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