {
592                         fsetpos(f, &off);
593                         ret = add_job(td, name, 0);
594                 } else {
595                         log_err("fio: job %s dropped\n", name);
596                         put_job(td);
597                 }
598         } while (!ret);
599
600         free(string);
601         free(name);
602         fclose(f);
603         return ret;
604 }
605
606 static int fill_def_thread(void)
607 {
608         memset(&def_thread, 0, sizeof(def_thread));
609
610         if (fio_getaffinity(getpid(), &def_thread.o.cpumask) == -1) {
611                 perror("sched_getaffinity");
612                 return 1;
613         }
614
615         /*
616          * fill default options
617          */
618         fio_fill_default_options(&def_thread);
619
620         def_thread.o.timeout = def_timeout;
621         def_thread.o.write_bw_log = write_bw_log;
622         def_thread.o.write_lat_log = write_lat_log;
623
624 #ifdef FIO_HAVE_DISK_UTIL
625         def_thread.o.do_disk_util = 1;
626 #endif
627
628         return 0;
629 }
630
631 static void free_shm(void)
632 {
633         struct shmid_ds sbuf;
634
635         if (threads) {
636                 shmdt((void *) threads);
637                 threads = NULL;
638                 shmctl(shm_id, IPC_RMID, &sbuf);
639         }
640 }
641
642 /*
643  * The thread area is shared between the main process and the job
644  * threads/processes. So setup a shared memory segment that will hold
645  * all the job info.
646  */
647 static int setup_thread_area(void)
648 {
649         /*
650          * 1024 is too much on some machines, scale max_jobs if
651          * we get a failure that looks like too large a shm segment
652          */
653         do {
654                 size_t size = max_jobs * sizeof(struct thread_data);
655
656                 shm_id = shmget(0, size, IPC_CREAT | 0600);
657                 if (shm_id != -1)
658                         break;
659                 if (errno != EINVAL) {
660                         perror("shmget");
661                         break;
662                 }
663
664                 max_jobs >>= 1;
665         } while (max_jobs);
666
667         if (shm_id == -1)
668                 return 1;
669
670         threads = shmat(shm_id, NULL, 0);
671         if (threads == (void *) -1) {
672                 perror("shmat");
673                 return 1;
674         }
675
676         atexit(free_shm);
677         return 0;
678 }
679
680 static void usage(void)
681 {
682         printf("%s\n", fio_version_string);
683         printf("\t--output\tWrite output to file\n");
684         printf("\t--timeout\tRuntime in seconds\n");
685         printf("\t--latency-log\tGenerate per-job latency logs\n");
686         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
687         printf("\t--minimal\tMinimal (terse) output\n");
688         printf("\t--version\tPrint version info and exit\n");
689         printf("\t--help\t\tPrint this page\n");
690         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
691 }
692
693 static int parse_cmd_line(int argc, char *argv[])
694 {
695         struct thread_data *td = NULL;
696         int c, ini_idx = 0, lidx, ret, dont_add_job = 0;
697
698         while ((c = getopt_long_only(argc, argv, "", long_options, &lidx)) != -1) {
699                 switch (c) {
700                 case 't':
701                         def_timeout = atoi(optarg);
702                         break;
703                 case 'l':
704                         write_lat_log = 1;
705                         break;
706                 case 'w':
707                         write_bw_log = 1;
708                         break;
709                 case 'o':
710                         f_out = fopen(optarg, "w+");
711                         if (!f_out) {
712                                 perror("fopen output");
713                                 exit(1);
714                         }
715                         f_err = f_out;
716                         break;
717                 case 'm':
718                         terse_output = 1;
719                         break;
720                 case 'h':
721                         usage();
722                         exit(0);
723                 case 'c':
724                         exit(fio_show_option_help(optarg));
725                 case 'v':
726                         printf("%s\n", fio_version_string);
727                         exit(0);
728                 case FIO_GETOPT_JOB: {
729                         const char *opt = long_options[lidx].name;
730                         char *val = optarg;
731
732                         if (!strncmp(opt, "name", 4) && td) {
733                                 ret = add_job(td, td->o.name ?: "fio", 0);
734                                 if (ret) {
735                                         put_job(td);
736                                         return 0;
737                                 }
738                                 td = NULL;
739                         }
740                         if (!td) {
741                                 int global = !strncmp(val, "global", 6);
742
743                                 td = get_new_job(global, &def_thread);
744                                 if (!td)
745                                         return 0;
746                         }
747
748                         ret = fio_cmd_option_parse(td, opt, val);
749                         if (ret)
750                                 dont_add_job = 1;
751                         break;
752                 }
753                 default:
754                         break;
755                 }
756         }
757
758         if (td) {
759                 if (dont_add_job)
760                         put_job(td);
761                 else {
762                         ret = add_job(td, td->o.name ?: "fio", 0);
763                         if (ret)
764                                 put_job(td);
765                 }
766         }
767
768         while (optind < argc) {
769                 ini_idx++;
770                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
771                 ini_file[ini_idx - 1] = strdup(argv[optind]);
772                 optind++;
773         }
774
775         return ini_idx;
776 }
777
778
779 int parse_options(int argc, char *argv[])
780 {
781         int job_files, i;
782
783         f_out = stdout;
784         f_err = stderr;
785
786         fio_options_dup_and_init(long_options);
787
788         if (setup_thread_area())
789                 return 1;
790         if (fill_def_thread())
791                 return 1;
792
793         job_files = parse_cmd_line(argc, argv);
794
795         for (i = 0; i < job_files; i++) {
796                 if (fill_def_thread())
797                         return 1;
798                 if (parse_jobs_ini(ini_file[i], i))
799                         return 1;
800                 free(ini_file[i]);
801         }
802
803         free(ini_file);
804
805         if (!thread_number) {
806                 log_err("No jobs defined(s)\n");
807                 return 1;
808         }
809
810         return 0;
811 }