Verify requeue handling
[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   = "description",
46                 .type   = FIO_OPT_STR_STORE,
47                 .off1   = td_var_offset(description),
48                 .help   = "Text job description",
49         },
50         {
51                 .name   = "name",
52                 .type   = FIO_OPT_STR_STORE,
53                 .off1   = td_var_offset(name),
54                 .help   = "Name of this job",
55         },
56         {
57                 .name   = "directory",
58                 .type   = FIO_OPT_STR_STORE,
59                 .off1   = td_var_offset(directory),
60                 .help   = "Directory to store files in",
61         },
62         {
63                 .name   = "filename",
64                 .type   = FIO_OPT_STR_STORE,
65                 .off1   = td_var_offset(filename),
66                 .help   = "Force the use of a specific file",
67         },
68         {
69                 .name   = "rw",
70                 .type   = FIO_OPT_STR,
71                 .cb     = str_rw_cb,
72                 .help   = "IO direction",
73                 .def    = "read",
74                 .posval = { "read", "write", "randwrite", "randread", "rw",
75                                 "randrw", },
76         },
77         {
78                 .name   = "ioengine",
79                 .type   = FIO_OPT_STR,
80                 .cb     = str_ioengine_cb,
81                 .help   = "IO engine to use",
82                 .def    = "sync",
83                 .posval = { "sync", "libaio", "posixaio", "mmap", "splice",
84                                 "sg", "null", "net", "syslet-rw" },
85         },
86         {
87                 .name   = "iodepth",
88                 .type   = FIO_OPT_INT,
89                 .off1   = td_var_offset(iodepth),
90                 .help   = "Amount of IO buffers to keep in flight",
91                 .def    = "1",
92         },
93         {
94                 .name   = "size",
95                 .type   = FIO_OPT_STR_VAL,
96                 .off1   = td_var_offset(total_file_size),
97                 .help   = "Size of device or file",
98         },
99         {
100                 .name   = "bs",
101                 .type   = FIO_OPT_STR_VAL_INT,
102                 .off1   = td_var_offset(bs[DDIR_READ]),
103                 .off2   = td_var_offset(bs[DDIR_WRITE]),
104                 .help   = "Block size unit",
105                 .def    = "4k",
106         },
107         {
108                 .name   = "bsrange",
109                 .type   = FIO_OPT_RANGE,
110                 .off1   = td_var_offset(min_bs[DDIR_READ]),
111                 .off2   = td_var_offset(max_bs[DDIR_READ]),
112                 .off3   = td_var_offset(min_bs[DDIR_WRITE]),
113                 .off4   = td_var_offset(max_bs[DDIR_WRITE]),
114                 .help   = "Set block size range (in more detail than bs)",
115         },
116         {
117                 .name   = "bs_unaligned",
118                 .type   = FIO_OPT_STR_SET,
119                 .off1   = td_var_offset(bs_unaligned),
120                 .help   = "Don't sector align IO buffer sizes",
121         },
122         {
123                 .name   = "offset",
124                 .type   = FIO_OPT_STR_VAL,
125                 .off1   = td_var_offset(start_offset),
126                 .help   = "Start IO from this offset",
127                 .def    = "0",
128         },
129         {
130                 .name   = "randrepeat",
131                 .type   = FIO_OPT_BOOL,
132                 .off1   = td_var_offset(rand_repeatable),
133                 .help   = "Use repeatable random IO pattern",
134                 .def    = "1",
135         },
136         {
137                 .name   = "norandommap",
138                 .type   = FIO_OPT_STR_SET,
139                 .off1   = td_var_offset(norandommap),
140                 .help   = "Accept potential duplicate random blocks",
141         },
142         {
143                 .name   = "nrfiles",
144                 .type   = FIO_OPT_INT,
145                 .off1   = td_var_offset(nr_files),
146                 .help   = "Split job workload between this number of files",
147                 .def    = "1",
148         },
149         {
150                 .name   = "fsync",
151                 .type   = FIO_OPT_INT,
152                 .off1   = td_var_offset(fsync_blocks),
153                 .help   = "Issue fsync for writes every given number of blocks",
154                 .def    = "0",
155         },
156         {
157                 .name   = "direct",
158                 .type   = FIO_OPT_BOOL,
159                 .off1   = td_var_offset(odirect),
160                 .help   = "Use O_DIRECT IO (negates buffered)",
161                 .def    = "0",
162         },
163         {
164                 .name   = "buffered",
165                 .type   = FIO_OPT_BOOL,
166                 .off1   = td_var_offset(odirect),
167                 .neg    = 1,
168                 .help   = "Use buffered IO (negates direct)",
169                 .def    = "1",
170         },
171         {
172                 .name   = "overwrite",
173                 .type   = FIO_OPT_BOOL,
174                 .off1   = td_var_offset(overwrite),
175                 .help   = "When writing, set whether to overwrite current data",
176                 .def    = "0",
177         },
178         {
179                 .name   = "loops",
180                 .type   = FIO_OPT_INT,
181                 .off1   = td_var_offset(loops),
182                 .help   = "Number of times to run the job",
183                 .def    = "1",
184         },
185         {
186                 .name   = "numjobs",
187                 .type   = FIO_OPT_INT,
188                 .off1   = td_var_offset(numjobs),
189                 .help   = "Duplicate this job this many times",
190                 .def    = "1",
191         },
192         {
193                 .name   = "startdelay",
194                 .type   = FIO_OPT_INT,
195                 .off1   = td_var_offset(start_delay),
196                 .help   = "Only start job when this period has passed",
197                 .def    = "0",
198         },
199         {
200                 .name   = "runtime",
201                 .alias  = "timeout",
202                 .type   = FIO_OPT_STR_VAL_TIME,
203                 .off1   = td_var_offset(timeout),
204                 .help   = "Stop workload when this amount of time has passed",
205                 .def    = "0",
206         },
207         {
208                 .name   = "mem",
209                 .type   = FIO_OPT_STR,
210                 .cb     = str_mem_cb,
211                 .help   = "Backing type for IO buffers",
212                 .def    = "malloc",
213                 .posval =  { "malloc", "shm", "shmhuge", "mmap", "mmaphuge", },
214         },
215         {
216                 .name   = "verify",
217                 .type   = FIO_OPT_STR,
218                 .cb     = str_verify_cb,
219                 .help   = "Verify sum function",
220                 .def    = "0",
221                 .posval = { "crc32", "md5", },
222         },
223         {
224                 .name   = "write_iolog",
225                 .type   = FIO_OPT_STR_STORE,
226                 .off1   = td_var_offset(write_iolog_file),
227                 .help   = "Store IO pattern to file",
228         },
229         {
230                 .name   = "read_iolog",
231                 .type   = FIO_OPT_STR_STORE,
232                 .off1   = td_var_offset(read_iolog_file),
233                 .help   = "Playback IO pattern from file",
234         },
235         {
236                 .name   = "exec_prerun",
237                 .type   = FIO_OPT_STR_STORE,
238                 .off1   = td_var_offset(exec_prerun),
239                 .help   = "Execute this file prior to running job",
240         },
241         {
242                 .name   = "exec_postrun",
243                 .type   = FIO_OPT_STR_STORE,
244                 .off1   = td_var_offset(exec_postrun),
245                 .help   = "Execute this file after running job",
246         },
247 #ifdef FIO_HAVE_IOSCHED_SWITCH
248         {
249                 .name   = "ioscheduler",
250                 .type   = FIO_OPT_STR_STORE,
251                 .off1   = td_var_offset(ioscheduler),
252                 .help   = "Use this IO scheduler on the backing device",
253         },
254 #endif
255         {
256                 .name   = "zonesize",
257                 .type   = FIO_OPT_STR_VAL,
258                 .off1   = td_var_offset(zone_size),
259                 .help   = "Give size of an IO zone",
260                 .def    = "0",
261         },
262         {
263                 .name   = "zoneskip",
264                 .type   = FIO_OPT_STR_VAL,
265                 .off1   = td_var_offset(zone_skip),
266                 .help   = "Space between IO zones",
267                 .def    = "0",
268         },
269         {
270                 .name   = "lockmem",
271                 .type   = FIO_OPT_STR_VAL,
272                 .cb     = str_lockmem_cb,
273                 .help   = "Lock down this amount of memory",
274                 .def    = "0",
275         },
276         {
277                 .name   = "rwmixcycle",
278                 .type   = FIO_OPT_INT,
279                 .off1   = td_var_offset(rwmixcycle),
280                 .help   = "Cycle period for mixed read/write workloads (msec)",
281                 .def    = "500",
282         },
283         {
284                 .name   = "rwmixread",
285                 .type   = FIO_OPT_INT,
286                 .off1   = td_var_offset(rwmixread),
287                 .maxval = 100,
288                 .help   = "Percentage of mixed workload that is reads",
289                 .def    = "50",
290         },
291         {
292                 .name   = "rwmixwrite",
293                 .type   = FIO_OPT_INT,
294                 .off1   = td_var_offset(rwmixwrite),
295                 .maxval = 100,
296                 .help   = "Percentage of mixed workload that is writes",
297                 .def    = "50",
298         },
299         {
300                 .name   = "nice",
301                 .type   = FIO_OPT_INT,
302                 .off1   = td_var_offset(nice),
303                 .help   = "Set job CPU nice value",
304                 .minval = -19,
305                 .maxval = 20,
306                 .def    = "0",
307         },
308 #ifdef FIO_HAVE_IOPRIO
309         {
310                 .name   = "prio",
311                 .type   = FIO_OPT_INT,
312                 .cb     = str_prio_cb,
313                 .help   = "Set job IO priority value",
314                 .minval = 0,
315                 .maxval = 7,
316         },
317         {
318                 .name   = "prioclass",
319                 .type   = FIO_OPT_INT,
320                 .cb     = str_prioclass_cb,
321                 .help   = "Set job IO priority class",
322                 .minval = 0,
323                 .maxval = 3,
324         },
325 #endif
326         {
327                 .name   = "thinktime",
328                 .type   = FIO_OPT_INT,
329                 .off1   = td_var_offset(thinktime),
330                 .help   = "Idle time between IO buffers (usec)",
331                 .def    = "0",
332         },
333         {
334                 .name   = "thinktime_spin",
335                 .type   = FIO_OPT_INT,
336                 .off1   = td_var_offset(thinktime_spin),
337                 .help   = "Start thinktime by spinning this amount (usec)",
338                 .def    = "0",
339         },
340         {
341                 .name   = "thinktime_blocks",
342                 .type   = FIO_OPT_INT,
343                 .off1   = td_var_offset(thinktime_blocks),
344                 .help   = "IO buffer period between 'thinktime'",
345                 .def    = "1",
346         },
347         {
348                 .name   = "rate",
349                 .type   = FIO_OPT_INT,
350                 .off1   = td_var_offset(rate),
351                 .help   = "Set bandwidth rate",
352         },
353         {
354                 .name   = "ratemin",
355                 .type   = FIO_OPT_INT,
356                 .off1   = td_var_offset(ratemin),
357                 .help   = "The bottom limit accepted",
358         },
359         {
360                 .name   = "ratecycle",
361                 .type   = FIO_OPT_INT,
362                 .off1   = td_var_offset(ratecycle),
363                 .help   = "Window average for rate limits (msec)",
364                 .def    = "1000",
365         },
366         {
367                 .name   = "invalidate",
368                 .type   = FIO_OPT_BOOL,
369                 .off1   = td_var_offset(invalidate_cache),
370                 .help   = "Invalidate buffer/page cache prior to running job",
371                 .def    = "1",
372         },
373         {
374                 .name   = "sync",
375                 .type   = FIO_OPT_BOOL,
376                 .off1   = td_var_offset(sync_io),
377                 .help   = "Use O_SYNC for buffered writes",
378                 .def    = "0",
379         },
380         {
381                 .name   = "bwavgtime",
382                 .type   = FIO_OPT_INT,
383                 .off1   = td_var_offset(bw_avg_time),
384                 .help   = "Time window over which to calculate bandwidth (msec)",
385                 .def    = "500",
386         },
387         {
388                 .name   = "create_serialize",
389                 .type   = FIO_OPT_BOOL,
390                 .off1   = td_var_offset(create_serialize),
391                 .help   = "Serialize creating of job files",
392                 .def    = "1",
393         },
394         {
395                 .name   = "create_fsync",
396                 .type   = FIO_OPT_BOOL,
397                 .off1   = td_var_offset(create_fsync),
398                 .help   = "Fsync file after creation",
399                 .def    = "1",
400         },
401         {
402                 .name   = "cpuload",
403                 .type   = FIO_OPT_INT,
404                 .off1   = td_var_offset(cpuload),
405                 .help   = "Use this percentage of CPU",
406         },
407         {
408                 .name   = "cpuchunks",
409                 .type   = FIO_OPT_INT,
410                 .off1   = td_var_offset(cpucycle),
411                 .help   = "Length of the CPU burn cycles",
412         },
413 #ifdef FIO_HAVE_CPU_AFFINITY
414         {
415                 .name   = "cpumask",
416                 .type   = FIO_OPT_INT,
417                 .cb     = str_cpumask_cb,
418                 .help   = "CPU affinity mask",
419         },
420 #endif
421         {
422                 .name   = "end_fsync",
423                 .type   = FIO_OPT_BOOL,
424                 .off1   = td_var_offset(end_fsync),
425                 .help   = "Include fsync at the end of job",
426                 .def    = "0",
427         },
428         {
429                 .name   = "unlink",
430                 .type   = FIO_OPT_BOOL,
431                 .off1   = td_var_offset(unlink),
432                 .help   = "Unlink created files after job has completed",
433                 .def    = "0",
434         },
435         {
436                 .name   = "exitall",
437                 .type   = FIO_OPT_STR_SET,
438                 .cb     = str_exitall_cb,
439                 .help   = "Terminate all jobs when one exits",
440         },
441         {
442                 .name   = "stonewall",
443                 .type   = FIO_OPT_STR_SET,
444                 .off1   = td_var_offset(stonewall),
445                 .help   = "Insert a hard barrier between this job and previous",
446         },
447         {
448                 .name   = "thread",
449                 .type   = FIO_OPT_STR_SET,
450                 .off1   = td_var_offset(use_thread),
451                 .help   = "Use threads instead of forks",
452         },
453         {
454                 .name   = "write_bw_log",
455                 .type   = FIO_OPT_STR_SET,
456                 .off1   = td_var_offset(write_bw_log),
457                 .help   = "Write log of bandwidth during run",
458         },
459         {
460                 .name   = "write_lat_log",
461                 .type   = FIO_OPT_STR_SET,
462                 .off1   = td_var_offset(write_lat_log),
463                 .help   = "Write log of latency during run",
464         },
465         {
466                 .name   = "hugepage-size",
467                 .type   = FIO_OPT_STR_VAL,
468                 .off1   = td_var_offset(hugepage_size),
469                 .help   = "When using hugepages, specify size of each page",
470                 .def    = __stringify(FIO_HUGE_PAGE),
471         },
472         {
473                 .name = NULL,
474         },
475 };
476
477 #define FIO_JOB_OPTS    (sizeof(options) / sizeof(struct fio_option))
478 #define FIO_CMD_OPTS    (16)
479 #define FIO_GETOPT_JOB  (0x89988998)
480
481 /*
482  * Command line options. These will contain the above, plus a few
483  * extra that only pertain to fio itself and not jobs.
484  */
485 static struct option long_options[FIO_JOB_OPTS + FIO_CMD_OPTS] = {
486         {
487                 .name           = "output",
488                 .has_arg        = required_argument,
489                 .val            = 'o',
490         },
491         {
492                 .name           = "timeout",
493                 .has_arg        = required_argument,
494                 .val            = 't',
495         },
496         {
497                 .name           = "latency-log",
498                 .has_arg        = required_argument,
499                 .val            = 'l',
500         },
501         {
502                 .name           = "bandwidth-log",
503                 .has_arg        = required_argument,
504                 .val            = 'b',
505         },
506         {
507                 .name           = "minimal",
508                 .has_arg        = optional_argument,
509                 .val            = 'm',
510         },
511         {
512                 .name           = "version",
513                 .has_arg        = no_argument,
514                 .val            = 'v',
515         },
516         {
517                 .name           = "help",
518                 .has_arg        = no_argument,
519                 .val            = 'h',
520         },
521         {
522                 .name           = "cmdhelp",
523                 .has_arg        = required_argument,
524                 .val            = 'c',
525         },
526         {
527                 .name           = NULL,
528         },
529 };
530
531 static int def_timeout = 0;
532
533 static char fio_version_string[] = "fio 1.11";
534
535 static char **ini_file;
536 static int max_jobs = MAX_JOBS;
537
538 struct thread_data def_thread;
539 struct thread_data *threads = NULL;
540
541 int exitall_on_terminate = 0;
542 int terse_output = 0;
543 unsigned long long mlock_size = 0;
544 FILE *f_out = NULL;
545 FILE *f_err = NULL;
546
547 static int write_lat_log = 0;
548 int write_bw_log = 0;
549
550 /*
551  * Return a free job structure.
552  */
553 static struct thread_data *get_new_job(int global, struct thread_data *parent)
554 {
555         struct thread_data *td;
556
557         if (global)
558                 return &def_thread;
559         if (thread_number >= max_jobs)
560                 return NULL;
561
562         td = &threads[thread_number++];
563         *td = *parent;
564
565         td->thread_number = thread_number;
566         return td;
567 }
568
569 static void put_job(struct thread_data *td)
570 {
571         if (td == &def_thread)
572                 return;
573
574         if (td->error)
575                 fprintf(f_out, "fio: %s\n", td->verror);
576
577         memset(&threads[td->thread_number - 1], 0, sizeof(*td));
578         thread_number--;
579 }
580
581 /*
582  * Lazy way of fixing up options that depend on each other. We could also
583  * define option callback handlers, but this is easier.
584  */
585 static void fixup_options(struct thread_data *td)
586 {
587         if (!td->rwmixread && td->rwmixwrite)
588                 td->rwmixread = 100 - td->rwmixwrite;
589
590         if (td->write_iolog_file && td->read_iolog_file) {
591                 log_err("fio: read iolog overrides write_iolog\n");
592                 free(td->write_iolog_file);
593                 td->write_iolog_file = NULL;
594         }
595
596         if (td->io_ops->flags & FIO_SYNCIO)
597                 td->iodepth = 1;
598         else {
599                 if (!td->iodepth)
600                         td->iodepth = td->nr_files;
601         }
602
603         /*
604          * only really works for sequential io for now, and with 1 file
605          */
606         if (td->zone_size && !td->sequential && td->nr_files == 1)
607                 td->zone_size = 0;
608
609         /*
610          * Reads can do overwrites, we always need to pre-create the file
611          */
612         if (td_read(td) || td_rw(td))
613                 td->overwrite = 1;
614
615         if (!td->min_bs[DDIR_READ])
616                 td->min_bs[DDIR_READ]= td->bs[DDIR_READ];
617         if (!td->max_bs[DDIR_READ])
618                 td->max_bs[DDIR_READ] = td->bs[DDIR_READ];
619         if (!td->min_bs[DDIR_WRITE])
620                 td->min_bs[DDIR_WRITE]= td->bs[DDIR_WRITE];
621         if (!td->max_bs[DDIR_WRITE])
622                 td->max_bs[DDIR_WRITE] = td->bs[DDIR_WRITE];
623
624         td->rw_min_bs = min(td->min_bs[DDIR_READ], td->min_bs[DDIR_WRITE]);
625
626         if (td_read(td) && !td_rw(td))
627                 td->verify = 0;
628
629         if (td->norandommap && td->verify != VERIFY_NONE) {
630                 log_err("fio: norandommap given, verify disabled\n");
631                 td->verify = VERIFY_NONE;
632         }
633         if (td->bs_unaligned && (td->odirect || td->io_ops->flags & FIO_RAWIO))
634                 log_err("fio: bs_unaligned may not work with raw io\n");
635
636         /*
637          * O_DIRECT and char doesn't mix, clear that flag if necessary.
638          */
639         if (td->filetype == FIO_TYPE_CHAR && td->odirect)
640                 td->odirect = 0;
641
642         /*
643          * thinktime_spin must be less than thinktime
644          */
645         if (td->thinktime_spin > td->thinktime)
646                 td->thinktime_spin = td->thinktime;
647 }
648
649 /*
650  * This function leaks the buffer
651  */
652 static char *to_kmg(unsigned int val)
653 {
654         char *buf = malloc(32);
655         char post[] = { 0, 'K', 'M', 'G', 'P', 'E', 0 };
656         char *p = post;
657
658         do {
659                 if (val & 1023)
660                         break;
661
662                 val >>= 10;
663                 p++;
664         } while (*p);
665
666         snprintf(buf, 31, "%u%c", val, *p);
667         return buf;
668 }
669
670 /*
671  * Adds a job to the list of things todo. Sanitizes the various options
672  * to make sure we don't have conflicts, and initializes various
673  * members of td.
674  */
675 static int add_job(struct thread_data *td, const char *jobname, int job_add_num)
676 {
677         const char *ddir_str[] = { "read", "write", "randread", "randwrite",
678                                    "rw", NULL, "randrw" };
679         struct stat sb;
680         int numjobs, ddir, i;
681         struct fio_file *f;
682
683         /*
684          * the def_thread is just for options, it's not a real job
685          */
686         if (td == &def_thread)
687                 return 0;
688
689         assert(td->io_ops);
690
691         if (td->odirect)
692                 td->io_ops->flags |= FIO_RAWIO;
693
694         td->filetype = FIO_TYPE_FILE;
695         if (td->filename && !lstat(td->filename, &sb)) {
696                 if (S_ISBLK(sb.st_mode))
697                         td->filetype = FIO_TYPE_BD;
698                 else if (S_ISCHR(sb.st_mode))
699                         td->filetype = FIO_TYPE_CHAR;
700         }
701
702         fixup_options(td);
703
704         if (td->filename)
705                 td->nr_uniq_files = 1;
706         else
707                 td->nr_uniq_files = td->nr_files;
708
709         if (td->filetype == FIO_TYPE_FILE || td->filename) {
710                 char tmp[PATH_MAX];
711                 int len = 0;
712
713                 if (td->directory && td->directory[0] != '\0') {
714                         if (lstat(td->directory, &sb) < 0) {
715                                 log_err("fio: %s is not a directory\n", td->directory);
716                                 td_verror(td, errno);
717                                 return 1;
718                         }
719                         if (!S_ISDIR(sb.st_mode)) {
720                                 log_err("fio: %s is not a directory\n", td->directory);
721                                 return 1;
722                         }
723                         len = sprintf(tmp, "%s/", td->directory);
724                 }
725
726                 td->files = malloc(sizeof(struct fio_file) * td->nr_files);
727
728                 for_each_file(td, f, i) {
729                         memset(f, 0, sizeof(*f));
730                         f->fd = -1;
731
732                         if (td->filename)
733                                 sprintf(tmp + len, "%s", td->filename);
734                         else
735                                 sprintf(tmp + len, "%s.%d.%d", jobname, td->thread_number, i);
736                         f->file_name = strdup(tmp);
737                 }
738         } else {
739                 td->nr_files = 1;
740                 td->files = malloc(sizeof(struct fio_file));
741                 f = &td->files[0];
742
743                 memset(f, 0, sizeof(*f));
744                 f->fd = -1;
745                 f->file_name = strdup(jobname);
746         }
747
748         for_each_file(td, f, i) {
749                 f->file_size = td->total_file_size / td->nr_files;
750                 f->file_offset = td->start_offset;
751         }
752                 
753         fio_sem_init(&td->mutex, 0);
754
755         td->clat_stat[0].min_val = td->clat_stat[1].min_val = ULONG_MAX;
756         td->slat_stat[0].min_val = td->slat_stat[1].min_val = ULONG_MAX;
757         td->bw_stat[0].min_val = td->bw_stat[1].min_val = ULONG_MAX;
758
759         if (td->stonewall && td->thread_number > 1)
760                 groupid++;
761
762         td->groupid = groupid;
763
764         if (setup_rate(td))
765                 goto err;
766
767         if (td->write_lat_log) {
768                 setup_log(&td->slat_log);
769                 setup_log(&td->clat_log);
770         }
771         if (td->write_bw_log)
772                 setup_log(&td->bw_log);
773
774         if (!td->name)
775                 td->name = strdup(jobname);
776
777         ddir = td->ddir + (!td->sequential << 1) + (td->iomix << 2);
778
779         if (!terse_output) {
780                 if (!job_add_num) {
781                         if (td->io_ops->flags & FIO_CPUIO)
782                                 fprintf(f_out, "%s: ioengine=cpu, cpuload=%u, cpucycle=%u\n", td->name, td->cpuload, td->cpucycle);
783                         else {
784                                 char *c1, *c2, *c3, *c4;
785
786                                 c1 = to_kmg(td->min_bs[DDIR_READ]);
787                                 c2 = to_kmg(td->max_bs[DDIR_READ]);
788                                 c3 = to_kmg(td->min_bs[DDIR_WRITE]);
789                                 c4 = to_kmg(td->max_bs[DDIR_WRITE]);
790
791                                 fprintf(f_out, "%s: (g=%d): rw=%s, bs=%s-%s/%s-%s, ioengine=%s, iodepth=%u\n", td->name, td->groupid, ddir_str[ddir], c1, c2, c3, c4, td->io_ops->name, td->iodepth);
792
793                                 free(c1);
794                                 free(c2);
795                                 free(c3);
796                                 free(c4);
797                         }
798                 } else if (job_add_num == 1)
799                         fprintf(f_out, "...\n");
800         }
801
802         /*
803          * recurse add identical jobs, clear numjobs and stonewall options
804          * as they don't apply to sub-jobs
805          */
806         numjobs = td->numjobs;
807         while (--numjobs) {
808                 struct thread_data *td_new = get_new_job(0, td);
809
810                 if (!td_new)
811                         goto err;
812
813                 td_new->numjobs = 1;
814                 td_new->stonewall = 0;
815                 job_add_num = numjobs - 1;
816
817                 if (add_job(td_new, jobname, job_add_num))
818                         goto err;
819         }
820         return 0;
821 err:
822         put_job(td);
823         return -1;
824 }
825
826 /*
827  * Initialize the various random states we need (random io, block size ranges,
828  * read/write mix, etc).
829  */
830 int init_random_state(struct thread_data *td)
831 {
832         unsigned long seeds[4];
833         int fd, num_maps, blocks, i;
834         struct fio_file *f;
835
836         if (td->io_ops->flags & FIO_CPUIO)
837                 return 0;
838
839         fd = open("/dev/urandom", O_RDONLY);
840         if (fd == -1) {
841                 td_verror(td, errno);
842                 return 1;
843         }
844
845         if (read(fd, seeds, sizeof(seeds)) < (int) sizeof(seeds)) {
846                 td_verror(td, EIO);
847                 close(fd);
848                 return 1;
849         }
850
851         close(fd);
852
853         os_random_seed(seeds[0], &td->bsrange_state);
854         os_random_seed(seeds[1], &td->verify_state);
855         os_random_seed(seeds[2], &td->rwmix_state);
856
857         if (td->sequential)
858                 return 0;
859
860         if (td->rand_repeatable)
861                 seeds[3] = FIO_RANDSEED * td->thread_number;
862
863         if (!td->norandommap) {
864                 for_each_file(td, f, i) {
865                         blocks = (f->real_file_size + td->rw_min_bs - 1) / td->rw_min_bs;
866                         num_maps = (blocks + BLOCKS_PER_MAP-1)/ BLOCKS_PER_MAP;
867                         f->file_map = malloc(num_maps * sizeof(long));
868                         f->num_maps = num_maps;
869                         memset(f->file_map, 0, num_maps * sizeof(long));
870                 }
871         }
872
873         os_random_seed(seeds[3], &td->random_state);
874         return 0;
875 }
876
877 static void fill_cpu_mask(os_cpu_mask_t cpumask, int cpu)
878 {
879 #ifdef FIO_HAVE_CPU_AFFINITY
880         unsigned int i;
881
882         CPU_ZERO(&cpumask);
883
884         for (i = 0; i < sizeof(int) * 8; i++) {
885                 if ((1 << i) & cpu)
886                         CPU_SET(i, &cpumask);
887         }
888 #endif
889 }
890
891 static int is_empty_or_comment(char *line)
892 {
893         unsigned int i;
894
895         for (i = 0; i < strlen(line); i++) {
896                 if (line[i] == ';')
897                         return 1;
898                 if (!isspace(line[i]) && !iscntrl(line[i]))
899                         return 0;
900         }
901
902         return 1;
903 }
904
905 static int str_rw_cb(void *data, const char *mem)
906 {
907         struct thread_data *td = data;
908
909         if (!strncmp(mem, "read", 4) || !strncmp(mem, "0", 1)) {
910                 td->ddir = DDIR_READ;
911                 td->sequential = 1;
912                 return 0;
913         } else if (!strncmp(mem, "randread", 8)) {
914                 td->ddir = DDIR_READ;
915                 td->sequential = 0;
916                 return 0;
917         } else if (!strncmp(mem, "write", 5) || !strncmp(mem, "1", 1)) {
918                 td->ddir = DDIR_WRITE;
919                 td->sequential = 1;
920                 return 0;
921         } else if (!strncmp(mem, "randwrite", 9)) {
922                 td->ddir = DDIR_WRITE;
923                 td->sequential = 0;
924                 return 0;
925         } else if (!strncmp(mem, "rw", 2)) {
926                 td->ddir = DDIR_READ;
927                 td->iomix = 1;
928                 td->sequential = 1;
929                 return 0;
930         } else if (!strncmp(mem, "randrw", 6)) {
931                 td->ddir = DDIR_READ;
932                 td->iomix = 1;
933                 td->sequential = 0;
934                 return 0;
935         }
936
937         log_err("fio: data direction: read, write, randread, randwrite, rw, randrw\n");
938         return 1;
939 }
940
941 static int str_verify_cb(void *data, const char *mem)
942 {
943         struct thread_data *td = data;
944
945         if (!strncmp(mem, "0", 1)) {
946                 td->verify = VERIFY_NONE;
947                 return 0;
948         } else if (!strncmp(mem, "md5", 3) || !strncmp(mem, "1", 1)) {
949                 td->verify = VERIFY_MD5;
950                 return 0;
951         } else if (!strncmp(mem, "crc32", 5)) {
952                 td->verify = VERIFY_CRC32;
953                 return 0;
954         }
955
956         log_err("fio: verify types: md5, crc32\n");
957         return 1;
958 }
959
960 /*
961  * Check if mmap/mmaphuge has a :/foo/bar/file at the end. If so, return that.
962  */
963 static char *get_mmap_file(const char *str)
964 {
965         char *p = strstr(str, ":");
966
967         if (!p)
968                 return NULL;
969
970         p++;
971         strip_blank_front(&p);
972         strip_blank_end(p);
973         return strdup(p);
974 }
975
976 static int str_mem_cb(void *data, const char *mem)
977 {
978         struct thread_data *td = data;
979
980         if (!strncmp(mem, "malloc", 6)) {
981                 td->mem_type = MEM_MALLOC;
982                 return 0;
983         } else if (!strncmp(mem, "mmaphuge", 8)) {
984 #ifdef FIO_HAVE_HUGETLB
985                 /*
986                  * mmaphuge must be appended with the actual file
987                  */
988                 td->mmapfile = get_mmap_file(mem);
989                 if (!td->mmapfile) {
990                         log_err("fio: mmaphuge:/path/to/file\n");
991                         return 1;
992                 }
993
994                 td->mem_type = MEM_MMAPHUGE;
995                 return 0;
996 #else
997                 log_err("fio: mmaphuge not available\n");
998                 return 1;
999 #endif
1000         } else if (!strncmp(mem, "mmap", 4)) {
1001                 /*
1002                  * Check if the user wants file backed memory. It's ok
1003                  * if there's no file given, we'll just use anon mamp then.
1004                  */
1005                 td->mmapfile = get_mmap_file(mem);
1006                 td->mem_type = MEM_MMAP;
1007                 return 0;
1008         } else if (!strncmp(mem, "shmhuge", 7)) {
1009 #ifdef FIO_HAVE_HUGETLB
1010                 td->mem_type = MEM_SHMHUGE;
1011                 return 0;
1012 #else
1013                 log_err("fio: shmhuge not available\n");
1014                 return 1;
1015 #endif
1016         } else if (!strncmp(mem, "shm", 3)) {
1017                 td->mem_type = MEM_SHM;
1018                 return 0;
1019         }
1020
1021         log_err("fio: mem type: malloc, shm, shmhuge, mmap, mmaphuge\n");
1022         return 1;
1023 }
1024
1025 static int str_ioengine_cb(void *data, const char *str)
1026 {
1027         struct thread_data *td = data;
1028
1029         td->io_ops = load_ioengine(td, str);
1030         if (td->io_ops)
1031                 return 0;
1032
1033         log_err("fio: ioengine= libaio, posixaio, sync, mmap, sgio, splice, cpu, null\n");
1034         log_err("fio: or specify path to dynamic ioengine module\n");
1035         return 1;
1036 }
1037
1038 static int str_lockmem_cb(void fio_unused *data, unsigned long *val)
1039 {
1040         mlock_size = *val;
1041         return 0;
1042 }
1043
1044 #ifdef FIO_HAVE_IOPRIO
1045 static int str_prioclass_cb(void *data, unsigned int *val)
1046 {
1047         struct thread_data *td = data;
1048
1049         td->ioprio |= *val << IOPRIO_CLASS_SHIFT;
1050         return 0;
1051 }
1052
1053 static int str_prio_cb(void *data, unsigned int *val)
1054 {
1055         struct thread_data *td = data;
1056
1057         td->ioprio |= *val;
1058         return 0;
1059 }
1060 #endif
1061
1062 static int str_exitall_cb(void)
1063 {
1064         exitall_on_terminate = 1;
1065         return 0;
1066 }
1067
1068 static int str_cpumask_cb(void *data, unsigned int *val)
1069 {
1070         struct thread_data *td = data;
1071
1072         fill_cpu_mask(td->cpumask, *val);
1073         return 0;
1074 }
1075
1076 /*
1077  * This is our [ini] type file parser.
1078  */
1079 static int parse_jobs_ini(char *file, int stonewall_flag)
1080 {
1081         unsigned int global;
1082         struct thread_data *td;
1083         char *string, *name;
1084         fpos_t off;
1085         FILE *f;
1086         char *p;
1087         int ret = 0, stonewall;
1088
1089         f = fopen(file, "r");
1090         if (!f) {
1091                 perror("fopen job file");
1092                 return 1;
1093         }
1094
1095         string = malloc(4096);
1096         name = malloc(256);
1097         memset(name, 0, 256);
1098
1099         stonewall = stonewall_flag;
1100         do {
1101                 p = fgets(string, 4095, f);
1102                 if (!p)
1103                         break;
1104                 if (is_empty_or_comment(p))
1105                         continue;
1106                 if (sscanf(p, "[%255s]", name) != 1)
1107                         continue;
1108
1109                 global = !strncmp(name, "global", 6);
1110
1111                 name[strlen(name) - 1] = '\0';
1112
1113                 td = get_new_job(global, &def_thread);
1114                 if (!td) {
1115                         ret = 1;
1116                         break;
1117                 }
1118
1119                 /*
1120                  * Seperate multiple job files by a stonewall
1121                  */
1122                 if (!global && stonewall) {
1123                         td->stonewall = stonewall;
1124                         stonewall = 0;
1125                 }
1126
1127                 fgetpos(f, &off);
1128                 while ((p = fgets(string, 4096, f)) != NULL) {
1129                         if (is_empty_or_comment(p))
1130                                 continue;
1131
1132                         strip_blank_front(&p);
1133
1134                         if (p[0] == '[')
1135                                 break;
1136
1137                         strip_blank_end(p);
1138
1139                         fgetpos(f, &off);
1140
1141                         /*
1142                          * Don't break here, continue parsing options so we
1143                          * dump all the bad ones. Makes trial/error fixups
1144                          * easier on the user.
1145                          */
1146                         ret |= parse_option(p, options, td);
1147                 }
1148
1149                 if (!ret) {
1150                         fsetpos(f, &off);
1151                         ret = add_job(td, name, 0);
1152                 } else {
1153                         log_err("fio: job %s dropped\n", name);
1154                         put_job(td);
1155                 }
1156         } while (!ret);
1157
1158         free(string);
1159         free(name);
1160         fclose(f);
1161         return ret;
1162 }
1163
1164 static int fill_def_thread(void)
1165 {
1166         memset(&def_thread, 0, sizeof(def_thread));
1167
1168         if (fio_getaffinity(getpid(), &def_thread.cpumask) == -1) {
1169                 perror("sched_getaffinity");
1170                 return 1;
1171         }
1172
1173         /*
1174          * fill default options
1175          */
1176         fill_default_options(&def_thread, options);
1177
1178         def_thread.timeout = def_timeout;
1179         def_thread.write_bw_log = write_bw_log;
1180         def_thread.write_lat_log = write_lat_log;
1181
1182 #ifdef FIO_HAVE_DISK_UTIL
1183         def_thread.do_disk_util = 1;
1184 #endif
1185
1186         return 0;
1187 }
1188
1189 static void usage(void)
1190 {
1191         printf("%s\n", fio_version_string);
1192         printf("\t--output\tWrite output to file\n");
1193         printf("\t--timeout\tRuntime in seconds\n");
1194         printf("\t--latency-log\tGenerate per-job latency logs\n");
1195         printf("\t--bandwidth-log\tGenerate per-job bandwidth logs\n");
1196         printf("\t--minimal\tMinimal (terse) output\n");
1197         printf("\t--version\tPrint version info and exit\n");
1198         printf("\t--help\t\tPrint this page\n");
1199         printf("\t--cmdhelp=cmd\tPrint command help, \"all\" for all of them\n");
1200 }
1201
1202 static int parse_cmd_line(int argc, char *argv[])
1203 {
1204         struct thread_data *td = NULL;
1205         int c, ini_idx = 0, lidx, ret;
1206
1207         while ((c = getopt_long(argc, argv, "", long_options, &lidx)) != -1) {
1208                 switch (c) {
1209                 case 't':
1210                         def_timeout = atoi(optarg);
1211                         break;
1212                 case 'l':
1213                         write_lat_log = 1;
1214                         break;
1215                 case 'w':
1216                         write_bw_log = 1;
1217                         break;
1218                 case 'o':
1219                         f_out = fopen(optarg, "w+");
1220                         if (!f_out) {
1221                                 perror("fopen output");
1222                                 exit(1);
1223                         }
1224                         f_err = f_out;
1225                         break;
1226                 case 'm':
1227                         terse_output = 1;
1228                         break;
1229                 case 'h':
1230                         usage();
1231                         exit(0);
1232                 case 'c':
1233                         ret = show_cmd_help(options, optarg);
1234                         exit(ret);
1235                 case 'v':
1236                         printf("%s\n", fio_version_string);
1237                         exit(0);
1238                 case FIO_GETOPT_JOB: {
1239                         const char *opt = long_options[lidx].name;
1240                         char *val = optarg;
1241
1242                         if (!strncmp(opt, "name", 4) && td) {
1243                                 ret = add_job(td, td->name ?: "fio", 0);
1244                                 if (ret) {
1245                                         put_job(td);
1246                                         return 0;
1247                                 }
1248                                 td = NULL;
1249                         }
1250                         if (!td) {
1251                                 int global = !strncmp(val, "global", 6);
1252
1253                                 td = get_new_job(global, &def_thread);
1254                                 if (!td)
1255                                         return 0;
1256                         }
1257
1258                         ret = parse_cmd_option(opt, val, options, td);
1259                         if (ret) {
1260                                 log_err("fio: job dropped\n");
1261                                 put_job(td);
1262                                 td = NULL;
1263                         }
1264                         break;
1265                 }
1266                 default:
1267                         break;
1268                 }
1269         }
1270
1271         if (td) {
1272                 ret = add_job(td, td->name ?: "fio", 0);
1273                 if (ret)
1274                         put_job(td);
1275         }
1276
1277         while (optind < argc) {
1278                 ini_idx++;
1279                 ini_file = realloc(ini_file, ini_idx * sizeof(char *));
1280                 ini_file[ini_idx - 1] = strdup(argv[optind]);
1281                 optind++;
1282         }
1283
1284         return ini_idx;
1285 }
1286
1287 static void free_shm(void)
1288 {
1289         struct shmid_ds sbuf;
1290
1291         if (threads) {
1292                 shmdt((void *) threads);
1293                 threads = NULL;
1294                 shmctl(shm_id, IPC_RMID, &sbuf);
1295         }
1296 }
1297
1298 /*
1299  * The thread area is shared between the main process and the job
1300  * threads/processes. So setup a shared memory segment that will hold
1301  * all the job info.
1302  */
1303 static int setup_thread_area(void)
1304 {
1305         /*
1306          * 1024 is too much on some machines, scale max_jobs if
1307          * we get a failure that looks like too large a shm segment
1308          */
1309         do {
1310                 size_t size = max_jobs * sizeof(struct thread_data);
1311
1312                 shm_id = shmget(0, size, IPC_CREAT | 0600);
1313                 if (shm_id != -1)
1314                         break;
1315                 if (errno != EINVAL) {
1316                         perror("shmget");
1317                         break;
1318                 }
1319
1320                 max_jobs >>= 1;
1321         } while (max_jobs);
1322
1323         if (shm_id == -1)
1324                 return 1;
1325
1326         threads = shmat(shm_id, NULL, 0);
1327         if (threads == (void *) -1) {
1328                 perror("shmat");
1329                 return 1;
1330         }
1331
1332         atexit(free_shm);
1333         return 0;
1334 }
1335
1336 /*
1337  * Copy the fio options into the long options map, so we mirror
1338  * job and cmd line options.
1339  */
1340 static void dupe_job_options(void)
1341 {
1342         struct fio_option *o;
1343         unsigned int i;
1344
1345         i = 0;
1346         while (long_options[i].name)
1347                 i++;
1348
1349         o = &options[0];
1350         while (o->name) {
1351                 long_options[i].name = o->name;
1352                 long_options[i].val = FIO_GETOPT_JOB;
1353                 if (o->type == FIO_OPT_STR_SET)
1354                         long_options[i].has_arg = no_argument;
1355                 else
1356                         long_options[i].has_arg = required_argument;
1357
1358                 i++;
1359                 o++;
1360                 assert(i < FIO_JOB_OPTS + FIO_CMD_OPTS);
1361         }
1362 }
1363
1364 int parse_options(int argc, char *argv[])
1365 {
1366         int job_files, i;
1367
1368         f_out = stdout;
1369         f_err = stderr;
1370
1371         options_init(options);
1372
1373         dupe_job_options();
1374
1375         if (setup_thread_area())
1376                 return 1;
1377         if (fill_def_thread())
1378                 return 1;
1379
1380         job_files = parse_cmd_line(argc, argv);
1381
1382         for (i = 0; i < job_files; i++) {
1383                 if (fill_def_thread())
1384                         return 1;
1385                 if (parse_jobs_ini(ini_file[i], i))
1386                         return 1;
1387                 free(ini_file[i]);
1388         }
1389
1390         free(ini_file);
1391
1392         if (!thread_number) {
1393                 log_err("No jobs defined(s)\n");
1394                 return 1;
1395         }
1396
1397         return 0;
1398 }