Speedup init_disk_util() when dealing with many jobs
[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 <assert.h>
13 #include <sys/ipc.h>
14 #include <sys/shm.h>
15 #include <sys/types.h>
16 #include <sys/stat.h>
17
18 #include "fio.h"
19 #include "parse.h"
20
21 #define FIO_RANDSEED            (0xb1899bedUL)
22
23 #define td_var_offset(var)      ((size_t) &((struct thread_data *)0)->var)
24
25 static int str_ioengine_cb(void *, const char *);
26 static int str_mem_cb(void *, const char *);
27 static int str_lockmem_cb(void *, unsigned long *);
28 #ifdef FIO_HAVE_IOPRIO
29 static int str_prio_cb(void *, unsigned int *);
30 static int str_prioclass_cb(void *, unsigned int *);
31 #endif
32 static int str_exitall_cb(void);
33 static int str_cpumask_cb(void *, unsigned int *);
34
35 #define __stringify_1(x)        #x
36 #define __stringify(x)          __stringify_1(x)
37
38 /*
39  * Map of job/command line options
40  */
41 static struct fio_option options[] = {
42         {
43                 .name   = "description",
44                 .type   = FIO_OPT_STR_STORE,
45                 .off1   = td_var_offset(description),
46                 .help   = "Text job description",
47         },
48         {
49                 .name   = "name",
50                 .type   = FIO_OPT_STR_STORE,
51                 .off1   = td_var_offset(name),
52                 .help   = "Name of this job",
53         },
54         {
55                 .name   = "directory",
56                 .type   = FIO_OPT_STR_STORE,
57                 .off1   = td_var_offset(directory),
58                 .help   = "Directory to store files in",
59         },
60         {
61                 .name   = "filename",
62                 .type   = FIO_OPT_STR_STORE,
63                 .off1   = td_var_offset(filename),
64                 .help   = "Force the use of a specific file",
65         },
66         {
67                 .name   = "rw",
68                 .type   = FIO_OPT_STR,
69                 .off1   = td_var_offset(td_ddir),
70                 .help   = "IO direction",
71                 .def    = "read",
72                 .posval = {
73                           { .ival = "read", .oval = TD_DDIR_READ },
74                           { .ival = "write", .oval = TD_DDIR_WRITE },
75                           { .ival = "randread", .oval = TD_DDIR_RANDREAD },
76                           { .ival = "randwrite", .oval = TD_DDIR_RANDWRITE },
77                           { .ival = "rw", .oval = TD_DDIR_RW },
78                           { .ival = "randrw", .oval = TD_DDIR_RANDRW },
79                           },
80         },
81         {
82                 .name   = "ioengine",
83                 .type   = FIO_OPT_STR,
84                 .cb     = str_ioengine_cb,
85                 .help   = "IO engine to use",
86                 .def    = "sync",
87                 .posval = {
88                           { .ival = "sync", },
89 #ifdef FIO_HAVE_LIBAIO
90                           { .ival = "libaio", },
91 #endif
92 #ifdef FIO_HAVE_POSIXAIO
93                           { .ival = "posixaio", },
94 #endif
95                           { .ival = "mmap", },
96 #ifdef FIO_HAVE_SPLICE
97                           { .ival = "splice", },
98 #endif
99 #ifdef FIO_HAVE_SGIO
100                           { .ival = "sg", },
101 #endif
102                           { .ival = "null", }, { .ival = "net", },
103 #ifdef FIO_HAVE_SYSLET
104                           { .ival = "syslet-rw", },
105 #endif
106                           },
107         },
108         {
109                 .name   = "iodepth",
110                 .type   = FIO_OPT_INT,
111                 .off1   = td_var_offset(iodepth),
112                 .help   = "Amount of IO buffers to keep in flight",
113                 .def    = "1",
114         },
115         {
116                 .name   = "iodepth_low",
117                 .type   = FIO_OPT_INT,
118                 .off1   = td_var_offset(iodepth_low),
119                 .help   = "Low water mark for queuing depth",
120         },
121         {
122                 .name   = "size",
123                 .type   = FIO_OPT_STR_VAL,
124                 .off1   = td_var_offset(total_file_size),
125                 .help   = "Size of device or file",
126         },
127         {
128                 .name   = "bs",
129                 .type   = FIO_OPT_STR_VAL_INT,
130                 .off1   = td_var_offset(bs[DDIR_READ]),
131                 .off2   = td_var_offset(bs[DDIR_WRITE]),
132                 .help   = "Block size unit",
133                 .def    = "4k",
134         },
135         {
136                 .name   = "bsrange",
137                 .type   = FIO_OPT_RANGE,
138                 .off1   = td_var_offset(min_bs[DDIR_READ]),
139                 .off2   = td_var_offset(max_bs[DDIR_READ]),
140                 .off3   = td_var_offset(min_bs[DDIR_WRITE]),
141                 .off4   = td_var_offset(max_bs[DDIR_WRITE]),
142                 .help   = "Set block size range (in more detail than bs)",
143         },
144         {
145                 .name   = "bs_unaligned",
146                 .type   = FIO_OPT_STR_SET,
147                 .off1   = td_var_offset(bs_unaligned),
148                 .help   = "Don't sector align IO buffer sizes",
149         },
150         {
151                 .name   = "offset",
152                 .type   = FIO_OPT_STR_VAL,
153                 .off1   = td_var_offset(start_offset),
154                 .help   = "Start IO from this offset",
155                 .def    = "0",
156         },
157         {
158                 .name   = "randrepeat",
159                 .type   = FIO_OPT_BOOL,
160                 .off1   = td_var_offset(rand_repeatable),
161                 .help   = "Use repeatable random IO pattern",
162                 .def    = "1",
163         },
164         {
165                 .name   = "norandommap",
166                 .type   = FIO_OPT_STR_SET,
167                 .off1   = td_var_offset(norandommap),
168                 .help   = "Accept potential duplicate random blocks",
169         },
170         {
171                 .name   = "nrfiles",
172                 .type   = FIO_OPT_INT,
173                 .off1   = td_var_offset(nr_files),
174                 .help   = "Split job workload between this number of files",
175                 .def    = "1",
176         },
177         {
178                 .name   = "file_service_type",
179                 .type   = FIO_OPT_STR,
180                 .off1   = td_var_offset(file_service_type),
181                 .help   = "How to select which file to service next",
182                 .def    = "roundrobin",
183                 .posval = {
184                           { .ival = "random", .oval = FIO_FSERVICE_RANDOM },
185                           { .ival = "roundrobin", .oval = FIO_FSERVICE_RR },
186                           },
187         },
188         {
189                 .name   = "fsync",
190                 .type   = FIO_OPT_INT,
191                 .off1   = td_var_offset(fsync_blocks),
192                 .help   = "Issue fsync for writes every given number of blocks",
193                 .def    = "0",
194         },
195         {
196                 .name   = "direct",
197                 .type   = FIO_OPT_BOOL,
198                 .off1   = td_var_offset(odirect),
199                 .help   = "Use O_DIRECT IO (negates buffered)",
200                 .def    = "0",
201         },
202         {
203                 .name   = "buffered",
204                 .type   = FIO_OPT_BOOL,
205                 .off1   = td_var_offset(odirect),
206                 .neg    = 1,
207                 .help   = "Use buffered IO (negates direct)",
208                 .def    = "1",
209         },
210         {
211                 .name   = "overwrite",
212                 .type   = FIO_OPT_BOOL,
213                 .off1   = td_var_offset(overwrite),
214                 .help   = "When writing, set whether to overwrite current data",
215                 .def    = "0",
216         },
217         {
218                 .name   = "loops",
219                 .type   = FIO_OPT_INT,
220                 .off1   = td_var_offset(loops),
221                 .help   = "Number of times to run the job",
222                 .def    = "1",
223         },
224         {
225                 .name   = "numjobs",
226                 .type   = FIO_OPT_INT,
227                 .off1   = td_var_offset(numjobs),
228                 .help   = "Duplicate this job this many times",
229                 .def    = "1",
230         },
231         {
232                 .name   = "startdelay",
233                 .type   = FIO_OPT_INT,
234                 .off1   = td_var_offset(start_delay),
235                 .help   = "Only start job when this period has passed",
236                 .def    = "0",
237         },
238         {
239                 .name   = "runtime",
240                 .alias  = "timeout",
241                 .type   = FIO_OPT_STR_VAL_TIME,
242                 .off1   = td_var_offset(timeout),
243                 .help   = "Stop workload when this amount of time has passed",
244                 .def    = "0",
245         },
246         {
247                 .name   = "mem",
248                 .type   = FIO_OPT_STR,
249                 .cb     = str_mem_cb,
250                 .off1   = td_var_offset(mem_type),
251                 .help   = "Backing type for IO buffers",
252                 .def    = "malloc",
253                 .posval = {
254                           { .ival = "malloc", .oval = MEM_MALLOC },
255                           { .ival = "shm", .oval = MEM_SHM },
256 #ifdef FIO_HAVE_HUGETLB
257                           { .ival = "shmhuge", .oval = MEM_SHMHUGE },
258 #endif
259                           { .ival = "mmap", .oval = MEM_MMAP },
260 #ifdef FIO_HAVE_HUGETLB
261                           { .ival = "mmaphuge", .oval = MEM_MMAPHUGE },
262 #endif
263                           },
264         },
265         {
266                 .name   = "verify",
267                 .type   = FIO_OPT_STR,
268                 .off1   = td_var_offset(verify),
269                 .help   = "Verify sum function",
270                 .def    = "0",
271                 .posval = {
272                           { .ival = "0", .oval = VERIFY_NONE },
273                           { .ival = "crc32", .oval = VERIFY_CRC32 },
274                           { .ival = "md5", .oval = VERIFY_MD5 },
275                           },
276         },
277         {
278                 .name   = "write_iolog",
279                 .type   = FIO_OPT_STR_STORE,
280                 .off1   = td_var_offset(write_iolog_file),
281                 .help   = "Store IO pattern to file",
282         },
283         {
284                 .name   = "read_iolog",
285                 .type   = FIO_OPT_STR_STORE,
286                 .off1   = td_var_offset(read_iolog_file),
287                 .help   = "Playback IO pattern from file",
288         },
289         {
290                 .name   = "exec_prerun",
291                 .type   = FIO_OPT_STR_STORE,
292                 .off1   = td_var_offset(exec_prerun),
293                 .help   = "Execute this file prior to running job",
294         },
295         {
296                 .name   = "exec_postrun",
297                 .type   = FIO_OPT_STR_STORE,
298                 .off1   = td_var_offset(exec_postrun),
299                 .help   = "Execute this file after running job",
300         },
301 #ifdef FIO_HAVE_IOSCHED_SWITCH
302         {
303                 .name   = "ioscheduler",
304                 .type   = FIO_OPT_STR_STORE,
305                 .off1   = td_var_offset(ioscheduler),
306                 .help   = "Use this IO scheduler on the backing device",
307         },
308 #endif
309         {
310                 .name   = "zonesize",
311                 .type   = FIO_OPT_STR_VAL,
312                 .off1   = td_var_offset(zone_size),
313                 .help   = "Give size of an IO zone",
314                 .def    = "0",
315         },
316         {
317                 .name   = "zoneskip",
318                 .type   = FIO_OPT_STR_VAL,
319                 .off1   = td_var_offset(zone_skip),
320                 .help   = "Space between IO zones",
321                 .def    = "0",
322         },
323         {
324                 .name   = "lockmem",
325                 .type   = FIO_OPT_STR_VAL,
326                 .cb     = str_lockmem_cb,
327                 .help   = "Lock down this amount of memory",
328                 .def    = "0",
329         },
330         {
331                 .name   = "rwmixcycle",
332                 .type   = FIO_OPT_INT,
333                 .off1   = td_var_offset(rwmixcycle),
334                 .help   = "Cycle period for mixed read/write workloads (msec)",
335                 .def    = "500",
336         },
337         {
338                 .name   = "rwmixread",
339                 .type   = FIO_OPT_INT,
340                 .off1   = td_var_offset(rwmixread),
341                 .maxval = 100,
342                 .help   = "Percentage of mixed workload that is reads",
343                 .def    = "50",
344         },
345         {
346                 .name   = "rwmixwrite",
347                 .type   = FIO_OPT_INT,
348                 .off1   = td_var_offset(rwmixwrite),
349                 .maxval = 100,
350                 .help   = "Percentage of mixed workload that is writes",
351                 .def    = "50",
352         },
353         {
354                 .name   = "nice",
355                 .type   = FIO_OPT_INT,
356                 .off1   = td_var_offset(nice),
357                 .help   = "Set job CPU nice value",
358                 .minval = -19,
359                 .maxval = 20,
360                 .def    = "0",
361         },
362 #ifdef FIO_HAVE_IOPRIO
363         {
364                 .name   = "prio",
365                 .type   = FIO_OPT_INT,
366                 .cb     = str_prio_cb,
367                 .help   = "Set job IO priority value",
368                 .minval = 0,
369                 .maxval = 7,
370         },
371         {
372                 .name   = "prioclass",
373                 .type   = FIO_OPT_INT,
374                 .cb     = str_prioclass_cb,
375                 .help   = "Set job IO priority class",
376                 .minval = 0,
377                 .maxval = 3,
378         },
379 #endif
380         {
381                 .name   = "thinktime",
382                 .type   = FIO_OPT_INT,
383                 .off1   = td_var_offset(thinktime),
384                 .help   = "Idle time between IO buffers (usec)",
385                 .def    = "0",
386         },
387         {
388                 .name   = "thinktime_spin",
389                 .type   = FIO_OPT_INT,
390                 .off1   = td_var_offset(thinktime_spin),
391                 .help   = "Start thinktime by spinning this amount (usec)",
392                 .def    = "0",
393         },
394         {
395                 .name   = "thinktime_blocks",
396                 .type   = FIO_OPT_INT,
397                 .off1   = td_var_offset(thinktime_blocks),
398                 .help   = "IO buffer period between 'thinktime'",
399                 .def    = "1",
400         },
401         {
402                 .name   = "rate",
403                 .type   = FIO_OPT_INT,
404                 .off1   = td_var_offset(rate),
405                 .help   = "Set bandwidth rate",
406         },
407         {
408                 .name   = "ratemin",
409                 .type   = FIO_OPT_INT,
410                 .off1   = td_var_offset(ratemin),
411                 .help   = "The bottom limit accepted",
412         },
413         {
414                 .name   = "ratecycle",
415                 .type   = FIO_OPT_INT,
416                 .off1   = td_var_offset(ratecycle),
417                 .help   = "Window average for rate limits (msec)",
418                 .def    = "1000",
419         },
420         {
421                 .name   = "invalidate",
422                 .type   = FIO_OPT_BOOL,
423                 .off1   = td_var_offset(invalidate_cache),
424                 .help   = "Invalidate buffer/page cache prior to running job",
425                 .def    = "1",
426         },
427         {
428                 .name   = "sync",
429                 .type   = FIO_OPT_BOOL,
430                 .off1   = td_var_offset(sync_io),
431                 .help   = "Use O_SYNC for buffered writes",
432                 .def    = "0",
433         },
434         {
435                 .name   = "bwavgtime",
436                 .type   = FIO_OPT_INT,
437                 .off1   = td_var_offset(bw_avg_time),
438                 .help   = "Time window over which to calculate bandwidth (msec)",
439                 .def    = "500",
440         },
441         {
442                 .name   = "create_serialize",
443                 .type   = FIO_OPT_BOOL,
444                 .off1   = td_var_offset(create_serialize),
445                 .help   = "Serialize creating of job files",
446                 .def    = "1",
447         },
448         {
449                 .name   = "create_fsync",
450                 .type   = FIO_OPT_BOOL,
451                 .off1   = td_var_offset(create_fsync),
452                 .help   = "Fsync file after creation",
453                 .def    = "1",
454         },
455         {
456                 .name   = "cpuload",
457                 .type   = FIO_OPT_INT,
458                 .off1   = td_var_offset(cpuload),
459                 .help   = "Use this percentage of CPU",
460         },
461         {
462                 .name   = "cpuchunks",
463                 .type   = FIO_OPT_INT,
464                 .off1   = td_var_offset(cpucycle),
465                 .help   = "Length of the CPU burn cycles",
466         },
467 #ifdef FIO_HAVE_CPU_AFFINITY
468         {
469                 .name   = "cpumask",
470                 .type   = FIO_OPT_INT,
471                 .cb     = str_cpumask_cb,
472                 .help   = "CPU affinity mask",
473         },
474 #endif
475         {
476                 .name   = "end_fsync",
477                 .type   = FIO_OPT_BOOL,
478                 .off1   = td_var_offset(end_fsync),
479                 .help   = "Include fsync at the end of job",
480                 .def    = "0",
481         },
482         {
483                 .name   = "unlink",
484                 .type   = FIO_OPT_BOOL,
485                 .off1   = td_var_offset(unlink),
486                 .help   = "Unlink created files after job has completed",
487                 .def    = "0",
488         },
489         {
490                 .name   = "exitall",
491                 .type   = FIO_OPT_STR_SET,
492                 .cb     = str_exitall_cb,
493                 .help   = "Terminate all jobs when one exits",
494         },
495         {
496                 .name   = "stonewall",
497                 .type   = FIO_OPT_STR_SET,
498                 .off1   = td_var_offset(stonewall),
499                 .help   = "Insert a hard barrier between this job and previous",
500         },
501         {
502                 .name   = "thread",
503                 .type   = FIO_OPT_STR_SET,
504                 .off1   = td_var_offset(use_thread),
505                 .help   = "Use threads instead of forks",
506         },
507         {
508                 .name   = "write_bw_log",
509                 .type   = FIO_OPT_STR_SET,
510                 .off1   = td_var_offset(write_bw_log),
511                 .help   = "Write log of bandwidth during run",
512         },
513         {
514                 .name   = "write_lat_log",
515                 .type   = FIO_OPT_STR_SET,
516                 .off1   = td_var_offset(write_lat_log),
517                 .help   = "Write log of latency during run",
518         },
519         {
520                 .name   = "hugepage-size",
521                 .type   = FIO_OPT_STR_VAL,
522                 .off1   = td_var_offset(hugepage_size),
523                 .help   = "When using hugepages, specify size of each page",
524                 .def    = __stringify(FIO_HUGE_PAGE),
525         },
526         {
527                 .name = NULL,
528         },
529 };
530
531 #define FIO_JOB_OPTS    (sizeof(options) / sizeof(struct fio_option))
532 #define FIO_CMD_OPTS    (16)
533 #define FIO_GETOPT_JOB  (0x89988998)
534
535 /*
536  * Command line options. These will contain the above, plus a few
537  * extra that only pertain to fio itself and not jobs.
538  */
539 static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
540         {
541                 .name           = "output",
542                 .has_arg        = required_argument,
543                 .val            = 'o',
544         },
545         {
546                 .name           = "timeout",
547                 .has_arg        = required_argument,
548                 .val            = 't',
549         },
550         {
551                 .name           = "latency-log",
552                 .has_arg        = required_argument,
553                 .val            = 'l',
554         },
555         {
556                 .name           = "bandwidth-log",
557                 .has_arg        = required_argument,
558                 .val            = 'b',
559         },
560         {
561                 .name           = "minimal",
562                 .has_arg        = optional_argument,
563                 .val            = 'm',
564         },
565         {
566                 .name           = "version",
567                 .has_arg        = no_argument,
568                 .val            = 'v',
569         },
570         {
571                 .name           = "help",
572                 .has_arg        = no_argument,
573                 .val            = 'h',
574         },
575         {
576                 .name           = "cmdhelp",
577                 .has_arg        = required_argument,
578                 .val            = 'c',
579         },
580         {
581                 .name           = NULL,
582         },
583 };
584
585 static int def_timeout = 0;
586
587 static char fio_version_string[] = "fio 1.11";
588
589 static char **ini_file;
590 static int max_jobs = MAX_JOBS;
591
592 struct thread_data def_thread;
593 struct thread_data *threads = NULL;
594
595 int exitall_on_terminate = 0;
596 int terse_output = 0;
597 unsigned long long mlock_size = 0;
598 FILE *f_out = NULL;
599 FILE *f_err = NULL;
600
601 static int write_lat_log = 0;
602 int write_bw_log = 0;
603
604 /*
605  * Return a free job structure.
606  */
607 static struct thread_data *get_new_job(int global, struct thread_data *parent)
608 {
609         struct thread_data *td;
610
611         if (global)
612                 return &def_thread;
613         if (thread_number >= max_jobs)
614                 return NULL;
615
616         td = &threads[thread_number++];
617         *td = *parent;
618
619         td->thread_number = thread_number;
620         return td;
621 }
622
623 static void put_job(struct thread_data *td)
624 {
625         if (td == &def_thread)
626                 return;
627
628         if (td->error)
629                 fprintf(f_out, "fio: %s\n", td->verror);
630
631         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
632         thread_number--;
633 }
634
635 /*
636  * Lazy way of fixing up options that depend on each other. We could also
637  * define option callback handlers, but this is easier.
638  */
639 static void fixup_options(struct thread_data *td)
640 {
641         if (!td->rwmixread && td->rwmixwrite)
642                 td->rwmixread = 100 - td->rwmixwrite;
643
644         if (td->write_iolog_file && td->read_iolog_file) {
645                 log_err("fio: read iolog overrides write_iolog\n");
646                 free(td->write_iolog_file);
647                 td->write_iolog_file = NULL;
648         }
649
650         if (td->io_ops->flags & FIO_SYNCIO)
651                 td->iodepth = 1;
652         else {
653                 if (!td->iodepth)
654                         td->iodepth = td->nr_files;
655         }
656
657         /*
658          * only really works for sequential io for now, and with 1 file
659          */
660         if (td->zone_size && td_random(td) && td->nr_files == 1)
661                 td->zone_size = 0;
662
663         /*
664          * Reads can do overwrites, we always need to pre-create the file
665          */
666         if (td_read(td) || td_rw(td))
667                 td->overwrite = 1;
668
669         if (!td->min_bs[DDIR_READ])
670                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
671         if (!td->max_bs[DDIR_READ])
672                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
673         if (!td->min_bs[DDIR_WRITE])
674                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
675         if (!td->max_bs[DDIR_WRITE])
676                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
677
678         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
679
680         if (td_read(td) && !td_rw(td))
681                 td->verify = 0;
682
683         if (td->norandommap && td->verify != VERIFY_NONE) {
684                 log_err("fio: norandommap given, verify disabled\n");
685                 td->verify = VERIFY_NONE;
686         }
687         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
688                 log_err("fio: bs_unaligned may not work with raw io\n");
689
690         /*
691          * O_DIRECT and char doesn't mix, clear that flag if necessary.
692          */
693         if (td->filetype == FIO_TYPE_CHAR && td->odirect)
694                 td->odirect = 0;
695
696         /*
697          * thinktime_spin must be less than thinktime
698          */
699         if (td->thinktime_spin > td->thinktime)
700                 td->thinktime_spin = td->thinktime;
701
702         /*
703          * The low water mark cannot be bigger than the iodepth
704          */
705         if (td->iodepth_low > td->iodepth || !td->iodepth_low)
706                 td->iodepth_low = td->iodepth;
707 }
708
709 /*
710  * This function leaks the buffer
711  */
712 static char *to_kmg(unsigned int val)
713 {
714         char *buf = malloc(32);
715         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
716         char *p = post;
717
718         do {
719                 if (val & 1023)
720                         break;
721
722                 val >>= 10;
723                 p++;
724         } while (*p);
725
726         snprintf(buf, 31, "%u%c", val, *p);
727         return buf;
728 }
729
730 /*
731  * Adds a job to the list of things todo. Sanitizes the various options
732  * to make sure we don't have conflicts, and initializes various
733  * members of td.
734  */
735 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
736 {
737         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
738                                    "randread", "randwrite", "randrw" };
739         struct stat sb;
740         int numjobs, i;
741         struct fio_file *f;
742
743         /*
744          * the def_thread is just for options, it's not a real job
745          */
746         if (td == &def_thread)
747                 return 0;
748
749         assert(td->io_ops);
750
751         if (td->odirect)
752                 td->io_ops->flags |= FIO_RAWIO;
753
754         td->filetype = FIO_TYPE_FILE;
755         if (td->filename && !lstat(td->filename, &sb)) {
756                 if (S_ISBLK(sb.st_mode))
757                         td->filetype = FIO_TYPE_BD;
758                 else if (S_ISCHR(sb.st_mode))
759                         td->filetype = FIO_TYPE_CHAR;
760         }
761
762         fixup_options(td);
763
764         if (td->filename)
765                 td->nr_uniq_files = 1;
766         else
767                 td->nr_uniq_files = td->nr_files;
768
769         if (td->filetype == FIO_TYPE_FILE || td->filename) {
770                 char tmp[PATH_MAX];
771                 int len = 0;
772
773                 if (td->directory && td->directory[0] != '\0') {
774                         if (lstat(td->directory, &sb) < 0) {
775                                 log_err("fio: %s is not a directory\n", td->directory);
776                                 td_verror(td, errno, "lstat");
777                                 return 1;
778                         }
779                         if (!S_ISDIR(sb.st_mode)) {
780                                 log_err("fio: %s is not a directory\n", td->directory);
781                                 return 1;
782                         }
783                         len = sprintf(tmp, "%s/", td->directory);
784                 }
785
786                 td->files = malloc(sizeof(struct fio_file) * td->nr_files);
787
788                 for_each_file(td, f, i) {
789                         memset(f, 0, sizeof(*f));
790                         f->fd = -1;
791
792                         if (td->filename)
793                                 sprintf(tmp + len, "%s", td->filename);
794                         else
795                                 sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
796                         f->file_name = strdup(tmp);
797                 }
798         } else {
799                 td->nr_files = 1;
800                 td->files = malloc(sizeof(struct fio_file));
801                 f = &td->files[0];
802
803                 memset(f, 0, sizeof(*f));
804                 f->fd = -1;
805                 f->file_name = strdup(jobname);
806         }
807
808         for_each_file(td, f, i) {
809                 f->file_size = td->total_file_size / td->nr_files;
810                 f->file_offset = td->start_offset;
811         }
812                 
813         fio_sem_init(&td->mutex, 0);
814
815         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
816         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
817         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
818
819         if (td->stonewall && td->thread_number > 1)
820                 groupid++;
821
822         td->groupid = groupid;
823
824         if (setup_rate(td))
825                 goto err;
826
827         if (td->write_lat_log) {
828                 setup_log(&td->ts.slat_log);
829                 setup_log(&td->ts.clat_log);
830         }
831         if (td->write_bw_log)
832                 setup_log(&td->ts.bw_log);
833
834         if (!td->name)
835                 td->name = strdup(jobname);
836
837         if (!terse_output) {
838                 if (!job_add_num) {
839                         if (td->io_ops->flags & FIO_CPUIO)
840                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
841                         else {
842                                 char *c1, *c2, *c3, *c4;
843
844                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
845                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
846                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
847                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
848
849                                 fprintf(f_out, "%s: (g=%d): rw=%s, bs=%s-%s/%s-%s, ioengine=%s, iodepth=%u\n", td->name, td->groupid, ddir_str[td->td_ddir], c1, c2, c3, c4, td->io_ops->name, td->iodepth);
850
851                                 free(c1);
852                                 free(c2);
853                                 free(c3);
854                                 free(c4);
855                         }
856                 } else if (job_add_num == 1)
857                         fprintf(f_out, "...\n");
858         }
859
860         /*
861          * recurse add identical jobs, clear numjobs and stonewall options
862          * as they don't apply to sub-jobs
863          */
864         numjobs = td->numjobs;
865         while (--numjobs) {
866                 struct thread_data *td_new = get_new_job(0, td);
867
868                 if (!td_new)
869                         goto err;
870
871                 td_new->numjobs = 1;
872                 td_new->stonewall = 0;
873                 job_add_num = numjobs - 1;
874
875                 if (add_job(td_new, jobname, job_add_num))
876                         goto err;
877         }
878         return 0;
879 err:
880         put_job(td);
881         return -1;
882 }
883
884 /*
885  * Initialize the various random states we need (random io, block size ranges,
886  * read/write mix, etc).
887  */
888 int init_random_state(struct thread_data *td)
889 {
890         unsigned long seeds[5];
891         int fd, num_maps, blocks, i;
892         struct fio_file *f;
893
894         if (td->io_ops->flags & FIO_CPUIO)
895                 return 0;
896
897         fd = open("/dev/urandom", O_RDONLY);
898         if (fd == -1) {
899                 td_verror(td, errno, "open");
900                 return 1;
901         }
902
903         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
904                 td_verror(td, EIO, "read");
905                 close(fd);
906                 return 1;
907         }
908
909         close(fd);
910
911         os_random_seed(seeds[0], &td->bsrange_state);
912         os_random_seed(seeds[1], &td->verify_state);
913         os_random_seed(seeds[2], &td->rwmix_state);
914
915         if (td->file_service_type == FIO_FSERVICE_RANDOM)
916                 os_random_seed(seeds[3], &td->next_file_state);
917
918         if (!td_random(td))
919                 return 0;
920
921         if (td->rand_repeatable)
922                 seeds[4] = FIO_RANDSEED * td->thread_number;
923
924         if (!td->norandommap) {
925                 for_each_file(td, f, i) {
926                         blocks = (f->real_file_size + td->rw_min_bs - 1) / td->rw_min_bs;
927                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
928                         f->file_map = malloc(num_maps * sizeof(long));
929                         f->num_maps = num_maps;
930                         memset(f->file_map, 0, num_maps * sizeof(long));
931                 }
932         }
933
934         os_random_seed(seeds[4], &td->random_state);
935         return 0;
936 }
937
938 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
939 {
940 #ifdef FIO_HAVE_CPU_AFFINITY
941         unsigned int i;
942
943         CPU_ZERO(&cpumask);
944
945         for (i = 0; i < sizeof(int) * 8; i++) {
946                 if ((1 << i) & cpu)
947                         CPU_SET(i, &cpumask);
948         }
949 #endif
950 }
951
952 static int is_empty_or_comment(char *line)
953 {
954         unsigned int i;
955
956         for (i = 0; i < strlen(line); i++) {
957                 if (line[i] == ';')
958                         return 1;
959                 if (line[i] == '#')
960                         return 1;
961                 if (!isspace(line[i]) && !iscntrl(line[i]))
962                         return 0;
963         }
964
965         return 1;
966 }
967
968 /*
969  * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
970  */
971 static char *get_mmap_file(const char *str)
972 {
973         char *p = strstr(str, ":");
974
975         if (!p)
976                 return NULL;
977
978         p++;
979         strip_blank_front(&p);
980         strip_blank_end(p);
981         return strdup(p);
982 }
983
984 static int str_mem_cb(void *data, const char *mem)
985 {
986         struct thread_data *td = data;
987
988         if (td->mem_type == MEM_MMAPHUGE || td->mem_type == MEM_MMAP) {
989                 td->mmapfile = get_mmap_file(mem);
990                 if (td->mem_type == MEM_MMAPHUGE && !td->mmapfile) {
991                         log_err("fio: mmaphuge:/path/to/file\n");
992                         return 1;
993                 }
994         }
995
996         return 0;
997 }
998
999 static int str_ioengine_cb(void *data, const char *str)
1000 {
1001         struct thread_data *td = data;
1002
1003         td->io_ops = load_ioengine(td, str);
1004         if (td->io_ops)
1005                 return 0;
1006
1007         return 1;
1008 }
1009
1010 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
1011 {
1012         mlock_size = *val;
1013         return 0;
1014 }
1015
1016 #ifdef FIO_HAVE_IOPRIO
1017 static int str_prioclass_cb(void *data, unsigned int *val)
1018 {
1019         struct thread_data *td = data;
1020
1021         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
1022         return 0;
1023 }
1024
1025 static int str_prio_cb(void *data, unsigned int *val)
1026 {
1027         struct thread_data *td = data;
1028
1029         td->ioprio |= *val;
1030         return 0;
1031 }
1032 #endif
1033
1034 static int str_exitall_cb(void)
1035 {
1036         exitall_on_terminate = 1;
1037         return 0;
1038 }
1039
1040 static int str_cpumask_cb(void *data, unsigned int *val)
1041 {
1042         struct thread_data *td = data;
1043
1044         fill_cpu_mask(td->cpumask, *val);
1045         return 0;
1046 }
1047
1048 /*
1049  * This is our [ini] type file parser.
1050  */
1051 static int parse_jobs_ini(char *file, int stonewall_flag)
1052 {
1053         unsigned int global;
1054         struct thread_data *td;
1055         char *string, *name;
1056         fpos_t off;
1057         FILE *f;
1058         char *p;
1059         int ret = 0, stonewall;
1060
1061         f = fopen(file, "r");
1062         if (!f) {
1063                 perror("fopen job file");
1064                 return 1;
1065         }
1066
1067         string = malloc(4096);
1068         name = malloc(256);
1069         memset(name, 0, 256);
1070
1071         stonewall = stonewall_flag;
1072         do {
1073                 p = fgets(string, 4095, f);
1074                 if (!p)
1075                         break;
1076                 if (is_empty_or_comment(p))
1077                         continue;
1078                 if (sscanf(p, "[%255s]", name) != 1)
1079                         continue;
1080
1081                 global = !strncmp(name, "global", 6);
1082
1083                 name[strlen(name) - 1] = '\0';
1084
1085                 td = get_new_job(global, &def_thread);
1086                 if (!td) {
1087                         ret = 1;
1088                         break;
1089                 }
1090
1091                 /*
1092                  * Seperate multiple job files by a stonewall
1093                  */
1094                 if (!global && stonewall) {
1095                         td->stonewall = stonewall;
1096                         stonewall = 0;
1097                 }
1098
1099                 fgetpos(f, &off);
1100                 while ((p = fgets(string, 4096, f)) != NULL) {
1101                         if (is_empty_or_comment(p))
1102                                 continue;
1103
1104                         strip_blank_front(&p);
1105
1106                         if (p[0] == '[')
1107                                 break;
1108
1109                         strip_blank_end(p);
1110
1111                         fgetpos(f, &off);
1112
1113                         /*
1114                          * Don't break here, continue parsing options so we
1115                          * dump all the bad ones. Makes trial/error fixups
1116                          * easier on the user.
1117                          */
1118                         ret |= parse_option(p, options, td);
1119                 }
1120
1121                 if (!ret) {
1122                         fsetpos(f, &off);
1123                         ret = add_job(td, name, 0);
1124                 } else {
1125                         log_err("fio: job %s dropped\n", name);
1126                         put_job(td);
1127                 }
1128         } while (!ret);
1129
1130         free(string);
1131         free(name);
1132         fclose(f);
1133         return ret;
1134 }
1135
1136 static int fill_def_thread(void)
1137 {
1138         memset(&def_thread, 0, sizeof(def_thread));
1139
1140         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1141                 perror("sched_getaffinity");
1142                 return 1;
1143         }
1144
1145         /*
1146          * fill default options
1147          */
1148         fill_default_options(&def_thread, options);
1149
1150         def_thread.timeout = def_timeout;
1151         def_thread.write_bw_log = write_bw_log;
1152         def_thread.write_lat_log = write_lat_log;
1153
1154 #ifdef FIO_HAVE_DISK_UTIL
1155         def_thread.do_disk_util = 1;
1156 #endif
1157
1158         return 0;
1159 }
1160
1161 static void usage(void)
1162 {
1163         printf("%s\n", fio_version_string);
1164         printf("\t--output\tWrite output to file\n");
1165         printf("\t--timeout\tRuntime in seconds\n");
1166         printf("\t--latency-log\tGenerate per-job latency logs\n");
1167         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1168         printf("\t--minimal\tMinimal (terse) output\n");
1169         printf("\t--version\tPrint version info and exit\n");
1170         printf("\t--help\t\tPrint this page\n");
1171         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1172 }
1173
1174 static int parse_cmd_line(int argc, char *argv[])
1175 {
1176         struct thread_data *td = NULL;
1177         int c, ini_idx = 0, lidx, ret, dont_add_job = 0;
1178
1179         while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1180                 switch (c) {
1181                 case 't':
1182                         def_timeout = atoi(optarg);
1183                         break;
1184                 case 'l':
1185                         write_lat_log = 1;
1186                         break;
1187                 case 'w':
1188                         write_bw_log = 1;
1189                         break;
1190                 case 'o':
1191                         f_out = fopen(optarg, "w+");
1192                         if (!f_out) {
1193                                 perror("fopen output");
1194                                 exit(1);
1195                         }
1196                         f_err = f_out;
1197                         break;
1198                 case 'm':
1199                         terse_output = 1;
1200                         break;
1201                 case 'h':
1202                         usage();
1203                         exit(0);
1204                 case 'c':
1205                         ret = show_cmd_help(options, optarg);
1206                         exit(ret);
1207                 case 'v':
1208                         printf("%s\n", fio_version_string);
1209                         exit(0);
1210                 case FIO_GETOPT_JOB: {
1211                         const char *opt = long_options[lidx].name;
1212                         char *val = optarg;
1213
1214                         if (!strncmp(opt, "name", 4) && td) {
1215                                 ret = add_job(td, td->name ?: "fio", 0);
1216                                 if (ret) {
1217                                         put_job(td);
1218                                         return 0;
1219                                 }
1220                                 td = NULL;
1221                         }
1222                         if (!td) {
1223                                 int global = !strncmp(val, "global", 6);
1224
1225                                 td = get_new_job(global, &def_thread);
1226                                 if (!td)
1227                                         return 0;
1228                         }
1229
1230                         ret = parse_cmd_option(opt, val, options, td);
1231                         if (ret)
1232                                 dont_add_job = 1;
1233                         break;
1234                 }
1235                 default:
1236                         break;
1237                 }
1238         }
1239
1240         if (td) {
1241                 if (dont_add_job)
1242                         put_job(td);
1243                 else {
1244                         ret = add_job(td, td->name ?: "fio", 0);
1245                         if (ret)
1246                                 put_job(td);
1247                 }
1248         }
1249
1250         while (optind < argc) {
1251                 ini_idx++;
1252                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1253                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1254                 optind++;
1255         }
1256
1257         return ini_idx;
1258 }
1259
1260 static void free_shm(void)
1261 {
1262         struct shmid_ds sbuf;
1263
1264         if (threads) {
1265                 shmdt((void *) threads);
1266                 threads = NULL;
1267                 shmctl(shm_id, IPC_RMID, &sbuf);
1268         }
1269 }
1270
1271 /*
1272  * The thread area is shared between the main process and the job
1273  * threads/processes. So setup a shared memory segment that will hold
1274  * all the job info.
1275  */
1276 static int setup_thread_area(void)
1277 {
1278         /*
1279          * 1024 is too much on some machines, scale max_jobs if
1280          * we get a failure that looks like too large a shm segment
1281          */
1282         do {
1283                 size_t size = max_jobs * sizeof(struct thread_data);
1284
1285                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1286                 if (shm_id != -1)
1287                         break;
1288                 if (errno != EINVAL) {
1289                         perror("shmget");
1290                         break;
1291                 }
1292
1293                 max_jobs >>= 1;
1294         } while (max_jobs);
1295
1296         if (shm_id == -1)
1297                 return 1;
1298
1299         threads = shmat(shm_id, NULL, 0);
1300         if (threads == (void *) -1) {
1301                 perror("shmat");
1302                 return 1;
1303         }
1304
1305         atexit(free_shm);
1306         return 0;
1307 }
1308
1309 /*
1310  * Copy the fio options into the long options map, so we mirror
1311  * job and cmd line options.
1312  */
1313 static void dupe_job_options(void)
1314 {
1315         struct fio_option *o;
1316         unsigned int i;
1317
1318         i = 0;
1319         while (long_options[i].name)
1320                 i++;
1321
1322         o = &options[0];
1323         while (o->name) {
1324                 long_options[i].name = o->name;
1325                 long_options[i].val = FIO_GETOPT_JOB;
1326                 if (o->type == FIO_OPT_STR_SET)
1327                         long_options[i].has_arg = no_argument;
1328                 else
1329                         long_options[i].has_arg = required_argument;
1330
1331                 i++;
1332                 o++;
1333                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1334         }
1335 }
1336
1337 int parse_options(int argc, char *argv[])
1338 {
1339         int job_files, i;
1340
1341         f_out = stdout;
1342         f_err = stderr;
1343
1344         options_init(options);
1345
1346         dupe_job_options();
1347
1348         if (setup_thread_area())
1349                 return 1;
1350         if (fill_def_thread())
1351                 return 1;
1352
1353         job_files = parse_cmd_line(argc, argv);
1354
1355         for (i = 0; i < job_files; i++) {
1356                 if (fill_def_thread())
1357                         return 1;
1358                 if (parse_jobs_ini(ini_file[i], i))
1359                         return 1;
1360                 free(ini_file[i]);
1361         }
1362
1363         free(ini_file);
1364
1365         if (!thread_number) {
1366                 log_err("No jobs defined(s)\n");
1367                 return 1;
1368         }
1369
1370         return 0;
1371 }