X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=backend.c;h=bc2e3eb50e37a501ae89f89dcac767f1023fa616;hp=1ff8b3f6e298adfa7f756453edf2cff1656b91f8;hb=5bb79f69c2d9dc8542c25af96f040d1884230688;hpb=3939fe853fd146378811a05b78731d7f20d0152f diff --git a/backend.c b/backend.c index 1ff8b3f6..bc2e3eb5 100644 --- a/backend.c +++ b/backend.c @@ -35,6 +35,7 @@ #include #include #include +#include #include "fio.h" #ifndef FIO_NO_HAVE_SHM_H @@ -48,16 +49,21 @@ #include "cgroup.h" #include "profile.h" #include "lib/rand.h" -#include "memalign.h" +#include "lib/memalign.h" #include "server.h" #include "lib/getrusage.h" #include "idletime.h" #include "err.h" +#include "lib/tp.h" +#include "workqueue.h" +#include "lib/mountcheck.h" + +static pthread_t helper_thread; +static pthread_mutex_t helper_lock; +pthread_cond_t helper_cond; +int helper_do_stat = 0; -static pthread_t disk_util_thread; -static struct fio_mutex *disk_thread_mutex; static struct fio_mutex *startup_mutex; -static struct fio_mutex *writeout_mutex; static struct flist_head *cgroup_list; static char *cgroup_mnt; static int exit_value; @@ -73,7 +79,7 @@ unsigned int stat_number = 0; int shm_id = 0; int temp_stall_ts; unsigned long done_secs = 0; -volatile int disk_util_exit = 0; +volatile int helper_exit = 0; #define PAGE_ALIGN(buf) \ (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask) @@ -87,7 +93,7 @@ static void sig_int(int sig) fio_server_got_signal(sig); else { log_info("\nfio: terminating on signal %d\n", sig); - fflush(stdout); + log_info_flush(); exit_value = 128; } @@ -95,7 +101,7 @@ static void sig_int(int sig) } } -static void sig_show_status(int sig) +void sig_show_status(int sig) { show_running_run_stats(); } @@ -138,8 +144,8 @@ static void set_sig_handlers(void) /* * Check if we are above the minimum rate given. */ -static int __check_min_rate(struct thread_data *td, struct timeval *now, - enum fio_ddir ddir) +static bool __check_min_rate(struct thread_data *td, struct timeval *now, + enum fio_ddir ddir) { unsigned long long bytes = 0; unsigned long iops = 0; @@ -152,13 +158,13 @@ static int __check_min_rate(struct thread_data *td, struct timeval *now, assert(ddir_rw(ddir)); if (!td->o.ratemin[ddir] && !td->o.rate_iops_min[ddir]) - return 0; + return false; /* * allow a 2 second settle period in the beginning */ if (mtime_since(&td->start, now) < 2000) - return 0; + return false; iops += td->this_io_blocks[ddir]; bytes += td->this_io_bytes[ddir]; @@ -172,24 +178,28 @@ static int __check_min_rate(struct thread_data *td, struct timeval *now, if (td->rate_bytes[ddir] || td->rate_blocks[ddir]) { spent = mtime_since(&td->lastrate[ddir], now); if (spent < td->o.ratecycle) - return 0; + return false; - if (td->o.rate[ddir]) { + if (td->o.rate[ddir] || td->o.ratemin[ddir]) { /* * check bandwidth specified rate */ if (bytes < td->rate_bytes[ddir]) { log_err("%s: min rate %u not met\n", td->o.name, ratemin); - return 1; + return true; } else { - rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent; + if (spent) + rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent; + else + rate = 0; + if (rate < ratemin || bytes < td->rate_bytes[ddir]) { log_err("%s: min rate %u not met, got" " %luKB/sec\n", td->o.name, ratemin, rate); - return 1; + return true; } } } else { @@ -199,14 +209,19 @@ static int __check_min_rate(struct thread_data *td, struct timeval *now, if (iops < rate_iops) { log_err("%s: min iops rate %u not met\n", td->o.name, rate_iops); - return 1; + return true; } else { - rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent; + if (spent) + rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent; + else + rate = 0; + if (rate < rate_iops_min || iops < td->rate_blocks[ddir]) { log_err("%s: min iops rate %u not met," " got %lu\n", td->o.name, rate_iops_min, rate); + return true; } } } @@ -215,19 +230,18 @@ static int __check_min_rate(struct thread_data *td, struct timeval *now, td->rate_bytes[ddir] = bytes; td->rate_blocks[ddir] = iops; memcpy(&td->lastrate[ddir], now, sizeof(*now)); - return 0; + return false; } -static int check_min_rate(struct thread_data *td, struct timeval *now, - uint64_t *bytes_done) +static bool check_min_rate(struct thread_data *td, struct timeval *now) { - int ret = 0; + bool ret = false; - if (bytes_done[DDIR_READ]) + if (td->bytes_done[DDIR_READ]) ret |= __check_min_rate(td, now, DDIR_READ); - if (bytes_done[DDIR_WRITE]) + if (td->bytes_done[DDIR_WRITE]) ret |= __check_min_rate(td, now, DDIR_WRITE); - if (bytes_done[DDIR_TRIM]) + if (td->bytes_done[DDIR_TRIM]) ret |= __check_min_rate(td, now, DDIR_TRIM); return ret; @@ -244,7 +258,7 @@ static void cleanup_pending_aio(struct thread_data *td) /* * get immediately available events, if any */ - r = io_u_queued_complete(td, 0, NULL); + r = io_u_queued_complete(td, 0); if (r < 0) return; @@ -265,27 +279,27 @@ static void cleanup_pending_aio(struct thread_data *td) } if (td->cur_depth) - r = io_u_queued_complete(td, td->cur_depth, NULL); + r = io_u_queued_complete(td, td->cur_depth); } /* * Helper to handle the final sync of a file. Works just like the normal * io path, just does everything sync. */ -static int fio_io_sync(struct thread_data *td, struct fio_file *f) +static bool fio_io_sync(struct thread_data *td, struct fio_file *f) { struct io_u *io_u = __get_io_u(td); int ret; if (!io_u) - return 1; + return true; io_u->ddir = DDIR_SYNC; io_u->file = f; if (td_io_prep(td, io_u)) { put_io_u(td, io_u); - return 1; + return true; } requeue: @@ -293,25 +307,25 @@ requeue: if (ret < 0) { td_verror(td, io_u->error, "td_io_queue"); put_io_u(td, io_u); - return 1; + return true; } else if (ret == FIO_Q_QUEUED) { - if (io_u_queued_complete(td, 1, NULL) < 0) - return 1; + if (io_u_queued_complete(td, 1) < 0) + return true; } else if (ret == FIO_Q_COMPLETED) { if (io_u->error) { td_verror(td, io_u->error, "td_io_queue"); - return 1; + return true; } - if (io_u_sync_complete(td, io_u, NULL) < 0) - return 1; + if (io_u_sync_complete(td, io_u) < 0) + return true; } else if (ret == FIO_Q_BUSY) { if (td_io_commit(td)) - return 1; + return true; goto requeue; } - return 0; + return false; } static int fio_file_fsync(struct thread_data *td, struct fio_file *f) @@ -340,20 +354,37 @@ static inline void update_tv_cache(struct thread_data *td) __update_tv_cache(td); } -static inline int runtime_exceeded(struct thread_data *td, struct timeval *t) +static inline bool runtime_exceeded(struct thread_data *td, struct timeval *t) { if (in_ramp_time(td)) - return 0; + return false; if (!td->o.timeout) - return 0; + return false; if (utime_since(&td->epoch, t) >= td->o.timeout) - return 1; + return true; - return 0; + return false; } -static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir, - int *retptr) +/* + * We need to update the runtime consistently in ms, but keep a running + * tally of the current elapsed time in microseconds for sub millisecond + * updates. + */ +static inline void update_runtime(struct thread_data *td, + unsigned long long *elapsed_us, + const enum fio_ddir ddir) +{ + if (ddir == DDIR_WRITE && td_write(td) && td->o.verify_only) + return; + + td->ts.runtime[ddir] -= (elapsed_us[ddir] + 999) / 1000; + elapsed_us[ddir] += utime_since_now(&td->start); + td->ts.runtime[ddir] += (elapsed_us[ddir] + 999) / 1000; +} + +static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir, + int *retptr) { int ret = *retptr; @@ -366,7 +397,7 @@ static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir, eb = td_error_type(ddir, err); if (!(td->o.continue_on_error & (1 << eb))) - return 1; + return true; if (td_non_fatal_error(td, eb, err)) { /* @@ -376,26 +407,26 @@ static int break_on_this_error(struct thread_data *td, enum fio_ddir ddir, update_error_count(td, err); td_clear_error(td); *retptr = 0; - return 0; + return false; } else if (td->o.fill_device && err == ENOSPC) { /* * We expect to hit this error if * fill_device option is set. */ td_clear_error(td); - td->terminate = 1; - return 1; + fio_mark_td_terminate(td); + return true; } else { /* * Stop the I/O in case of a fatal * error. */ update_error_count(td, err); - return 1; + return true; } } - return 0; + return false; } static void check_update_rusage(struct thread_data *td) @@ -407,13 +438,132 @@ static void check_update_rusage(struct thread_data *td) } } +static int wait_for_completions(struct thread_data *td, struct timeval *time) +{ + const int full = queue_full(td); + int min_evts = 0; + int ret; + + /* + * if the queue is full, we MUST reap at least 1 event + */ + min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth); + if ((full && !min_evts) || !td->o.iodepth_batch_complete_min) + min_evts = 1; + + if (time && (__should_check_rate(td, DDIR_READ) || + __should_check_rate(td, DDIR_WRITE) || + __should_check_rate(td, DDIR_TRIM))) + fio_gettime(time, NULL); + + do { + ret = io_u_queued_complete(td, min_evts); + if (ret < 0) + break; + } while (full && (td->cur_depth > td->o.iodepth_low)); + + return ret; +} + +int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret, + enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify, + struct timeval *comp_time) +{ + int ret2; + + switch (*ret) { + case FIO_Q_COMPLETED: + if (io_u->error) { + *ret = -io_u->error; + clear_io_u(td, io_u); + } else if (io_u->resid) { + int bytes = io_u->xfer_buflen - io_u->resid; + struct fio_file *f = io_u->file; + + if (bytes_issued) + *bytes_issued += bytes; + + if (!from_verify) + trim_io_piece(td, io_u); + + /* + * zero read, fail + */ + if (!bytes) { + if (!from_verify) + unlog_io_piece(td, io_u); + td_verror(td, EIO, "full resid"); + put_io_u(td, io_u); + break; + } + + io_u->xfer_buflen = io_u->resid; + io_u->xfer_buf += bytes; + io_u->offset += bytes; + + if (ddir_rw(io_u->ddir)) + td->ts.short_io_u[io_u->ddir]++; + + f = io_u->file; + if (io_u->offset == f->real_file_size) + goto sync_done; + + requeue_io_u(td, &io_u); + } else { +sync_done: + if (comp_time && (__should_check_rate(td, DDIR_READ) || + __should_check_rate(td, DDIR_WRITE) || + __should_check_rate(td, DDIR_TRIM))) + fio_gettime(comp_time, NULL); + + *ret = io_u_sync_complete(td, io_u); + if (*ret < 0) + break; + } + return 0; + case FIO_Q_QUEUED: + /* + * if the engine doesn't have a commit hook, + * the io_u is really queued. if it does have such + * a hook, it has to call io_u_queued() itself. + */ + if (td->io_ops->commit == NULL) + io_u_queued(td, io_u); + if (bytes_issued) + *bytes_issued += io_u->xfer_buflen; + break; + case FIO_Q_BUSY: + if (!from_verify) + unlog_io_piece(td, io_u); + requeue_io_u(td, &io_u); + ret2 = td_io_commit(td); + if (ret2 < 0) + *ret = ret2; + break; + default: + assert(*ret < 0); + td_verror(td, -(*ret), "td_io_queue"); + break; + } + + if (break_on_this_error(td, ddir, ret)) + return 1; + + return 0; +} + +static inline bool io_in_polling(struct thread_data *td) +{ + return !td->o.iodepth_batch_complete_min && + !td->o.iodepth_batch_complete_max; +} + /* * The main verify engine. Runs over the writes we previously submitted, * reads the blocks back in, and checks the crc/md5 of the data. */ static void do_verify(struct thread_data *td, uint64_t verify_bytes) { - uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 }; struct fio_file *f; struct io_u *io_u; int ret, min_events; @@ -444,7 +594,7 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes) io_u = NULL; while (!td->terminate) { enum fio_ddir ddir; - int ret2, full; + int full; update_tv_cache(td); check_update_rusage(td); @@ -452,7 +602,7 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes) if (runtime_exceeded(td, &td->tv_cache)) { __update_tv_cache(td); if (runtime_exceeded(td, &td->tv_cache)) { - td->terminate = 1; + fio_mark_td_terminate(td); break; } } @@ -475,7 +625,7 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes) break; } } else { - if (ddir_rw_sum(bytes_done) + td->o.rw_min_bs > verify_bytes) + if (ddir_rw_sum(td->bytes_done) + td->o.rw_min_bs > verify_bytes) break; while ((io_u = get_io_u(td)) != NULL) { @@ -500,7 +650,7 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes) continue; } else if (io_u->ddir == DDIR_TRIM) { io_u->ddir = DDIR_READ; - io_u->flags |= IO_U_F_TRIMMED; + io_u_set(io_u, IO_U_F_TRIMMED); break; } else if (io_u->ddir == DDIR_WRITE) { io_u->ddir = DDIR_READ; @@ -515,65 +665,23 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes) break; } + if (verify_state_should_stop(td, io_u)) { + put_io_u(td, io_u); + break; + } + if (td->o.verify_async) io_u->end_io = verify_io_u_async; else io_u->end_io = verify_io_u; ddir = io_u->ddir; + if (!td->o.disable_slat) + fio_gettime(&io_u->start_time, NULL); ret = td_io_queue(td, io_u); - switch (ret) { - case FIO_Q_COMPLETED: - if (io_u->error) { - ret = -io_u->error; - clear_io_u(td, io_u); - } else if (io_u->resid) { - int bytes = io_u->xfer_buflen - io_u->resid; - - /* - * zero read, fail - */ - if (!bytes) { - td_verror(td, EIO, "full resid"); - put_io_u(td, io_u); - break; - } - - io_u->xfer_buflen = io_u->resid; - io_u->xfer_buf += bytes; - io_u->offset += bytes; - if (ddir_rw(io_u->ddir)) - td->ts.short_io_u[io_u->ddir]++; - - f = io_u->file; - if (io_u->offset == f->real_file_size) - goto sync_done; - - requeue_io_u(td, &io_u); - } else { -sync_done: - ret = io_u_sync_complete(td, io_u, bytes_done); - if (ret < 0) - break; - } - continue; - case FIO_Q_QUEUED: - break; - case FIO_Q_BUSY: - requeue_io_u(td, &io_u); - ret2 = td_io_commit(td); - if (ret2 < 0) - ret = ret2; - break; - default: - assert(ret < 0); - td_verror(td, -ret, "td_io_queue"); - break; - } - - if (break_on_this_error(td, ddir, &ret)) + if (io_queue_event(td, io_u, &ret, ddir, NULL, 1, NULL)) break; /* @@ -583,27 +691,9 @@ sync_done: */ reap: full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth); - if (full || !td->o.iodepth_batch_complete) { - min_events = min(td->o.iodepth_batch_complete, - td->cur_depth); - /* - * if the queue is full, we MUST reap at least 1 event - */ - if (full && !min_events) - min_events = 1; + if (full || io_in_polling(td)) + ret = wait_for_completions(td, NULL); - do { - /* - * Reap required number of io units, if any, - * and do the verification on them through - * the callback handler - */ - if (io_u_queued_complete(td, min_events, bytes_done) < 0) { - ret = -1; - break; - } - } while (full && (td->cur_depth > td->o.iodepth_low)); - } if (ret < 0) break; } @@ -614,7 +704,7 @@ reap: min_events = td->cur_depth; if (min_events) - ret = io_u_queued_complete(td, min_events, NULL); + ret = io_u_queued_complete(td, min_events); } else cleanup_pending_aio(td); @@ -623,22 +713,44 @@ reap: dprint(FD_VERIFY, "exiting loop\n"); } -static unsigned int exceeds_number_ios(struct thread_data *td) +static bool exceeds_number_ios(struct thread_data *td) { unsigned long long number_ios; if (!td->o.number_ios) - return 0; + return false; - number_ios = ddir_rw_sum(td->this_io_blocks); + number_ios = ddir_rw_sum(td->io_blocks); number_ios += td->io_u_queued + td->io_u_in_flight; - return number_ios >= td->o.number_ios; + return number_ios >= (td->o.number_ios * td->loops); +} + +static bool io_issue_bytes_exceeded(struct thread_data *td) +{ + unsigned long long bytes, limit; + + if (td_rw(td)) + bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE]; + else if (td_write(td)) + bytes = td->io_issue_bytes[DDIR_WRITE]; + else if (td_read(td)) + bytes = td->io_issue_bytes[DDIR_READ]; + else + bytes = td->io_issue_bytes[DDIR_TRIM]; + + if (td->o.io_limit) + limit = td->o.io_limit; + else + limit = td->o.size; + + limit *= td->loops; + return bytes >= limit || exceeds_number_ios(td); } -static int io_bytes_exceeded(struct thread_data *td) +static bool io_complete_bytes_exceeded(struct thread_data *td) { - unsigned long long bytes; + unsigned long long bytes, limit; if (td_rw(td)) bytes = td->this_io_bytes[DDIR_READ] + td->this_io_bytes[DDIR_WRITE]; @@ -649,7 +761,45 @@ static int io_bytes_exceeded(struct thread_data *td) else bytes = td->this_io_bytes[DDIR_TRIM]; - return bytes >= td->o.size || exceeds_number_ios(td); + if (td->o.io_limit) + limit = td->o.io_limit; + else + limit = td->o.size; + + limit *= td->loops; + return bytes >= limit || exceeds_number_ios(td); +} + +/* + * used to calculate the next io time for rate control + * + */ +static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir) +{ + uint64_t secs, remainder, bps, bytes, iops; + + assert(!(td->flags & TD_F_CHILD)); + bytes = td->rate_io_issue_bytes[ddir]; + bps = td->rate_bps[ddir]; + + if (td->o.rate_process == RATE_PROCESS_POISSON) { + uint64_t val; + iops = bps / td->o.bs[ddir]; + val = (int64_t) (1000000 / iops) * + -logf(__rand_0_1(&td->poisson_state)); + if (val) { + dprint(FD_RATE, "poisson rate iops=%llu\n", + (unsigned long long) 1000000 / val); + } + td->last_usec += val; + return td->last_usec; + } else if (bps) { + secs = bytes / bps; + remainder = bytes % bps; + return remainder * 1000000 / bps + secs * 1000000; + } + + return 0; } /* @@ -660,7 +810,6 @@ static int io_bytes_exceeded(struct thread_data *td) */ static uint64_t do_io(struct thread_data *td) { - uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 }; unsigned int i; int ret = 0; uint64_t total_bytes, bytes_issued = 0; @@ -672,23 +821,33 @@ static uint64_t do_io(struct thread_data *td) lat_target_init(td); + total_bytes = td->o.size; + /* + * Allow random overwrite workloads to write up to io_limit + * before starting verification phase as 'size' doesn't apply. + */ + if (td_write(td) && td_random(td) && td->o.norandommap) + total_bytes = max(total_bytes, (uint64_t) td->o.io_limit); /* * If verify_backlog is enabled, we'll run the verify in this * handler as well. For that case, we may need up to twice the * amount of bytes. */ - total_bytes = td->o.size; if (td->o.verify != VERIFY_NONE && (td_write(td) && td->o.verify_backlog)) total_bytes += td->o.size; + /* In trimwrite mode, each byte is trimmed and then written, so + * allow total_bytes to be twice as big */ + if (td_trimwrite(td)) + total_bytes += td->total_io_size; + while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) || - (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) || + (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) || td->o.time_based) { struct timeval comp_time; - int min_evts = 0; struct io_u *io_u; - int ret2, full; + int full; enum fio_ddir ddir; check_update_rusage(td); @@ -701,7 +860,7 @@ static uint64_t do_io(struct thread_data *td) if (runtime_exceeded(td, &td->tv_cache)) { __update_tv_cache(td); if (runtime_exceeded(td, &td->tv_cache)) { - td->terminate = 1; + fio_mark_td_terminate(td); break; } } @@ -709,7 +868,7 @@ static uint64_t do_io(struct thread_data *td) if (flow_threshold_exceeded(td)) continue; - if (bytes_issued >= total_bytes) + if (!td->o.time_based && bytes_issued >= total_bytes) break; io_u = get_io_u(td); @@ -737,9 +896,14 @@ static uint64_t do_io(struct thread_data *td) ((io_u->flags & IO_U_F_VER_LIST) || !td_rw(td))) { if (!td->o.verify_pattern_bytes) { - io_u->rand_seed = __rand(&td->__verify_state); + io_u->rand_seed = __rand(&td->verify_state); if (sizeof(int) != sizeof(long *)) - io_u->rand_seed *= __rand(&td->__verify_state); + io_u->rand_seed *= __rand(&td->verify_state); + } + + if (verify_state_should_stop(td, io_u)) { + put_io_u(td, io_u); + break; } if (td->o.verify_async) @@ -763,111 +927,56 @@ static uint64_t do_io(struct thread_data *td) !td->o.experimental_verify) log_io_piece(td, io_u); - ret = td_io_queue(td, io_u); - switch (ret) { - case FIO_Q_COMPLETED: - if (io_u->error) { - ret = -io_u->error; - clear_io_u(td, io_u); - } else if (io_u->resid) { - int bytes = io_u->xfer_buflen - io_u->resid; - struct fio_file *f = io_u->file; - - bytes_issued += bytes; - /* - * zero read, fail - */ - if (!bytes) { - td_verror(td, EIO, "full resid"); - put_io_u(td, io_u); - break; - } - - io_u->xfer_buflen = io_u->resid; - io_u->xfer_buf += bytes; - io_u->offset += bytes; - - if (ddir_rw(io_u->ddir)) - td->ts.short_io_u[io_u->ddir]++; + if (td->o.io_submit_mode == IO_MODE_OFFLOAD) { + const unsigned long blen = io_u->xfer_buflen; + const enum fio_ddir ddir = acct_ddir(io_u); - if (io_u->offset == f->real_file_size) - goto sync_done; + if (td->error) + break; - requeue_io_u(td, &io_u); - } else { -sync_done: - if (__should_check_rate(td, DDIR_READ) || - __should_check_rate(td, DDIR_WRITE) || - __should_check_rate(td, DDIR_TRIM)) - fio_gettime(&comp_time, NULL); + ret = workqueue_enqueue(&td->io_wq, &io_u->work); + if (ret) + ret = FIO_Q_QUEUED; + else + ret = FIO_Q_BUSY; - ret = io_u_sync_complete(td, io_u, bytes_done); - if (ret < 0) - break; - bytes_issued += io_u->xfer_buflen; + if (ret == FIO_Q_QUEUED && ddir_rw(ddir)) { + td->io_issues[ddir]++; + td->io_issue_bytes[ddir] += blen; + td->rate_io_issue_bytes[ddir] += blen; } - break; - case FIO_Q_QUEUED: - /* - * if the engine doesn't have a commit hook, - * the io_u is really queued. if it does have such - * a hook, it has to call io_u_queued() itself. - */ - if (td->io_ops->commit == NULL) - io_u_queued(td, io_u); - bytes_issued += io_u->xfer_buflen; - break; - case FIO_Q_BUSY: - requeue_io_u(td, &io_u); - ret2 = td_io_commit(td); - if (ret2 < 0) - ret = ret2; - break; - default: - assert(ret < 0); - put_io_u(td, io_u); - break; - } - if (break_on_this_error(td, ddir, &ret)) - break; + if (should_check_rate(td)) + td->rate_next_io_time[ddir] = usec_for_io(td, ddir); - /* - * See if we need to complete some commands. Note that we - * can get BUSY even without IO queued, if the system is - * resource starved. - */ -reap: - full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth); - if (full || !td->o.iodepth_batch_complete) { - min_evts = min(td->o.iodepth_batch_complete, - td->cur_depth); - /* - * if the queue is full, we MUST reap at least 1 event - */ - if (full && !min_evts) - min_evts = 1; + } else { + ret = td_io_queue(td, io_u); - if (__should_check_rate(td, DDIR_READ) || - __should_check_rate(td, DDIR_WRITE) || - __should_check_rate(td, DDIR_TRIM)) - fio_gettime(&comp_time, NULL); + if (should_check_rate(td)) + td->rate_next_io_time[ddir] = usec_for_io(td, ddir); - do { - ret = io_u_queued_complete(td, min_evts, bytes_done); - if (ret < 0) - break; + if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time)) + break; - } while (full && (td->cur_depth > td->o.iodepth_low)); + /* + * See if we need to complete some commands. Note that + * we can get BUSY even without IO queued, if the + * system is resource starved. + */ +reap: + full = queue_full(td) || + (ret == FIO_Q_BUSY && td->cur_depth); + if (full || io_in_polling(td)) + ret = wait_for_completions(td, &comp_time); } - if (ret < 0) break; - if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO)) + if (!ddir_rw_sum(td->bytes_done) && + !(td->io_ops->flags & FIO_NOIO)) continue; - if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) { - if (check_min_rate(td, &comp_time, bytes_done)) { + if (!in_ramp_time(td) && should_check_rate(td)) { + if (check_min_rate(td, &comp_time)) { if (exitall_on_terminate) fio_terminate_threads(td->groupid); td_verror(td, EIO, "check_min_rate"); @@ -903,14 +1012,19 @@ reap: if (td->o.fill_device && td->error == ENOSPC) { td->error = 0; - td->terminate = 1; + fio_mark_td_terminate(td); } if (!td->error) { struct fio_file *f; - i = td->cur_depth; + if (td->o.io_submit_mode == IO_MODE_OFFLOAD) { + workqueue_flush(&td->io_wq); + i = 0; + } else + i = td->cur_depth; + if (i) { - ret = io_u_queued_complete(td, i, bytes_done); + ret = io_u_queued_complete(td, i); if (td->o.fill_device && td->error == ENOSPC) td->error = 0; } @@ -935,7 +1049,7 @@ reap: if (!ddir_rw_sum(td->this_io_bytes)) td->done = 1; - return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM]; + return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM]; } static void cleanup_io_u(struct thread_data *td) @@ -955,6 +1069,9 @@ static void cleanup_io_u(struct thread_data *td) io_u_rexit(&td->io_u_requeues); io_u_qexit(&td->io_u_freelist); io_u_qexit(&td->io_u_all); + + if (td->last_write_comp) + sfree(td->last_write_comp); } static int init_io_u(struct thread_data *td) @@ -1071,6 +1188,14 @@ static int init_io_u(struct thread_data *td) p += max_bs; } + if (td->o.verify != VERIFY_NONE) { + td->last_write_comp = scalloc(max_units, sizeof(uint64_t)); + if (!td->last_write_comp) { + log_err("fio: failed to alloc write comp data\n"); + return 1; + } + } + return 0; } @@ -1111,12 +1236,18 @@ static int switch_ioscheduler(struct thread_data *td) /* * Read back and check that the selected scheduler is now the default. */ - ret = fread(tmp, 1, sizeof(tmp), f); + memset(tmp, 0, sizeof(tmp)); + ret = fread(tmp, sizeof(tmp), 1, f); if (ferror(f) || ret < 0) { td_verror(td, errno, "fread"); fclose(f); return 1; } + /* + * either a list of io schedulers or "none\n" is expected. + */ + tmp[strlen(tmp) - 1] = '\0'; + sprintf(tmp2, "[%s]", td->o.ioscheduler); if (!strstr(tmp, tmp2)) { @@ -1130,42 +1261,50 @@ static int switch_ioscheduler(struct thread_data *td) return 0; } -static int keep_running(struct thread_data *td) +static bool keep_running(struct thread_data *td) { + unsigned long long limit; + if (td->done) - return 0; + return false; if (td->o.time_based) - return 1; + return true; if (td->o.loops) { td->o.loops--; - return 1; + return true; } if (exceeds_number_ios(td)) - return 0; + return false; + + if (td->o.io_limit) + limit = td->o.io_limit; + else + limit = td->o.size; - if (td->o.size != -1ULL && ddir_rw_sum(td->io_bytes) < td->o.size) { + if (limit != -1ULL && ddir_rw_sum(td->io_bytes) < limit) { uint64_t diff; /* * If the difference is less than the minimum IO size, we * are done. */ - diff = td->o.size - ddir_rw_sum(td->io_bytes); + diff = limit - ddir_rw_sum(td->io_bytes); if (diff < td_max_bs(td)) - return 0; + return false; if (fio_files_done(td)) - return 0; + return false; - return 1; + return true; } - return 0; + return false; } static int exec_string(struct thread_options *o, const char *string, const char *mode) { - int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1; + size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1; + int ret; char *str; str = malloc(newlen); @@ -1185,12 +1324,10 @@ static int exec_string(struct thread_options *o, const char *string, const char */ static uint64_t do_dry_run(struct thread_data *td) { - uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 }; - td_set_runstate(td, TD_RUNNING); while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) || - (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) { + (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) { struct io_u *io_u; int ret; @@ -1201,7 +1338,7 @@ static uint64_t do_dry_run(struct thread_data *td) if (!io_u) break; - io_u->flags |= IO_U_F_FLIGHT; + io_u_set(io_u, IO_U_F_FLIGHT); io_u->error = 0; io_u->resid = 0; if (ddir_rw(acct_ddir(io_u))) @@ -1217,20 +1354,89 @@ static uint64_t do_dry_run(struct thread_data *td) !td->o.experimental_verify) log_io_piece(td, io_u); - ret = io_u_sync_complete(td, io_u, bytes_done); + ret = io_u_sync_complete(td, io_u); (void) ret; } - return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM]; + return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM]; +} + +static void io_workqueue_fn(struct thread_data *td, struct workqueue_work *work) +{ + struct io_u *io_u = container_of(work, struct io_u, work); + const enum fio_ddir ddir = io_u->ddir; + int ret; + + dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid()); + + io_u_set(io_u, IO_U_F_NO_FILE_PUT); + + td->cur_depth++; + + do { + ret = td_io_queue(td, io_u); + if (ret != FIO_Q_BUSY) + break; + ret = io_u_queued_complete(td, 1); + if (ret > 0) + td->cur_depth -= ret; + io_u_clear(io_u, IO_U_F_FLIGHT); + } while (1); + + dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid()); + + io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL); + + if (ret == FIO_Q_COMPLETED) + td->cur_depth--; + else if (ret == FIO_Q_QUEUED) { + unsigned int min_evts; + + if (td->o.iodepth == 1) + min_evts = 1; + else + min_evts = 0; + + ret = io_u_queued_complete(td, min_evts); + if (ret > 0) + td->cur_depth -= ret; + } else if (ret == FIO_Q_BUSY) { + ret = io_u_queued_complete(td, td->cur_depth); + if (ret > 0) + td->cur_depth -= ret; + } +} + +static bool io_workqueue_pre_sleep_flush_fn(struct thread_data *td) +{ + if (td->io_u_queued || td->cur_depth || td->io_u_in_flight) + return true; + + return false; +} + +static void io_workqueue_pre_sleep_fn(struct thread_data *td) +{ + int ret; + + ret = io_u_quiesce(td); + if (ret > 0) + td->cur_depth -= ret; } +struct workqueue_ops rated_wq_ops = { + .fn = io_workqueue_fn, + .pre_sleep_flush_fn = io_workqueue_pre_sleep_flush_fn, + .pre_sleep_fn = io_workqueue_pre_sleep_fn, +}; + /* * Entry point for the thread based jobs. The process based jobs end up * here as well, after a little setup. */ static void *thread_main(void *data) { - unsigned long long elapsed; + unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, }; struct thread_data *td = data; struct thread_options *o = &td->o; pthread_condattr_t attr; @@ -1243,11 +1449,6 @@ static void *thread_main(void *data) } else td->pid = gettid(); - /* - * fio_time_init() may not have been called yet if running as a server - */ - fio_time_init(); - fio_local_clock_init(o->use_thread); dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid); @@ -1298,7 +1499,7 @@ static void *thread_main(void *data) * Set affinity first, in case it has an impact on the memory * allocations. */ - if (o->cpumask_set) { + if (fio_option_is_set(o, cpumask)) { if (o->cpus_allowed_policy == FIO_CPUS_SPLIT) { ret = fio_cpus_split(&o->cpumask, td->thread_number - 1); if (!ret) { @@ -1317,16 +1518,19 @@ static void *thread_main(void *data) #ifdef CONFIG_LIBNUMA /* numa node setup */ - if (o->numa_cpumask_set || o->numa_memmask_set) { - int ret; + if (fio_option_is_set(o, numa_cpunodes) || + fio_option_is_set(o, numa_memnodes)) { + struct bitmask *mask; if (numa_available() < 0) { td_verror(td, errno, "Does not support NUMA API\n"); goto err; } - if (o->numa_cpumask_set) { - ret = numa_run_on_node_mask(o->numa_cpunodesmask); + if (fio_option_is_set(o, numa_cpunodes)) { + mask = numa_parse_nodestring(o->numa_cpunodes); + ret = numa_run_on_node_mask(mask); + numa_free_nodemask(mask); if (ret == -1) { td_verror(td, errno, \ "numa_run_on_node_mask failed\n"); @@ -1334,14 +1538,17 @@ static void *thread_main(void *data) } } - if (o->numa_memmask_set) { + if (fio_option_is_set(o, numa_memnodes)) { + mask = NULL; + if (o->numa_memnodes) + mask = numa_parse_nodestring(o->numa_memnodes); switch (o->numa_mem_mode) { case MPOL_INTERLEAVE: - numa_set_interleave_mask(o->numa_memnodesmask); + numa_set_interleave_mask(mask); break; case MPOL_BIND: - numa_set_membind(o->numa_memnodesmask); + numa_set_membind(mask); break; case MPOL_LOCAL: numa_set_localalloc(); @@ -1354,6 +1561,9 @@ static void *thread_main(void *data) break; } + if (mask) + numa_free_nodemask(mask); + } } #endif @@ -1374,7 +1584,8 @@ static void *thread_main(void *data) if (o->verify_async && verify_async_init(td)) goto err; - if (o->ioprio) { + if (fio_option_is_set(o, ioprio) || + fio_option_is_set(o, ioprio_class)) { ret = ioprio_set(IOPRIO_WHO_PROCESS, 0, o->ioprio_class, o->ioprio); if (ret == -1) { td_verror(td, errno, "ioprio_set"); @@ -1411,31 +1622,39 @@ static void *thread_main(void *data) goto err; } + if (td->flags & TD_F_COMPRESS_LOG) + tp_init(&td->tp_data); + fio_verify_init(td); + if ((o->io_submit_mode == IO_MODE_OFFLOAD) && + workqueue_init(td, &td->io_wq, &rated_wq_ops, td->o.iodepth)) + goto err; + fio_gettime(&td->epoch, NULL); fio_getrusage(&td->ru_start); + memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch)); + memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch)); + + if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] || + o->ratemin[DDIR_TRIM]) { + memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time, + sizeof(td->bw_sample_time)); + memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time, + sizeof(td->bw_sample_time)); + memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time, + sizeof(td->bw_sample_time)); + } + clear_state = 0; while (keep_running(td)) { uint64_t verify_bytes; fio_gettime(&td->start, NULL); - memcpy(&td->bw_sample_time, &td->start, sizeof(td->start)); - memcpy(&td->iops_sample_time, &td->start, sizeof(td->start)); memcpy(&td->tv_cache, &td->start, sizeof(td->start)); - if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] || - o->ratemin[DDIR_TRIM]) { - memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time, - sizeof(td->bw_sample_time)); - memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time, - sizeof(td->bw_sample_time)); - memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time, - sizeof(td->bw_sample_time)); - } - if (clear_state) - clear_io_state(td); + clear_io_state(td, 0); prune_io_piece_log(td); @@ -1446,18 +1665,24 @@ static void *thread_main(void *data) clear_state = 1; - if (td_read(td) && td->io_bytes[DDIR_READ]) { - elapsed = utime_since_now(&td->start); - td->ts.runtime[DDIR_READ] += elapsed; - } - if (td_write(td) && td->io_bytes[DDIR_WRITE]) { - elapsed = utime_since_now(&td->start); - td->ts.runtime[DDIR_WRITE] += elapsed; - } - if (td_trim(td) && td->io_bytes[DDIR_TRIM]) { - elapsed = utime_since_now(&td->start); - td->ts.runtime[DDIR_TRIM] += elapsed; - } + /* + * Make sure we've successfully updated the rusage stats + * before waiting on the stat mutex. Otherwise we could have + * the stat thread holding stat mutex and waiting for + * the rusage_sem, which would never get upped because + * this thread is waiting for the stat mutex. + */ + check_update_rusage(td); + + fio_mutex_down(stat_mutex); + if (td_read(td) && td->io_bytes[DDIR_READ]) + update_runtime(td, elapsed_us, DDIR_READ); + if (td_write(td) && td->io_bytes[DDIR_WRITE]) + update_runtime(td, elapsed_us, DDIR_WRITE); + if (td_trim(td) && td->io_bytes[DDIR_TRIM]) + update_runtime(td, elapsed_us, DDIR_TRIM); + fio_gettime(&td->start, NULL); + fio_mutex_up(stat_mutex); if (td->error || td->terminate) break; @@ -1467,68 +1692,46 @@ static void *thread_main(void *data) (td->io_ops->flags & FIO_UNIDIR)) continue; - clear_io_state(td); + clear_io_state(td, 0); fio_gettime(&td->start, NULL); do_verify(td, verify_bytes); - td->ts.runtime[DDIR_READ] += utime_since_now(&td->start); + /* + * See comment further up for why this is done here. + */ + check_update_rusage(td); + + fio_mutex_down(stat_mutex); + update_runtime(td, elapsed_us, DDIR_READ); + fio_gettime(&td->start, NULL); + fio_mutex_up(stat_mutex); if (td->error || td->terminate) break; } update_rusage_stat(td); - td->ts.runtime[DDIR_READ] = (td->ts.runtime[DDIR_READ] + 999) / 1000; - td->ts.runtime[DDIR_WRITE] = (td->ts.runtime[DDIR_WRITE] + 999) / 1000; - td->ts.runtime[DDIR_TRIM] = (td->ts.runtime[DDIR_TRIM] + 999) / 1000; td->ts.total_run_time = mtime_since_now(&td->epoch); td->ts.io_bytes[DDIR_READ] = td->io_bytes[DDIR_READ]; td->ts.io_bytes[DDIR_WRITE] = td->io_bytes[DDIR_WRITE]; td->ts.io_bytes[DDIR_TRIM] = td->io_bytes[DDIR_TRIM]; + if (td->o.verify_state_save && !(td->flags & TD_F_VSTATE_SAVED) && + (td->o.verify != VERIFY_NONE && td_write(td))) + verify_save_state(td->thread_number); + fio_unpin_memory(td); - fio_mutex_down(writeout_mutex); - finalize_logs(td); - if (td->bw_log) { - if (o->bw_log_file) { - finish_log_named(td, td->bw_log, - o->bw_log_file, "bw"); - } else - finish_log(td, td->bw_log, "bw"); - } - if (td->lat_log) { - if (o->lat_log_file) { - finish_log_named(td, td->lat_log, - o->lat_log_file, "lat"); - } else - finish_log(td, td->lat_log, "lat"); - } - if (td->slat_log) { - if (o->lat_log_file) { - finish_log_named(td, td->slat_log, - o->lat_log_file, "slat"); - } else - finish_log(td, td->slat_log, "slat"); - } - if (td->clat_log) { - if (o->lat_log_file) { - finish_log_named(td, td->clat_log, - o->lat_log_file, "clat"); - } else - finish_log(td, td->clat_log, "clat"); - } - if (td->iops_log) { - if (o->iops_log_file) { - finish_log_named(td, td->iops_log, - o->iops_log_file, "iops"); - } else - finish_log(td, td->iops_log, "iops"); - } + fio_writeout_logs(td); + + if (o->io_submit_mode == IO_MODE_OFFLOAD) + workqueue_exit(&td->io_wq); + + if (td->flags & TD_F_COMPRESS_LOG) + tp_exit(&td->tp_data); - fio_mutex_up(writeout_mutex); if (o->exec_postrun) exec_string(o, o->exec_postrun, (const char *)"postrun"); @@ -1547,11 +1750,12 @@ err: cleanup_io_u(td); close_ioengine(td); cgroup_shutdown(td, &cgroup_mnt); + verify_free_state(td); - if (o->cpumask_set) { - int ret = fio_cpuset_exit(&o->cpumask); - - td_verror(td, ret, "fio_cpuset_exit"); + if (fio_option_is_set(o, cpumask)) { + ret = fio_cpuset_exit(&o->cpumask); + if (ret) + td_verror(td, ret, "fio_cpuset_exit"); } /* @@ -1560,13 +1764,17 @@ err: if (o->write_iolog_file) write_iolog_close(td); - fio_mutex_remove(td->rusage_sem); - td->rusage_sem = NULL; - fio_mutex_remove(td->mutex); td->mutex = NULL; td_set_runstate(td, TD_EXITED); + + /* + * Do this last after setting our runstate to exited, so we + * know that the stat thread is signaled. + */ + check_update_rusage(td); + return (void *) (uintptr_t) td->error; } @@ -1580,7 +1788,7 @@ static int fork_main(int shmid, int offset) struct thread_data *td; void *data, *ret; -#ifndef __hpux +#if !defined(__hpux) && !defined(CONFIG_NO_SHM) data = shmat(shmid, NULL, 0); if (data == (void *) -1) { int __err = errno; @@ -1601,6 +1809,13 @@ static int fork_main(int shmid, int offset) return (int) (uintptr_t) ret; } +static void dump_td_info(struct thread_data *td) +{ + log_err("fio: job '%s' hasn't exited in %lu seconds, it appears to " + "be stuck. Doing forceful exit of this job.\n", td->o.name, + (unsigned long) time_since_now(&td->terminate_time)); +} + /* * Run over the job map and reap the threads that have exited, if any. */ @@ -1678,6 +1893,17 @@ static void reap_threads(unsigned int *nr_running, unsigned int *t_rate, } } + /* + * If the job is stuck, do a forceful timeout of it and + * move on. + */ + if (td->terminate && + time_since_now(&td->terminate_time) >= FIO_REAP_TIMEOUT) { + dump_td_info(td); + td_set_runstate(td, TD_REAPED); + goto reaped; + } + /* * thread is not dead, continue */ @@ -1701,12 +1927,105 @@ reaped: fio_terminate_threads(TERMINATE_ALL); } +static bool __check_trigger_file(void) +{ + struct stat sb; + + if (!trigger_file) + return false; + + if (stat(trigger_file, &sb)) + return false; + + if (unlink(trigger_file) < 0) + log_err("fio: failed to unlink %s: %s\n", trigger_file, + strerror(errno)); + + return true; +} + +static bool trigger_timedout(void) +{ + if (trigger_timeout) + return time_since_genesis() >= trigger_timeout; + + return false; +} + +void exec_trigger(const char *cmd) +{ + int ret; + + if (!cmd) + return; + + ret = system(cmd); + if (ret == -1) + log_err("fio: failed executing %s trigger\n", cmd); +} + +void check_trigger_file(void) +{ + if (__check_trigger_file() || trigger_timedout()) { + if (nr_clients) + fio_clients_send_trigger(trigger_remote_cmd); + else { + verify_save_state(IO_LIST_ALL); + fio_terminate_threads(TERMINATE_ALL); + exec_trigger(trigger_cmd); + } + } +} + +static int fio_verify_load_state(struct thread_data *td) +{ + int ret; + + if (!td->o.verify_state) + return 0; + + if (is_backend) { + void *data; + int ver; + + ret = fio_server_get_verify_state(td->o.name, + td->thread_number - 1, &data, &ver); + if (!ret) + verify_convert_assign_state(td, data, ver); + } else + ret = verify_load_state(td, "local"); + + return ret; +} + static void do_usleep(unsigned int usecs) { check_for_running_stats(); + check_trigger_file(); usleep(usecs); } +static bool check_mount_writes(struct thread_data *td) +{ + struct fio_file *f; + unsigned int i; + + if (!td_write(td) || td->o.allow_mounted_write) + return false; + + for_each_file(td, f, i) { + if (f->filetype != FIO_TYPE_BD) + continue; + if (device_is_mounted(f->file_name)) + goto mounted; + } + + return false; +mounted: + log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name); + return true; +} + /* * Main function for kicking off and reaping jobs, as needed. */ @@ -1725,13 +2044,15 @@ static void run_threads(void) nr_thread = nr_process = 0; for_each_td(td, i) { + if (check_mount_writes(td)) + return; if (td->o.use_thread) nr_thread++; else nr_process++; } - if (output_format == FIO_OUTPUT_NORMAL) { + if (output_format & FIO_OUTPUT_NORMAL) { log_info("Starting "); if (nr_thread) log_info("%d thread%s", nr_thread, @@ -1743,7 +2064,7 @@ static void run_threads(void) nr_process > 1 ? "es" : ""); } log_info("\n"); - fflush(stdout); + log_info_flush(); } todo = thread_number; @@ -1757,12 +2078,16 @@ static void run_threads(void) if (!td->o.create_serialize) continue; + if (fio_verify_load_state(td)) + goto reap; + /* * do file setup here so it happens sequentially, * we don't want X number of threads getting their * client data interspersed on disk */ if (setup_files(td)) { +reap: exit_value++; if (td->error) log_err("fio: pid=%d, err=%d/%s\n", @@ -1950,57 +2275,71 @@ static void run_threads(void) update_io_ticks(); } -void wait_for_disk_thread_exit(void) +static void wait_for_helper_thread_exit(void) { - fio_mutex_down(disk_thread_mutex); + void *ret; + + helper_exit = 1; + pthread_cond_signal(&helper_cond); + pthread_join(helper_thread, &ret); } static void free_disk_util(void) { - disk_util_start_exit(); - wait_for_disk_thread_exit(); disk_util_prune_entries(); + + pthread_cond_destroy(&helper_cond); } -static void *disk_thread_main(void *data) +static void *helper_thread_main(void *data) { int ret = 0; fio_mutex_up(startup_mutex); - while (threads && !ret) { - usleep(DISK_UTIL_MSEC * 1000); - if (!threads) - break; + while (!ret) { + uint64_t sec = DISK_UTIL_MSEC / 1000; + uint64_t nsec = (DISK_UTIL_MSEC % 1000) * 1000000; + struct timespec ts; + struct timeval tv; + + gettimeofday(&tv, NULL); + ts.tv_sec = tv.tv_sec + sec; + ts.tv_nsec = (tv.tv_usec * 1000) + nsec; + + if (ts.tv_nsec >= 1000000000ULL) { + ts.tv_nsec -= 1000000000ULL; + ts.tv_sec++; + } + + pthread_cond_timedwait(&helper_cond, &helper_lock, &ts); + ret = update_io_ticks(); + if (helper_do_stat) { + helper_do_stat = 0; + __show_running_run_stats(); + } + if (!is_backend) print_thread_status(); } - fio_mutex_up(disk_thread_mutex); return NULL; } -static int create_disk_util_thread(void) +static int create_helper_thread(void) { int ret; setup_disk_util(); - disk_thread_mutex = fio_mutex_init(FIO_MUTEX_LOCKED); - - ret = pthread_create(&disk_util_thread, NULL, disk_thread_main, NULL); - if (ret) { - fio_mutex_remove(disk_thread_mutex); - log_err("Can't create disk util thread: %s\n", strerror(ret)); - return 1; - } + pthread_cond_init(&helper_cond, NULL); + pthread_mutex_init(&helper_lock, NULL); - ret = pthread_detach(disk_util_thread); + ret = pthread_create(&helper_thread, NULL, helper_thread_main, NULL); if (ret) { - fio_mutex_remove(disk_thread_mutex); - log_err("Can't detatch disk util thread: %s\n", strerror(ret)); + log_err("Can't create helper thread: %s\n", strerror(ret)); return 1; } @@ -2025,40 +2364,49 @@ int fio_backend(void) return 0; if (write_bw_log) { - setup_log(&agg_io_log[DDIR_READ], 0, IO_LOG_TYPE_BW); - setup_log(&agg_io_log[DDIR_WRITE], 0, IO_LOG_TYPE_BW); - setup_log(&agg_io_log[DDIR_TRIM], 0, IO_LOG_TYPE_BW); + struct log_params p = { + .log_type = IO_LOG_TYPE_BW, + }; + + setup_log(&agg_io_log[DDIR_READ], &p, "agg-read_bw.log"); + setup_log(&agg_io_log[DDIR_WRITE], &p, "agg-write_bw.log"); + setup_log(&agg_io_log[DDIR_TRIM], &p, "agg-trim_bw.log"); } startup_mutex = fio_mutex_init(FIO_MUTEX_LOCKED); if (startup_mutex == NULL) return 1; - writeout_mutex = fio_mutex_init(FIO_MUTEX_UNLOCKED); - if (writeout_mutex == NULL) - return 1; set_genesis_time(); stat_init(); - create_disk_util_thread(); + create_helper_thread(); cgroup_list = smalloc(sizeof(*cgroup_list)); INIT_FLIST_HEAD(cgroup_list); run_threads(); + wait_for_helper_thread_exit(); + if (!fio_abort) { - show_run_stats(); + __show_run_stats(); if (write_bw_log) { - __finish_log(agg_io_log[DDIR_READ], "agg-read_bw.log"); - __finish_log(agg_io_log[DDIR_WRITE], - "agg-write_bw.log"); - __finish_log(agg_io_log[DDIR_TRIM], - "agg-write_bw.log"); + for (i = 0; i < DDIR_RWDIR_CNT; i++) { + struct io_log *log = agg_io_log[i]; + + flush_log(log, 0); + free_log(log); + } } } - for_each_td(td, i) + for_each_td(td, i) { fio_options_free(td); + if (td->rusage_sem) { + fio_mutex_remove(td->rusage_sem); + td->rusage_sem = NULL; + } + } free_disk_util(); cgroup_kill(cgroup_list); @@ -2066,8 +2414,6 @@ int fio_backend(void) sfree(cgroup_mnt); fio_mutex_remove(startup_mutex); - fio_mutex_remove(writeout_mutex); - fio_mutex_remove(disk_thread_mutex); stat_exit(); return exit_value; }