8 #include <sys/resource.h>
19 #include "compiler/compiler.h"
23 #include "arch/arch.h"
37 #include "lib/getopt.h"
47 #ifdef FIO_HAVE_SOLARISAIO
48 #include <sys/asynch.h>
51 #ifdef FIO_HAVE_LIBNUMA
52 #include <linux/mempolicy.h>
56 * "local" is pseudo-policy
58 #define MPOL_LOCAL MPOL_MAX
62 * What type of allocation to use for io buffers
65 MEM_MALLOC = 0, /* ordinary malloc */
66 MEM_SHM, /* use shared memory segments */
67 MEM_SHMHUGE, /* use shared memory segments with huge pages */
68 MEM_MMAP, /* use anonynomous mmap */
69 MEM_MMAPHUGE, /* memory mapped huge file */
73 * offset generator types
81 * What type of errors to continue on when continue_on_error is used
84 ERROR_TYPE_READ_BIT = 0,
85 ERROR_TYPE_WRITE_BIT = 1,
86 ERROR_TYPE_VERIFY_BIT = 2,
92 ERROR_TYPE_READ = 1 << ERROR_TYPE_READ_BIT,
93 ERROR_TYPE_WRITE = 1 << ERROR_TYPE_WRITE_BIT,
94 ERROR_TYPE_VERIFY = 1 << ERROR_TYPE_VERIFY_BIT,
95 ERROR_TYPE_ANY = 0xffff,
103 struct thread_options {
111 enum td_ddir td_ddir;
113 unsigned int kb_base;
114 unsigned int ddir_seq_nr;
116 unsigned int iodepth;
117 unsigned int iodepth_low;
118 unsigned int iodepth_batch;
119 unsigned int iodepth_batch_complete;
121 unsigned long long size;
122 unsigned int size_percent;
123 unsigned int fill_device;
124 unsigned long long file_size_low;
125 unsigned long long file_size_high;
126 unsigned long long start_offset;
128 unsigned int bs[DDIR_RWDIR_CNT];
129 unsigned int ba[DDIR_RWDIR_CNT];
130 unsigned int min_bs[DDIR_RWDIR_CNT];
131 unsigned int max_bs[DDIR_RWDIR_CNT];
132 struct bssplit *bssplit[DDIR_RWDIR_CNT];
133 unsigned int bssplit_nr[DDIR_RWDIR_CNT];
135 int *ignore_error[ERROR_TYPE_CNT];
136 unsigned int ignore_error_nr[ERROR_TYPE_CNT];
137 unsigned int error_dump;
139 unsigned int nr_files;
140 unsigned int open_files;
141 enum file_lock_mode file_lock_mode;
142 unsigned int lockfile_batch;
144 unsigned int odirect;
145 unsigned int invalidate_cache;
146 unsigned int create_serialize;
147 unsigned int create_fsync;
148 unsigned int create_on_open;
149 unsigned int create_only;
150 unsigned int end_fsync;
151 unsigned int pre_read;
152 unsigned int sync_io;
154 unsigned int do_verify;
155 unsigned int verifysort;
156 unsigned int verify_interval;
157 unsigned int verify_offset;
158 char verify_pattern[MAX_PATTERN_SIZE];
159 unsigned int verify_pattern_bytes;
160 unsigned int verify_fatal;
161 unsigned int verify_dump;
162 unsigned int verify_async;
163 unsigned long long verify_backlog;
164 unsigned int verify_batch;
165 unsigned int use_thread;
167 unsigned int do_disk_util;
168 unsigned int override_sync;
169 unsigned int rand_repeatable;
170 unsigned int use_os_rand;
171 unsigned int write_lat_log;
172 unsigned int write_bw_log;
173 unsigned int write_iops_log;
174 unsigned int log_avg_msec;
175 unsigned int norandommap;
176 unsigned int softrandommap;
177 unsigned int bs_unaligned;
178 unsigned int fsync_on_close;
180 unsigned int random_distribution;
184 unsigned int random_generator;
186 unsigned int hugepage_size;
187 unsigned int rw_min_bs;
188 unsigned int thinktime;
189 unsigned int thinktime_spin;
190 unsigned int thinktime_blocks;
191 unsigned int fsync_blocks;
192 unsigned int fdatasync_blocks;
193 unsigned int barrier_blocks;
194 unsigned long long start_delay;
195 unsigned long long timeout;
196 unsigned long long ramp_time;
197 unsigned int overwrite;
198 unsigned int bw_avg_time;
199 unsigned int iops_avg_time;
201 unsigned long long zone_range;
202 unsigned long long zone_size;
203 unsigned long long zone_skip;
204 enum fio_memtype mem_type;
205 unsigned int mem_align;
207 unsigned int max_latency;
209 unsigned int stonewall;
210 unsigned int new_group;
211 unsigned int numjobs;
212 os_cpu_mask_t cpumask;
213 unsigned int cpumask_set;
214 os_cpu_mask_t verify_cpumask;
215 unsigned int verify_cpumask_set;
216 #ifdef FIO_HAVE_LIBNUMA
217 struct bitmask *numa_cpunodesmask;
218 unsigned int numa_cpumask_set;
219 unsigned short numa_mem_mode;
220 unsigned int numa_mem_prefer_node;
221 struct bitmask *numa_memnodesmask;
222 unsigned int numa_memmask_set;
225 unsigned int rwmixcycle;
226 unsigned int rwmix[2];
228 unsigned int file_service_type;
229 unsigned int group_reporting;
230 unsigned int fadvise_hint;
231 enum fio_fallocate_mode fallocate_mode;
232 unsigned int zero_buffers;
233 unsigned int refill_buffers;
234 unsigned int scramble_buffers;
235 unsigned int compress_percentage;
236 unsigned int compress_chunk;
237 unsigned int time_based;
238 unsigned int disable_lat;
239 unsigned int disable_clat;
240 unsigned int disable_slat;
241 unsigned int disable_bw;
242 unsigned int gtod_reduce;
243 unsigned int gtod_cpu;
244 unsigned int gtod_offload;
245 enum fio_cs clocksource;
246 unsigned int no_stall;
247 unsigned int trim_percentage;
248 unsigned int trim_batch;
249 unsigned int trim_zero;
250 unsigned long long trim_backlog;
251 unsigned int clat_percentiles;
252 unsigned int overwrite_plist;
253 fio_fp64_t percentile_list[FIO_IO_U_LIST_MAX_LEN];
255 char *read_iolog_file;
256 char *write_iolog_file;
260 char *replay_redirect;
263 * Pre-run and post-run shell
268 unsigned int rate[DDIR_RWDIR_CNT];
269 unsigned int ratemin[DDIR_RWDIR_CNT];
270 unsigned int ratecycle;
271 unsigned int rate_iops[DDIR_RWDIR_CNT];
272 unsigned int rate_iops_min[DDIR_RWDIR_CNT];
277 * CPU "io" cycle burner
279 unsigned int cpuload;
280 unsigned int cpucycle;
285 enum error_type continue_on_error;
288 * Benchmark profile type
293 * blkio cgroup support
296 unsigned int cgroup_weight;
297 unsigned int cgroup_nodelete;
305 unsigned int flow_sleep;
307 unsigned long long offset_increment;
309 unsigned int sync_file_range;
313 TD_F_VER_BACKLOG = 1,
314 TD_F_TRIM_BACKLOG = 2,
316 TD_F_REFILL_BUFFERS = 8,
317 TD_F_SCRAMBLE_BUFFERS = 16,
319 TD_F_PROFILE_OPS = 64,
323 * This describes a single thread/process executing a fio job.
326 struct thread_options o;
329 char verror[FIO_VERROR_SIZE];
333 struct thread_stat ts;
335 struct io_log *slat_log;
336 struct io_log *clat_log;
337 struct io_log *lat_log;
338 struct io_log *bw_log;
339 struct io_log *iops_log;
341 uint64_t stat_io_bytes[DDIR_RWDIR_CNT];
342 struct timeval bw_sample_time;
344 uint64_t stat_io_blocks[DDIR_RWDIR_CNT];
345 struct timeval iops_sample_time;
347 struct rusage ru_start;
348 struct rusage ru_end;
350 struct fio_file **files;
351 unsigned int files_size;
352 unsigned int files_index;
353 unsigned int nr_open_files;
354 unsigned int nr_done_files;
355 unsigned int nr_normal_files;
357 unsigned int next_file;
358 os_random_state_t next_file_state;
359 struct frand_state __next_file_state;
366 size_t orig_buffer_size;
367 volatile int terminate;
368 volatile int runstate;
370 unsigned int ioprio_set;
371 unsigned int last_was_sync;
372 enum fio_ddir last_ddir;
382 unsigned long rand_seeds[8];
385 os_random_state_t bsrange_state;
386 struct frand_state __bsrange_state;
389 os_random_state_t verify_state;
390 struct frand_state __verify_state;
393 os_random_state_t trim_state;
394 struct frand_state __trim_state;
397 struct frand_state buf_state;
399 unsigned int verify_batch;
400 unsigned int trim_batch;
405 * IO engine hooks, contains everything needed to submit an io_u
406 * to any of the available IO engines.
408 struct ioengine_ops *io_ops;
411 * Queue depth of io_u's that fio MIGHT do
413 unsigned int cur_depth;
416 * io_u's about to be committed
418 unsigned int io_u_queued;
421 * io_u's submitted but not completed yet
423 unsigned int io_u_in_flight;
426 * List of free and busy io_u's
428 struct flist_head io_u_freelist;
429 struct flist_head io_u_busylist;
430 struct flist_head io_u_requeues;
431 pthread_mutex_t io_u_lock;
432 pthread_cond_t free_cond;
435 * async verify offload
437 struct flist_head verify_list;
438 pthread_t *verify_threads;
439 unsigned int nr_verify_threads;
440 pthread_cond_t verify_cond;
441 int verify_thread_exit;
446 unsigned long long rate_bps[DDIR_RWDIR_CNT];
447 long rate_pending_usleep[DDIR_RWDIR_CNT];
448 unsigned long rate_bytes[DDIR_RWDIR_CNT];
449 unsigned long rate_blocks[DDIR_RWDIR_CNT];
450 struct timeval lastrate[DDIR_RWDIR_CNT];
452 unsigned long long total_io_size;
453 unsigned long long fill_device_size;
455 unsigned long io_issues[DDIR_RWDIR_CNT];
456 unsigned long long io_blocks[DDIR_RWDIR_CNT];
457 unsigned long long this_io_blocks[DDIR_RWDIR_CNT];
458 unsigned long long io_bytes[DDIR_RWDIR_CNT];
459 unsigned long long io_skip_bytes;
460 unsigned long long this_io_bytes[DDIR_RWDIR_CNT];
461 unsigned long long zone_bytes;
462 struct fio_mutex *mutex;
465 * State for random io, a bitmap of blocks done vs not done
468 os_random_state_t random_state;
469 struct frand_state __random_state;
472 struct timeval start; /* start of this loop */
473 struct timeval epoch; /* time job was started */
474 struct timeval last_issue;
475 struct timeval tv_cache;
476 unsigned int tv_cache_nr;
477 unsigned int tv_cache_mask;
478 unsigned int ramp_time_over;
481 * read/write mixed workload state
484 os_random_state_t rwmix_state;
485 struct frand_state __rwmix_state;
487 unsigned long rwmix_issues;
488 enum fio_ddir rwmix_ddir;
489 unsigned int ddir_seq_nr;
492 * IO history logs for verification. We use a tree for sorting,
493 * if we are overwriting. Otherwise just use a fifo.
495 struct rb_root io_hist_tree;
496 struct flist_head io_hist_list;
497 unsigned long io_hist_len;
502 struct flist_head io_log_list;
505 * For tracking/handling discards
507 struct flist_head trim_list;
508 unsigned long trim_entries;
511 * for fileservice, how often to switch to a new file
513 unsigned int file_service_nr;
514 unsigned int file_service_left;
515 struct fio_file *file_service_file;
517 unsigned int sync_file_range_nr;
520 * For generating file sizes
523 os_random_state_t file_size_state;
524 struct frand_state __file_size_state;
530 unsigned int total_err_count;
533 struct fio_flow *flow;
536 * Can be overloaded by profiles
538 struct prof_io_ops prof_io_ops;
543 * when should interactive ETA output be generated
551 #define __td_verror(td, err, msg, func) \
557 if (!(td)->first_error) \
558 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, func=%s, error=%s", __FILE__, __LINE__, (func), (msg)); \
562 #define td_clear_error(td) \
564 #define td_verror(td, err, func) \
565 __td_verror((td), (err), strerror((err)), (func))
566 #define td_vmsg(td, err, msg, func) \
567 __td_verror((td), (err), (msg), (func))
569 #define __fio_stringify_1(x) #x
570 #define __fio_stringify(x) __fio_stringify_1(x)
572 extern int exitall_on_terminate;
573 extern unsigned int thread_number;
574 extern unsigned int stat_number;
575 extern unsigned int nr_process, nr_thread;
578 extern int output_format;
579 extern int temp_stall_ts;
580 extern unsigned long long mlock_size;
581 extern uintptr_t page_mask, page_size;
582 extern int read_only;
583 extern int eta_print;
584 extern unsigned long done_secs;
585 extern char *job_section;
586 extern int fio_gtod_offload;
587 extern int fio_gtod_cpu;
588 extern enum fio_cs fio_clock_source;
589 extern int warnings_fatal;
590 extern int terse_version;
591 extern int is_backend;
592 extern int nr_clients;
593 extern int log_syslog;
594 extern const char fio_version_string[];
595 extern const fio_fp64_t def_percentile_list[FIO_IO_U_LIST_MAX_LEN];
597 extern struct thread_data *threads;
599 static inline void fio_ro_check(struct thread_data *td, struct io_u *io_u)
601 assert(!(io_u->ddir == DDIR_WRITE && !td_write(td)));
604 #define REAL_MAX_JOBS 2048
606 static inline enum error_type_bit td_error_type(enum fio_ddir ddir, int err)
609 return ERROR_TYPE_VERIFY_BIT;
610 if (ddir == DDIR_READ)
611 return ERROR_TYPE_READ_BIT;
612 return ERROR_TYPE_WRITE_BIT;
615 static int __NON_FATAL_ERR[] = {EIO, EILSEQ};
616 static inline int td_non_fatal_error(struct thread_data *td,
617 enum error_type_bit etype, int err)
620 if (!td->o.ignore_error[etype]) {
621 td->o.ignore_error[etype] = __NON_FATAL_ERR;
622 td->o.ignore_error_nr[etype] = sizeof(__NON_FATAL_ERR)
626 if (!(td->o.continue_on_error & (1 << etype)))
628 for (i = 0; i < td->o.ignore_error_nr[etype]; i++)
629 if (td->o.ignore_error[etype][i] == err)
634 static inline void update_error_count(struct thread_data *td, int err)
636 td->total_err_count++;
637 if (td->total_err_count == 1)
638 td->first_error = err;
641 static inline int should_fsync(struct thread_data *td)
643 if (td->last_was_sync)
647 if (td_write(td) || td_rw(td) || td->o.override_sync)
654 * Init/option functions
656 extern int __must_check parse_options(int, char **);
657 extern int parse_jobs_ini(char *, int, int);
658 extern int parse_cmd_line(int, char **);
659 extern int fio_backend(void);
660 extern void reset_fio_state(void);
661 extern void clear_io_state(struct thread_data *);
662 extern int fio_options_parse(struct thread_data *, char **, int);
663 extern void fio_keywords_init(void);
664 extern int fio_cmd_option_parse(struct thread_data *, const char *, char *);
665 extern int fio_cmd_ioengine_option_parse(struct thread_data *, const char *, char *);
666 extern void fio_fill_default_options(struct thread_data *);
667 extern int fio_show_option_help(const char *);
668 extern void fio_options_set_ioengine_opts(struct option *long_options, struct thread_data *td);
669 extern void fio_options_dup_and_init(struct option *);
670 extern void fio_options_mem_dupe(struct thread_data *);
671 extern void options_mem_dupe(void *data, struct fio_option *options);
672 extern void td_fill_rand_seeds(struct thread_data *);
673 extern void add_job_opts(const char **);
674 extern char *num2str(unsigned long, int, int, int);
675 extern int ioengine_load(struct thread_data *);
677 #define FIO_GETOPT_JOB 0x89000000
678 #define FIO_GETOPT_IOENGINE 0x98000000
679 #define FIO_NR_OPTIONS (FIO_MAX_OPTS + 128)
684 extern void print_thread_status(void);
685 extern void print_status_init(int);
688 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
689 * will never back again. It may cycle between running/verififying/fsyncing.
690 * Once the thread reaches TD_EXITED, it is just waiting for the core to
707 extern void td_set_runstate(struct thread_data *, int);
708 #define TERMINATE_ALL (-1)
709 extern void fio_terminate_threads(int);
714 extern int __must_check fio_pin_memory(void);
715 extern void fio_unpin_memory(void);
716 extern int __must_check allocate_io_mem(struct thread_data *);
717 extern void free_io_mem(struct thread_data *);
720 * Reset stats after ramp time completes
722 extern void reset_all_stats(struct thread_data *);
727 #ifdef FIO_HAVE_BLKTRACE
728 extern int is_blktrace(const char *);
729 extern int load_blktrace(struct thread_data *, const char *);
733 * Mark unused variables passed to ops functions as unused, to silence gcc
735 #define fio_unused __attribute((__unused__))
736 #define fio_init __attribute__((constructor))
737 #define fio_exit __attribute__((destructor))
739 #define for_each_td(td, i) \
740 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
741 #define for_each_file(td, f, i) \
742 if ((td)->files_index) \
743 for ((i) = 0, (f) = (td)->files[0]; \
744 (i) < (td)->o.nr_files && ((f) = (td)->files[i]) != NULL; \
747 #define fio_assert(td, cond) do { \
750 fprintf(stderr, "file:%s:%d, assert %s failed\n", __FILE__, __LINE__, #cond); \
751 td_set_runstate((td), TD_EXITED); \
752 (td)->error = EFAULT; \
757 static inline int fio_fill_issue_time(struct thread_data *td)
759 if (td->o.read_iolog_file ||
760 !td->o.disable_clat || !td->o.disable_slat || !td->o.disable_bw)
766 static inline int __should_check_rate(struct thread_data *td,
769 struct thread_options *o = &td->o;
772 * If some rate setting was given, we need to check it
774 if (o->rate[ddir] || o->ratemin[ddir] || o->rate_iops[ddir] ||
775 o->rate_iops_min[ddir])
781 static inline int should_check_rate(struct thread_data *td,
782 unsigned long *bytes_done)
786 if (bytes_done[DDIR_READ])
787 ret |= __should_check_rate(td, DDIR_READ);
788 if (bytes_done[DDIR_WRITE])
789 ret |= __should_check_rate(td, DDIR_WRITE);
790 if (bytes_done[DDIR_TRIM])
791 ret |= __should_check_rate(td, DDIR_TRIM);
796 static inline int is_power_of_2(unsigned int val)
798 return (val != 0 && ((val & (val - 1)) == 0));
802 * We currently only need to do locking if we have verifier threads
803 * accessing our internal structures too
805 static inline void td_io_u_lock(struct thread_data *td)
807 if (td->o.verify_async)
808 pthread_mutex_lock(&td->io_u_lock);
811 static inline void td_io_u_unlock(struct thread_data *td)
813 if (td->o.verify_async)
814 pthread_mutex_unlock(&td->io_u_lock);
817 static inline void td_io_u_free_notify(struct thread_data *td)
819 if (td->o.verify_async)
820 pthread_cond_signal(&td->free_cond);
823 extern const char *fio_get_arch_string(int);
824 extern const char *fio_get_os_string(int);
827 FIO_OUTPUT_TERSE = 0,
833 FIO_RAND_DIST_RANDOM = 0,
835 FIO_RAND_DIST_PARETO,
839 FIO_RAND_GEN_TAUSWORTHE = 0,