Fix file unlinking
[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  * Adds a job to the list of things todo. Sanitizes the various options
363  * to make sure we don't have conflicts, and initializes various
364  * members of td.
365  */
366 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
367 {
368         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
369                                    "randread", "randwrite", "randrw" };
370         unsigned int i;
371         const char *engine;
372         char fname[PATH_MAX];
373         int numjobs, file_alloced;
374
375         /*
376          * the def_thread is just for options, it's not a real job
377          */
378         if (td == &def_thread)
379                 return 0;
380
381         engine = get_engine_name(td->o.ioengine);
382         td->io_ops = load_ioengine(td, engine);
383         if (!td->io_ops) {
384                 log_err("fio: failed to load engine %s\n", engine);
385                 goto err;
386         }
387
388         if (td->o.use_thread)
389                 nr_thread++;
390         else
391                 nr_process++;
392
393         if (td->o.odirect)
394                 td->io_ops->flags |= FIO_RAWIO;
395
396         file_alloced = 0;
397         if (!td->o.filename && !td->files_index) {
398                 file_alloced = 1;
399
400                 if (td->o.nr_files == 1 && exists_and_not_file(jobname))
401                         add_file(td, jobname);
402                 else {
403                         for (i = 0; i < td->o.nr_files; i++) {
404                                 sprintf(fname, "%s.%d.%d", jobname, td->thread_number, i);
405                                 add_file(td, fname);
406                         }
407                 }
408         }
409
410         if (fixup_options(td))
411                 goto err;
412
413         td->mutex = fio_sem_init(0);
414
415         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
416         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
417         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
418         td->ddir_nr = td->o.ddir_nr;
419
420         if ((td->o.stonewall || td->o.numjobs > 1 || td->o.new_group)
421              && prev_group_jobs) {
422                 prev_group_jobs = 0;
423                 groupid++;
424         }
425
426         td->groupid = groupid;
427         prev_group_jobs++;
428
429         if (init_random_state(td))
430                 goto err;
431
432         if (setup_rate(td))
433                 goto err;
434
435         if (td->o.write_lat_log) {
436                 setup_log(&td->ts.slat_log);
437                 setup_log(&td->ts.clat_log);
438         }
439         if (td->o.write_bw_log)
440                 setup_log(&td->ts.bw_log);
441
442         if (!td->o.name)
443                 td->o.name = strdup(jobname);
444
445         if (!terse_output) {
446                 if (!job_add_num) {
447                         if (!strcmp(td->io_ops->name, "cpuio"))
448                                 log_info("%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->o.name, td->o.cpuload, td->o.cpucycle);
449                         else {
450                                 char *c1, *c2, *c3, *c4;
451
452                                 c1 = to_kmg(td->o.min_bs[DDIR_READ]);
453                                 c2 = to_kmg(td->o.max_bs[DDIR_READ]);
454                                 c3 = to_kmg(td->o.min_bs[DDIR_WRITE]);
455                                 c4 = to_kmg(td->o.max_bs[DDIR_WRITE]);
456
457                                 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);
458
459                                 free(c1);
460                                 free(c2);
461                                 free(c3);
462                                 free(c4);
463                         }
464                 } else if (job_add_num == 1)
465                         log_info("...\n");
466         }
467
468         /*
469          * recurse add identical jobs, clear numjobs and stonewall options
470          * as they don't apply to sub-jobs
471          */
472         numjobs = td->o.numjobs;
473         while (--numjobs) {
474                 struct thread_data *td_new = get_new_job(0, td);
475
476                 if (!td_new)
477                         goto err;
478
479                 td_new->o.numjobs = 1;
480                 td_new->o.stonewall = 0;
481
482                 if (file_alloced) {
483                         td_new->o.filename = NULL;
484                         td_new->files_index = 0;
485                         td_new->files = NULL;
486                 }
487
488                 job_add_num = numjobs - 1;
489
490                 if (add_job(td_new, jobname, job_add_num))
491                         goto err;
492         }
493
494         return 0;
495 err:
496         put_job(td);
497         return -1;
498 }
499
500 static int is_empty_or_comment(char *line)
501 {
502         unsigned int i;
503
504         for (i = 0; i < strlen(line); i++) {
505                 if (line[i] == ';')
506                         return 1;
507                 if (line[i] == '#')
508                         return 1;
509                 if (!isspace(line[i]) && !iscntrl(line[i]))
510                         return 0;
511         }
512
513         return 1;
514 }
515
516 /*
517  * This is our [ini] type file parser.
518  */
519 static int parse_jobs_ini(char *file, int stonewall_flag)
520 {
521         unsigned int global;
522         struct thread_data *td;
523         char *string, *name;
524         fpos_t off;
525         FILE *f;
526         char *p;
527         int ret = 0, stonewall;
528
529         f = fopen(file, "r");
530         if (!f) {
531                 perror("fopen job file");
532                 return 1;
533         }
534
535         string = malloc(4096);
536         name = malloc(256);
537         memset(name, 0, 256);
538
539         stonewall = stonewall_flag;
540         do {
541                 p = fgets(string, 4095, f);
542                 if (!p)
543                         break;
544                 if (is_empty_or_comment(p))
545                         continue;
546                 if (sscanf(p, "[%255s]", name) != 1)
547                         continue;
548
549                 global = !strncmp(name, "global", 6);
550
551                 name[strlen(name) - 1] = '\0';
552
553                 td = get_new_job(global, &def_thread);
554                 if (!td) {
555                         ret = 1;
556                         break;
557                 }
558
559                 /*
560                  * Seperate multiple job files by a stonewall
561                  */
562                 if (!global && stonewall) {
563                         td->o.stonewall = stonewall;
564                         stonewall = 0;
565                 }
566
567                 fgetpos(f, &off);
568                 while ((p = fgets(string, 4096, f)) != NULL) {
569                         if (is_empty_or_comment(p))
570                                 continue;
571
572                         strip_blank_front(&p);
573
574                         if (p[0] == '[')
575                                 break;
576
577                         strip_blank_end(p);
578
579                         fgetpos(f, &off);
580
581                         /*
582                          * Don't break here, continue parsing options so we
583                          * dump all the bad ones. Makes trial/error fixups
584                          * easier on the user.
585                          */
586                         ret |= fio_option_parse(td, p);
587                 }
588
589                 if (!ret) {
590                         fsetpos(f, &off);
591                         ret = add_job(td, name, 0);
592                 } else {
593                         log_err("fio: job %s dropped\n", name);
594                         put_job(td);
595                 }
596         } while (!ret);
597
598         free(string);
599         free(name);
600         fclose(f);
601         return ret;
602 }
603
604 static int fill_def_thread(void)
605 {
606         memset(&def_thread, 0, sizeof(def_thread));
607
608         if (fio_getaffinity(getpid(), &def_thread.o.cpumask) == -1) {
609                 perror("sched_getaffinity");
610                 return 1;
611         }
612
613         /*
614          * fill default options
615          */
616         fio_fill_default_options(&def_thread);
617
618         def_thread.o.timeout = def_timeout;
619         def_thread.o.write_bw_log = write_bw_log;
620         def_thread.o.write_lat_log = write_lat_log;
621
622 #ifdef FIO_HAVE_DISK_UTIL
623         def_thread.o.do_disk_util = 1;
624 #endif
625
626         return 0;
627 }
628
629 static void free_shm(void)
630 {
631         struct shmid_ds sbuf;
632
633         if (threads) {
634                 shmdt((void *) threads);
635                 threads = NULL;
636                 shmctl(shm_id, IPC_RMID, &sbuf);
637         }
638 }
639
640 /*
641  * The thread area is shared between the main process and the job
642  * threads/processes. So setup a shared memory segment that will hold
643  * all the job info.
644  */
645 static int setup_thread_area(void)
646 {
647         /*
648          * 1024 is too much on some machines, scale max_jobs if
649          * we get a failure that looks like too large a shm segment
650          */
651         do {
652                 size_t size = max_jobs * sizeof(struct thread_data);
653
654                 shm_id = shmget(0, size, IPC_CREAT | 0600);
655                 if (shm_id != -1)
656                         break;
657                 if (errno != EINVAL) {
658                         perror("shmget");
659                         break;
660                 }
661
662                 max_jobs >>= 1;
663         } while (max_jobs);
664
665         if (shm_id == -1)
666                 return 1;
667
668         threads = shmat(shm_id, NULL, 0);
669         if (threads == (void *) -1) {
670                 perror("shmat");
671                 return 1;
672         }
673
674         atexit(free_shm);
675         return 0;
676 }
677
678 static void usage(void)
679 {
680         printf("%s\n", fio_version_string);
681         printf("\t--output\tWrite output to file\n");
682         printf("\t--timeout\tRuntime in seconds\n");
683         printf("\t--latency-log\tGenerate per-job latency logs\n");
684         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
685         printf("\t--minimal\tMinimal (terse) output\n");
686         printf("\t--version\tPrint version info and exit\n");
687         printf("\t--help\t\tPrint this page\n");
688         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
689 }
690
691 static int parse_cmd_line(int argc, char *argv[])
692 {
693         struct thread_data *td = NULL;
694         int c, ini_idx = 0, lidx, ret, dont_add_job = 0;
695
696         while ((c = getopt_long_only(argc, argv, "", long_options, &lidx)) != -1) {
697                 switch (c) {
698                 case 't':
699                         def_timeout = atoi(optarg);
700                         break;
701                 case 'l':
702                         write_lat_log = 1;
703                         break;
704                 case 'w':
705                         write_bw_log = 1;
706                         break;
707                 case 'o':
708                         f_out = fopen(optarg, "w+");
709                         if (!f_out) {
710                                 perror("fopen output");
711                                 exit(1);
712                         }
713                         f_err = f_out;
714                         break;
715                 case 'm':
716                         terse_output = 1;
717                         break;
718                 case 'h':
719                         usage();
720                         exit(0);
721                 case 'c':
722                         exit(fio_show_option_help(optarg));
723                 case 'v':
724                         printf("%s\n", fio_version_string);
725                         exit(0);
726                 case FIO_GETOPT_JOB: {
727                         const char *opt = long_options[lidx].name;
728                         char *val = optarg;
729
730                         if (!strncmp(opt, "name", 4) && td) {
731                                 ret = add_job(td, td->o.name ?: "fio", 0);
732                                 if (ret) {
733                                         put_job(td);
734                                         return 0;
735                                 }
736                                 td = NULL;
737                         }
738                         if (!td) {
739                                 int global = !strncmp(val, "global", 6);
740
741                                 td = get_new_job(global, &def_thread);
742                                 if (!td)
743                                         return 0;
744                         }
745
746                         ret = fio_cmd_option_parse(td, opt, val);
747                         if (ret)
748                                 dont_add_job = 1;
749                         break;
750                 }
751                 default:
752                         break;
753                 }
754         }
755
756         if (td) {
757                 if (dont_add_job)
758                         put_job(td);
759                 else {
760                         ret = add_job(td, td->o.name ?: "fio", 0);
761                         if (ret)
762                                 put_job(td);
763                 }
764         }
765
766         while (optind < argc) {
767                 ini_idx++;
768                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
769                 ini_file[ini_idx - 1] = strdup(argv[optind]);
770                 optind++;
771         }
772
773         return ini_idx;
774 }
775
776
777 int parse_options(int argc, char *argv[])
778 {
779         int job_files, i;
780
781         f_out = stdout;
782         f_err = stderr;
783
784         fio_options_dup_and_init(long_options);
785
786         if (setup_thread_area())
787                 return 1;
788         if (fill_def_thread())
789                 return 1;
790
791         job_files = parse_cmd_line(argc, argv);
792
793         for (i = 0; i < job_files; i++) {
794                 if (fill_def_thread())
795                         return 1;
796                 if (parse_jobs_ini(ini_file[i], i))
797                         return 1;
798                 free(ini_file[i]);
799         }
800
801         free(ini_file);
802
803         if (!thread_number) {
804                 log_err("No jobs defined(s)\n");
805                 return 1;
806         }
807
808         return 0;
809 }