5d6b0a904f8348af3ccd649054b24132e29de20d
[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_rw_cb(void *, const char *);
26 static int str_ioengine_cb(void *, const char *);
27 static int str_mem_cb(void *, const char *);
28 static int str_verify_cb(void *, const char *);
29 static int str_lockmem_cb(void *, unsigned long *);
30 #ifdef FIO_HAVE_IOPRIO
31 static int str_prio_cb(void *, unsigned int *);
32 static int str_prioclass_cb(void *, unsigned int *);
33 #endif
34 static int str_exitall_cb(void);
35 static int str_cpumask_cb(void *, unsigned int *);
36
37 #define __stringify_1(x)        #x
38 #define __stringify(x)          __stringify_1(x)
39
40 /*
41  * Map of job/command line options
42  */
43 static struct fio_option options[] = {
44         {
45                 .name   = "name",
46                 .type   = FIO_OPT_STR_STORE,
47                 .off1   = td_var_offset(name),
48                 .help   = "Name of this job",
49         },
50         {
51                 .name   = "directory",
52                 .type   = FIO_OPT_STR_STORE,
53                 .off1   = td_var_offset(directory),
54                 .help   = "Directory to store files in",
55         },
56         {
57                 .name   = "filename",
58                 .type   = FIO_OPT_STR_STORE,
59                 .off1   = td_var_offset(filename),
60                 .help   = "Force the use of a specific file",
61         },
62         {
63                 .name   = "rw",
64                 .type   = FIO_OPT_STR,
65                 .cb     = str_rw_cb,
66                 .help   = "IO direction",
67                 .def    = "read",
68                 .posval = { "read", "write", "randwrite", "randread", "rw",
69                                 "randrw", },
70         },
71         {
72                 .name   = "ioengine",
73                 .type   = FIO_OPT_STR,
74                 .cb     = str_ioengine_cb,
75                 .help   = "IO engine to use",
76                 .def    = "sync",
77                 .posval = { "sync", "libaio", "posixaio", "mmap", "splice",
78                                 "sg", "null", },
79         },
80         {
81                 .name   = "iodepth",
82                 .type   = FIO_OPT_INT,
83                 .off1   = td_var_offset(iodepth),
84                 .help   = "Amount of IO buffers to keep in flight",
85                 .def    = "1",
86         },
87         {
88                 .name   = "size",
89                 .type   = FIO_OPT_STR_VAL,
90                 .off1   = td_var_offset(total_file_size),
91                 .help   = "Size of device or file",
92         },
93         {
94                 .name   = "bs",
95                 .type   = FIO_OPT_STR_VAL_INT,
96                 .off1   = td_var_offset(bs[DDIR_READ]),
97                 .off2   = td_var_offset(bs[DDIR_WRITE]),
98                 .help   = "Block size unit",
99                 .def    = "4k",
100         },
101         {
102                 .name   = "bsrange",
103                 .type   = FIO_OPT_RANGE,
104                 .off1   = td_var_offset(min_bs[DDIR_READ]),
105                 .off2   = td_var_offset(max_bs[DDIR_READ]),
106                 .off3   = td_var_offset(min_bs[DDIR_WRITE]),
107                 .off4   = td_var_offset(max_bs[DDIR_WRITE]),
108                 .help   = "Set block size range (in more detail than bs)",
109         },
110         {
111                 .name   = "bs_unaligned",
112                 .type   = FIO_OPT_STR_SET,
113                 .off1   = td_var_offset(bs_unaligned),
114                 .help   = "Don't sector align IO buffer sizes",
115         },
116         {
117                 .name   = "offset",
118                 .type   = FIO_OPT_STR_VAL,
119                 .off1   = td_var_offset(start_offset),
120                 .help   = "Start IO from this offset",
121                 .def    = "0",
122         },
123         {
124                 .name   = "randrepeat",
125                 .type   = FIO_OPT_BOOL,
126                 .off1   = td_var_offset(rand_repeatable),
127                 .help   = "Use repeatable random IO pattern",
128                 .def    = "1",
129         },
130         {
131                 .name   = "norandommap",
132                 .type   = FIO_OPT_STR_SET,
133                 .off1   = td_var_offset(norandommap),
134                 .help   = "Accept potential duplicate random blocks",
135         },
136         {
137                 .name   = "nrfiles",
138                 .type   = FIO_OPT_INT,
139                 .off1   = td_var_offset(nr_files),
140                 .help   = "Split job workload between this number of files",
141                 .def    = "1",
142         },
143         {
144                 .name   = "fsync",
145                 .type   = FIO_OPT_INT,
146                 .off1   = td_var_offset(fsync_blocks),
147                 .help   = "Issue fsync for writes every given number of blocks",
148                 .def    = "0",
149         },
150         {
151                 .name   = "direct",
152                 .type   = FIO_OPT_BOOL,
153                 .off1   = td_var_offset(odirect),
154                 .help   = "Use O_DIRECT IO (negates buffered)",
155                 .def    = "0",
156         },
157         {
158                 .name   = "buffered",
159                 .type   = FIO_OPT_BOOL,
160                 .off1   = td_var_offset(odirect),
161                 .neg    = 1,
162                 .help   = "Use buffered IO (negates direct)",
163                 .def    = "1",
164         },
165         {
166                 .name   = "overwrite",
167                 .type   = FIO_OPT_BOOL,
168                 .off1   = td_var_offset(overwrite),
169                 .help   = "When writing, set whether to overwrite current data",
170                 .def    = "0",
171         },
172         {
173                 .name   = "loops",
174                 .type   = FIO_OPT_INT,
175                 .off1   = td_var_offset(loops),
176                 .help   = "Number of times to run the job",
177                 .def    = "1",
178         },
179         {
180                 .name   = "numjobs",
181                 .type   = FIO_OPT_INT,
182                 .off1   = td_var_offset(numjobs),
183                 .help   = "Duplicate this job this many times",
184                 .def    = "1",
185         },
186         {
187                 .name   = "startdelay",
188                 .type   = FIO_OPT_INT,
189                 .off1   = td_var_offset(start_delay),
190                 .help   = "Only start job when this period has passed",
191                 .def    = "0",
192         },
193         {
194                 .name   = "runtime",
195                 .alias  = "timeout",
196                 .type   = FIO_OPT_STR_VAL_TIME,
197                 .off1   = td_var_offset(timeout),
198                 .help   = "Stop workload when this amount of time has passed",
199                 .def    = "0",
200         },
201         {
202                 .name   = "mem",
203                 .type   = FIO_OPT_STR,
204                 .cb     = str_mem_cb,
205                 .help   = "Backing type for IO buffers",
206                 .def    = "malloc",
207                 .posval =  { "malloc", "shm", "shmhuge", "mmap", "mmaphuge", },
208         },
209         {
210                 .name   = "verify",
211                 .type   = FIO_OPT_STR,
212                 .cb     = str_verify_cb,
213                 .help   = "Verify sum function",
214                 .def    = "0",
215                 .posval = { "crc32", "md5", },
216         },
217         {
218                 .name   = "write_iolog",
219                 .type   = FIO_OPT_STR_STORE,
220                 .off1   = td_var_offset(write_iolog_file),
221                 .help   = "Store IO pattern to file",
222         },
223         {
224                 .name   = "read_iolog",
225                 .type   = FIO_OPT_STR_STORE,
226                 .off1   = td_var_offset(read_iolog_file),
227                 .help   = "Playback IO pattern from file",
228         },
229         {
230                 .name   = "exec_prerun",
231                 .type   = FIO_OPT_STR_STORE,
232                 .off1   = td_var_offset(exec_prerun),
233                 .help   = "Execute this file prior to running job",
234         },
235         {
236                 .name   = "exec_postrun",
237                 .type   = FIO_OPT_STR_STORE,
238                 .off1   = td_var_offset(exec_postrun),
239                 .help   = "Execute this file after running job",
240         },
241 #ifdef FIO_HAVE_IOSCHED_SWITCH
242         {
243                 .name   = "ioscheduler",
244                 .type   = FIO_OPT_STR_STORE,
245                 .off1   = td_var_offset(ioscheduler),
246                 .help   = "Use this IO scheduler on the backing device",
247         },
248 #endif
249         {
250                 .name   = "zonesize",
251                 .type   = FIO_OPT_STR_VAL,
252                 .off1   = td_var_offset(zone_size),
253                 .help   = "Give size of an IO zone",
254                 .def    = "0",
255         },
256         {
257                 .name   = "zoneskip",
258                 .type   = FIO_OPT_STR_VAL,
259                 .off1   = td_var_offset(zone_skip),
260                 .help   = "Space between IO zones",
261                 .def    = "0",
262         },
263         {
264                 .name   = "lockmem",
265                 .type   = FIO_OPT_STR_VAL,
266                 .cb     = str_lockmem_cb,
267                 .help   = "Lock down this amount of memory",
268                 .def    = "0",
269         },
270         {
271                 .name   = "rwmixcycle",
272                 .type   = FIO_OPT_INT,
273                 .off1   = td_var_offset(rwmixcycle),
274                 .help   = "Cycle period for mixed read/write workloads (msec)",
275                 .def    = "500",
276         },
277         {
278                 .name   = "rwmixread",
279                 .type   = FIO_OPT_INT,
280                 .off1   = td_var_offset(rwmixread),
281                 .maxval = 100,
282                 .help   = "Percentage of mixed workload that is reads",
283                 .def    = "50",
284         },
285         {
286                 .name   = "rwmixwrite",
287                 .type   = FIO_OPT_INT,
288                 .off1   = td_var_offset(rwmixwrite),
289                 .maxval = 100,
290                 .help   = "Percentage of mixed workload that is writes",
291                 .def    = "50",
292         },
293         {
294                 .name   = "nice",
295                 .type   = FIO_OPT_INT,
296                 .off1   = td_var_offset(nice),
297                 .help   = "Set job CPU nice value",
298                 .minval = -19,
299                 .maxval = 20,
300                 .def    = "0",
301         },
302 #ifdef FIO_HAVE_IOPRIO
303         {
304                 .name   = "prio",
305                 .type   = FIO_OPT_INT,
306                 .cb     = str_prio_cb,
307                 .help   = "Set job IO priority value",
308                 .minval = 0,
309                 .maxval = 7,
310         },
311         {
312                 .name   = "prioclass",
313                 .type   = FIO_OPT_INT,
314                 .cb     = str_prioclass_cb,
315                 .help   = "Set job IO priority class",
316                 .minval = 0,
317                 .maxval = 3,
318         },
319 #endif
320         {
321                 .name   = "thinktime",
322                 .type   = FIO_OPT_INT,
323                 .off1   = td_var_offset(thinktime),
324                 .help   = "Idle time between IO buffers",
325                 .def    = "0",
326         },
327         {
328                 .name   = "thinktime_blocks",
329                 .type   = FIO_OPT_INT,
330                 .off1   = td_var_offset(thinktime_blocks),
331                 .help   = "IO buffer period between 'thinktime'",
332                 .def    = "1",
333         },
334         {
335                 .name   = "rate",
336                 .type   = FIO_OPT_INT,
337                 .off1   = td_var_offset(rate),
338                 .help   = "Set bandwidth rate",
339         },
340         {
341                 .name   = "ratemin",
342                 .type   = FIO_OPT_INT,
343                 .off1   = td_var_offset(ratemin),
344                 .help   = "The bottom limit accepted",
345         },
346         {
347                 .name   = "ratecycle",
348                 .type   = FIO_OPT_INT,
349                 .off1   = td_var_offset(ratecycle),
350                 .help   = "Window average for rate limits (msec)",
351                 .def    = "1000",
352         },
353         {
354                 .name   = "invalidate",
355                 .type   = FIO_OPT_BOOL,
356                 .off1   = td_var_offset(invalidate_cache),
357                 .help   = "Invalidate buffer/page cache prior to running job",
358                 .def    = "1",
359         },
360         {
361                 .name   = "sync",
362                 .type   = FIO_OPT_BOOL,
363                 .off1   = td_var_offset(sync_io),
364                 .help   = "Use O_SYNC for buffered writes",
365                 .def    = "0",
366         },
367         {
368                 .name   = "bwavgtime",
369                 .type   = FIO_OPT_INT,
370                 .off1   = td_var_offset(bw_avg_time),
371                 .help   = "Time window over which to calculate bandwidth (msec)",
372                 .def    = "500",
373         },
374         {
375                 .name   = "create_serialize",
376                 .type   = FIO_OPT_BOOL,
377                 .off1   = td_var_offset(create_serialize),
378                 .help   = "Serialize creating of job files",
379                 .def    = "1",
380         },
381         {
382                 .name   = "create_fsync",
383                 .type   = FIO_OPT_BOOL,
384                 .off1   = td_var_offset(create_fsync),
385                 .help   = "Fsync file after creation",
386                 .def    = "1",
387         },
388         {
389                 .name   = "cpuload",
390                 .type   = FIO_OPT_INT,
391                 .off1   = td_var_offset(cpuload),
392                 .help   = "Use this percentage of CPU",
393         },
394         {
395                 .name   = "cpuchunks",
396                 .type   = FIO_OPT_INT,
397                 .off1   = td_var_offset(cpucycle),
398                 .help   = "Length of the CPU burn cycles",
399         },
400 #ifdef FIO_HAVE_CPU_AFFINITY
401         {
402                 .name   = "cpumask",
403                 .type   = FIO_OPT_INT,
404                 .cb     = str_cpumask_cb,
405                 .help   = "CPU affinity mask",
406         },
407 #endif
408         {
409                 .name   = "end_fsync",
410                 .type   = FIO_OPT_BOOL,
411                 .off1   = td_var_offset(end_fsync),
412                 .help   = "Include fsync at the end of job",
413                 .def    = "0",
414         },
415         {
416                 .name   = "unlink",
417                 .type   = FIO_OPT_BOOL,
418                 .off1   = td_var_offset(unlink),
419                 .help   = "Unlink created files after job has completed",
420                 .def    = "1",
421         },
422         {
423                 .name   = "exitall",
424                 .type   = FIO_OPT_STR_SET,
425                 .cb     = str_exitall_cb,
426                 .help   = "Terminate all jobs when one exits",
427         },
428         {
429                 .name   = "stonewall",
430                 .type   = FIO_OPT_STR_SET,
431                 .off1   = td_var_offset(stonewall),
432                 .help   = "Insert a hard barrier between this job and previous",
433         },
434         {
435                 .name   = "thread",
436                 .type   = FIO_OPT_STR_SET,
437                 .off1   = td_var_offset(use_thread),
438                 .help   = "Use threads instead of forks",
439         },
440         {
441                 .name   = "write_bw_log",
442                 .type   = FIO_OPT_STR_SET,
443                 .off1   = td_var_offset(write_bw_log),
444                 .help   = "Write log of bandwidth during run",
445         },
446         {
447                 .name   = "write_lat_log",
448                 .type   = FIO_OPT_STR_SET,
449                 .off1   = td_var_offset(write_lat_log),
450                 .help   = "Write log of latency during run",
451         },
452         {
453                 .name   = "hugepage-size",
454                 .type   = FIO_OPT_STR_VAL,
455                 .off1   = td_var_offset(hugepage_size),
456                 .help   = "When using hugepages, specify size of each page",
457                 .def    = __stringify(FIO_HUGE_PAGE),
458         },
459         {
460                 .name = NULL,
461         },
462 };
463
464 #define FIO_JOB_OPTS    (sizeof(options) / sizeof(struct fio_option))
465 #define FIO_CMD_OPTS    (16)
466 #define FIO_GETOPT_JOB  (0x89988998)
467
468 /*
469  * Command line options. These will contain the above, plus a few
470  * extra that only pertain to fio itself and not jobs.
471  */
472 static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
473         {
474                 .name           = "output",
475                 .has_arg        = required_argument,
476                 .val            = 'o',
477         },
478         {
479                 .name           = "timeout",
480                 .has_arg        = required_argument,
481                 .val            = 't',
482         },
483         {
484                 .name           = "latency-log",
485                 .has_arg        = required_argument,
486                 .val            = 'l',
487         },
488         {
489                 .name           = "bandwidth-log",
490                 .has_arg        = required_argument,
491                 .val            = 'b',
492         },
493         {
494                 .name           = "minimal",
495                 .has_arg        = optional_argument,
496                 .val            = 'm',
497         },
498         {
499                 .name           = "version",
500                 .has_arg        = no_argument,
501                 .val            = 'v',
502         },
503         {
504                 .name           = "help",
505                 .has_arg        = no_argument,
506                 .val            = 'h',
507         },
508         {
509                 .name           = "cmdhelp",
510                 .has_arg        = required_argument,
511                 .val            = 'c',
512         },
513         {
514                 .name           = NULL,
515         },
516 };
517
518 static int def_timeout = 0;
519
520 static char fio_version_string[] = "fio 1.11";
521
522 static char **ini_file;
523 static int max_jobs = MAX_JOBS;
524
525 struct thread_data def_thread;
526 struct thread_data *threads = NULL;
527
528 int exitall_on_terminate = 0;
529 int terse_output = 0;
530 unsigned long long mlock_size = 0;
531 FILE *f_out = NULL;
532 FILE *f_err = NULL;
533
534 static int write_lat_log = 0;
535 static int write_bw_log = 0;
536
537 /*
538  * Return a free job structure.
539  */
540 static struct thread_data *get_new_job(int global, struct thread_data *parent)
541 {
542         struct thread_data *td;
543
544         if (global)
545                 return &def_thread;
546         if (thread_number >= max_jobs)
547                 return NULL;
548
549         td = &threads[thread_number++];
550         *td = *parent;
551
552         td->thread_number = thread_number;
553         return td;
554 }
555
556 static void put_job(struct thread_data *td)
557 {
558         if (td == &def_thread)
559                 return;
560
561         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
562         thread_number--;
563 }
564
565 /*
566  * Lazy way of fixing up options that depend on each other. We could also
567  * define option callback handlers, but this is easier.
568  */
569 static void fixup_options(struct thread_data *td)
570 {
571         if (!td->rwmixread && td->rwmixwrite)
572                 td->rwmixread = 100 - td->rwmixwrite;
573
574         if (td->write_iolog_file && td->read_iolog_file) {
575                 log_err("fio: read iolog overrides write_iolog\n");
576                 free(td->write_iolog_file);
577                 td->write_iolog_file = NULL;
578         }
579
580         if (td->io_ops->flags & FIO_SYNCIO)
581                 td->iodepth = 1;
582         else {
583                 if (!td->iodepth)
584                         td->iodepth = td->nr_files;
585         }
586
587         /*
588          * only really works for sequential io for now, and with 1 file
589          */
590         if (td->zone_size && !td->sequential && td->nr_files == 1)
591                 td->zone_size = 0;
592
593         /*
594          * Reads can do overwrites, we always need to pre-create the file
595          */
596         if (td_read(td) || td_rw(td))
597                 td->overwrite = 1;
598
599         if (!td->min_bs[DDIR_READ])
600                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
601         if (!td->max_bs[DDIR_READ])
602                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
603         if (!td->min_bs[DDIR_WRITE])
604                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
605         if (!td->max_bs[DDIR_WRITE])
606                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
607
608         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
609
610         if (td_read(td) && !td_rw(td))
611                 td->verify = 0;
612
613         if (td->norandommap && td->verify != VERIFY_NONE) {
614                 log_err("fio: norandommap given, verify disabled\n");
615                 td->verify = VERIFY_NONE;
616         }
617         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
618                 log_err("fio: bs_unaligned may not work with raw io\n");
619
620         /*
621          * O_DIRECT and char doesn't mix, clear that flag if necessary.
622          */
623         if (td->filetype == FIO_TYPE_CHAR && td->odirect)
624                 td->odirect = 0;
625 }
626
627 /*
628  * This function leaks the buffer
629  */
630 static char *to_kmg(unsigned int val)
631 {
632         char *buf = malloc(32);
633         char post[] = { 0, 'K', 'M', 'G', 'P', 0 };
634         char *p = post;
635
636         do {
637                 if (val & 1023)
638                         break;
639
640                 val >>= 10;
641                 p++;
642         } while (*p);
643
644         snprintf(buf, 31, "%u%c", val, *p);
645         return buf;
646 }
647
648 /*
649  * Adds a job to the list of things todo. Sanitizes the various options
650  * to make sure we don't have conflicts, and initializes various
651  * members of td.
652  */
653 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
654 {
655         const char *ddir_str[] = { "read", "write", "randread", "randwrite",
656                                    "rw", NULL, "randrw" };
657         struct stat sb;
658         int numjobs, ddir, i;
659         struct fio_file *f;
660
661         /*
662          * the def_thread is just for options, it's not a real job
663          */
664         if (td == &def_thread)
665                 return 0;
666
667         assert(td->io_ops);
668
669         if (td->odirect)
670                 td->io_ops->flags |= FIO_RAWIO;
671
672         td->filetype = FIO_TYPE_FILE;
673         if (!stat(jobname, &sb)) {
674                 if (S_ISBLK(sb.st_mode))
675                         td->filetype = FIO_TYPE_BD;
676                 else if (S_ISCHR(sb.st_mode))
677                         td->filetype = FIO_TYPE_CHAR;
678         }
679
680         fixup_options(td);
681
682         if (td->filename)
683                 td->nr_uniq_files = 1;
684         else
685                 td->nr_uniq_files = td->nr_files;
686
687         if (td->filetype == FIO_TYPE_FILE || td->filename) {
688                 char tmp[PATH_MAX];
689                 int len = 0;
690
691                 if (td->directory && td->directory[0] != '\0')
692                         len = sprintf(tmp, "%s/", td->directory);
693
694                 td->files = malloc(sizeof(struct fio_file) * td->nr_files);
695
696                 for_each_file(td, f, i) {
697                         memset(f, 0, sizeof(*f));
698                         f->fd = -1;
699
700                         if (td->filename)
701                                 sprintf(tmp + len, "%s", td->filename);
702                         else
703                                 sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
704                         f->file_name = strdup(tmp);
705                 }
706         } else {
707                 td->nr_files = 1;
708                 td->files = malloc(sizeof(struct fio_file));
709                 f = &td->files[0];
710
711                 memset(f, 0, sizeof(*f));
712                 f->fd = -1;
713                 f->file_name = strdup(jobname);
714         }
715
716         for_each_file(td, f, i) {
717                 f->file_size = td->total_file_size / td->nr_files;
718                 f->file_offset = td->start_offset;
719         }
720                 
721         fio_sem_init(&td->mutex, 0);
722
723         td->clat_stat[0].min_val = td->clat_stat[1].min_val = ULONG_MAX;
724         td->slat_stat[0].min_val = td->slat_stat[1].min_val = ULONG_MAX;
725         td->bw_stat[0].min_val = td->bw_stat[1].min_val = ULONG_MAX;
726
727         if (td->stonewall && td->thread_number > 1)
728                 groupid++;
729
730         td->groupid = groupid;
731
732         if (setup_rate(td))
733                 goto err;
734
735         if (td->write_lat_log) {
736                 setup_log(&td->slat_log);
737                 setup_log(&td->clat_log);
738         }
739         if (td->write_bw_log)
740                 setup_log(&td->bw_log);
741
742         if (!td->name)
743                 td->name = strdup(jobname);
744
745         ddir = td->ddir + (!td->sequential << 1) + (td->iomix << 2);
746
747         if (!terse_output) {
748                 if (!job_add_num) {
749                         if (td->io_ops->flags & FIO_CPUIO)
750                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
751                         else {
752                                 char *c1, *c2, *c3, *c4;
753
754                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
755                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
756                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
757                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
758
759                                 fprintf(f_out, "%s: (g=%d): rw=%s, odir=%u, bs=%s-%s/%s-%s, rate=%u, ioengine=%s, iodepth=%u\n", td->name, td->groupid, ddir_str[ddir], td->odirect, c1, c2, c3, c4, td->rate, td->io_ops->name, td->iodepth);
760
761                                 free(c1);
762                                 free(c2);
763                                 free(c3);
764                                 free(c4);
765                         }
766                 } else if (job_add_num == 1)
767                         fprintf(f_out, "...\n");
768         }
769
770         /*
771          * recurse add identical jobs, clear numjobs and stonewall options
772          * as they don't apply to sub-jobs
773          */
774         numjobs = td->numjobs;
775         while (--numjobs) {
776                 struct thread_data *td_new = get_new_job(0, td);
777
778                 if (!td_new)
779                         goto err;
780
781                 td_new->numjobs = 1;
782                 td_new->stonewall = 0;
783                 job_add_num = numjobs - 1;
784
785                 if (add_job(td_new, jobname, job_add_num))
786                         goto err;
787         }
788         return 0;
789 err:
790         put_job(td);
791         return -1;
792 }
793
794 /*
795  * Initialize the various random states we need (random io, block size ranges,
796  * read/write mix, etc).
797  */
798 int init_random_state(struct thread_data *td)
799 {
800         unsigned long seeds[4];
801         int fd, num_maps, blocks, i;
802         struct fio_file *f;
803
804         if (td->io_ops->flags & FIO_CPUIO)
805                 return 0;
806
807         fd = open("/dev/urandom", O_RDONLY);
808         if (fd == -1) {
809                 td_verror(td, errno);
810                 return 1;
811         }
812
813         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
814                 td_verror(td, EIO);
815                 close(fd);
816                 return 1;
817         }
818
819         close(fd);
820
821         os_random_seed(seeds[0], &td->bsrange_state);
822         os_random_seed(seeds[1], &td->verify_state);
823         os_random_seed(seeds[2], &td->rwmix_state);
824
825         if (td->sequential)
826                 return 0;
827
828         if (td->rand_repeatable)
829                 seeds[3] = FIO_RANDSEED;
830
831         if (!td->norandommap) {
832                 for_each_file(td, f, i) {
833                         blocks = (f->file_size + td->rw_min_bs - 1) / td->rw_min_bs;
834                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
835                         f->file_map = malloc(num_maps * sizeof(long));
836                         f->num_maps = num_maps;
837                         memset(f->file_map, 0, num_maps * sizeof(long));
838                 }
839         }
840
841         os_random_seed(seeds[3], &td->random_state);
842         return 0;
843 }
844
845 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
846 {
847 #ifdef FIO_HAVE_CPU_AFFINITY
848         unsigned int i;
849
850         CPU_ZERO(&cpumask);
851
852         for (i = 0; i < sizeof(int) * 8; i++) {
853                 if ((1 << i) & cpu)
854                         CPU_SET(i, &cpumask);
855         }
856 #endif
857 }
858
859 static int is_empty_or_comment(char *line)
860 {
861         unsigned int i;
862
863         for (i = 0; i < strlen(line); i++) {
864                 if (line[i] == ';')
865                         return 1;
866                 if (!isspace(line[i]) && !iscntrl(line[i]))
867                         return 0;
868         }
869
870         return 1;
871 }
872
873 static int str_rw_cb(void *data, const char *mem)
874 {
875         struct thread_data *td = data;
876
877         if (!strncmp(mem, "read", 4) || !strncmp(mem, "0", 1)) {
878                 td->ddir = DDIR_READ;
879                 td->sequential = 1;
880                 return 0;
881         } else if (!strncmp(mem, "randread", 8)) {
882                 td->ddir = DDIR_READ;
883                 td->sequential = 0;
884                 return 0;
885         } else if (!strncmp(mem, "write", 5) || !strncmp(mem, "1", 1)) {
886                 td->ddir = DDIR_WRITE;
887                 td->sequential = 1;
888                 return 0;
889         } else if (!strncmp(mem, "randwrite", 9)) {
890                 td->ddir = DDIR_WRITE;
891                 td->sequential = 0;
892                 return 0;
893         } else if (!strncmp(mem, "rw", 2)) {
894                 td->ddir = DDIR_READ;
895                 td->iomix = 1;
896                 td->sequential = 1;
897                 return 0;
898         } else if (!strncmp(mem, "randrw", 6)) {
899                 td->ddir = DDIR_READ;
900                 td->iomix = 1;
901                 td->sequential = 0;
902                 return 0;
903         }
904
905         log_err("fio: data direction: read, write, randread, randwrite, rw, randrw\n");
906         return 1;
907 }
908
909 static int str_verify_cb(void *data, const char *mem)
910 {
911         struct thread_data *td = data;
912
913         if (!strncmp(mem, "0", 1)) {
914                 td->verify = VERIFY_NONE;
915                 return 0;
916         } else if (!strncmp(mem, "md5", 3) || !strncmp(mem, "1", 1)) {
917                 td->verify = VERIFY_MD5;
918                 return 0;
919         } else if (!strncmp(mem, "crc32", 5)) {
920                 td->verify = VERIFY_CRC32;
921                 return 0;
922         }
923
924         log_err("fio: verify types: md5, crc32\n");
925         return 1;
926 }
927
928 /*
929  * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
930  */
931 static char *get_mmap_file(const char *str)
932 {
933         char *p = strstr(str, ":");
934
935         if (!p)
936                 return NULL;
937
938         p++;
939         strip_blank_front(&p);
940         strip_blank_end(p);
941         return strdup(p);
942 }
943
944 static int str_mem_cb(void *data, const char *mem)
945 {
946         struct thread_data *td = data;
947
948         if (!strncmp(mem, "malloc", 6)) {
949                 td->mem_type = MEM_MALLOC;
950                 return 0;
951         } else if (!strncmp(mem, "mmaphuge", 8)) {
952 #ifdef FIO_HAVE_HUGETLB
953                 /*
954                  * mmaphuge must be appended with the actual file
955                  */
956                 td->mmapfile = get_mmap_file(mem);
957                 if (!td->mmapfile) {
958                         log_err("fio: mmaphuge:/path/to/file\n");
959                         return 1;
960                 }
961
962                 td->mem_type = MEM_MMAPHUGE;
963                 return 0;
964 #else
965                 log_err("fio: mmaphuge not available\n");
966                 return 1;
967 #endif
968         } else if (!strncmp(mem, "mmap", 4)) {
969                 /*
970                  * Check if the user wants file backed memory. It's ok
971                  * if there's no file given, we'll just use anon mamp then.
972                  */
973                 td->mmapfile = get_mmap_file(mem);
974                 td->mem_type = MEM_MMAP;
975                 return 0;
976         } else if (!strncmp(mem, "shmhuge", 7)) {
977 #ifdef FIO_HAVE_HUGETLB
978                 td->mem_type = MEM_SHMHUGE;
979                 return 0;
980 #else
981                 log_err("fio: shmhuge not available\n");
982                 return 1;
983 #endif
984         } else if (!strncmp(mem, "shm", 3)) {
985                 td->mem_type = MEM_SHM;
986                 return 0;
987         }
988
989         log_err("fio: mem type: malloc, shm, shmhuge, mmap, mmaphuge\n");
990         return 1;
991 }
992
993 static int str_ioengine_cb(void *data, const char *str)
994 {
995         struct thread_data *td = data;
996
997         td->io_ops = load_ioengine(td, str);
998         if (td->io_ops)
999                 return 0;
1000
1001         log_err("fio: ioengine= libaio, posixaio, sync, mmap, sgio, splice, cpu, null\n");
1002         log_err("fio: or specify path to dynamic ioengine module\n");
1003         return 1;
1004 }
1005
1006 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
1007 {
1008         mlock_size = *val;
1009         return 0;
1010 }
1011
1012 #ifdef FIO_HAVE_IOPRIO
1013 static int str_prioclass_cb(void *data, unsigned int *val)
1014 {
1015         struct thread_data *td = data;
1016
1017         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
1018         return 0;
1019 }
1020
1021 static int str_prio_cb(void *data, unsigned int *val)
1022 {
1023         struct thread_data *td = data;
1024
1025         td->ioprio |= *val;
1026         return 0;
1027 }
1028 #endif
1029
1030 static int str_exitall_cb(void)
1031 {
1032         exitall_on_terminate = 1;
1033         return 0;
1034 }
1035
1036 static int str_cpumask_cb(void *data, unsigned int *val)
1037 {
1038         struct thread_data *td = data;
1039
1040         fill_cpu_mask(td->cpumask, *val);
1041         return 0;
1042 }
1043
1044 /*
1045  * This is our [ini] type file parser.
1046  */
1047 static int parse_jobs_ini(char *file, int stonewall_flag)
1048 {
1049         unsigned int global;
1050         struct thread_data *td;
1051         char *string, *name;
1052         fpos_t off;
1053         FILE *f;
1054         char *p;
1055         int ret = 0, stonewall;
1056
1057         f = fopen(file, "r");
1058         if (!f) {
1059                 perror("fopen job file");
1060                 return 1;
1061         }
1062
1063         string = malloc(4096);
1064         name = malloc(256);
1065         memset(name, 0, 256);
1066
1067         stonewall = stonewall_flag;
1068         do {
1069                 p = fgets(string, 4095, f);
1070                 if (!p)
1071                         break;
1072                 if (is_empty_or_comment(p))
1073                         continue;
1074                 if (sscanf(p, "[%255s]", name) != 1)
1075                         continue;
1076
1077                 global = !strncmp(name, "global", 6);
1078
1079                 name[strlen(name) - 1] = '\0';
1080
1081                 td = get_new_job(global, &def_thread);
1082                 if (!td) {
1083                         ret = 1;
1084                         break;
1085                 }
1086
1087                 /*
1088                  * Seperate multiple job files by a stonewall
1089                  */
1090                 if (!global && stonewall) {
1091                         td->stonewall = stonewall;
1092                         stonewall = 0;
1093                 }
1094
1095                 fgetpos(f, &off);
1096                 while ((p = fgets(string, 4096, f)) != NULL) {
1097                         if (is_empty_or_comment(p))
1098                                 continue;
1099
1100                         strip_blank_front(&p);
1101
1102                         if (p[0] == '[')
1103                                 break;
1104
1105                         strip_blank_end(p);
1106
1107                         fgetpos(f, &off);
1108
1109                         /*
1110                          * Don't break here, continue parsing options so we
1111                          * dump all the bad ones. Makes trial/error fixups
1112                          * easier on the user.
1113                          */
1114                         ret |= parse_option(p, options, td);
1115                 }
1116
1117                 if (!ret) {
1118                         fsetpos(f, &off);
1119                         ret = add_job(td, name, 0);
1120                 } else {
1121                         log_err("fio: job %s dropped\n", name);
1122                         put_job(td);
1123                 }
1124         } while (!ret);
1125
1126         free(string);
1127         free(name);
1128         fclose(f);
1129         return ret;
1130 }
1131
1132 static int fill_def_thread(void)
1133 {
1134         memset(&def_thread, 0, sizeof(def_thread));
1135
1136         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1137                 perror("sched_getaffinity");
1138                 return 1;
1139         }
1140
1141         /*
1142          * fill default options
1143          */
1144         fill_default_options(&def_thread, options);
1145
1146         def_thread.timeout = def_timeout;
1147         def_thread.write_bw_log = write_bw_log;
1148         def_thread.write_lat_log = write_lat_log;
1149
1150 #ifdef FIO_HAVE_DISK_UTIL
1151         def_thread.do_disk_util = 1;
1152 #endif
1153
1154         return 0;
1155 }
1156
1157 static void usage(void)
1158 {
1159         printf("%s\n", fio_version_string);
1160         printf("\t--output\tWrite output to file\n");
1161         printf("\t--timeout\tRuntime in seconds\n");
1162         printf("\t--latency-log\tGenerate per-job latency logs\n");
1163         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1164         printf("\t--minimal\tMinimal (terse) output\n");
1165         printf("\t--version\tPrint version info and exit\n");
1166         printf("\t--help\t\tPrint this page\n");
1167         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1168 }
1169
1170 static int parse_cmd_line(int argc, char *argv[])
1171 {
1172         struct thread_data *td = NULL;
1173         int c, ini_idx = 0, lidx, ret;
1174
1175         while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1176                 switch (c) {
1177                 case 't':
1178                         def_timeout = atoi(optarg);
1179                         break;
1180                 case 'l':
1181                         write_lat_log = 1;
1182                         break;
1183                 case 'w':
1184                         write_bw_log = 1;
1185                         break;
1186                 case 'o':
1187                         f_out = fopen(optarg, "w+");
1188                         if (!f_out) {
1189                                 perror("fopen output");
1190                                 exit(1);
1191                         }
1192                         f_err = f_out;
1193                         break;
1194                 case 'm':
1195                         terse_output = 1;
1196                         break;
1197                 case 'h':
1198                         usage();
1199                         exit(0);
1200                 case 'c':
1201                         ret = show_cmd_help(options, optarg);
1202                         exit(ret);
1203                 case 'v':
1204                         printf("%s\n", fio_version_string);
1205                         exit(0);
1206                 case FIO_GETOPT_JOB: {
1207                         const char *opt = long_options[lidx].name;
1208                         char *val = optarg;
1209
1210                         if (!strncmp(opt, "name", 4) && td) {
1211                                 ret = add_job(td, td->name ?: "fio", 0);
1212                                 if (ret) {
1213                                         put_job(td);
1214                                         return 0;
1215                                 }
1216                                 td = NULL;
1217                         }
1218                         if (!td) {
1219                                 int global = !strncmp(val, "global", 6);
1220
1221                                 td = get_new_job(global, &def_thread);
1222                                 if (!td)
1223                                         return 0;
1224                         }
1225
1226                         ret = parse_cmd_option(opt, val, options, td);
1227                         if (ret) {
1228                                 log_err("fio: job dropped\n");
1229                                 put_job(td);
1230                                 td = NULL;
1231                         }
1232                         break;
1233                 }
1234                 default:
1235                         break;
1236                 }
1237         }
1238
1239         if (td) {
1240                 ret = add_job(td, td->name ?: "fio", 0);
1241                 if (ret)
1242                         put_job(td);
1243         }
1244
1245         while (optind < argc) {
1246                 ini_idx++;
1247                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1248                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1249                 optind++;
1250         }
1251
1252         return ini_idx;
1253 }
1254
1255 static void free_shm(void)
1256 {
1257         struct shmid_ds sbuf;
1258
1259         if (threads) {
1260                 shmdt((void *) threads);
1261                 threads = NULL;
1262                 shmctl(shm_id, IPC_RMID, &sbuf);
1263         }
1264 }
1265
1266 /*
1267  * The thread area is shared between the main process and the job
1268  * threads/processes. So setup a shared memory segment that will hold
1269  * all the job info.
1270  */
1271 static int setup_thread_area(void)
1272 {
1273         /*
1274          * 1024 is too much on some machines, scale max_jobs if
1275          * we get a failure that looks like too large a shm segment
1276          */
1277         do {
1278                 size_t size = max_jobs * sizeof(struct thread_data);
1279
1280                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1281                 if (shm_id != -1)
1282                         break;
1283                 if (errno != EINVAL) {
1284                         perror("shmget");
1285                         break;
1286                 }
1287
1288                 max_jobs >>= 1;
1289         } while (max_jobs);
1290
1291         if (shm_id == -1)
1292                 return 1;
1293
1294         threads = shmat(shm_id, NULL, 0);
1295         if (threads == (void *) -1) {
1296                 perror("shmat");
1297                 return 1;
1298         }
1299
1300         atexit(free_shm);
1301         return 0;
1302 }
1303
1304 /*
1305  * Copy the fio options into the long options map, so we mirror
1306  * job and cmd line options.
1307  */
1308 static void dupe_job_options(void)
1309 {
1310         struct fio_option *o;
1311         unsigned int i;
1312
1313         i = 0;
1314         while (long_options[i].name)
1315                 i++;
1316
1317         o = &options[0];
1318         while (o->name) {
1319                 long_options[i].name = o->name;
1320                 long_options[i].val = FIO_GETOPT_JOB;
1321                 if (o->type == FIO_OPT_STR_SET)
1322                         long_options[i].has_arg = no_argument;
1323                 else
1324                         long_options[i].has_arg = required_argument;
1325
1326                 i++;
1327                 o++;
1328                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1329         }
1330 }
1331
1332 int parse_options(int argc, char *argv[])
1333 {
1334         int job_files, i;
1335
1336         f_out = stdout;
1337         f_err = stderr;
1338
1339         options_init(options);
1340
1341         dupe_job_options();
1342
1343         if (setup_thread_area())
1344                 return 1;
1345         if (fill_def_thread())
1346                 return 1;
1347
1348         job_files = parse_cmd_line(argc, argv);
1349
1350         for (i = 0; i < job_files; i++) {
1351                 if (fill_def_thread())
1352                         return 1;
1353                 if (parse_jobs_ini(ini_file[i], i))
1354                         return 1;
1355                 free(ini_file[i]);
1356         }
1357
1358         free(ini_file);
1359
1360         if (!thread_number) {
1361                 log_err("No jobs defined(s)\n");
1362                 return 1;
1363         }
1364
1365         return 0;
1366 }