8 #include <sys/resource.h>
17 #include "compiler/compiler.h"
18 #include "thread_options.h"
21 #include "arch/arch.h"
27 #include "ioengines.h"
35 #include "oslib/getopt.h"
37 #include "lib/rbtree.h"
38 #include "lib/num2str.h"
39 #include "lib/memalign.h"
46 #include "io_u_queue.h"
47 #include "workqueue.h"
48 #include "steadystate.h"
49 #include "lib/nowarn_snprintf.h"
52 #ifdef CONFIG_SOLARISAIO
53 #include <sys/asynch.h>
57 #include <linux/mempolicy.h>
61 * "local" is pseudo-policy
74 #define MAX_TRIM_RANGE 256
77 * Range for trim command
80 unsigned long long start;
81 unsigned long long len;
85 * offset generator types
93 __TD_F_VER_BACKLOG = 0,
96 __TD_F_REFILL_BUFFERS,
97 __TD_F_SCRAMBLE_BUFFERS,
111 __TD_F_LAST, /* not a real bit, keep last */
115 TD_F_VER_BACKLOG = 1U << __TD_F_VER_BACKLOG,
116 TD_F_TRIM_BACKLOG = 1U << __TD_F_TRIM_BACKLOG,
117 TD_F_READ_IOLOG = 1U << __TD_F_READ_IOLOG,
118 TD_F_REFILL_BUFFERS = 1U << __TD_F_REFILL_BUFFERS,
119 TD_F_SCRAMBLE_BUFFERS = 1U << __TD_F_SCRAMBLE_BUFFERS,
120 TD_F_DO_VERIFY = 1U << __TD_F_DO_VERIFY,
121 TD_F_PROFILE_OPS = 1U << __TD_F_PROFILE_OPS,
122 TD_F_COMPRESS = 1U << __TD_F_COMPRESS,
123 TD_F_COMPRESS_LOG = 1U << __TD_F_COMPRESS_LOG,
124 TD_F_VSTATE_SAVED = 1U << __TD_F_VSTATE_SAVED,
125 TD_F_NEED_LOCK = 1U << __TD_F_NEED_LOCK,
126 TD_F_CHILD = 1U << __TD_F_CHILD,
127 TD_F_NO_PROGRESS = 1U << __TD_F_NO_PROGRESS,
128 TD_F_REGROW_LOGS = 1U << __TD_F_REGROW_LOGS,
129 TD_F_MMAP_KEEP = 1U << __TD_F_MMAP_KEEP,
130 TD_F_DIRS_CREATED = 1U << __TD_F_DIRS_CREATED,
131 TD_F_CHECK_RATE = 1U << __TD_F_CHECK_RATE,
132 TD_F_SYNCS = 1U << __TD_F_SYNCS,
143 FIO_RAND_FILE_SIZE_OFF,
146 FIO_RAND_SEQ_RAND_READ_OFF,
147 FIO_RAND_SEQ_RAND_WRITE_OFF,
148 FIO_RAND_SEQ_RAND_TRIM_OFF,
149 FIO_RAND_START_DELAY,
151 FIO_RAND_POISSON_OFF,
153 FIO_RAND_POISSON2_OFF,
154 FIO_RAND_POISSON3_OFF,
156 FIO_RAND_DEDUPE_WORKING_SET_IX,
165 RATE_PROCESS_LINEAR = 0,
166 RATE_PROCESS_POISSON = 1,
168 THINKTIME_BLOCKS_TYPE_COMPLETE = 0,
169 THINKTIME_BLOCKS_TYPE_ISSUE = 1,
181 * Per-thread/process specific data. Only used for the network client
184 void sk_out_assign(struct sk_out *);
185 void sk_out_drop(void);
187 struct zone_split_index {
189 uint8_t size_perc_prev;
195 * This describes a single thread/process executing a fio job.
198 struct flist_head opt_list;
199 unsigned long long flags;
200 struct thread_options o;
203 unsigned int thread_number;
204 unsigned int subjob_number;
205 unsigned int groupid;
206 struct thread_stat ts __attribute__ ((aligned(8)));
210 struct io_log *slat_log;
211 struct io_log *clat_log;
212 struct io_log *clat_hist_log;
213 struct io_log *lat_log;
214 struct io_log *bw_log;
215 struct io_log *iops_log;
217 struct workqueue log_compress_wq;
219 struct thread_data *parent;
221 uint64_t stat_io_bytes[DDIR_RWDIR_CNT];
222 struct timespec bw_sample_time;
224 uint64_t stat_io_blocks[DDIR_RWDIR_CNT];
225 struct timespec iops_sample_time;
227 volatile int update_rusage;
228 struct fio_sem *rusage_sem;
229 struct rusage ru_start;
230 struct rusage ru_end;
232 struct fio_file **files;
233 unsigned char *file_locks;
234 unsigned int files_size;
235 unsigned int files_index;
236 unsigned int nr_open_files;
237 unsigned int nr_done_files;
239 unsigned int next_file;
240 struct frand_state next_file_state;
243 struct zipf_state next_file_zipf;
244 struct gauss_state next_file_gauss;
251 double random_center;
258 size_t orig_buffer_size;
259 volatile int runstate;
260 volatile bool terminate;
262 enum fio_ddir last_ddir_completed;
263 enum fio_ddir last_ddir_issued;
270 uint64_t rand_seeds[FIO_RAND_NR_OFFS];
272 struct frand_state bsrange_state[DDIR_RWDIR_CNT];
273 struct frand_state verify_state;
274 struct frand_state verify_state_last_do_io;
275 struct frand_state trim_state;
276 struct frand_state delay_state;
277 struct frand_state fdp_state;
279 struct frand_state buf_state;
280 struct frand_state buf_state_prev;
281 struct frand_state buf_state_ret;
282 struct frand_state dedupe_state;
283 struct frand_state zone_state;
284 struct frand_state prio_state;
285 struct frand_state dedupe_working_set_index_state;
286 struct frand_state *dedupe_working_set_states;
288 unsigned long long num_unique_pages;
290 struct zone_split_index **zone_state_index;
291 unsigned int num_write_zones;
293 unsigned int verify_batch;
294 unsigned int trim_batch;
297 struct thread_io_list *vstate;
302 * Job default IO priority set with prioclass and prio options.
307 * IO engine hooks, contains everything needed to submit an io_u
308 * to any of the available IO engines.
310 struct ioengine_ops *io_ops;
314 * IO engine private data and dlhandle.
319 * Queue depth of io_u's that fio MIGHT do
321 unsigned int cur_depth;
324 * io_u's about to be committed
326 unsigned int io_u_queued;
329 * io_u's submitted but not completed yet
331 unsigned int io_u_in_flight;
334 * List of free and busy io_u's
336 struct io_u_ring io_u_requeues;
337 struct io_u_queue io_u_freelist;
338 struct io_u_queue io_u_all;
339 pthread_mutex_t io_u_lock;
340 pthread_cond_t free_cond;
343 * async verify offload
345 struct flist_head verify_list;
346 pthread_t *verify_threads;
347 unsigned int nr_verify_threads;
348 pthread_cond_t verify_cond;
349 int verify_thread_exit;
354 uint64_t rate_bps[DDIR_RWDIR_CNT];
355 uint64_t rate_next_io_time[DDIR_RWDIR_CNT];
356 unsigned long long last_rate_check_bytes[DDIR_RWDIR_CNT];
357 unsigned long last_rate_check_blocks[DDIR_RWDIR_CNT];
358 unsigned long long rate_io_issue_bytes[DDIR_RWDIR_CNT];
359 struct timespec last_rate_check_time[DDIR_RWDIR_CNT];
360 int64_t last_usec[DDIR_RWDIR_CNT];
361 struct frand_state poisson_state[DDIR_RWDIR_CNT];
364 * Enforced rate submission/completion workqueue
366 struct workqueue io_wq;
368 uint64_t total_io_size;
369 uint64_t fill_device_size;
374 uint64_t io_issues[DDIR_RWDIR_CNT];
375 uint64_t verify_read_issues;
376 uint64_t io_issue_bytes[DDIR_RWDIR_CNT];
382 uint64_t io_blocks[DDIR_RWDIR_CNT];
383 uint64_t this_io_blocks[DDIR_RWDIR_CNT];
384 uint64_t io_bytes[DDIR_RWDIR_CNT];
385 uint64_t this_io_bytes[DDIR_RWDIR_CNT];
386 uint64_t io_skip_bytes;
389 uint64_t bytes_done[DDIR_RWDIR_CNT];
390 uint64_t bytes_verified;
391 uint32_t last_write_comp_depth;
393 uint64_t *thinktime_blocks_counter;
394 struct timespec last_thinktime;
395 int64_t last_thinktime_blocks;
398 * State for random io, a bitmap of blocks done vs not done
400 struct frand_state random_state;
402 struct timespec start; /* start of this loop */
403 struct timespec epoch; /* time job was started */
404 unsigned long long alternate_epoch; /* Time job was started, as clock_gettime(log_alternate_epoch_clock_id) */
405 unsigned long long job_start; /* Time job was started, as clock_gettime(job_start_clock_id) */
406 struct timespec last_issue;
408 struct timespec ts_cache;
409 struct timespec terminate_time;
410 unsigned int ts_cache_nr;
411 unsigned int ts_cache_mask;
415 * Time since last latency_window was started
417 struct timespec latency_ts;
418 unsigned int latency_qd;
419 unsigned int latency_qd_high;
420 unsigned int latency_qd_low;
421 unsigned int latency_failed;
422 unsigned int latency_stable_count;
423 uint64_t latency_ios;
427 * read/write mixed workload state
429 struct frand_state rwmix_state;
430 unsigned long rwmix_issues;
431 enum fio_ddir rwmix_ddir;
432 unsigned int ddir_seq_nr;
435 * rand/seq mixed workload state
437 struct frand_state seq_rand_state[DDIR_RWDIR_CNT];
440 * IO history logs for verification. We use a tree for sorting,
441 * if we are overwriting. Otherwise just use a fifo.
443 struct rb_root io_hist_tree;
444 struct flist_head io_hist_list;
445 unsigned long io_hist_len;
450 struct flist_head io_log_list;
452 unsigned int io_log_blktrace;
453 unsigned int io_log_blktrace_swap;
454 unsigned long long io_log_last_ttime;
455 struct timespec io_log_start_time;
456 unsigned int io_log_current;
457 unsigned int io_log_checkmark;
458 unsigned int io_log_highmark;
459 unsigned int io_log_version;
460 struct timespec io_log_highmark_time;
463 * For tracking/handling discards
465 struct flist_head trim_list;
466 unsigned long trim_entries;
469 * for fileservice, how often to switch to a new file
471 unsigned int file_service_nr;
472 unsigned int file_service_left;
473 struct fio_file *file_service_file;
475 unsigned int sync_file_range_nr;
478 * For generating file sizes
480 struct frand_state file_size_state;
485 unsigned int total_err_count;
488 struct fio_flow *flow;
489 unsigned long long flow_counter;
492 * Can be overloaded by profiles
494 struct prof_io_ops prof_io_ops;
499 struct steadystate_data ss;
501 char verror[FIO_VERROR_SIZE];
505 * for GPU memory management
511 CUdeviceptr dev_mem_ptr;
516 struct thread_segment {
517 struct thread_data *threads;
523 * when should interactive ETA output be generated
531 #define __td_verror(td, err, msg, func) \
533 unsigned int ____e = (err); \
536 (td)->error = ____e; \
537 if (!(td)->first_error) \
538 nowarn_snprintf(td->verror, sizeof(td->verror), \
539 "file:%s:%d, func=%s, error=%s", \
540 __FILE__, __LINE__, (func), (msg)); \
544 #define td_clear_error(td) do { \
547 (td)->parent->error = 0; \
550 #define td_verror(td, err, func) do { \
551 __td_verror((td), (err), strerror((err)), (func)); \
553 __td_verror((td)->parent, (err), strerror((err)), (func)); \
556 #define td_vmsg(td, err, msg, func) do { \
557 __td_verror((td), (err), (msg), (func)); \
559 __td_verror((td)->parent, (err), (msg), (func)); \
562 #define __fio_stringify_1(x) #x
563 #define __fio_stringify(x) __fio_stringify_1(x)
565 #define REAL_MAX_JOBS 4096
566 #define JOBS_PER_SEG 8
567 #define REAL_MAX_SEG (REAL_MAX_JOBS / JOBS_PER_SEG)
569 extern bool exitall_on_terminate;
570 extern unsigned int thread_number;
571 extern unsigned int stat_number;
572 extern unsigned int nr_segments;
573 extern unsigned int cur_segment;
575 extern int output_format;
576 extern int append_terse_output;
577 extern int temp_stall_ts;
578 extern uintptr_t page_mask, page_size;
579 extern bool read_only;
580 extern int eta_print;
581 extern int eta_new_line;
582 extern unsigned int eta_interval_msec;
583 extern unsigned long done_secs;
584 extern int fio_gtod_offload;
585 extern int fio_gtod_cpu;
586 extern enum fio_cs fio_clock_source;
587 extern int fio_clock_source_set;
588 extern int warnings_fatal;
589 extern int terse_version;
590 extern bool is_backend;
591 extern bool is_local_backend;
592 extern int nr_clients;
593 extern bool log_syslog;
594 extern int status_interval;
595 extern const char fio_version_string[];
596 extern char *trigger_file;
597 extern char *trigger_cmd;
598 extern char *trigger_remote_cmd;
599 extern long long trigger_timeout;
600 extern char *aux_path;
602 extern struct thread_segment segments[REAL_MAX_SEG];
604 static inline struct thread_data *tnumber_to_td(unsigned int tnumber)
606 struct thread_segment *seg;
608 seg = &segments[tnumber / JOBS_PER_SEG];
609 return &seg->threads[tnumber & (JOBS_PER_SEG - 1)];
612 static inline bool is_running_backend(void)
614 return is_backend || is_local_backend;
617 extern bool eta_time_within_slack(unsigned int time);
619 static inline void fio_ro_check(const struct thread_data *td, struct io_u *io_u)
621 assert(!(io_u->ddir == DDIR_WRITE && !td_write(td)) &&
622 !(io_u->ddir == DDIR_TRIM && !(td_trim(td) || td->trim_verify)));
625 * The last line above allows trim operations during trim/verify
626 * workloads. For these workloads we cannot simply set the trim bit for
627 * the thread's ddir because then fio would assume that
628 * ddir={trimewrite, randtrimwrite}.
632 static inline bool multi_range_trim(struct thread_data *td, struct io_u *io_u)
634 if (io_u->ddir == DDIR_TRIM && td->o.num_range > 1)
640 static inline bool should_fsync(struct thread_data *td)
642 if (ddir_sync(td->last_ddir_issued))
644 if (td_write(td) || td->o.override_sync)
651 * Init/option functions
653 extern int __must_check fio_init_options(void);
654 extern int __must_check parse_options(int, char **);
655 extern int parse_jobs_ini(char *, int, int, int);
656 extern int parse_cmd_line(int, char **, int);
657 extern int fio_backend(struct sk_out *);
658 extern void reset_fio_state(void);
659 extern void clear_io_state(struct thread_data *, int);
660 extern int fio_options_parse(struct thread_data *, char **, int);
661 extern void fio_keywords_init(void);
662 extern void fio_keywords_exit(void);
663 extern int fio_cmd_option_parse(struct thread_data *, const char *, char *);
664 extern int fio_cmd_ioengine_option_parse(struct thread_data *, const char *, char *);
665 extern void fio_fill_default_options(struct thread_data *);
666 extern int fio_show_option_help(const char *);
667 extern void fio_options_set_ioengine_opts(struct option *long_options, struct thread_data *td);
668 extern void fio_options_dup_and_init(struct option *);
669 extern char *fio_option_dup_subs(const char *);
670 extern void fio_options_mem_dupe(struct thread_data *);
671 extern void td_fill_rand_seeds(struct thread_data *);
672 extern void add_job_opts(const char **, int);
673 extern int ioengine_load(struct thread_data *);
674 extern bool parse_dryrun(void);
675 extern int fio_running_or_pending_io_threads(void);
676 extern int fio_set_fd_nonblocking(int, const char *);
677 extern void sig_show_status(int sig);
678 extern struct thread_data *get_global_options(void);
680 extern uintptr_t page_mask;
681 extern uintptr_t page_size;
682 extern int initialize_fio(char *envp[]);
683 extern void deinitialize_fio(void);
685 #define FIO_GETOPT_JOB 0x89000000
686 #define FIO_GETOPT_IOENGINE 0x98000000
687 #define FIO_NR_OPTIONS (FIO_MAX_OPTS + 128)
692 extern void print_thread_status(void);
693 extern void print_status_init(int);
694 extern char *fio_uint_to_kmg(unsigned int val);
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
719 #define TD_ENG_FLAG_SHIFT (__TD_F_LAST)
720 #define TD_ENG_FLAG_MASK ((1ULL << (__TD_F_LAST)) - 1)
722 static inline void td_set_ioengine_flags(struct thread_data *td)
724 td->flags = (~(TD_ENG_FLAG_MASK << TD_ENG_FLAG_SHIFT) & td->flags) |
725 ((unsigned long long)td->io_ops->flags << TD_ENG_FLAG_SHIFT);
728 static inline bool td_ioengine_flagged(struct thread_data *td,
729 enum fio_ioengine_flags flags)
731 return ((td->flags >> TD_ENG_FLAG_SHIFT) & flags) != 0;
734 extern void td_set_runstate(struct thread_data *, int);
735 extern int td_bump_runstate(struct thread_data *, int);
736 extern void td_restore_runstate(struct thread_data *, int);
737 extern const char *runstate_to_name(int runstate);
740 * Allow 60 seconds for a job to quit on its own, otherwise reap with
743 #define FIO_REAP_TIMEOUT 300
748 TERMINATE_STONEWALL = 2,
752 extern void fio_terminate_threads(unsigned int, unsigned int);
753 extern void fio_mark_td_terminate(struct thread_data *);
758 extern int __must_check fio_pin_memory(struct thread_data *);
759 extern void fio_unpin_memory(struct thread_data *);
760 extern int __must_check allocate_io_mem(struct thread_data *);
761 extern void free_io_mem(struct thread_data *);
762 extern void free_threads_shm(void);
765 #define PTR_ALIGN(ptr, mask) \
766 (char *) (((uintptr_t) (ptr) + (mask)) & ~(mask))
770 * Reset stats after ramp time completes
772 extern void reset_all_stats(struct thread_data *);
774 extern int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
775 enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
776 struct timespec *comp_time);
779 * Latency target helpers
781 extern void lat_target_check(struct thread_data *);
782 extern void lat_target_init(struct thread_data *);
783 extern void lat_target_reset(struct thread_data *);
786 * Iterates all threads/processes within all the defined jobs
788 * for_each_td(var_name_for_td) {
789 * << bodoy of your loop >>
790 * Note: internally-scoped loop index availble as __td_index
791 * } end_for_each_td()
793 #define for_each_td(td) \
796 struct thread_data *(td); \
797 for (__td_index = 0, (td) = &segments[0].threads[0];\
798 __td_index < (int) thread_number; __td_index++, (td) = tnumber_to_td(__td_index))
799 #define for_each_td_index() \
802 for (__td_index = 0; __td_index < (int) thread_number; __td_index++)
803 #define end_for_each() }
805 #define for_each_file(td, f, i) \
806 if ((td)->files_index) \
807 for ((i) = 0, (f) = (td)->files[0]; \
808 (i) < (td)->o.nr_files && ((f) = (td)->files[i]) != NULL; \
811 static inline bool fio_offset_overlap_risk(struct thread_data *td)
813 if (td->o.norandommap || td->o.softrandommap ||
814 td->o.ddir_seq_add || (td->o.ddir_seq_nr > 1))
820 static inline bool fio_fill_issue_time(struct thread_data *td)
822 if (td->o.read_iolog_file ||
823 !td->o.disable_clat || !td->o.disable_slat || !td->o.disable_bw)
829 static inline bool option_check_rate(struct thread_data *td, enum fio_ddir ddir)
831 struct thread_options *o = &td->o;
834 * If some rate setting was given, we need to check it
836 if (o->rate[ddir] || o->ratemin[ddir] || o->rate_iops[ddir] ||
837 o->rate_iops_min[ddir])
843 static inline bool should_check_rate(struct thread_data *td)
845 return (td->flags & TD_F_CHECK_RATE) != 0;
848 static inline unsigned long long td_max_bs(struct thread_data *td)
850 unsigned long long max_bs;
852 max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]);
853 return max(td->o.max_bs[DDIR_TRIM], max_bs);
856 static inline unsigned long long td_min_bs(struct thread_data *td)
858 unsigned long long min_bs;
860 min_bs = min(td->o.min_bs[DDIR_READ], td->o.min_bs[DDIR_WRITE]);
861 return min(td->o.min_bs[DDIR_TRIM], min_bs);
864 static inline bool td_async_processing(struct thread_data *td)
866 return (td->flags & TD_F_NEED_LOCK) != 0;
869 static inline bool td_offload_overlap(struct thread_data *td)
871 return td->o.serialize_overlap && td->o.io_submit_mode == IO_MODE_OFFLOAD;
875 * We currently only need to do locking if we have verifier threads
876 * accessing our internal structures too
878 static inline void __td_io_u_lock(struct thread_data *td)
880 pthread_mutex_lock(&td->io_u_lock);
883 static inline void __td_io_u_unlock(struct thread_data *td)
885 pthread_mutex_unlock(&td->io_u_lock);
888 static inline void td_io_u_free_notify(struct thread_data *td)
890 if (td_async_processing(td))
891 pthread_cond_signal(&td->free_cond);
894 static inline void td_flags_clear(struct thread_data *td, unsigned int *flags,
897 if (!td_async_processing(td))
900 __sync_fetch_and_and(flags, ~value);
903 static inline void td_flags_set(struct thread_data *td, unsigned int *flags,
906 if (!td_async_processing(td))
909 __sync_fetch_and_or(flags, value);
912 extern const char *fio_get_arch_string(int);
913 extern const char *fio_get_os_string(int);
916 __FIO_OUTPUT_TERSE = 0,
917 __FIO_OUTPUT_JSON = 1,
918 __FIO_OUTPUT_NORMAL = 2,
919 __FIO_OUTPUT_JSON_PLUS = 3,
922 FIO_OUTPUT_TERSE = 1U << __FIO_OUTPUT_TERSE,
923 FIO_OUTPUT_JSON = 1U << __FIO_OUTPUT_JSON,
924 FIO_OUTPUT_NORMAL = 1U << __FIO_OUTPUT_NORMAL,
925 FIO_OUTPUT_JSON_PLUS = 1U << __FIO_OUTPUT_JSON_PLUS,
929 FIO_RAND_DIST_RANDOM = 0,
931 FIO_RAND_DIST_PARETO,
934 FIO_RAND_DIST_ZONED_ABS,
937 #define FIO_DEF_ZIPF 1.1
938 #define FIO_DEF_PARETO 0.2
941 FIO_RAND_GEN_TAUSWORTHE = 0,
943 FIO_RAND_GEN_TAUSWORTHE64,
951 extern void exec_trigger(const char *);
952 extern void check_trigger_file(void);
954 extern bool in_flight_overlap(struct io_u_queue *q, struct io_u *io_u);
955 extern pthread_mutex_t overlap_check;
957 static inline void *fio_memalign(size_t alignment, size_t size, bool shared)
959 return __fio_memalign(alignment, size, shared ? smalloc : malloc);
962 static inline void fio_memfree(void *ptr, size_t size, bool shared)
964 return __fio_memfree(ptr, size, shared ? sfree : free);