8 #include <sys/resource.h>
19 #include "compiler/compiler.h"
23 #include "arch/arch.h"
37 #include "lib/getopt.h"
43 #ifdef FIO_HAVE_SOLARISAIO
44 #include <sys/asynch.h>
48 #include <linux/mempolicy.h>
52 * "local" is pseudo-policy
54 #define MPOL_LOCAL MPOL_MAX
58 * What type of allocation to use for io buffers
61 MEM_MALLOC = 0, /* ordinary malloc */
62 MEM_SHM, /* use shared memory segments */
63 MEM_SHMHUGE, /* use shared memory segments with huge pages */
64 MEM_MMAP, /* use anonynomous mmap */
65 MEM_MMAPHUGE, /* memory mapped huge file */
69 * offset generator types
77 * What type of errors to continue on when continue_on_error is used
80 ERROR_TYPE_READ_BIT = 0,
81 ERROR_TYPE_WRITE_BIT = 1,
82 ERROR_TYPE_VERIFY_BIT = 2,
88 ERROR_TYPE_READ = 1 << ERROR_TYPE_READ_BIT,
89 ERROR_TYPE_WRITE = 1 << ERROR_TYPE_WRITE_BIT,
90 ERROR_TYPE_VERIFY = 1 << ERROR_TYPE_VERIFY_BIT,
91 ERROR_TYPE_ANY = 0xffff,
99 struct thread_options {
107 enum td_ddir td_ddir;
109 unsigned int kb_base;
110 unsigned int ddir_seq_nr;
112 unsigned int iodepth;
113 unsigned int iodepth_low;
114 unsigned int iodepth_batch;
115 unsigned int iodepth_batch_complete;
117 unsigned long long size;
118 unsigned int size_percent;
119 unsigned int fill_device;
120 unsigned long long file_size_low;
121 unsigned long long file_size_high;
122 unsigned long long start_offset;
124 unsigned int bs[DDIR_RWDIR_CNT];
125 unsigned int ba[DDIR_RWDIR_CNT];
126 unsigned int min_bs[DDIR_RWDIR_CNT];
127 unsigned int max_bs[DDIR_RWDIR_CNT];
128 struct bssplit *bssplit[DDIR_RWDIR_CNT];
129 unsigned int bssplit_nr[DDIR_RWDIR_CNT];
131 int *ignore_error[ERROR_TYPE_CNT];
132 unsigned int ignore_error_nr[ERROR_TYPE_CNT];
133 unsigned int error_dump;
135 unsigned int nr_files;
136 unsigned int open_files;
137 enum file_lock_mode file_lock_mode;
138 unsigned int lockfile_batch;
140 unsigned int odirect;
141 unsigned int invalidate_cache;
142 unsigned int create_serialize;
143 unsigned int create_fsync;
144 unsigned int create_on_open;
145 unsigned int create_only;
146 unsigned int end_fsync;
147 unsigned int pre_read;
148 unsigned int sync_io;
150 unsigned int do_verify;
151 unsigned int verifysort;
152 unsigned int verify_interval;
153 unsigned int verify_offset;
154 char verify_pattern[MAX_PATTERN_SIZE];
155 unsigned int verify_pattern_bytes;
156 unsigned int verify_fatal;
157 unsigned int verify_dump;
158 unsigned int verify_async;
159 unsigned long long verify_backlog;
160 unsigned int verify_batch;
161 unsigned int use_thread;
163 unsigned int do_disk_util;
164 unsigned int override_sync;
165 unsigned int rand_repeatable;
166 unsigned int use_os_rand;
167 unsigned int write_lat_log;
168 unsigned int write_bw_log;
169 unsigned int write_iops_log;
170 unsigned int log_avg_msec;
171 unsigned int norandommap;
172 unsigned int softrandommap;
173 unsigned int bs_unaligned;
174 unsigned int fsync_on_close;
176 unsigned int random_distribution;
180 unsigned int random_generator;
182 unsigned int hugepage_size;
183 unsigned int rw_min_bs;
184 unsigned int thinktime;
185 unsigned int thinktime_spin;
186 unsigned int thinktime_blocks;
187 unsigned int fsync_blocks;
188 unsigned int fdatasync_blocks;
189 unsigned int barrier_blocks;
190 unsigned long long start_delay;
191 unsigned long long timeout;
192 unsigned long long ramp_time;
193 unsigned int overwrite;
194 unsigned int bw_avg_time;
195 unsigned int iops_avg_time;
197 unsigned long long zone_range;
198 unsigned long long zone_size;
199 unsigned long long zone_skip;
200 enum fio_memtype mem_type;
201 unsigned int mem_align;
203 unsigned int max_latency;
205 unsigned int stonewall;
206 unsigned int new_group;
207 unsigned int numjobs;
208 os_cpu_mask_t cpumask;
209 unsigned int cpumask_set;
210 os_cpu_mask_t verify_cpumask;
211 unsigned int verify_cpumask_set;
212 #ifdef CONFIG_LIBNUMA
213 struct bitmask *numa_cpunodesmask;
214 unsigned int numa_cpumask_set;
215 unsigned short numa_mem_mode;
216 unsigned int numa_mem_prefer_node;
217 struct bitmask *numa_memnodesmask;
218 unsigned int numa_memmask_set;
221 unsigned int rwmixcycle;
222 unsigned int rwmix[2];
224 unsigned int file_service_type;
225 unsigned int group_reporting;
226 unsigned int fadvise_hint;
227 enum fio_fallocate_mode fallocate_mode;
228 unsigned int zero_buffers;
229 unsigned int refill_buffers;
230 unsigned int scramble_buffers;
231 unsigned int compress_percentage;
232 unsigned int compress_chunk;
233 unsigned int time_based;
234 unsigned int disable_lat;
235 unsigned int disable_clat;
236 unsigned int disable_slat;
237 unsigned int disable_bw;
238 unsigned int gtod_reduce;
239 unsigned int gtod_cpu;
240 unsigned int gtod_offload;
241 enum fio_cs clocksource;
242 unsigned int no_stall;
243 unsigned int trim_percentage;
244 unsigned int trim_batch;
245 unsigned int trim_zero;
246 unsigned long long trim_backlog;
247 unsigned int clat_percentiles;
248 unsigned int overwrite_plist;
249 fio_fp64_t percentile_list[FIO_IO_U_LIST_MAX_LEN];
251 char *read_iolog_file;
252 char *write_iolog_file;
256 char *replay_redirect;
259 * Pre-run and post-run shell
264 unsigned int rate[DDIR_RWDIR_CNT];
265 unsigned int ratemin[DDIR_RWDIR_CNT];
266 unsigned int ratecycle;
267 unsigned int rate_iops[DDIR_RWDIR_CNT];
268 unsigned int rate_iops_min[DDIR_RWDIR_CNT];
273 * CPU "io" cycle burner
275 unsigned int cpuload;
276 unsigned int cpucycle;
281 enum error_type continue_on_error;
284 * Benchmark profile type
289 * blkio cgroup support
292 unsigned int cgroup_weight;
293 unsigned int cgroup_nodelete;
301 unsigned int flow_sleep;
303 unsigned long long offset_increment;
305 unsigned int sync_file_range;
309 TD_F_VER_BACKLOG = 1,
310 TD_F_TRIM_BACKLOG = 2,
312 TD_F_REFILL_BUFFERS = 8,
313 TD_F_SCRAMBLE_BUFFERS = 16,
315 TD_F_PROFILE_OPS = 64,
324 FIO_RAND_FILE_SIZE_OFF,
331 * This describes a single thread/process executing a fio job.
334 struct thread_options o;
337 char verror[FIO_VERROR_SIZE];
341 struct thread_stat ts;
343 struct io_log *slat_log;
344 struct io_log *clat_log;
345 struct io_log *lat_log;
346 struct io_log *bw_log;
347 struct io_log *iops_log;
349 uint64_t stat_io_bytes[DDIR_RWDIR_CNT];
350 struct timeval bw_sample_time;
352 uint64_t stat_io_blocks[DDIR_RWDIR_CNT];
353 struct timeval iops_sample_time;
355 struct rusage ru_start;
356 struct rusage ru_end;
358 struct fio_file **files;
359 unsigned int files_size;
360 unsigned int files_index;
361 unsigned int nr_open_files;
362 unsigned int nr_done_files;
363 unsigned int nr_normal_files;
365 unsigned int next_file;
366 os_random_state_t next_file_state;
367 struct frand_state __next_file_state;
374 size_t orig_buffer_size;
375 volatile int terminate;
376 volatile int runstate;
378 unsigned int ioprio_set;
379 unsigned int last_was_sync;
380 enum fio_ddir last_ddir;
390 unsigned long rand_seeds[FIO_RAND_NR_OFFS];
393 os_random_state_t bsrange_state;
394 struct frand_state __bsrange_state;
397 os_random_state_t verify_state;
398 struct frand_state __verify_state;
401 os_random_state_t trim_state;
402 struct frand_state __trim_state;
405 struct frand_state buf_state;
407 unsigned int verify_batch;
408 unsigned int trim_batch;
413 * IO engine hooks, contains everything needed to submit an io_u
414 * to any of the available IO engines.
416 struct ioengine_ops *io_ops;
419 * Queue depth of io_u's that fio MIGHT do
421 unsigned int cur_depth;
424 * io_u's about to be committed
426 unsigned int io_u_queued;
429 * io_u's submitted but not completed yet
431 unsigned int io_u_in_flight;
434 * List of free and busy io_u's
436 struct flist_head io_u_freelist;
437 struct flist_head io_u_busylist;
438 struct flist_head io_u_requeues;
439 pthread_mutex_t io_u_lock;
440 pthread_cond_t free_cond;
443 * async verify offload
445 struct flist_head verify_list;
446 pthread_t *verify_threads;
447 unsigned int nr_verify_threads;
448 pthread_cond_t verify_cond;
449 int verify_thread_exit;
454 unsigned long long rate_bps[DDIR_RWDIR_CNT];
455 long rate_pending_usleep[DDIR_RWDIR_CNT];
456 unsigned long rate_bytes[DDIR_RWDIR_CNT];
457 unsigned long rate_blocks[DDIR_RWDIR_CNT];
458 struct timeval lastrate[DDIR_RWDIR_CNT];
460 unsigned long long total_io_size;
461 unsigned long long fill_device_size;
463 unsigned long io_issues[DDIR_RWDIR_CNT];
464 unsigned long long io_blocks[DDIR_RWDIR_CNT];
465 unsigned long long this_io_blocks[DDIR_RWDIR_CNT];
466 unsigned long long io_bytes[DDIR_RWDIR_CNT];
467 unsigned long long io_skip_bytes;
468 unsigned long long this_io_bytes[DDIR_RWDIR_CNT];
469 unsigned long long zone_bytes;
470 struct fio_mutex *mutex;
473 * State for random io, a bitmap of blocks done vs not done
476 os_random_state_t random_state;
477 struct frand_state __random_state;
480 struct timeval start; /* start of this loop */
481 struct timeval epoch; /* time job was started */
482 struct timeval last_issue;
483 struct timeval tv_cache;
484 unsigned int tv_cache_nr;
485 unsigned int tv_cache_mask;
486 unsigned int ramp_time_over;
489 * read/write mixed workload state
492 os_random_state_t rwmix_state;
493 struct frand_state __rwmix_state;
495 unsigned long rwmix_issues;
496 enum fio_ddir rwmix_ddir;
497 unsigned int ddir_seq_nr;
500 * IO history logs for verification. We use a tree for sorting,
501 * if we are overwriting. Otherwise just use a fifo.
503 struct rb_root io_hist_tree;
504 struct flist_head io_hist_list;
505 unsigned long io_hist_len;
510 struct flist_head io_log_list;
513 * For tracking/handling discards
515 struct flist_head trim_list;
516 unsigned long trim_entries;
519 * for fileservice, how often to switch to a new file
521 unsigned int file_service_nr;
522 unsigned int file_service_left;
523 struct fio_file *file_service_file;
525 unsigned int sync_file_range_nr;
528 * For generating file sizes
531 os_random_state_t file_size_state;
532 struct frand_state __file_size_state;
538 unsigned int total_err_count;
541 struct fio_flow *flow;
544 * Can be overloaded by profiles
546 struct prof_io_ops prof_io_ops;
551 * when should interactive ETA output be generated
559 #define __td_verror(td, err, msg, func) \
565 if (!(td)->first_error) \
566 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, func=%s, error=%s", __FILE__, __LINE__, (func), (msg)); \
570 #define td_clear_error(td) \
572 #define td_verror(td, err, func) \
573 __td_verror((td), (err), strerror((err)), (func))
574 #define td_vmsg(td, err, msg, func) \
575 __td_verror((td), (err), (msg), (func))
577 #define __fio_stringify_1(x) #x
578 #define __fio_stringify(x) __fio_stringify_1(x)
580 extern int exitall_on_terminate;
581 extern unsigned int thread_number;
582 extern unsigned int stat_number;
583 extern unsigned int nr_process, nr_thread;
586 extern int output_format;
587 extern int temp_stall_ts;
588 extern unsigned long long mlock_size;
589 extern uintptr_t page_mask, page_size;
590 extern int read_only;
591 extern int eta_print;
592 extern unsigned long done_secs;
593 extern char *job_section;
594 extern int fio_gtod_offload;
595 extern int fio_gtod_cpu;
596 extern enum fio_cs fio_clock_source;
597 extern int fio_clock_source_set;
598 extern int warnings_fatal;
599 extern int terse_version;
600 extern int is_backend;
601 extern int nr_clients;
602 extern int log_syslog;
603 extern const char fio_version_string[];
604 extern const fio_fp64_t def_percentile_list[FIO_IO_U_LIST_MAX_LEN];
606 extern struct thread_data *threads;
608 static inline void fio_ro_check(struct thread_data *td, struct io_u *io_u)
610 assert(!(io_u->ddir == DDIR_WRITE && !td_write(td)));
613 #define REAL_MAX_JOBS 2048
615 static inline enum error_type_bit td_error_type(enum fio_ddir ddir, int err)
618 return ERROR_TYPE_VERIFY_BIT;
619 if (ddir == DDIR_READ)
620 return ERROR_TYPE_READ_BIT;
621 return ERROR_TYPE_WRITE_BIT;
624 static int __NON_FATAL_ERR[] = {EIO, EILSEQ};
625 static inline int td_non_fatal_error(struct thread_data *td,
626 enum error_type_bit etype, int err)
629 if (!td->o.ignore_error[etype]) {
630 td->o.ignore_error[etype] = __NON_FATAL_ERR;
631 td->o.ignore_error_nr[etype] = sizeof(__NON_FATAL_ERR)
635 if (!(td->o.continue_on_error & (1 << etype)))
637 for (i = 0; i < td->o.ignore_error_nr[etype]; i++)
638 if (td->o.ignore_error[etype][i] == err)
643 static inline void update_error_count(struct thread_data *td, int err)
645 td->total_err_count++;
646 if (td->total_err_count == 1)
647 td->first_error = err;
650 static inline int should_fsync(struct thread_data *td)
652 if (td->last_was_sync)
656 if (td_write(td) || td_rw(td) || td->o.override_sync)
663 * Init/option functions
665 extern int __must_check parse_options(int, char **);
666 extern int parse_jobs_ini(char *, int, int);
667 extern int parse_cmd_line(int, char **);
668 extern int fio_backend(void);
669 extern void reset_fio_state(void);
670 extern void clear_io_state(struct thread_data *);
671 extern int fio_options_parse(struct thread_data *, char **, int);
672 extern void fio_keywords_init(void);
673 extern int fio_cmd_option_parse(struct thread_data *, const char *, char *);
674 extern int fio_cmd_ioengine_option_parse(struct thread_data *, const char *, char *);
675 extern void fio_fill_default_options(struct thread_data *);
676 extern int fio_show_option_help(const char *);
677 extern void fio_options_set_ioengine_opts(struct option *long_options, struct thread_data *td);
678 extern void fio_options_dup_and_init(struct option *);
679 extern void fio_options_mem_dupe(struct thread_data *);
680 extern void options_mem_dupe(void *data, struct fio_option *options);
681 extern void td_fill_rand_seeds(struct thread_data *);
682 extern void add_job_opts(const char **);
683 extern char *num2str(unsigned long, int, int, int);
684 extern int ioengine_load(struct thread_data *);
686 #define FIO_GETOPT_JOB 0x89000000
687 #define FIO_GETOPT_IOENGINE 0x98000000
688 #define FIO_NR_OPTIONS (FIO_MAX_OPTS + 128)
693 extern void print_thread_status(void);
694 extern void print_status_init(int);
697 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
698 * will never back again. It may cycle between running/verififying/fsyncing.
699 * Once the thread reaches TD_EXITED, it is just waiting for the core to
716 extern void td_set_runstate(struct thread_data *, int);
717 #define TERMINATE_ALL (-1)
718 extern void fio_terminate_threads(int);
723 extern int __must_check fio_pin_memory(void);
724 extern void fio_unpin_memory(void);
725 extern int __must_check allocate_io_mem(struct thread_data *);
726 extern void free_io_mem(struct thread_data *);
729 * Reset stats after ramp time completes
731 extern void reset_all_stats(struct thread_data *);
736 #ifdef FIO_HAVE_BLKTRACE
737 extern int is_blktrace(const char *);
738 extern int load_blktrace(struct thread_data *, const char *);
742 * Mark unused variables passed to ops functions as unused, to silence gcc
744 #define fio_unused __attribute((__unused__))
745 #define fio_init __attribute__((constructor))
746 #define fio_exit __attribute__((destructor))
748 #define for_each_td(td, i) \
749 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
750 #define for_each_file(td, f, i) \
751 if ((td)->files_index) \
752 for ((i) = 0, (f) = (td)->files[0]; \
753 (i) < (td)->o.nr_files && ((f) = (td)->files[i]) != NULL; \
756 #define fio_assert(td, cond) do { \
759 fprintf(stderr, "file:%s:%d, assert %s failed\n", __FILE__, __LINE__, #cond); \
760 td_set_runstate((td), TD_EXITED); \
761 (td)->error = EFAULT; \
766 static inline int fio_fill_issue_time(struct thread_data *td)
768 if (td->o.read_iolog_file ||
769 !td->o.disable_clat || !td->o.disable_slat || !td->o.disable_bw)
775 static inline int __should_check_rate(struct thread_data *td,
778 struct thread_options *o = &td->o;
781 * If some rate setting was given, we need to check it
783 if (o->rate[ddir] || o->ratemin[ddir] || o->rate_iops[ddir] ||
784 o->rate_iops_min[ddir])
790 static inline int should_check_rate(struct thread_data *td,
791 unsigned long *bytes_done)
795 if (bytes_done[DDIR_READ])
796 ret |= __should_check_rate(td, DDIR_READ);
797 if (bytes_done[DDIR_WRITE])
798 ret |= __should_check_rate(td, DDIR_WRITE);
799 if (bytes_done[DDIR_TRIM])
800 ret |= __should_check_rate(td, DDIR_TRIM);
805 static inline int is_power_of_2(unsigned int val)
807 return (val != 0 && ((val & (val - 1)) == 0));
811 * We currently only need to do locking if we have verifier threads
812 * accessing our internal structures too
814 static inline void td_io_u_lock(struct thread_data *td)
816 if (td->o.verify_async)
817 pthread_mutex_lock(&td->io_u_lock);
820 static inline void td_io_u_unlock(struct thread_data *td)
822 if (td->o.verify_async)
823 pthread_mutex_unlock(&td->io_u_lock);
826 static inline void td_io_u_free_notify(struct thread_data *td)
828 if (td->o.verify_async)
829 pthread_cond_signal(&td->free_cond);
832 extern const char *fio_get_arch_string(int);
833 extern const char *fio_get_os_string(int);
836 FIO_OUTPUT_TERSE = 0,
842 FIO_RAND_DIST_RANDOM = 0,
844 FIO_RAND_DIST_PARETO,
848 FIO_RAND_GEN_TAUSWORTHE = 0,