33a1c281c696607dfea6c3124bec80d8372141cb
[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_batch",
117                 .type   = FIO_OPT_INT,
118                 .off1   = td_var_offset(iodepth_batch),
119                 .help   = "Number of IO to submit in one go",
120         },
121         {
122                 .name   = "iodepth_low",
123                 .type   = FIO_OPT_INT,
124                 .off1   = td_var_offset(iodepth_low),
125                 .help   = "Low water mark for queuing depth",
126         },
127         {
128                 .name   = "size",
129                 .type   = FIO_OPT_STR_VAL,
130                 .off1   = td_var_offset(total_file_size),
131                 .help   = "Size of device or file",
132         },
133         {
134                 .name   = "bs",
135                 .type   = FIO_OPT_STR_VAL_INT,
136                 .off1   = td_var_offset(bs[DDIR_READ]),
137                 .off2   = td_var_offset(bs[DDIR_WRITE]),
138                 .help   = "Block size unit",
139                 .def    = "4k",
140         },
141         {
142                 .name   = "bsrange",
143                 .type   = FIO_OPT_RANGE,
144                 .off1   = td_var_offset(min_bs[DDIR_READ]),
145                 .off2   = td_var_offset(max_bs[DDIR_READ]),
146                 .off3   = td_var_offset(min_bs[DDIR_WRITE]),
147                 .off4   = td_var_offset(max_bs[DDIR_WRITE]),
148                 .help   = "Set block size range (in more detail than bs)",
149         },
150         {
151                 .name   = "bs_unaligned",
152                 .type   = FIO_OPT_STR_SET,
153                 .off1   = td_var_offset(bs_unaligned),
154                 .help   = "Don't sector align IO buffer sizes",
155         },
156         {
157                 .name   = "offset",
158                 .type   = FIO_OPT_STR_VAL,
159                 .off1   = td_var_offset(start_offset),
160                 .help   = "Start IO from this offset",
161                 .def    = "0",
162         },
163         {
164                 .name   = "randrepeat",
165                 .type   = FIO_OPT_BOOL,
166                 .off1   = td_var_offset(rand_repeatable),
167                 .help   = "Use repeatable random IO pattern",
168                 .def    = "1",
169         },
170         {
171                 .name   = "norandommap",
172                 .type   = FIO_OPT_STR_SET,
173                 .off1   = td_var_offset(norandommap),
174                 .help   = "Accept potential duplicate random blocks",
175         },
176         {
177                 .name   = "nrfiles",
178                 .type   = FIO_OPT_INT,
179                 .off1   = td_var_offset(nr_files),
180                 .help   = "Split job workload between this number of files",
181                 .def    = "1",
182         },
183         {
184                 .name   = "file_service_type",
185                 .type   = FIO_OPT_STR,
186                 .off1   = td_var_offset(file_service_type),
187                 .help   = "How to select which file to service next",
188                 .def    = "roundrobin",
189                 .posval = {
190                           { .ival = "random", .oval = FIO_FSERVICE_RANDOM },
191                           { .ival = "roundrobin", .oval = FIO_FSERVICE_RR },
192                           },
193         },
194         {
195                 .name   = "fsync",
196                 .type   = FIO_OPT_INT,
197                 .off1   = td_var_offset(fsync_blocks),
198                 .help   = "Issue fsync for writes every given number of blocks",
199                 .def    = "0",
200         },
201         {
202                 .name   = "direct",
203                 .type   = FIO_OPT_BOOL,
204                 .off1   = td_var_offset(odirect),
205                 .help   = "Use O_DIRECT IO (negates buffered)",
206                 .def    = "0",
207         },
208         {
209                 .name   = "buffered",
210                 .type   = FIO_OPT_BOOL,
211                 .off1   = td_var_offset(odirect),
212                 .neg    = 1,
213                 .help   = "Use buffered IO (negates direct)",
214                 .def    = "1",
215         },
216         {
217                 .name   = "overwrite",
218                 .type   = FIO_OPT_BOOL,
219                 .off1   = td_var_offset(overwrite),
220                 .help   = "When writing, set whether to overwrite current data",
221                 .def    = "0",
222         },
223         {
224                 .name   = "loops",
225                 .type   = FIO_OPT_INT,
226                 .off1   = td_var_offset(loops),
227                 .help   = "Number of times to run the job",
228                 .def    = "1",
229         },
230         {
231                 .name   = "numjobs",
232                 .type   = FIO_OPT_INT,
233                 .off1   = td_var_offset(numjobs),
234                 .help   = "Duplicate this job this many times",
235                 .def    = "1",
236         },
237         {
238                 .name   = "startdelay",
239                 .type   = FIO_OPT_INT,
240                 .off1   = td_var_offset(start_delay),
241                 .help   = "Only start job when this period has passed",
242                 .def    = "0",
243         },
244         {
245                 .name   = "runtime",
246                 .alias  = "timeout",
247                 .type   = FIO_OPT_STR_VAL_TIME,
248                 .off1   = td_var_offset(timeout),
249                 .help   = "Stop workload when this amount of time has passed",
250                 .def    = "0",
251         },
252         {
253                 .name   = "mem",
254                 .type   = FIO_OPT_STR,
255                 .cb     = str_mem_cb,
256                 .off1   = td_var_offset(mem_type),
257                 .help   = "Backing type for IO buffers",
258                 .def    = "malloc",
259                 .posval = {
260                           { .ival = "malloc", .oval = MEM_MALLOC },
261                           { .ival = "shm", .oval = MEM_SHM },
262 #ifdef FIO_HAVE_HUGETLB
263                           { .ival = "shmhuge", .oval = MEM_SHMHUGE },
264 #endif
265                           { .ival = "mmap", .oval = MEM_MMAP },
266 #ifdef FIO_HAVE_HUGETLB
267                           { .ival = "mmaphuge", .oval = MEM_MMAPHUGE },
268 #endif
269                           },
270         },
271         {
272                 .name   = "verify",
273                 .type   = FIO_OPT_STR,
274                 .off1   = td_var_offset(verify),
275                 .help   = "Verify sum function",
276                 .def    = "0",
277                 .posval = {
278                           { .ival = "0", .oval = VERIFY_NONE },
279                           { .ival = "crc32", .oval = VERIFY_CRC32 },
280                           { .ival = "md5", .oval = VERIFY_MD5 },
281                           },
282         },
283         {
284                 .name   = "write_iolog",
285                 .type   = FIO_OPT_STR_STORE,
286                 .off1   = td_var_offset(write_iolog_file),
287                 .help   = "Store IO pattern to file",
288         },
289         {
290                 .name   = "read_iolog",
291                 .type   = FIO_OPT_STR_STORE,
292                 .off1   = td_var_offset(read_iolog_file),
293                 .help   = "Playback IO pattern from file",
294         },
295         {
296                 .name   = "exec_prerun",
297                 .type   = FIO_OPT_STR_STORE,
298                 .off1   = td_var_offset(exec_prerun),
299                 .help   = "Execute this file prior to running job",
300         },
301         {
302                 .name   = "exec_postrun",
303                 .type   = FIO_OPT_STR_STORE,
304                 .off1   = td_var_offset(exec_postrun),
305                 .help   = "Execute this file after running job",
306         },
307 #ifdef FIO_HAVE_IOSCHED_SWITCH
308         {
309                 .name   = "ioscheduler",
310                 .type   = FIO_OPT_STR_STORE,
311                 .off1   = td_var_offset(ioscheduler),
312                 .help   = "Use this IO scheduler on the backing device",
313         },
314 #endif
315         {
316                 .name   = "zonesize",
317                 .type   = FIO_OPT_STR_VAL,
318                 .off1   = td_var_offset(zone_size),
319                 .help   = "Give size of an IO zone",
320                 .def    = "0",
321         },
322         {
323                 .name   = "zoneskip",
324                 .type   = FIO_OPT_STR_VAL,
325                 .off1   = td_var_offset(zone_skip),
326                 .help   = "Space between IO zones",
327                 .def    = "0",
328         },
329         {
330                 .name   = "lockmem",
331                 .type   = FIO_OPT_STR_VAL,
332                 .cb     = str_lockmem_cb,
333                 .help   = "Lock down this amount of memory",
334                 .def    = "0",
335         },
336         {
337                 .name   = "rwmixcycle",
338                 .type   = FIO_OPT_INT,
339                 .off1   = td_var_offset(rwmixcycle),
340                 .help   = "Cycle period for mixed read/write workloads (msec)",
341                 .def    = "500",
342         },
343         {
344                 .name   = "rwmixread",
345                 .type   = FIO_OPT_INT,
346                 .off1   = td_var_offset(rwmixread),
347                 .maxval = 100,
348                 .help   = "Percentage of mixed workload that is reads",
349                 .def    = "50",
350         },
351         {
352                 .name   = "rwmixwrite",
353                 .type   = FIO_OPT_INT,
354                 .off1   = td_var_offset(rwmixwrite),
355                 .maxval = 100,
356                 .help   = "Percentage of mixed workload that is writes",
357                 .def    = "50",
358         },
359         {
360                 .name   = "nice",
361                 .type   = FIO_OPT_INT,
362                 .off1   = td_var_offset(nice),
363                 .help   = "Set job CPU nice value",
364                 .minval = -19,
365                 .maxval = 20,
366                 .def    = "0",
367         },
368 #ifdef FIO_HAVE_IOPRIO
369         {
370                 .name   = "prio",
371                 .type   = FIO_OPT_INT,
372                 .cb     = str_prio_cb,
373                 .help   = "Set job IO priority value",
374                 .minval = 0,
375                 .maxval = 7,
376         },
377         {
378                 .name   = "prioclass",
379                 .type   = FIO_OPT_INT,
380                 .cb     = str_prioclass_cb,
381                 .help   = "Set job IO priority class",
382                 .minval = 0,
383                 .maxval = 3,
384         },
385 #endif
386         {
387                 .name   = "thinktime",
388                 .type   = FIO_OPT_INT,
389                 .off1   = td_var_offset(thinktime),
390                 .help   = "Idle time between IO buffers (usec)",
391                 .def    = "0",
392         },
393         {
394                 .name   = "thinktime_spin",
395                 .type   = FIO_OPT_INT,
396                 .off1   = td_var_offset(thinktime_spin),
397                 .help   = "Start thinktime by spinning this amount (usec)",
398                 .def    = "0",
399         },
400         {
401                 .name   = "thinktime_blocks",
402                 .type   = FIO_OPT_INT,
403                 .off1   = td_var_offset(thinktime_blocks),
404                 .help   = "IO buffer period between 'thinktime'",
405                 .def    = "1",
406         },
407         {
408                 .name   = "rate",
409                 .type   = FIO_OPT_INT,
410                 .off1   = td_var_offset(rate),
411                 .help   = "Set bandwidth rate",
412         },
413         {
414                 .name   = "ratemin",
415                 .type   = FIO_OPT_INT,
416                 .off1   = td_var_offset(ratemin),
417                 .help   = "The bottom limit accepted",
418         },
419         {
420                 .name   = "ratecycle",
421                 .type   = FIO_OPT_INT,
422                 .off1   = td_var_offset(ratecycle),
423                 .help   = "Window average for rate limits (msec)",
424                 .def    = "1000",
425         },
426         {
427                 .name   = "invalidate",
428                 .type   = FIO_OPT_BOOL,
429                 .off1   = td_var_offset(invalidate_cache),
430                 .help   = "Invalidate buffer/page cache prior to running job",
431                 .def    = "1",
432         },
433         {
434                 .name   = "sync",
435                 .type   = FIO_OPT_BOOL,
436                 .off1   = td_var_offset(sync_io),
437                 .help   = "Use O_SYNC for buffered writes",
438                 .def    = "0",
439         },
440         {
441                 .name   = "bwavgtime",
442                 .type   = FIO_OPT_INT,
443                 .off1   = td_var_offset(bw_avg_time),
444                 .help   = "Time window over which to calculate bandwidth (msec)",
445                 .def    = "500",
446         },
447         {
448                 .name   = "create_serialize",
449                 .type   = FIO_OPT_BOOL,
450                 .off1   = td_var_offset(create_serialize),
451                 .help   = "Serialize creating of job files",
452                 .def    = "1",
453         },
454         {
455                 .name   = "create_fsync",
456                 .type   = FIO_OPT_BOOL,
457                 .off1   = td_var_offset(create_fsync),
458                 .help   = "Fsync file after creation",
459                 .def    = "1",
460         },
461         {
462                 .name   = "cpuload",
463                 .type   = FIO_OPT_INT,
464                 .off1   = td_var_offset(cpuload),
465                 .help   = "Use this percentage of CPU",
466         },
467         {
468                 .name   = "cpuchunks",
469                 .type   = FIO_OPT_INT,
470                 .off1   = td_var_offset(cpucycle),
471                 .help   = "Length of the CPU burn cycles",
472         },
473 #ifdef FIO_HAVE_CPU_AFFINITY
474         {
475                 .name   = "cpumask",
476                 .type   = FIO_OPT_INT,
477                 .cb     = str_cpumask_cb,
478                 .help   = "CPU affinity mask",
479         },
480 #endif
481         {
482                 .name   = "end_fsync",
483                 .type   = FIO_OPT_BOOL,
484                 .off1   = td_var_offset(end_fsync),
485                 .help   = "Include fsync at the end of job",
486                 .def    = "0",
487         },
488         {
489                 .name   = "unlink",
490                 .type   = FIO_OPT_BOOL,
491                 .off1   = td_var_offset(unlink),
492                 .help   = "Unlink created files after job has completed",
493                 .def    = "0",
494         },
495         {
496                 .name   = "exitall",
497                 .type   = FIO_OPT_STR_SET,
498                 .cb     = str_exitall_cb,
499                 .help   = "Terminate all jobs when one exits",
500         },
501         {
502                 .name   = "stonewall",
503                 .type   = FIO_OPT_STR_SET,
504                 .off1   = td_var_offset(stonewall),
505                 .help   = "Insert a hard barrier between this job and previous",
506         },
507         {
508                 .name   = "thread",
509                 .type   = FIO_OPT_STR_SET,
510                 .off1   = td_var_offset(use_thread),
511                 .help   = "Use threads instead of forks",
512         },
513         {
514                 .name   = "write_bw_log",
515                 .type   = FIO_OPT_STR_SET,
516                 .off1   = td_var_offset(write_bw_log),
517                 .help   = "Write log of bandwidth during run",
518         },
519         {
520                 .name   = "write_lat_log",
521                 .type   = FIO_OPT_STR_SET,
522                 .off1   = td_var_offset(write_lat_log),
523                 .help   = "Write log of latency during run",
524         },
525         {
526                 .name   = "hugepage-size",
527                 .type   = FIO_OPT_STR_VAL,
528                 .off1   = td_var_offset(hugepage_size),
529                 .help   = "When using hugepages, specify size of each page",
530                 .def    = __stringify(FIO_HUGE_PAGE),
531         },
532         {
533                 .name = NULL,
534         },
535 };
536
537 #define FIO_JOB_OPTS    (sizeof(options) / sizeof(struct fio_option))
538 #define FIO_CMD_OPTS    (16)
539 #define FIO_GETOPT_JOB  (0x89988998)
540
541 /*
542  * Command line options. These will contain the above, plus a few
543  * extra that only pertain to fio itself and not jobs.
544  */
545 static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
546         {
547                 .name           = "output",
548                 .has_arg        = required_argument,
549                 .val            = 'o',
550         },
551         {
552                 .name           = "timeout",
553                 .has_arg        = required_argument,
554                 .val            = 't',
555         },
556         {
557                 .name           = "latency-log",
558                 .has_arg        = required_argument,
559                 .val            = 'l',
560         },
561         {
562                 .name           = "bandwidth-log",
563                 .has_arg        = required_argument,
564                 .val            = 'b',
565         },
566         {
567                 .name           = "minimal",
568                 .has_arg        = optional_argument,
569                 .val            = 'm',
570         },
571         {
572                 .name           = "version",
573                 .has_arg        = no_argument,
574                 .val            = 'v',
575         },
576         {
577                 .name           = "help",
578                 .has_arg        = no_argument,
579                 .val            = 'h',
580         },
581         {
582                 .name           = "cmdhelp",
583                 .has_arg        = required_argument,
584                 .val            = 'c',
585         },
586         {
587                 .name           = NULL,
588         },
589 };
590
591 static int def_timeout = 0;
592
593 static char fio_version_string[] = "fio 1.11";
594
595 static char **ini_file;
596 static int max_jobs = MAX_JOBS;
597
598 struct thread_data def_thread;
599 struct thread_data *threads = NULL;
600
601 int exitall_on_terminate = 0;
602 int terse_output = 0;
603 unsigned long long mlock_size = 0;
604 FILE *f_out = NULL;
605 FILE *f_err = NULL;
606
607 static int write_lat_log = 0;
608 int write_bw_log = 0;
609
610 /*
611  * Return a free job structure.
612  */
613 static struct thread_data *get_new_job(int global, struct thread_data *parent)
614 {
615         struct thread_data *td;
616
617         if (global)
618                 return &def_thread;
619         if (thread_number >= max_jobs)
620                 return NULL;
621
622         td = &threads[thread_number++];
623         *td = *parent;
624
625         td->thread_number = thread_number;
626         return td;
627 }
628
629 static void put_job(struct thread_data *td)
630 {
631         if (td == &def_thread)
632                 return;
633
634         if (td->error)
635                 fprintf(f_out, "fio: %s\n", td->verror);
636
637         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
638         thread_number--;
639 }
640
641 /*
642  * Lazy way of fixing up options that depend on each other. We could also
643  * define option callback handlers, but this is easier.
644  */
645 static void fixup_options(struct thread_data *td)
646 {
647         if (!td->rwmixread && td->rwmixwrite)
648                 td->rwmixread = 100 - td->rwmixwrite;
649
650         if (td->write_iolog_file && td->read_iolog_file) {
651                 log_err("fio: read iolog overrides write_iolog\n");
652                 free(td->write_iolog_file);
653                 td->write_iolog_file = NULL;
654         }
655
656         if (td->io_ops->flags & FIO_SYNCIO)
657                 td->iodepth = 1;
658         else {
659                 if (!td->iodepth)
660                         td->iodepth = td->nr_files;
661         }
662
663         /*
664          * only really works for sequential io for now, and with 1 file
665          */
666         if (td->zone_size && td_random(td) && td->nr_files == 1)
667                 td->zone_size = 0;
668
669         /*
670          * Reads can do overwrites, we always need to pre-create the file
671          */
672         if (td_read(td) || td_rw(td))
673                 td->overwrite = 1;
674
675         if (!td->min_bs[DDIR_READ])
676                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
677         if (!td->max_bs[DDIR_READ])
678                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
679         if (!td->min_bs[DDIR_WRITE])
680                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
681         if (!td->max_bs[DDIR_WRITE])
682                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
683
684         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
685
686         if (td_read(td) && !td_rw(td))
687                 td->verify = 0;
688
689         if (td->norandommap && td->verify != VERIFY_NONE) {
690                 log_err("fio: norandommap given, verify disabled\n");
691                 td->verify = VERIFY_NONE;
692         }
693         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
694                 log_err("fio: bs_unaligned may not work with raw io\n");
695
696         /*
697          * O_DIRECT and char doesn't mix, clear that flag if necessary.
698          */
699         if (td->filetype == FIO_TYPE_CHAR && td->odirect)
700                 td->odirect = 0;
701
702         /*
703          * thinktime_spin must be less than thinktime
704          */
705         if (td->thinktime_spin > td->thinktime)
706                 td->thinktime_spin = td->thinktime;
707
708         /*
709          * The low water mark cannot be bigger than the iodepth
710          */
711         if (td->iodepth_low > td->iodepth || !td->iodepth_low) {
712                 /*
713                  * syslet work around - if the workload is sequential,
714                  * we want to let the queue drain all the way down to
715                  * avoid seeking between async threads
716                  */
717                 if (!strcmp(td->io_ops->name, "syslet-rw") && !td_random(td))
718                         td->iodepth_low = 1;
719                 else
720                         td->iodepth_low = td->iodepth;
721         }
722
723         /*
724          * If batch number isn't set, default to the same as iodepth
725          */
726         if (td->iodepth_batch > td->iodepth || !td->iodepth_batch)
727                 td->iodepth_batch = td->iodepth;
728 }
729
730 /*
731  * This function leaks the buffer
732  */
733 static char *to_kmg(unsigned int val)
734 {
735         char *buf = malloc(32);
736         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
737         char *p = post;
738
739         do {
740                 if (val & 1023)
741                         break;
742
743                 val >>= 10;
744                 p++;
745         } while (*p);
746
747         snprintf(buf, 31, "%u%c", val, *p);
748         return buf;
749 }
750
751 /*
752  * Adds a job to the list of things todo. Sanitizes the various options
753  * to make sure we don't have conflicts, and initializes various
754  * members of td.
755  */
756 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
757 {
758         const char *ddir_str[] = { NULL, "read", "write", "rw", NULL,
759                                    "randread", "randwrite", "randrw" };
760         struct stat sb;
761         int numjobs, i;
762         struct fio_file *f;
763
764         /*
765          * the def_thread is just for options, it's not a real job
766          */
767         if (td == &def_thread)
768                 return 0;
769
770         assert(td->io_ops);
771
772         if (td->odirect)
773                 td->io_ops->flags |= FIO_RAWIO;
774
775         td->filetype = FIO_TYPE_FILE;
776         if (td->filename && !lstat(td->filename, &sb)) {
777                 if (S_ISBLK(sb.st_mode))
778                         td->filetype = FIO_TYPE_BD;
779                 else if (S_ISCHR(sb.st_mode))
780                         td->filetype = FIO_TYPE_CHAR;
781         }
782
783         fixup_options(td);
784
785         if (td->filename)
786                 td->nr_uniq_files = 1;
787         else
788                 td->nr_uniq_files = td->nr_files;
789
790         if (td->filetype == FIO_TYPE_FILE || td->filename) {
791                 char tmp[PATH_MAX];
792                 int len = 0;
793
794                 if (td->directory && td->directory[0] != '\0') {
795                         if (lstat(td->directory, &sb) < 0) {
796                                 log_err("fio: %s is not a directory\n", td->directory);
797                                 td_verror(td, errno, "lstat");
798                                 return 1;
799                         }
800                         if (!S_ISDIR(sb.st_mode)) {
801                                 log_err("fio: %s is not a directory\n", td->directory);
802                                 return 1;
803                         }
804                         len = sprintf(tmp, "%s/", td->directory);
805                 }
806
807                 td->files = malloc(sizeof(struct fio_file) * td->nr_files);
808
809                 for_each_file(td, f, i) {
810                         memset(f, 0, sizeof(*f));
811                         f->fd = -1;
812
813                         if (td->filename)
814                                 sprintf(tmp + len, "%s", td->filename);
815                         else
816                                 sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
817                         f->file_name = strdup(tmp);
818                 }
819         } else {
820                 td->nr_files = 1;
821                 td->files = malloc(sizeof(struct fio_file));
822                 f = &td->files[0];
823
824                 memset(f, 0, sizeof(*f));
825                 f->fd = -1;
826                 f->file_name = strdup(jobname);
827         }
828
829         for_each_file(td, f, i) {
830                 f->file_size = td->total_file_size / td->nr_files;
831                 f->file_offset = td->start_offset;
832         }
833                 
834         fio_sem_init(&td->mutex, 0);
835
836         td->ts.clat_stat[0].min_val = td->ts.clat_stat[1].min_val = ULONG_MAX;
837         td->ts.slat_stat[0].min_val = td->ts.slat_stat[1].min_val = ULONG_MAX;
838         td->ts.bw_stat[0].min_val = td->ts.bw_stat[1].min_val = ULONG_MAX;
839
840         if (td->stonewall && td->thread_number > 1)
841                 groupid++;
842
843         td->groupid = groupid;
844
845         if (setup_rate(td))
846                 goto err;
847
848         if (td->write_lat_log) {
849                 setup_log(&td->ts.slat_log);
850                 setup_log(&td->ts.clat_log);
851         }
852         if (td->write_bw_log)
853                 setup_log(&td->ts.bw_log);
854
855         if (!td->name)
856                 td->name = strdup(jobname);
857
858         if (!terse_output) {
859                 if (!job_add_num) {
860                         if (td->io_ops->flags & FIO_CPUIO)
861                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
862                         else {
863                                 char *c1, *c2, *c3, *c4;
864
865                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
866                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
867                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
868                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
869
870                                 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);
871
872                                 free(c1);
873                                 free(c2);
874                                 free(c3);
875                                 free(c4);
876                         }
877                 } else if (job_add_num == 1)
878                         fprintf(f_out, "...\n");
879         }
880
881         /*
882          * recurse add identical jobs, clear numjobs and stonewall options
883          * as they don't apply to sub-jobs
884          */
885         numjobs = td->numjobs;
886         while (--numjobs) {
887                 struct thread_data *td_new = get_new_job(0, td);
888
889                 if (!td_new)
890                         goto err;
891
892                 td_new->numjobs = 1;
893                 td_new->stonewall = 0;
894                 job_add_num = numjobs - 1;
895
896                 if (add_job(td_new, jobname, job_add_num))
897                         goto err;
898         }
899         return 0;
900 err:
901         put_job(td);
902         return -1;
903 }
904
905 /*
906  * Initialize the various random states we need (random io, block size ranges,
907  * read/write mix, etc).
908  */
909 int init_random_state(struct thread_data *td)
910 {
911         unsigned long seeds[5];
912         int fd, num_maps, blocks, i;
913         struct fio_file *f;
914
915         if (td->io_ops->flags & FIO_CPUIO)
916                 return 0;
917
918         fd = open("/dev/urandom", O_RDONLY);
919         if (fd == -1) {
920                 td_verror(td, errno, "open");
921                 return 1;
922         }
923
924         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
925                 td_verror(td, EIO, "read");
926                 close(fd);
927                 return 1;
928         }
929
930         close(fd);
931
932         os_random_seed(seeds[0], &td->bsrange_state);
933         os_random_seed(seeds[1], &td->verify_state);
934         os_random_seed(seeds[2], &td->rwmix_state);
935
936         if (td->file_service_type == FIO_FSERVICE_RANDOM)
937                 os_random_seed(seeds[3], &td->next_file_state);
938
939         if (!td_random(td))
940                 return 0;
941
942         if (td->rand_repeatable)
943                 seeds[4] = FIO_RANDSEED * td->thread_number;
944
945         if (!td->norandommap) {
946                 for_each_file(td, f, i) {
947                         blocks = (f->real_file_size + td->rw_min_bs - 1) / td->rw_min_bs;
948                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
949                         f->file_map = malloc(num_maps * sizeof(long));
950                         f->num_maps = num_maps;
951                         memset(f->file_map, 0, num_maps * sizeof(long));
952                 }
953         }
954
955         os_random_seed(seeds[4], &td->random_state);
956         return 0;
957 }
958
959 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
960 {
961 #ifdef FIO_HAVE_CPU_AFFINITY
962         unsigned int i;
963
964         CPU_ZERO(&cpumask);
965
966         for (i = 0; i < sizeof(int) * 8; i++) {
967                 if ((1 << i) & cpu)
968                         CPU_SET(i, &cpumask);
969         }
970 #endif
971 }
972
973 static int is_empty_or_comment(char *line)
974 {
975         unsigned int i;
976
977         for (i = 0; i < strlen(line); i++) {
978                 if (line[i] == ';')
979                         return 1;
980                 if (line[i] == '#')
981                         return 1;
982                 if (!isspace(line[i]) && !iscntrl(line[i]))
983                         return 0;
984         }
985
986         return 1;
987 }
988
989 /*
990  * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
991  */
992 static char *get_mmap_file(const char *str)
993 {
994         char *p = strstr(str, ":");
995
996         if (!p)
997                 return NULL;
998
999         p++;
1000         strip_blank_front(&p);
1001         strip_blank_end(p);
1002         return strdup(p);
1003 }
1004
1005 static int str_mem_cb(void *data, const char *mem)
1006 {
1007         struct thread_data *td = data;
1008
1009         if (td->mem_type == MEM_MMAPHUGE || td->mem_type == MEM_MMAP) {
1010                 td->mmapfile = get_mmap_file(mem);
1011                 if (td->mem_type == MEM_MMAPHUGE && !td->mmapfile) {
1012                         log_err("fio: mmaphuge:/path/to/file\n");
1013                         return 1;
1014                 }
1015         }
1016
1017         return 0;
1018 }
1019
1020 static int str_ioengine_cb(void *data, const char *str)
1021 {
1022         struct thread_data *td = data;
1023
1024         td->io_ops = load_ioengine(td, str);
1025         if (td->io_ops)
1026                 return 0;
1027
1028         return 1;
1029 }
1030
1031 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
1032 {
1033         mlock_size = *val;
1034         return 0;
1035 }
1036
1037 #ifdef FIO_HAVE_IOPRIO
1038 static int str_prioclass_cb(void *data, unsigned int *val)
1039 {
1040         struct thread_data *td = data;
1041
1042         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
1043         return 0;
1044 }
1045
1046 static int str_prio_cb(void *data, unsigned int *val)
1047 {
1048         struct thread_data *td = data;
1049
1050         td->ioprio |= *val;
1051         return 0;
1052 }
1053 #endif
1054
1055 static int str_exitall_cb(void)
1056 {
1057         exitall_on_terminate = 1;
1058         return 0;
1059 }
1060
1061 static int str_cpumask_cb(void *data, unsigned int *val)
1062 {
1063         struct thread_data *td = data;
1064
1065         fill_cpu_mask(td->cpumask, *val);
1066         return 0;
1067 }
1068
1069 /*
1070  * This is our [ini] type file parser.
1071  */
1072 static int parse_jobs_ini(char *file, int stonewall_flag)
1073 {
1074         unsigned int global;
1075         struct thread_data *td;
1076         char *string, *name;
1077         fpos_t off;
1078         FILE *f;
1079         char *p;
1080         int ret = 0, stonewall;
1081
1082         f = fopen(file, "r");
1083         if (!f) {
1084                 perror("fopen job file");
1085                 return 1;
1086         }
1087
1088         string = malloc(4096);
1089         name = malloc(256);
1090         memset(name, 0, 256);
1091
1092         stonewall = stonewall_flag;
1093         do {
1094                 p = fgets(string, 4095, f);
1095                 if (!p)
1096                         break;
1097                 if (is_empty_or_comment(p))
1098                         continue;
1099                 if (sscanf(p, "[%255s]", name) != 1)
1100                         continue;
1101
1102                 global = !strncmp(name, "global", 6);
1103
1104                 name[strlen(name) - 1] = '\0';
1105
1106                 td = get_new_job(global, &def_thread);
1107                 if (!td) {
1108                         ret = 1;
1109                         break;
1110                 }
1111
1112                 /*
1113                  * Seperate multiple job files by a stonewall
1114                  */
1115                 if (!global && stonewall) {
1116                         td->stonewall = stonewall;
1117                         stonewall = 0;
1118                 }
1119
1120                 fgetpos(f, &off);
1121                 while ((p = fgets(string, 4096, f)) != NULL) {
1122                         if (is_empty_or_comment(p))
1123                                 continue;
1124
1125                         strip_blank_front(&p);
1126
1127                         if (p[0] == '[')
1128                                 break;
1129
1130                         strip_blank_end(p);
1131
1132                         fgetpos(f, &off);
1133
1134                         /*
1135                          * Don't break here, continue parsing options so we
1136                          * dump all the bad ones. Makes trial/error fixups
1137                          * easier on the user.
1138                          */
1139                         ret |= parse_option(p, options, td);
1140                 }
1141
1142                 if (!ret) {
1143                         fsetpos(f, &off);
1144                         ret = add_job(td, name, 0);
1145                 } else {
1146                         log_err("fio: job %s dropped\n", name);
1147                         put_job(td);
1148                 }
1149         } while (!ret);
1150
1151         free(string);
1152         free(name);
1153         fclose(f);
1154         return ret;
1155 }
1156
1157 static int fill_def_thread(void)
1158 {
1159         memset(&def_thread, 0, sizeof(def_thread));
1160
1161         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1162                 perror("sched_getaffinity");
1163                 return 1;
1164         }
1165
1166         /*
1167          * fill default options
1168          */
1169         fill_default_options(&def_thread, options);
1170
1171         def_thread.timeout = def_timeout;
1172         def_thread.write_bw_log = write_bw_log;
1173         def_thread.write_lat_log = write_lat_log;
1174
1175 #ifdef FIO_HAVE_DISK_UTIL
1176         def_thread.do_disk_util = 1;
1177 #endif
1178
1179         return 0;
1180 }
1181
1182 static void usage(void)
1183 {
1184         printf("%s\n", fio_version_string);
1185         printf("\t--output\tWrite output to file\n");
1186         printf("\t--timeout\tRuntime in seconds\n");
1187         printf("\t--latency-log\tGenerate per-job latency logs\n");
1188         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1189         printf("\t--minimal\tMinimal (terse) output\n");
1190         printf("\t--version\tPrint version info and exit\n");
1191         printf("\t--help\t\tPrint this page\n");
1192         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1193 }
1194
1195 static int parse_cmd_line(int argc, char *argv[])
1196 {
1197         struct thread_data *td = NULL;
1198         int c, ini_idx = 0, lidx, ret, dont_add_job = 0;
1199
1200         while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1201                 switch (c) {
1202                 case 't':
1203                         def_timeout = atoi(optarg);
1204                         break;
1205                 case 'l':
1206                         write_lat_log = 1;
1207                         break;
1208                 case 'w':
1209                         write_bw_log = 1;
1210                         break;
1211                 case 'o':
1212                         f_out = fopen(optarg, "w+");
1213                         if (!f_out) {
1214                                 perror("fopen output");
1215                                 exit(1);
1216                         }
1217                         f_err = f_out;
1218                         break;
1219                 case 'm':
1220                         terse_output = 1;
1221                         break;
1222                 case 'h':
1223                         usage();
1224                         exit(0);
1225                 case 'c':
1226                         ret = show_cmd_help(options, optarg);
1227                         exit(ret);
1228                 case 'v':
1229                         printf("%s\n", fio_version_string);
1230                         exit(0);
1231                 case FIO_GETOPT_JOB: {
1232                         const char *opt = long_options[lidx].name;
1233                         char *val = optarg;
1234
1235                         if (!strncmp(opt, "name", 4) && td) {
1236                                 ret = add_job(td, td->name ?: "fio", 0);
1237                                 if (ret) {
1238                                         put_job(td);
1239                                         return 0;
1240                                 }
1241                                 td = NULL;
1242                         }
1243                         if (!td) {
1244                                 int global = !strncmp(val, "global", 6);
1245
1246                                 td = get_new_job(global, &def_thread);
1247                                 if (!td)
1248                                         return 0;
1249                         }
1250
1251                         ret = parse_cmd_option(opt, val, options, td);
1252                         if (ret)
1253                                 dont_add_job = 1;
1254                         break;
1255                 }
1256                 default:
1257                         break;
1258                 }
1259         }
1260
1261         if (td) {
1262                 if (dont_add_job)
1263                         put_job(td);
1264                 else {
1265                         ret = add_job(td, td->name ?: "fio", 0);
1266                         if (ret)
1267                                 put_job(td);
1268                 }
1269         }
1270
1271         while (optind < argc) {
1272                 ini_idx++;
1273                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1274                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1275                 optind++;
1276         }
1277
1278         return ini_idx;
1279 }
1280
1281 static void free_shm(void)
1282 {
1283         struct shmid_ds sbuf;
1284
1285         if (threads) {
1286                 shmdt((void *) threads);
1287                 threads = NULL;
1288                 shmctl(shm_id, IPC_RMID, &sbuf);
1289         }
1290 }
1291
1292 /*
1293  * The thread area is shared between the main process and the job
1294  * threads/processes. So setup a shared memory segment that will hold
1295  * all the job info.
1296  */
1297 static int setup_thread_area(void)
1298 {
1299         /*
1300          * 1024 is too much on some machines, scale max_jobs if
1301          * we get a failure that looks like too large a shm segment
1302          */
1303         do {
1304                 size_t size = max_jobs * sizeof(struct thread_data);
1305
1306                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1307                 if (shm_id != -1)
1308                         break;
1309                 if (errno != EINVAL) {
1310                         perror("shmget");
1311                         break;
1312                 }
1313
1314                 max_jobs >>= 1;
1315         } while (max_jobs);
1316
1317         if (shm_id == -1)
1318                 return 1;
1319
1320         threads = shmat(shm_id, NULL, 0);
1321         if (threads == (void *) -1) {
1322                 perror("shmat");
1323                 return 1;
1324         }
1325
1326         atexit(free_shm);
1327         return 0;
1328 }
1329
1330 /*
1331  * Copy the fio options into the long options map, so we mirror
1332  * job and cmd line options.
1333  */
1334 static void dupe_job_options(void)
1335 {
1336         struct fio_option *o;
1337         unsigned int i;
1338
1339         i = 0;
1340         while (long_options[i].name)
1341                 i++;
1342
1343         o = &options[0];
1344         while (o->name) {
1345                 long_options[i].name = o->name;
1346                 long_options[i].val = FIO_GETOPT_JOB;
1347                 if (o->type == FIO_OPT_STR_SET)
1348                         long_options[i].has_arg = no_argument;
1349                 else
1350                         long_options[i].has_arg = required_argument;
1351
1352                 i++;
1353                 o++;
1354                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1355         }
1356 }
1357
1358 int parse_options(int argc, char *argv[])
1359 {
1360         int job_files, i;
1361
1362         f_out = stdout;
1363         f_err = stderr;
1364
1365         options_init(options);
1366
1367         dupe_job_options();
1368
1369         if (setup_thread_area())
1370                 return 1;
1371         if (fill_def_thread())
1372                 return 1;
1373
1374         job_files = parse_cmd_line(argc, argv);
1375
1376         for (i = 0; i < job_files; i++) {
1377                 if (fill_def_thread())
1378                         return 1;
1379                 if (parse_jobs_ini(ini_file[i], i))
1380                         return 1;
1381                 free(ini_file[i]);
1382         }
1383
1384         free(ini_file);
1385
1386         if (!thread_number) {
1387                 log_err("No jobs defined(s)\n");
1388                 return 1;
1389         }
1390
1391         return 0;
1392 }