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 hugepage_size;
181 unsigned int rw_min_bs;
182 unsigned int thinktime;
183 unsigned int thinktime_spin;
184 unsigned int thinktime_blocks;
185 unsigned int fsync_blocks;
186 unsigned int fdatasync_blocks;
187 unsigned int barrier_blocks;
188 unsigned long long start_delay;
189 unsigned long long timeout;
190 unsigned long long ramp_time;
191 unsigned int overwrite;
192 unsigned int bw_avg_time;
193 unsigned int iops_avg_time;
195 unsigned long long zone_range;
196 unsigned long long zone_size;
197 unsigned long long zone_skip;
198 enum fio_memtype mem_type;
199 unsigned int mem_align;
201 unsigned int max_latency;
203 unsigned int stonewall;
204 unsigned int new_group;
205 unsigned int numjobs;
206 os_cpu_mask_t cpumask;
207 unsigned int cpumask_set;
208 os_cpu_mask_t verify_cpumask;
209 unsigned int verify_cpumask_set;
210 #ifdef FIO_HAVE_LIBNUMA
211 struct bitmask *numa_cpunodesmask;
212 unsigned int numa_cpumask_set;
213 unsigned short numa_mem_mode;
214 unsigned int numa_mem_prefer_node;
215 struct bitmask *numa_memnodesmask;
216 unsigned int numa_memmask_set;
219 unsigned int rwmixcycle;
220 unsigned int rwmix[2];
222 unsigned int file_service_type;
223 unsigned int group_reporting;
224 unsigned int fadvise_hint;
225 enum fio_fallocate_mode fallocate_mode;
226 unsigned int zero_buffers;
227 unsigned int refill_buffers;
228 unsigned int scramble_buffers;
229 unsigned int compress_percentage;
230 unsigned int compress_chunk;
231 unsigned int time_based;
232 unsigned int disable_lat;
233 unsigned int disable_clat;
234 unsigned int disable_slat;
235 unsigned int disable_bw;
236 unsigned int gtod_reduce;
237 unsigned int gtod_cpu;
238 unsigned int gtod_offload;
239 enum fio_cs clocksource;
240 unsigned int no_stall;
241 unsigned int trim_percentage;
242 unsigned int trim_batch;
243 unsigned int trim_zero;
244 unsigned long long trim_backlog;
245 unsigned int clat_percentiles;
246 unsigned int overwrite_plist;
247 fio_fp64_t percentile_list[FIO_IO_U_LIST_MAX_LEN];
249 char *read_iolog_file;
250 char *write_iolog_file;
254 char *replay_redirect;
257 * Pre-run and post-run shell
262 unsigned int rate[DDIR_RWDIR_CNT];
263 unsigned int ratemin[DDIR_RWDIR_CNT];
264 unsigned int ratecycle;
265 unsigned int rate_iops[DDIR_RWDIR_CNT];
266 unsigned int rate_iops_min[DDIR_RWDIR_CNT];
271 * CPU "io" cycle burner
273 unsigned int cpuload;
274 unsigned int cpucycle;
279 enum error_type continue_on_error;
282 * Benchmark profile type
287 * blkio cgroup support
290 unsigned int cgroup_weight;
291 unsigned int cgroup_nodelete;
299 unsigned int flow_sleep;
301 unsigned long long offset_increment;
303 unsigned int sync_file_range;
307 * This describes a single thread/process executing a fio job.
310 struct thread_options o;
312 char verror[FIO_VERROR_SIZE];
316 struct thread_stat ts;
318 struct io_log *slat_log;
319 struct io_log *clat_log;
320 struct io_log *lat_log;
321 struct io_log *bw_log;
322 struct io_log *iops_log;
324 uint64_t stat_io_bytes[DDIR_RWDIR_CNT];
325 struct timeval bw_sample_time;
327 uint64_t stat_io_blocks[DDIR_RWDIR_CNT];
328 struct timeval iops_sample_time;
330 struct rusage ru_start;
331 struct rusage ru_end;
333 struct fio_file **files;
334 unsigned int files_size;
335 unsigned int files_index;
336 unsigned int nr_open_files;
337 unsigned int nr_done_files;
338 unsigned int nr_normal_files;
340 unsigned int next_file;
341 os_random_state_t next_file_state;
342 struct frand_state __next_file_state;
349 size_t orig_buffer_size;
350 volatile int terminate;
351 volatile int runstate;
353 unsigned int ioprio_set;
354 unsigned int last_was_sync;
355 enum fio_ddir last_ddir;
365 unsigned long rand_seeds[8];
368 os_random_state_t bsrange_state;
369 struct frand_state __bsrange_state;
372 os_random_state_t verify_state;
373 struct frand_state __verify_state;
376 os_random_state_t trim_state;
377 struct frand_state __trim_state;
380 struct frand_state buf_state;
382 unsigned int verify_batch;
383 unsigned int trim_batch;
388 * IO engine hooks, contains everything needed to submit an io_u
389 * to any of the available IO engines.
391 struct ioengine_ops *io_ops;
394 * Queue depth of io_u's that fio MIGHT do
396 unsigned int cur_depth;
399 * io_u's about to be committed
401 unsigned int io_u_queued;
404 * io_u's submitted but not completed yet
406 unsigned int io_u_in_flight;
409 * List of free and busy io_u's
411 struct flist_head io_u_freelist;
412 struct flist_head io_u_busylist;
413 struct flist_head io_u_requeues;
414 pthread_mutex_t io_u_lock;
415 pthread_cond_t free_cond;
418 * async verify offload
420 struct flist_head verify_list;
421 pthread_t *verify_threads;
422 unsigned int nr_verify_threads;
423 pthread_cond_t verify_cond;
424 int verify_thread_exit;
429 unsigned long long rate_bps[DDIR_RWDIR_CNT];
430 long rate_pending_usleep[DDIR_RWDIR_CNT];
431 unsigned long rate_bytes[DDIR_RWDIR_CNT];
432 unsigned long rate_blocks[DDIR_RWDIR_CNT];
433 struct timeval lastrate[DDIR_RWDIR_CNT];
435 unsigned long long total_io_size;
436 unsigned long long fill_device_size;
438 unsigned long io_issues[DDIR_RWDIR_CNT];
439 unsigned long long io_blocks[DDIR_RWDIR_CNT];
440 unsigned long long this_io_blocks[DDIR_RWDIR_CNT];
441 unsigned long long io_bytes[DDIR_RWDIR_CNT];
442 unsigned long long io_skip_bytes;
443 unsigned long long this_io_bytes[DDIR_RWDIR_CNT];
444 unsigned long long zone_bytes;
445 struct fio_mutex *mutex;
448 * State for random io, a bitmap of blocks done vs not done
451 os_random_state_t random_state;
452 struct frand_state __random_state;
455 struct timeval start; /* start of this loop */
456 struct timeval epoch; /* time job was started */
457 struct timeval last_issue;
458 struct timeval tv_cache;
459 unsigned int tv_cache_nr;
460 unsigned int tv_cache_mask;
461 unsigned int ramp_time_over;
464 * read/write mixed workload state
467 os_random_state_t rwmix_state;
468 struct frand_state __rwmix_state;
470 unsigned long rwmix_issues;
471 enum fio_ddir rwmix_ddir;
472 unsigned int ddir_seq_nr;
475 * IO history logs for verification. We use a tree for sorting,
476 * if we are overwriting. Otherwise just use a fifo.
478 struct rb_root io_hist_tree;
479 struct flist_head io_hist_list;
480 unsigned long io_hist_len;
485 struct flist_head io_log_list;
488 * For tracking/handling discards
490 struct flist_head trim_list;
491 unsigned long trim_entries;
494 * for fileservice, how often to switch to a new file
496 unsigned int file_service_nr;
497 unsigned int file_service_left;
498 struct fio_file *file_service_file;
500 unsigned int sync_file_range_nr;
503 * For generating file sizes
506 os_random_state_t file_size_state;
507 struct frand_state __file_size_state;
513 unsigned int total_err_count;
516 struct fio_flow *flow;
519 * Can be overloaded by profiles
521 struct prof_io_ops prof_io_ops;
526 * when should interactive ETA output be generated
534 #define __td_verror(td, err, msg, func) \
540 if (!(td)->first_error) \
541 snprintf(td->verror, sizeof(td->verror) - 1, "file:%s:%d, func=%s, error=%s", __FILE__, __LINE__, (func), (msg)); \
545 #define td_clear_error(td) \
547 #define td_verror(td, err, func) \
548 __td_verror((td), (err), strerror((err)), (func))
549 #define td_vmsg(td, err, msg, func) \
550 __td_verror((td), (err), (msg), (func))
552 #define __fio_stringify_1(x) #x
553 #define __fio_stringify(x) __fio_stringify_1(x)
555 extern int exitall_on_terminate;
556 extern unsigned int thread_number;
557 extern unsigned int nr_process, nr_thread;
560 extern int output_format;
561 extern int temp_stall_ts;
562 extern unsigned long long mlock_size;
563 extern uintptr_t page_mask, page_size;
564 extern int read_only;
565 extern int eta_print;
566 extern unsigned long done_secs;
567 extern char *job_section;
568 extern int fio_gtod_offload;
569 extern int fio_gtod_cpu;
570 extern enum fio_cs fio_clock_source;
571 extern int warnings_fatal;
572 extern int terse_version;
573 extern int is_backend;
574 extern int nr_clients;
575 extern int log_syslog;
576 extern const char fio_version_string[];
577 extern const fio_fp64_t def_percentile_list[FIO_IO_U_LIST_MAX_LEN];
579 extern struct thread_data *threads;
581 static inline void fio_ro_check(struct thread_data *td, struct io_u *io_u)
583 assert(!(io_u->ddir == DDIR_WRITE && !td_write(td)));
586 #define BLOCKS_PER_MAP (8 * sizeof(unsigned long))
587 #define TO_MAP_BLOCK(f, b) (b)
588 #define RAND_MAP_IDX(f, b) (TO_MAP_BLOCK(f, b) / BLOCKS_PER_MAP)
589 #define RAND_MAP_BIT(f, b) (TO_MAP_BLOCK(f, b) & (BLOCKS_PER_MAP - 1))
591 #define REAL_MAX_JOBS 2048
593 static inline enum error_type td_error_type(enum fio_ddir ddir, int err)
596 return ERROR_TYPE_VERIFY_BIT;
597 if (ddir == DDIR_READ)
598 return ERROR_TYPE_READ_BIT;
599 return ERROR_TYPE_WRITE_BIT;
602 static int __NON_FATAL_ERR[] = {EIO, EILSEQ};
603 static inline int td_non_fatal_error(struct thread_data *td,
604 enum error_type_bit etype, int err)
607 if (!td->o.ignore_error[etype]) {
608 td->o.ignore_error[etype] = __NON_FATAL_ERR;
609 td->o.ignore_error_nr[etype] = sizeof(__NON_FATAL_ERR)
613 if (!(td->o.continue_on_error & (1 << etype)))
615 for (i = 0; i < td->o.ignore_error_nr[etype]; i++)
616 if (td->o.ignore_error[etype][i] == err)
621 static inline void update_error_count(struct thread_data *td, int err)
623 td->total_err_count++;
624 if (td->total_err_count == 1)
625 td->first_error = err;
628 static inline int should_fsync(struct thread_data *td)
630 if (td->last_was_sync)
634 if (td_write(td) || td_rw(td) || td->o.override_sync)
641 * Init/option functions
643 extern int __must_check parse_options(int, char **);
644 extern int parse_jobs_ini(char *, int, int);
645 extern int parse_cmd_line(int, char **);
646 extern int fio_backend(void);
647 extern void reset_fio_state(void);
648 extern void clear_io_state(struct thread_data *);
649 extern int fio_options_parse(struct thread_data *, char **, int);
650 extern void fio_keywords_init(void);
651 extern int fio_cmd_option_parse(struct thread_data *, const char *, char *);
652 extern int fio_cmd_ioengine_option_parse(struct thread_data *, const char *, char *);
653 extern void fio_fill_default_options(struct thread_data *);
654 extern int fio_show_option_help(const char *);
655 extern void fio_options_set_ioengine_opts(struct option *long_options, struct thread_data *td);
656 extern void fio_options_dup_and_init(struct option *);
657 extern void fio_options_mem_dupe(struct thread_data *);
658 extern void options_mem_dupe(void *data, struct fio_option *options);
659 extern void td_fill_rand_seeds(struct thread_data *);
660 extern void add_job_opts(const char **);
661 extern char *num2str(unsigned long, int, int, int);
662 extern int ioengine_load(struct thread_data *);
664 #define FIO_GETOPT_JOB 0x89000000
665 #define FIO_GETOPT_IOENGINE 0x98000000
666 #define FIO_NR_OPTIONS (FIO_MAX_OPTS + 128)
671 extern void print_thread_status(void);
672 extern void print_status_init(int);
675 * Thread life cycle. Once a thread has a runstate beyond TD_INITIALIZED, it
676 * will never back again. It may cycle between running/verififying/fsyncing.
677 * Once the thread reaches TD_EXITED, it is just waiting for the core to
693 extern void td_set_runstate(struct thread_data *, int);
694 #define TERMINATE_ALL (-1)
695 extern void fio_terminate_threads(int);
700 extern int __must_check fio_pin_memory(void);
701 extern void fio_unpin_memory(void);
702 extern int __must_check allocate_io_mem(struct thread_data *);
703 extern void free_io_mem(struct thread_data *);
706 * Reset stats after ramp time completes
708 extern void reset_all_stats(struct thread_data *);
713 #ifdef FIO_HAVE_BLKTRACE
714 extern int is_blktrace(const char *);
715 extern int load_blktrace(struct thread_data *, const char *);
719 * Mark unused variables passed to ops functions as unused, to silence gcc
721 #define fio_unused __attribute((__unused__))
722 #define fio_init __attribute__((constructor))
723 #define fio_exit __attribute__((destructor))
725 #define for_each_td(td, i) \
726 for ((i) = 0, (td) = &threads[0]; (i) < (int) thread_number; (i)++, (td)++)
727 #define for_each_file(td, f, i) \
728 if ((td)->files_index) \
729 for ((i) = 0, (f) = (td)->files[0]; \
730 (i) < (td)->o.nr_files && ((f) = (td)->files[i]) != NULL; \
733 #define fio_assert(td, cond) do { \
736 fprintf(stderr, "file:%s:%d, assert %s failed\n", __FILE__, __LINE__, #cond); \
737 td_set_runstate((td), TD_EXITED); \
738 (td)->error = EFAULT; \
743 static inline int fio_fill_issue_time(struct thread_data *td)
745 if (td->o.read_iolog_file ||
746 !td->o.disable_clat || !td->o.disable_slat || !td->o.disable_bw)
752 static inline int __should_check_rate(struct thread_data *td,
755 struct thread_options *o = &td->o;
758 * If some rate setting was given, we need to check it
760 if (o->rate[ddir] || o->ratemin[ddir] || o->rate_iops[ddir] ||
761 o->rate_iops_min[ddir])
767 static inline int should_check_rate(struct thread_data *td,
768 unsigned long *bytes_done)
772 if (bytes_done[DDIR_READ])
773 ret |= __should_check_rate(td, DDIR_READ);
774 if (bytes_done[DDIR_WRITE])
775 ret |= __should_check_rate(td, DDIR_WRITE);
776 if (bytes_done[DDIR_TRIM])
777 ret |= __should_check_rate(td, DDIR_TRIM);
782 static inline int is_power_of_2(unsigned int val)
784 return (val != 0 && ((val & (val - 1)) == 0));
788 * We currently only need to do locking if we have verifier threads
789 * accessing our internal structures too
791 static inline void td_io_u_lock(struct thread_data *td)
793 if (td->o.verify_async)
794 pthread_mutex_lock(&td->io_u_lock);
797 static inline void td_io_u_unlock(struct thread_data *td)
799 if (td->o.verify_async)
800 pthread_mutex_unlock(&td->io_u_lock);
803 static inline void td_io_u_free_notify(struct thread_data *td)
805 if (td->o.verify_async)
806 pthread_cond_signal(&td->free_cond);
809 extern const char *fio_get_arch_string(int);
810 extern const char *fio_get_os_string(int);
813 FIO_OUTPUT_TERSE = 0,