9fc8ff740fd85013f57e94c217addc4b6a28214b
[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_mem_cb(void *, const char *);
26 static int str_lockmem_cb(void *, unsigned long *);
27 #ifdef FIO_HAVE_IOPRIO
28 static int str_prio_cb(void *, unsigned int *);
29 static int str_prioclass_cb(void *, unsigned int *);
30 #endif
31 static int str_exitall_cb(void);
32 static int str_cpumask_cb(void *, unsigned int *);
33 static int str_fst_cb(void *, const char *);
34 static int str_filename_cb(void *, const char *);
35 static int str_directory_cb(void *, const char *);
36 static int str_opendir_cb(void *, const char *);
37
38 #define __stringify_1(x)        #x
39 #define __stringify(x)          __stringify_1(x)
40
41 /*
42  * Map of job/command line options
43  */
44 static struct fio_option options[] = {
45         {
46                 .name   = "description",
47                 .type   = FIO_OPT_STR_STORE,
48                 .off1   = td_var_offset(description),
49                 .help   = "Text job description",
50         },
51         {
52                 .name   = "name",
53                 .type   = FIO_OPT_STR_STORE,
54                 .off1   = td_var_offset(name),
55                 .help   = "Name of this job",
56         },
57         {
58                 .name   = "directory",
59                 .type   = FIO_OPT_STR_STORE,
60                 .off1   = td_var_offset(directory),
61                 .cb     = str_directory_cb,
62                 .help   = "Directory to store files in",
63         },
64         {
65                 .name   = "filename",
66                 .type   = FIO_OPT_STR_STORE,
67                 .off1   = td_var_offset(filename),
68                 .cb     = str_filename_cb,
69                 .help   = "File(s) to use for the workload",
70         },
71         {
72                 .name   = "opendir",
73                 .type   = FIO_OPT_STR_STORE,
74                 .off1   = td_var_offset(opendir),
75                 .cb     = str_opendir_cb,
76                 .help   = "Recursively add files from this directory and down",
77         },
78         {
79                 .name   = "rw",
80                 .type   = FIO_OPT_STR,
81                 .off1   = td_var_offset(td_ddir),
82                 .help   = "IO direction",
83                 .def    = "read",
84                 .posval = {
85                           { .ival = "read", .oval = TD_DDIR_READ },
86                           { .ival = "write", .oval = TD_DDIR_WRITE },
87                           { .ival = "randread", .oval = TD_DDIR_RANDREAD },
88                           { .ival = "randwrite", .oval = TD_DDIR_RANDWRITE },
89                           { .ival = "rw", .oval = TD_DDIR_RW },
90                           { .ival = "randrw", .oval = TD_DDIR_RANDRW },
91                           },
92         },
93         {
94                 .name   = "ioengine",
95                 .type   = FIO_OPT_STR_STORE,
96                 .off1   = td_var_offset(ioengine),
97                 .help   = "IO engine to use",
98                 .def    = "sync",
99                 .posval = {
100                           { .ival = "sync", },
101 #ifdef FIO_HAVE_LIBAIO
102                           { .ival = "libaio", },
103 #endif
104 #ifdef FIO_HAVE_POSIXAIO
105                           { .ival = "posixaio", },
106 #endif
107                           { .ival = "mmap", },
108 #ifdef FIO_HAVE_SPLICE
109                           { .ival = "splice", },
110 #endif
111 #ifdef FIO_HAVE_SGIO
112                           { .ival = "sg", },
113 #endif
114                           { .ival = "null", }, { .ival = "net", },
115 #ifdef FIO_HAVE_SYSLET
116                           { .ival = "syslet-rw", },
117 #endif
118                           { .ival = "cpuio", },
119                           { .ival = "external", },
120                           },
121         },
122         {
123                 .name   = "iodepth",
124                 .type   = FIO_OPT_INT,
125                 .off1   = td_var_offset(iodepth),
126                 .help   = "Amount of IO buffers to keep in flight",
127                 .def    = "1",
128         },
129         {
130                 .name   = "iodepth_batch",
131                 .type   = FIO_OPT_INT,
132                 .off1   = td_var_offset(iodepth_batch),
133                 .help   = "Number of IO to submit in one go",
134         },
135         {
136                 .name   = "iodepth_low",
137                 .type   = FIO_OPT_INT,
138                 .off1   = td_var_offset(iodepth_low),
139                 .help   = "Low water mark for queuing depth",
140         },
141         {
142                 .name   = "size",
143                 .type   = FIO_OPT_STR_VAL,
144                 .off1   = td_var_offset(total_file_size),
145                 .help   = "Size of device or file",
146         },
147         {
148                 .name   = "bs",
149                 .type   = FIO_OPT_STR_VAL_INT,
150                 .off1   = td_var_offset(bs[DDIR_READ]),
151                 .off2   = td_var_offset(bs[DDIR_WRITE]),
152                 .help   = "Block size unit",
153                 .def    = "4k",
154         },
155         {
156                 .name   = "bsrange",
157                 .type   = FIO_OPT_RANGE,
158                 .off1   = td_var_offset(min_bs[DDIR_READ]),
159                 .off2   = td_var_offset(max_bs[DDIR_READ]),
160                 .off3   = td_var_offset(min_bs[DDIR_WRITE]),
161                 .off4   = td_var_offset(max_bs[DDIR_WRITE]),
162                 .help   = "Set block size range (in more detail than bs)",
163         },
164         {
165                 .name   = "bs_unaligned",
166                 .type   = FIO_OPT_STR_SET,
167                 .off1   = td_var_offset(bs_unaligned),
168                 .help   = "Don't sector align IO buffer sizes",
169         },
170         {
171                 .name   = "offset",
172                 .type   = FIO_OPT_STR_VAL,
173                 .off1   = td_var_offset(start_offset),
174                 .help   = "Start IO from this offset",
175                 .def    = "0",
176         },
177         {
178                 .name   = "randrepeat",
179                 .type   = FIO_OPT_BOOL,
180                 .off1   = td_var_offset(rand_repeatable),
181                 .help   = "Use repeatable random IO pattern",
182                 .def    = "1",
183         },
184         {
185                 .name   = "norandommap",
186                 .type   = FIO_OPT_STR_SET,
187                 .off1   = td_var_offset(norandommap),
188                 .help   = "Accept potential duplicate random blocks",
189         },
190         {
191                 .name   = "nrfiles",
192                 .type   = FIO_OPT_INT,
193                 .off1   = td_var_offset(nr_files),
194                 .help   = "Split job workload between this number of files",
195                 .def    = "1",
196         },
197         {
198                 .name   = "openfiles",
199                 .type   = FIO_OPT_INT,
200                 .off1   = td_var_offset(open_files),
201                 .help   = "Number of files to keep open at the same time",
202         },
203         {
204                 .name   = "file_service_type",
205                 .type   = FIO_OPT_STR,
206                 .cb     = str_fst_cb,
207                 .off1   = td_var_offset(file_service_type),
208                 .help   = "How to select which file to service next",
209                 .def    = "roundrobin",
210                 .posval = {
211                           { .ival = "random", .oval = FIO_FSERVICE_RANDOM },
212                           { .ival = "roundrobin", .oval = FIO_FSERVICE_RR },
213                           },
214         },
215         {
216                 .name   = "fsync",
217                 .type   = FIO_OPT_INT,
218                 .off1   = td_var_offset(fsync_blocks),
219                 .help   = "Issue fsync for writes every given number of blocks",
220                 .def    = "0",
221         },
222         {
223                 .name   = "direct",
224                 .type   = FIO_OPT_BOOL,
225                 .off1   = td_var_offset(odirect),
226                 .help   = "Use O_DIRECT IO (negates buffered)",
227                 .def    = "0",
228         },
229         {
230                 .name   = "buffered",
231                 .type   = FIO_OPT_BOOL,
232                 .off1   = td_var_offset(odirect),
233                 .neg    = 1,
234                 .help   = "Use buffered IO (negates direct)",
235                 .def    = "1",
236         },
237         {
238                 .name   = "overwrite",
239                 .type   = FIO_OPT_BOOL,
240                 .off1   = td_var_offset(overwrite),
241                 .help   = "When writing, set whether to overwrite current data",
242                 .def    = "0",
243         },
244         {
245                 .name   = "loops",
246                 .type   = FIO_OPT_INT,
247                 .off1   = td_var_offset(loops),
248                 .help   = "Number of times to run the job",
249                 .def    = "1",
250         },
251         {
252                 .name   = "numjobs",
253                 .type   = FIO_OPT_INT,
254                 .off1   = td_var_offset(numjobs),
255                 .help   = "Duplicate this job this many times",
256                 .def    = "1",
257         },
258         {
259                 .name   = "startdelay",
260                 .type   = FIO_OPT_INT,
261                 .off1   = td_var_offset(start_delay),
262                 .help   = "Only start job when this period has passed",
263                 .def    = "0",
264         },
265         {
266                 .name   = "runtime",
267                 .alias  = "timeout",
268                 .type   = FIO_OPT_STR_VAL_TIME,
269                 .off1   = td_var_offset(timeout),
270                 .help   = "Stop workload when this amount of time has passed",
271                 .def    = "0",
272         },
273         {
274                 .name   = "mem",
275                 .type   = FIO_OPT_STR,
276                 .cb     = str_mem_cb,
277                 .off1   = td_var_offset(mem_type),
278                 .help   = "Backing type for IO buffers",
279                 .def    = "malloc",
280                 .posval = {
281                           { .ival = "malloc", .oval = MEM_MALLOC },
282                           { .ival = "shm", .oval = MEM_SHM },
283 #ifdef FIO_HAVE_HUGETLB
284                           { .ival = "shmhuge", .oval = MEM_SHMHUGE },
285 #endif
286                           { .ival = "mmap", .oval = MEM_MMAP },
287 #ifdef FIO_HAVE_HUGETLB
288                           { .ival = "mmaphuge", .oval = MEM_MMAPHUGE },
289 #endif
290                           },
291         },
292         {
293                 .name   = "verify",
294                 .type   = FIO_OPT_STR,
295                 .off1   = td_var_offset(verify),
296                 .help   = "Verify sum function",
297                 .def    = "0",
298                 .posval = {
299                           { .ival = "0", .oval = VERIFY_NONE },
300                           { .ival = "crc32", .oval = VERIFY_CRC32 },
301                           { .ival = "md5", .oval = VERIFY_MD5 },
302                           },
303         },
304         {
305                 .name   = "write_iolog",
306                 .type   = FIO_OPT_STR_STORE,
307                 .off1   = td_var_offset(write_iolog_file),
308                 .help   = "Store IO pattern to file",
309         },
310         {
311                 .name   = "read_iolog",
312                 .type   = FIO_OPT_STR_STORE,
313                 .off1   = td_var_offset(read_iolog_file),
314                 .help   = "Playback IO pattern from file",
315         },
316         {
317                 .name   = "exec_prerun",
318                 .type   = FIO_OPT_STR_STORE,
319                 .off1   = td_var_offset(exec_prerun),
320                 .help   = "Execute this file prior to running job",
321         },
322         {
323                 .name   = "exec_postrun",
324                 .type   = FIO_OPT_STR_STORE,
325                 .off1   = td_var_offset(exec_postrun),
326                 .help   = "Execute this file after running job",
327         },
328 #ifdef FIO_HAVE_IOSCHED_SWITCH
329         {
330                 .name   = "ioscheduler",
331                 .type   = FIO_OPT_STR_STORE,
332                 .off1   = td_var_offset(ioscheduler),
333                 .help   = "Use this IO scheduler on the backing device",
334         },
335 #endif
336         {
337                 .name   = "zonesize",
338                 .type   = FIO_OPT_STR_VAL,
339                 .off1   = td_var_offset(zone_size),
340                 .help   = "Give size of an IO zone",
341                 .def    = "0",
342         },
343         {
344                 .name   = "zoneskip",
345                 .type   = FIO_OPT_STR_VAL,
346                 .off1   = td_var_offset(zone_skip),
347                 .help   = "Space between IO zones",
348                 .def    = "0",
349         },
350         {
351                 .name   = "lockmem",
352                 .type   = FIO_OPT_STR_VAL,
353                 .cb     = str_lockmem_cb,
354                 .help   = "Lock down this amount of memory",
355                 .def    = "0",
356         },
357         {
358                 .name   = "rwmixcycle",
359                 .type   = FIO_OPT_INT,
360                 .off1   = td_var_offset(rwmixcycle),
361                 .help   = "Cycle period for mixed read/write workloads (msec)",
362                 .def    = "500",
363         },
364         {
365                 .name   = "rwmixread",
366                 .type   = FIO_OPT_INT,
367                 .off1   = td_var_offset(rwmixread),
368                 .maxval = 100,
369                 .help   = "Percentage of mixed workload that is reads",
370                 .def    = "50",
371         },
372         {
373                 .name   = "rwmixwrite",
374                 .type   = FIO_OPT_INT,
375                 .off1   = td_var_offset(rwmixwrite),
376                 .maxval = 100,
377                 .help   = "Percentage of mixed workload that is writes",
378                 .def    = "50",
379         },
380         {
381                 .name   = "nice",
382                 .type   = FIO_OPT_INT,
383                 .off1   = td_var_offset(nice),
384                 .help   = "Set job CPU nice value",
385                 .minval = -19,
386                 .maxval = 20,
387                 .def    = "0",
388         },
389 #ifdef FIO_HAVE_IOPRIO
390         {
391                 .name   = "prio",
392                 .type   = FIO_OPT_INT,
393                 .cb     = str_prio_cb,
394                 .help   = "Set job IO priority value",
395                 .minval = 0,
396                 .maxval = 7,
397         },
398         {
399                 .name   = "prioclass",
400                 .type   = FIO_OPT_INT,
401                 .cb     = str_prioclass_cb,
402                 .help   = "Set job IO priority class",
403                 .minval = 0,
404                 .maxval = 3,
405         },
406 #endif
407         {
408                 .name   = "thinktime",
409                 .type   = FIO_OPT_INT,
410                 .off1   = td_var_offset(thinktime),
411                 .help   = "Idle time between IO buffers (usec)",
412                 .def    = "0",
413         },
414         {
415                 .name   = "thinktime_spin",
416                 .type   = FIO_OPT_INT,
417                 .off1   = td_var_offset(thinktime_spin),
418                 .help   = "Start think time by spinning this amount (usec)",
419                 .def    = "0",
420         },
421         {
422                 .name   = "thinktime_blocks",
423                 .type   = FIO_OPT_INT,
424                 .off1   = td_var_offset(thinktime_blocks),
425                 .help   = "IO buffer period between 'thinktime'",
426                 .def    = "1",
427         },
428         {
429                 .name   = "rate",
430                 .type   = FIO_OPT_INT,
431                 .off1   = td_var_offset(rate),
432                 .help   = "Set bandwidth rate",
433         },
434         {
435                 .name   = "ratemin",
436                 .type   = FIO_OPT_INT,
437                 .off1   = td_var_offset(ratemin),
438                 .help   = "The bottom limit accepted",
439         },
440         {
441                 .name   = "ratecycle",
442                 .type   = FIO_OPT_INT,
443                 .off1   = td_var_offset(ratecycle),
444                 .help   = "Window average for rate limits (msec)",
445                 .def    = "1000",
446         },
447         {
448                 .name   = "invalidate",
449                 .type   = FIO_OPT_BOOL,
450                 .off1   = td_var_offset(invalidate_cache),
451                 .help   = "Invalidate buffer/page cache prior to running job",
452                 .def    = "1",
453         },
454         {
455                 .name   = "sync",
456                 .type   = FIO_OPT_BOOL,
457                 .off1   = td_var_offset(sync_io),
458                 .help   = "Use O_SYNC for buffered writes",
459                 .def    = "0",
460         },
461         {
462                 .name   = "bwavgtime",
463                 .type   = FIO_OPT_INT,
464                 .off1   = td_var_offset(bw_avg_time),
465                 .help   = "Time window over which to calculate bandwidth (msec)",
466                 .def    = "500",
467         },
468         {
469                 .name   = "create_serialize",
470                 .type   = FIO_OPT_BOOL,
471                 .off1   = td_var_offset(create_serialize),
472                 .help   = "Serialize creating of job files",
473                 .def    = "1",
474         },
475         {
476                 .name   = "create_fsync",
477                 .type   = FIO_OPT_BOOL,
478                 .off1   = td_var_offset(create_fsync),
479                 .help   = "Fsync file after creation",
480                 .def    = "1",
481         },
482         {
483                 .name   = "cpuload",
484                 .type   = FIO_OPT_INT,
485                 .off1   = td_var_offset(cpuload),
486                 .help   = "Use this percentage of CPU",
487         },
488         {
489                 .name   = "cpuchunks",
490                 .type   = FIO_OPT_INT,
491                 .off1   = td_var_offset(cpucycle),
492                 .help   = "Length of the CPU burn cycles (usecs)",
493                 .def    = "50000",
494         },
495 #ifdef FIO_HAVE_CPU_AFFINITY
496         {
497                 .name   = "cpumask",
498                 .type   = FIO_OPT_INT,
499                 .cb     = str_cpumask_cb,
500                 .help   = "CPU affinity mask",
501         },
502 #endif
503         {
504                 .name   = "end_fsync",
505                 .type   = FIO_OPT_BOOL,
506                 .off1   = td_var_offset(end_fsync),
507                 .help   = "Include fsync at the end of job",
508                 .def    = "0",
509         },
510         {
511                 .name   = "fsync_on_close",
512                 .type   = FIO_OPT_BOOL,
513                 .off1   = td_var_offset(fsync_on_close),
514                 .help   = "fsync files on close",
515                 .def    = "0",
516         },
517         {
518                 .name   = "unlink",
519                 .type   = FIO_OPT_BOOL,
520                 .off1   = td_var_offset(unlink),
521                 .help   = "Unlink created files after job has completed",
522                 .def    = "0",
523         },
524         {
525                 .name   = "exitall",
526                 .type   = FIO_OPT_STR_SET,
527                 .cb     = str_exitall_cb,
528                 .help   = "Terminate all jobs when one exits",
529         },
530         {
531                 .name   = "stonewall",
532                 .type   = FIO_OPT_STR_SET,
533                 .off1   = td_var_offset(stonewall),
534                 .help   = "Insert a hard barrier between this job and previous",
535         },
536         {
537                 .name   = "thread",
538                 .type   = FIO_OPT_STR_SET,
539                 .off1   = td_var_offset(use_thread),
540                 .help   = "Use threads instead of forks",
541         },
542         {
543                 .name   = "write_bw_log",
544                 .type   = FIO_OPT_STR_SET,
545                 .off1   = td_var_offset(write_bw_log),
546                 .help   = "Write log of bandwidth during run",
547         },
548         {
549                 .name   = "write_lat_log",
550                 .type   = FIO_OPT_STR_SET,
551                 .off1   = td_var_offset(write_lat_log),
552                 .help   = "Write log of latency during run",
553         },
554         {
555                 .name   = "hugepage-size",
556                 .type   = FIO_OPT_STR_VAL,
557                 .off1   = td_var_offset(hugepage_size),
558                 .help   = "When using hugepages, specify size of each page",
559                 .def    = __stringify(FIO_HUGE_PAGE),
560         },
561         {
562                 .name   = "group_reporting",
563                 .type   = FIO_OPT_STR_SET,
564                 .off1   = td_var_offset(group_reporting),
565                 .help   = "Do reporting on a per-group basis",
566         },
567         {
568                 .name = NULL,
569         },
570 };
571
572 #define FIO_JOB_OPTS    (sizeof(options) / sizeof(struct fio_option))
573 #define FIO_CMD_OPTS    (16)
574 #define FIO_GETOPT_JOB  (0x89988998)
575
576 /*
577  * Command line options. These will contain the above, plus a few
578  * extra that only pertain to fio itself and not jobs.
579  */
580 static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
581         {
582                 .name           = "output",
583                 .has_arg        = required_argument,
584                 .val            = 'o',
585         },
586         {
587                 .name           = "timeout",
588                 .has_arg        = required_argument,
589                 .val            = 't',
590         },
591         {
592                 .name           = "latency-log",
593                 .has_arg        = required_argument,
594                 .val            = 'l',
595         },
596         {
597                 .name           = "bandwidth-log",
598                 .has_arg        = required_argument,
599                 .val            = 'b',
600         },
601         {
602                 .name           = "minimal",
603                 .has_arg        = optional_argument,
604                 .val            = 'm',
605         },
606         {
607                 .name           = "version",
608                 .has_arg        = no_argument,
609                 .val            = 'v',
610         },
611         {
612                 .name           = "help",
613                 .has_arg        = no_argument,
614                 .val            = 'h',
615         },
616         {
617                 .name           = "cmdhelp",
618                 .has_arg        = optional_argument,
619                 .val            = 'c',
620         },
621         {
622                 .name           = NULL,
623         },
624 };
625
626 static int def_timeout = 0;
627
628 static char fio_version_string[] = "fio 1.14";
629
630 static char **ini_file;
631 static int max_jobs = MAX_JOBS;
632
633 struct thread_data def_thread;
634 struct thread_data *threads = NULL;
635
636 int exitall_on_terminate = 0;
637 int terse_output = 0;
638 unsigned long long mlock_size = 0;
639 FILE *f_out = NULL;
640 FILE *f_err = NULL;
641
642 static int write_lat_log = 0;
643 int write_bw_log = 0;
644
645 static int prev_group_jobs;
646
647 FILE *get_f_out()
648 {
649         return f_out;
650 }
651
652 FILE *get_f_err()
653 {
654         return f_err;
655 }
656
657 /*
658  * Return a free job structure.
659  */
660 static struct thread_data *get_new_job(int global, struct thread_data *parent)
661 {
662         struct thread_data *td;
663
664         if (global)
665                 return &def_thread;
666         if (thread_number >= max_jobs)
667                 return NULL;
668
669         td = &threads[thread_number++];
670         *td = *parent;
671
672         td->thread_number = thread_number;
673         return td;
674 }
675
676 static void put_job(struct thread_data *td)
677 {
678         if (td == &def_thread)
679                 return;
680
681         if (td->error)
682                 fprintf(f_out, "fio: %s\n", td->verror);
683
684         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
685         thread_number--;
686 }
687
688 /*
689  * Lazy way of fixing up options that depend on each other. We could also
690  * define option callback handlers, but this is easier.
691  */
692 static void fixup_options(struct thread_data *td)
693 {
694         if (!td->rwmixread && td->rwmixwrite)
695                 td->rwmixread = 100 - td->rwmixwrite;
696
697         if (td->write_iolog_file && td->read_iolog_file) {
698                 log_err("fio: read iolog overrides write_iolog\n");
699                 free(td->write_iolog_file);
700                 td->write_iolog_file = NULL;
701         }
702
703         if (td->io_ops->flags & FIO_SYNCIO)
704                 td->iodepth = 1;
705         else {
706                 if (!td->iodepth)
707                         td->iodepth = td->open_files;
708         }
709
710         /*
711          * only really works for sequential io for now, and with 1 file
712          */
713         if (td->zone_size && td_random(td) && td->open_files == 1)
714                 td->zone_size = 0;
715
716         /*
717          * Reads can do overwrites, we always need to pre-create the file
718          */
719         if (td_read(td) || td_rw(td))
720                 td->overwrite = 1;
721
722         if (!td->min_bs[DDIR_READ])
723                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
724         if (!td->max_bs[DDIR_READ])
725                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
726         if (!td->min_bs[DDIR_WRITE])
727                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
728         if (!td->max_bs[DDIR_WRITE])
729                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
730
731         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
732
733         if (td_read(td) && !td_rw(td))
734                 td->verify = 0;
735
736         if (td->norandommap && td->verify != VERIFY_NONE) {
737                 log_err("fio: norandommap given, verify disabled\n");
738                 td->verify = VERIFY_NONE;
739         }
740         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
741                 log_err("fio: bs_unaligned may not work with raw io\n");
742
743         /*
744          * thinktime_spin must be less than thinktime
745          */
746         if (td->thinktime_spin > td->thinktime)
747                 td->thinktime_spin = td->thinktime;
748
749         /*
750          * The low water mark cannot be bigger than the iodepth
751          */
752         if (td->iodepth_low > td->iodepth || !td->iodepth_low) {
753                 /*
754                  * syslet work around - if the workload is sequential,
755                  * we want to let the queue drain all the way down to
756                  * avoid seeking between async threads
757                  */
758                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
759                         td->iodepth_low = 1;
760                 else
761                         td->iodepth_low = td->iodepth;
762         }
763
764         /*
765          * If batch number isn't set, default to the same as iodepth
766          */
767         if (td->iodepth_batch > td->iodepth || !td->iodepth_batch)
768                 td->iodepth_batch = td->iodepth;
769
770         if (td->nr_files > td->files_index)
771                 td->nr_files = td->files_index;
772
773         if (td->open_files > td->nr_files || !td->open_files)
774                 td->open_files = td->nr_files;
775 }
776
777 /*
778  * This function leaks the buffer
779  */
780 static char *to_kmg(unsigned int val)
781 {
782         char *buf = malloc(32);
783         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
784         char *p = post;
785
786         do {
787                 if (val & 1023)
788                         break;
789
790                 val >>= 10;
791                 p++;
792         } while (*p);
793
794         snprintf(buf, 31, "%u%c", val, *p);
795         return buf;
796 }
797
798 /* External engines are specified by "external:name.o") */
799 static const char *get_engine_name(const char *str)
800 {
801         char *p = strstr(str, ":");
802
803         if (!p)
804                 return str;
805
806         p++;
807         strip_blank_front(&p);
808         strip_blank_end(p);
809         return p;
810 }
811
812 /*
813  * Adds a job to the list of things todo. Sanitizes the various options
814  * to make sure we don't have conflicts, and initializes various
815  * members of td.
816  */
817 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
818 {
819         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
820                                    "randread", "randwrite", "randrw" };
821         unsigned int i;
822         struct fio_file *f;
823         const char *engine;
824         char fname[PATH_MAX];
825         int numjobs;
826
827         /*
828          * the def_thread is just for options, it's not a real job
829          */
830         if (td == &def_thread)
831                 return 0;
832
833         engine = get_engine_name(td->ioengine);
834         td->io_ops = load_ioengine(td, engine);
835         if (!td->io_ops) {
836                 log_err("fio: failed to load engine %s\n", engine);
837                 return 1;
838         }
839
840         if (td->use_thread)
841                 nr_thread++;
842         else
843                 nr_process++;
844
845         if (td->odirect)
846                 td->io_ops->flags |= FIO_RAWIO;
847
848         if (!td->filename && !td->files_index) {
849                 td->filename = strdup(jobname);
850
851                 if (td->nr_files == 1)
852                         add_file(td, td->filename);
853                 else {
854                         for (i = 0; i < td->nr_files; i++) {
855                                 sprintf(fname, "%s.%d.%d", td->filename, td->thread_number, i);
856                                 add_file(td, fname);
857                         }
858                 }
859         }
860
861         fixup_options(td);
862
863         for_each_file(td, f, i) {
864                 if (td->directory && f->filetype == FIO_TYPE_FILE) {
865                         sprintf(fname, "%s/%s", td->directory, f->file_name);
866                         f->file_name = strdup(fname);
867                 }
868         }
869                 
870         td->mutex = fio_sem_init(0);
871
872         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
873         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
874         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
875
876         if ((td->stonewall || td->numjobs > 1) && prev_group_jobs) {
877                 prev_group_jobs = 0;
878                 groupid++;
879         }
880
881         td->groupid = groupid;
882         prev_group_jobs++;
883
884         if (setup_rate(td))
885                 goto err;
886
887         if (td->write_lat_log) {
888                 setup_log(&td->ts.slat_log);
889                 setup_log(&td->ts.clat_log);
890         }
891         if (td->write_bw_log)
892                 setup_log(&td->ts.bw_log);
893
894         if (!td->name)
895                 td->name = strdup(jobname);
896
897         if (!terse_output) {
898                 if (!job_add_num) {
899                         if (!strcmp(td->io_ops->name, "cpuio"))
900                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
901                         else {
902                                 char *c1, *c2, *c3, *c4;
903
904                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
905                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
906                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
907                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
908
909                                 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);
910
911                                 free(c1);
912                                 free(c2);
913                                 free(c3);
914                                 free(c4);
915                         }
916                 } else if (job_add_num == 1)
917                         fprintf(f_out, "...\n");
918         }
919
920         /*
921          * recurse add identical jobs, clear numjobs and stonewall options
922          * as they don't apply to sub-jobs
923          */
924         numjobs = td->numjobs;
925         while (--numjobs) {
926                 struct thread_data *td_new = get_new_job(0, td);
927
928                 if (!td_new)
929                         goto err;
930
931                 td_new->numjobs = 1;
932                 td_new->stonewall = 0;
933                 job_add_num = numjobs - 1;
934
935                 if (add_job(td_new, jobname, job_add_num))
936                         goto err;
937         }
938
939         if (td->numjobs > 1) {
940                 groupid++;
941                 prev_group_jobs = 0;
942         }
943
944         return 0;
945 err:
946         put_job(td);
947         return -1;
948 }
949
950 /*
951  * Initialize the various random states we need (random io, block size ranges,
952  * read/write mix, etc).
953  */
954 int init_random_state(struct thread_data *td)
955 {
956         unsigned long seeds[5];
957         int fd, num_maps, blocks;
958         struct fio_file *f;
959         unsigned int i;
960
961         if (td->io_ops->flags & FIO_DISKLESSIO)
962                 return 0;
963
964         fd = open("/dev/urandom", O_RDONLY);
965         if (fd == -1) {
966                 td_verror(td, errno, "open");
967                 return 1;
968         }
969
970         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
971                 td_verror(td, EIO, "read");
972                 close(fd);
973                 return 1;
974         }
975
976         close(fd);
977
978         os_random_seed(seeds[0], &td->bsrange_state);
979         os_random_seed(seeds[1], &td->verify_state);
980         os_random_seed(seeds[2], &td->rwmix_state);
981
982         if (td->file_service_type == FIO_FSERVICE_RANDOM)
983                 os_random_seed(seeds[3], &td->next_file_state);
984
985         if (!td_random(td))
986                 return 0;
987
988         if (td->rand_repeatable)
989                 seeds[4] = FIO_RANDSEED * td->thread_number;
990
991         if (!td->norandommap) {
992                 for_each_file(td, f, i) {
993                         blocks = (f->real_file_size + td->rw_min_bs - 1) / td->rw_min_bs;
994                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
995                         f->file_map = malloc(num_maps * sizeof(long));
996                         if (!f->file_map) {
997                                 log_err("fio: failed allocating random map. If running a large number of jobs, try the 'norandommap' option\n");
998                                 return 1;
999                         }
1000                         f->num_maps = num_maps;
1001                         memset(f->file_map, 0, num_maps * sizeof(long));
1002                 }
1003         }
1004
1005         os_random_seed(seeds[4], &td->random_state);
1006         return 0;
1007 }
1008
1009 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
1010 {
1011 #ifdef FIO_HAVE_CPU_AFFINITY
1012         unsigned int i;
1013
1014         CPU_ZERO(&cpumask);
1015
1016         for (i = 0; i < sizeof(int) * 8; i++) {
1017                 if ((1 << i) & cpu)
1018                         CPU_SET(i, &cpumask);
1019         }
1020 #endif
1021 }
1022
1023 static int is_empty_or_comment(char *line)
1024 {
1025         unsigned int i;
1026
1027         for (i = 0; i < strlen(line); i++) {
1028                 if (line[i] == ';')
1029                         return 1;
1030                 if (line[i] == '#')
1031                         return 1;
1032                 if (!isspace(line[i]) && !iscntrl(line[i]))
1033                         return 0;
1034         }
1035
1036         return 1;
1037 }
1038
1039 /*
1040  * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
1041  */
1042 static char *get_opt_postfix(const char *str)
1043 {
1044         char *p = strstr(str, ":");
1045
1046         if (!p)
1047                 return NULL;
1048
1049         p++;
1050         strip_blank_front(&p);
1051         strip_blank_end(p);
1052         return strdup(p);
1053 }
1054
1055 static int str_mem_cb(void *data, const char *mem)
1056 {
1057         struct thread_data *td = data;
1058
1059         if (td->mem_type == MEM_MMAPHUGE || td->mem_type == MEM_MMAP) {
1060                 td->mmapfile = get_opt_postfix(mem);
1061                 if (td->mem_type == MEM_MMAPHUGE && !td->mmapfile) {
1062                         log_err("fio: mmaphuge:/path/to/file\n");
1063                         return 1;
1064                 }
1065         }
1066
1067         return 0;
1068 }
1069
1070 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
1071 {
1072         mlock_size = *val;
1073         return 0;
1074 }
1075
1076 #ifdef FIO_HAVE_IOPRIO
1077 static int str_prioclass_cb(void *data, unsigned int *val)
1078 {
1079         struct thread_data *td = data;
1080
1081         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
1082         return 0;
1083 }
1084
1085 static int str_prio_cb(void *data, unsigned int *val)
1086 {
1087         struct thread_data *td = data;
1088
1089         td->ioprio |= *val;
1090         return 0;
1091 }
1092 #endif
1093
1094 static int str_exitall_cb(void)
1095 {
1096         exitall_on_terminate = 1;
1097         return 0;
1098 }
1099
1100 static int str_cpumask_cb(void *data, unsigned int *val)
1101 {
1102         struct thread_data *td = data;
1103
1104         fill_cpu_mask(td->cpumask, *val);
1105         return 0;
1106 }
1107
1108 static int str_fst_cb(void *data, const char *str)
1109 {
1110         struct thread_data *td = data;
1111         char *nr = get_opt_postfix(str);
1112
1113         td->file_service_nr = 1;
1114         if (nr)
1115                 td->file_service_nr = atoi(nr);
1116
1117         return 0;
1118 }
1119
1120 static int str_filename_cb(void *data, const char *input)
1121 {
1122         struct thread_data *td = data;
1123         char *fname, *str, *p;
1124
1125         p = str = strdup(input);
1126
1127         strip_blank_front(&str);
1128         strip_blank_end(str);
1129
1130         if (!td->files_index)
1131                 td->nr_files = 0;
1132
1133         while ((fname = strsep(&str, ":")) != NULL) {
1134                 if (!strlen(fname))
1135                         break;
1136                 add_file(td, fname);
1137                 td->nr_files++;
1138         }
1139
1140         free(p);
1141         return 0;
1142 }
1143
1144 static int str_directory_cb(void *data, const char fio_unused *str)
1145 {
1146         struct thread_data *td = data;
1147         struct stat sb;
1148
1149         if (lstat(td->directory, &sb) < 0) {
1150                 log_err("fio: %s is not a directory\n", td->directory);
1151                 td_verror(td, errno, "lstat");
1152                 return 1;
1153         }
1154         if (!S_ISDIR(sb.st_mode)) {
1155                 log_err("fio: %s is not a directory\n", td->directory);
1156                 return 1;
1157         }
1158
1159         return 0;
1160 }
1161
1162 static int str_opendir_cb(void *data, const char fio_unused *str)
1163 {
1164         struct thread_data *td = data;
1165
1166         if (!td->files_index)
1167                 td->nr_files = 0;
1168
1169         return add_dir_files(td, td->opendir);
1170 }
1171
1172 /*
1173  * This is our [ini] type file parser.
1174  */
1175 static int parse_jobs_ini(char *file, int stonewall_flag)
1176 {
1177         unsigned int global;
1178         struct thread_data *td;
1179         char *string, *name;
1180         fpos_t off;
1181         FILE *f;
1182         char *p;
1183         int ret = 0, stonewall;
1184
1185         f = fopen(file, "r");
1186         if (!f) {
1187                 perror("fopen job file");
1188                 return 1;
1189         }
1190
1191         string = malloc(4096);
1192         name = malloc(256);
1193         memset(name, 0, 256);
1194
1195         stonewall = stonewall_flag;
1196         do {
1197                 p = fgets(string, 4095, f);
1198                 if (!p)
1199                         break;
1200                 if (is_empty_or_comment(p))
1201                         continue;
1202                 if (sscanf(p, "[%255s]", name) != 1)
1203                         continue;
1204
1205                 global = !strncmp(name, "global", 6);
1206
1207                 name[strlen(name) - 1] = '\0';
1208
1209                 td = get_new_job(global, &def_thread);
1210                 if (!td) {
1211                         ret = 1;
1212                         break;
1213                 }
1214
1215                 /*
1216                  * Seperate multiple job files by a stonewall
1217                  */
1218                 if (!global && stonewall) {
1219                         td->stonewall = stonewall;
1220                         stonewall = 0;
1221                 }
1222
1223                 fgetpos(f, &off);
1224                 while ((p = fgets(string, 4096, f)) != NULL) {
1225                         if (is_empty_or_comment(p))
1226                                 continue;
1227
1228                         strip_blank_front(&p);
1229
1230                         if (p[0] == '[')
1231                                 break;
1232
1233                         strip_blank_end(p);
1234
1235                         fgetpos(f, &off);
1236
1237                         /*
1238                          * Don't break here, continue parsing options so we
1239                          * dump all the bad ones. Makes trial/error fixups
1240                          * easier on the user.
1241                          */
1242                         ret |= parse_option(p, options, td);
1243                 }
1244
1245                 if (!ret) {
1246                         fsetpos(f, &off);
1247                         ret = add_job(td, name, 0);
1248                 } else {
1249                         log_err("fio: job %s dropped\n", name);
1250                         put_job(td);
1251                 }
1252         } while (!ret);
1253
1254         free(string);
1255         free(name);
1256         fclose(f);
1257         return ret;
1258 }
1259
1260 static int fill_def_thread(void)
1261 {
1262         memset(&def_thread, 0, sizeof(def_thread));
1263
1264         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1265                 perror("sched_getaffinity");
1266                 return 1;
1267         }
1268
1269         /*
1270          * fill default options
1271          */
1272         fill_default_options(&def_thread, options);
1273
1274         def_thread.timeout = def_timeout;
1275         def_thread.write_bw_log = write_bw_log;
1276         def_thread.write_lat_log = write_lat_log;
1277
1278 #ifdef FIO_HAVE_DISK_UTIL
1279         def_thread.do_disk_util = 1;
1280 #endif
1281
1282         return 0;
1283 }
1284
1285 static void usage(void)
1286 {
1287         printf("%s\n", fio_version_string);
1288         printf("\t--output\tWrite output to file\n");
1289         printf("\t--timeout\tRuntime in seconds\n");
1290         printf("\t--latency-log\tGenerate per-job latency logs\n");
1291         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1292         printf("\t--minimal\tMinimal (terse) output\n");
1293         printf("\t--version\tPrint version info and exit\n");
1294         printf("\t--help\t\tPrint this page\n");
1295         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1296 }
1297
1298 static int parse_cmd_line(int argc, char *argv[])
1299 {
1300         struct thread_data *td = NULL;
1301         int c, ini_idx = 0, lidx, ret, dont_add_job = 0;
1302
1303         while ((c = getopt_long_only(argc, argv, "", long_options, &lidx)) != -1) {
1304                 switch (c) {
1305                 case 't':
1306                         def_timeout = atoi(optarg);
1307                         break;
1308                 case 'l':
1309                         write_lat_log = 1;
1310                         break;
1311                 case 'w':
1312                         write_bw_log = 1;
1313                         break;
1314                 case 'o':
1315                         f_out = fopen(optarg, "w+");
1316                         if (!f_out) {
1317                                 perror("fopen output");
1318                                 exit(1);
1319                         }
1320                         f_err = f_out;
1321                         break;
1322                 case 'm':
1323                         terse_output = 1;
1324                         break;
1325                 case 'h':
1326                         usage();
1327                         exit(0);
1328                 case 'c':
1329                         ret = show_cmd_help(options, optarg);
1330                         exit(ret);
1331                 case 'v':
1332                         printf("%s\n", fio_version_string);
1333                         exit(0);
1334                 case FIO_GETOPT_JOB: {
1335                         const char *opt = long_options[lidx].name;
1336                         char *val = optarg;
1337
1338                         if (!strncmp(opt, "name", 4) && td) {
1339                                 ret = add_job(td, td->name ?: "fio", 0);
1340                                 if (ret) {
1341                                         put_job(td);
1342                                         return 0;
1343                                 }
1344                                 td = NULL;
1345                         }
1346                         if (!td) {
1347                                 int global = !strncmp(val, "global", 6);
1348
1349                                 td = get_new_job(global, &def_thread);
1350                                 if (!td)
1351                                         return 0;
1352                         }
1353
1354                         ret = parse_cmd_option(opt, val, options, td);
1355                         if (ret)
1356                                 dont_add_job = 1;
1357                         break;
1358                 }
1359                 default:
1360                         break;
1361                 }
1362         }
1363
1364         if (td) {
1365                 if (dont_add_job)
1366                         put_job(td);
1367                 else {
1368                         ret = add_job(td, td->name ?: "fio", 0);
1369                         if (ret)
1370                                 put_job(td);
1371                 }
1372         }
1373
1374         while (optind < argc) {
1375                 ini_idx++;
1376                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1377                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1378                 optind++;
1379         }
1380
1381         return ini_idx;
1382 }
1383
1384 static void free_shm(void)
1385 {
1386         struct shmid_ds sbuf;
1387
1388         if (threads) {
1389                 shmdt((void *) threads);
1390                 threads = NULL;
1391                 shmctl(shm_id, IPC_RMID, &sbuf);
1392         }
1393 }
1394
1395 /*
1396  * The thread area is shared between the main process and the job
1397  * threads/processes. So setup a shared memory segment that will hold
1398  * all the job info.
1399  */
1400 static int setup_thread_area(void)
1401 {
1402         /*
1403          * 1024 is too much on some machines, scale max_jobs if
1404          * we get a failure that looks like too large a shm segment
1405          */
1406         do {
1407                 size_t size = max_jobs * sizeof(struct thread_data);
1408
1409                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1410                 if (shm_id != -1)
1411                         break;
1412                 if (errno != EINVAL) {
1413                         perror("shmget");
1414                         break;
1415                 }
1416
1417                 max_jobs >>= 1;
1418         } while (max_jobs);
1419
1420         if (shm_id == -1)
1421                 return 1;
1422
1423         threads = shmat(shm_id, NULL, 0);
1424         if (threads == (void *) -1) {
1425                 perror("shmat");
1426                 return 1;
1427         }
1428
1429         atexit(free_shm);
1430         return 0;
1431 }
1432
1433 /*
1434  * Copy the fio options into the long options map, so we mirror
1435  * job and cmd line options.
1436  */
1437 static void dupe_job_options(void)
1438 {
1439         struct fio_option *o;
1440         unsigned int i;
1441
1442         i = 0;
1443         while (long_options[i].name)
1444                 i++;
1445
1446         o = &options[0];
1447         while (o->name) {
1448                 long_options[i].name = o->name;
1449                 long_options[i].val = FIO_GETOPT_JOB;
1450                 if (o->type == FIO_OPT_STR_SET)
1451                         long_options[i].has_arg = no_argument;
1452                 else
1453                         long_options[i].has_arg = required_argument;
1454
1455                 i++;
1456                 o++;
1457                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1458         }
1459 }
1460
1461 int parse_options(int argc, char *argv[])
1462 {
1463         int job_files, i;
1464
1465         f_out = stdout;
1466         f_err = stderr;
1467
1468         options_init(options);
1469
1470         dupe_job_options();
1471
1472         if (setup_thread_area())
1473                 return 1;
1474         if (fill_def_thread())
1475                 return 1;
1476
1477         job_files = parse_cmd_line(argc, argv);
1478
1479         for (i = 0; i < job_files; i++) {
1480                 if (fill_def_thread())
1481                         return 1;
1482                 if (parse_jobs_ini(ini_file[i], i))
1483                         return 1;
1484                 free(ini_file[i]);
1485         }
1486
1487         free(ini_file);
1488
1489         if (!thread_number) {
1490                 log_err("No jobs defined(s)\n");
1491                 return 1;
1492         }
1493
1494         return 0;
1495 }