X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=backend.c;h=3df0133f3e56d1c2648580fccaf1b870847992f4;hp=180a48747607294d8a14dd8552677b18bbad6bef;hb=67f28a8cf74bf64c1c35fde49a425d62bb143fa8;hpb=20876c53b5d32f2da9049af5e7fb102133946981 diff --git a/backend.c b/backend.c index 180a4874..6bf5d679 100644 --- a/backend.c +++ b/backend.c @@ -35,6 +35,7 @@ #include #include #include +#include #include "fio.h" #ifndef FIO_NO_HAVE_SHM_H @@ -48,15 +49,17 @@ #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 "workqueue.h" +#include "lib/mountcheck.h" +#include "rate-submit.h" +#include "helper_thread.h" -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; @@ -72,7 +75,6 @@ unsigned int stat_number = 0; int shm_id = 0; int temp_stall_ts; unsigned long done_secs = 0; -volatile int disk_util_exit = 0; #define PAGE_ALIGN(buf) \ (char *) (((uintptr_t) (buf) + page_mask) & ~page_mask) @@ -86,7 +88,7 @@ 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; } @@ -94,7 +96,7 @@ static void sig_int(int sig) } } -static void sig_show_status(int sig) +void sig_show_status(int sig) { show_running_run_stats(); } @@ -137,8 +139,8 @@ 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; @@ -151,13 +153,13 @@ 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]; @@ -171,24 +173,28 @@ 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 { @@ -198,14 +204,19 @@ 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; } } } @@ -214,19 +225,18 @@ 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; @@ -243,7 +253,7 @@ 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; @@ -264,27 +274,27 @@ 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: @@ -292,25 +302,27 @@ 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 (td_io_commit(td)) + return true; + 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) @@ -339,20 +351,37 @@ 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; - if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000) - return 1; + return false; + if (utime_since(&td->epoch, t) >= td->o.timeout) + 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; @@ -365,7 +394,7 @@ 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)) { /* @@ -375,26 +404,26 @@ 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) @@ -406,13 +435,171 @@ 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 (td->flags & TD_F_REGROW_LOGS) { + ret = io_u_quiesce(td); + regrow_logs(td); + return 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; + } + + if (td->flags & TD_F_REGROW_LOGS) + regrow_logs(td); + + /* + * when doing I/O (not when verifying), + * check for any errors that are to be ignored + */ + if (!from_verify) + 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; +} +/* + * Unlinks files from thread data fio_file structure + */ +static int unlink_all_files(struct thread_data *td) +{ + struct fio_file *f; + unsigned int i; + int ret = 0; + + for_each_file(td, f, i) { + if (f->filetype != FIO_TYPE_FILE) + continue; + ret = td_io_unlink_file(td, f); + if (ret) + break; + } + + if (ret) + td_verror(td, ret, "unlink_all_files"); + + return ret; +} + /* * 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; @@ -438,12 +625,21 @@ static void do_verify(struct thread_data *td, uint64_t verify_bytes) if (td->error) return; + /* + * verify_state needs to be reset before verification + * proceeds so that expected random seeds match actual + * random seeds in headers. The main loop will reset + * all random number generators if randrepeat is set. + */ + if (!td->o.rand_repeatable) + td_fill_verify_state_seed(td); + td_set_runstate(td, TD_VERIFYING); io_u = NULL; while (!td->terminate) { enum fio_ddir ddir; - int ret2, full; + int full; update_tv_cache(td); check_update_rusage(td); @@ -451,7 +647,7 @@ 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; } } @@ -474,10 +670,16 @@ 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) { + if (IS_ERR_OR_NULL(io_u)) { + io_u = NULL; + ret = FIO_Q_BUSY; + goto reap; + } + /* * We are only interested in the places where * we wrote or trimmed IOs. Turn those into @@ -493,7 +695,7 @@ 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; @@ -508,65 +710,23 @@ 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; /* @@ -574,28 +734,11 @@ sync_done: * completed io_u's first. 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_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; } @@ -606,7 +749,7 @@ sync_done: 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); @@ -615,9 +758,44 @@ sync_done: dprint(FD_VERIFY, "exiting loop\n"); } -static int io_bytes_exceeded(struct thread_data *td) +static bool exceeds_number_ios(struct thread_data *td) +{ + unsigned long long number_ios; + + if (!td->o.number_ios) + return false; + + 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 * 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 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]; @@ -628,7 +806,45 @@ static int io_bytes_exceeded(struct thread_data *td) else bytes = td->this_io_bytes[DDIR_TRIM]; - return bytes >= td->o.size; + 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; } /* @@ -637,25 +853,49 @@ static int io_bytes_exceeded(struct thread_data *td) * * Returns number of bytes written and trimmed. */ -static uint64_t do_io(struct thread_data *td) +static void do_io(struct thread_data *td, uint64_t *bytes_done) { - uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 }; unsigned int i; int ret = 0; - uint64_t bytes_issued = 0; + uint64_t total_bytes, bytes_issued = 0; + + for (i = 0; i < DDIR_RWDIR_CNT; i++) + bytes_done[i] = td->bytes_done[i]; if (in_ramp_time(td)) td_set_runstate(td, TD_RAMP); else td_set_runstate(td, TD_RUNNING); + 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. + */ + 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); @@ -668,7 +908,7 @@ 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; } } @@ -676,12 +916,29 @@ static uint64_t do_io(struct thread_data *td) if (flow_threshold_exceeded(td)) continue; - if (bytes_issued >= (uint64_t) td->o.size) + /* + * Break if we exceeded the bytes. The exception is time + * based runs, but we still need to break out of the loop + * for those to run verification, if enabled. + */ + if (bytes_issued >= total_bytes && + (!td->o.time_based || + (td->o.time_based && td->o.verify != VERIFY_NONE))) break; io_u = get_io_u(td); - if (!io_u) + if (IS_ERR_OR_NULL(io_u)) { + int err = PTR_ERR(io_u); + + io_u = NULL; + if (err == -EBUSY) { + ret = FIO_Q_BUSY; + goto reap; + } + if (td->o.latency_target) + goto reap; break; + } ddir = io_u->ddir; @@ -692,6 +949,18 @@ static uint64_t do_io(struct thread_data *td) */ if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ && ((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); + if (sizeof(int) != sizeof(long *)) + 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) io_u->end_io = verify_io_u_async; else @@ -702,116 +971,72 @@ static uint64_t do_io(struct thread_data *td) else td_set_runstate(td, TD_RUNNING); - 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; - } + /* + * Always log IO before it's issued, so we know the specific + * order of it. The logged unit will track when the IO has + * completed. + */ + if (td_write(td) && io_u->ddir == DDIR_WRITE && + td->o.do_verify && + td->o.verify != VERIFY_NONE && + !td->o.experimental_verify) + log_io_piece(td, io_u); - io_u->xfer_buflen = io_u->resid; - io_u->xfer_buf += bytes; - io_u->offset += bytes; + 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 (ddir_rw(io_u->ddir)) - td->ts.short_io_u[io_u->ddir]++; + if (td->error) + break; - if (io_u->offset == f->real_file_size) - goto sync_done; + workqueue_enqueue(&td->io_wq, &io_u->work); + ret = FIO_Q_QUEUED; - 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 = io_u_sync_complete(td, io_u, bytes_done); - if (ret < 0) - break; - bytes_issued += io_u->xfer_buflen; + if (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. - */ - 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 (exitall_on_terminate) + if (!in_ramp_time(td) && should_check_rate(td)) { + if (check_min_rate(td, &comp_time)) { + if (exitall_on_terminate || td->o.exitall_error) fio_terminate_threads(td->groupid); td_verror(td, EIO, "check_min_rate"); break; } } + if (!in_ramp_time(td) && td->o.latency_target) + lat_target_check(td); if (td->o.thinktime) { unsigned long long b; @@ -839,14 +1064,19 @@ sync_done: 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; } @@ -871,7 +1101,43 @@ sync_done: if (!ddir_rw_sum(td->this_io_bytes)) td->done = 1; - return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM]; + for (i = 0; i < DDIR_RWDIR_CNT; i++) + bytes_done[i] = td->bytes_done[i] - bytes_done[i]; +} + +static void free_file_completion_logging(struct thread_data *td) +{ + struct fio_file *f; + unsigned int i; + + for_each_file(td, f, i) { + if (!f->last_write_comp) + break; + sfree(f->last_write_comp); + } +} + +static int init_file_completion_logging(struct thread_data *td, + unsigned int depth) +{ + struct fio_file *f; + unsigned int i; + + if (td->o.verify == VERIFY_NONE || !td->o.verify_state_save) + return 0; + + for_each_file(td, f, i) { + f->last_write_comp = scalloc(depth, sizeof(uint64_t)); + if (!f->last_write_comp) + goto cleanup; + } + + return 0; + +cleanup: + free_file_completion_logging(td); + log_err("fio: failed to alloc write comp data\n"); + return 1; } static void cleanup_io_u(struct thread_data *td) @@ -891,6 +1157,8 @@ 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); + + free_file_completion_logging(td); } static int init_io_u(struct thread_data *td) @@ -926,7 +1194,8 @@ static int init_io_u(struct thread_data *td) * overflow later. this adjustment may be too much if we get * lucky and the allocator gives us an aligned address. */ - if (td->o.odirect || td->o.mem_align || (td->io_ops->flags & FIO_RAWIO)) + if (td->o.odirect || td->o.mem_align || td->o.oatomic || + (td->io_ops->flags & FIO_RAWIO)) td->orig_buffer_size += page_mask + td->o.mem_align; if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) { @@ -944,7 +1213,7 @@ static int init_io_u(struct thread_data *td) if (data_xfer && allocate_io_mem(td)) return 1; - if (td->o.odirect || td->o.mem_align || + if (td->o.odirect || td->o.mem_align || td->o.oatomic || (td->io_ops->flags & FIO_RAWIO)) p = PAGE_ALIGN(td->orig_buffer) + td->o.mem_align; else @@ -980,7 +1249,7 @@ static int init_io_u(struct thread_data *td) * Fill the buffer with the pattern if we are * going to be doing writes. */ - fill_pattern(td, io_u->buf, max_bs, io_u, 0, 0); + fill_verify_pattern(td, io_u->buf, max_bs, io_u, 0, 0); } } @@ -1006,11 +1275,15 @@ static int init_io_u(struct thread_data *td) p += max_bs; } + if (init_file_completion_logging(td, max_units)) + return 1; + return 0; } static int switch_ioscheduler(struct thread_data *td) { +#ifdef FIO_HAVE_IOSCHED_SWITCH char tmp[256], tmp2[128]; FILE *f; int ret; @@ -1046,12 +1319,26 @@ 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'; + + /* + * Write to "none" entry doesn't fail, so check the result here. + */ + if (!strcmp(tmp, "none")) { + log_err("fio: io scheduler is not tunable\n"); + fclose(f); + return 0; + } sprintf(tmp2, "[%s]", td->o.ioscheduler); if (!strstr(tmp, tmp2)) { @@ -1063,39 +1350,55 @@ static int switch_ioscheduler(struct thread_data *td) fclose(f); return 0; +#else + return 0; +#endif } -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 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; - return 1; + if (fio_files_done(td) && !td->o.io_limit) + return false; + + 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); @@ -1110,19 +1413,70 @@ static int exec_string(struct thread_options *o, const char *string, const char return ret; } +/* + * Dry run to compute correct state of numberio for verification. + */ +static uint64_t do_dry_run(struct thread_data *td) +{ + td_set_runstate(td, TD_RUNNING); + + while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) || + (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) { + struct io_u *io_u; + int ret; + + if (td->terminate || td->done) + break; + + io_u = get_io_u(td); + if (IS_ERR_OR_NULL(io_u)) + break; + + io_u_set(io_u, IO_U_F_FLIGHT); + io_u->error = 0; + io_u->resid = 0; + if (ddir_rw(acct_ddir(io_u))) + td->io_issues[acct_ddir(io_u)]++; + if (ddir_rw(io_u->ddir)) { + io_u_mark_depth(td, 1); + td->ts.total_io_u[io_u->ddir]++; + } + + if (td_write(td) && io_u->ddir == DDIR_WRITE && + td->o.do_verify && + td->o.verify != VERIFY_NONE && + !td->o.experimental_verify) + log_io_piece(td, io_u); + + ret = io_u_sync_complete(td, io_u); + (void) ret; + } + + return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM]; +} + +struct fork_data { + struct thread_data *td; + struct sk_out *sk_out; +}; + /* * 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; - struct thread_data *td = data; + struct fork_data *fd = data; + unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, }; + struct thread_data *td = fd->td; struct thread_options *o = &td->o; - pthread_condattr_t attr; + struct sk_out *sk_out = fd->sk_out; int clear_state; int ret; + sk_out_assign(sk_out); + free(fd); + if (!o->use_thread) { setsid(); td->pid = getpid(); @@ -1141,12 +1495,18 @@ static void *thread_main(void *data) INIT_FLIST_HEAD(&td->verify_list); INIT_FLIST_HEAD(&td->trim_list); INIT_FLIST_HEAD(&td->next_rand_list); - pthread_mutex_init(&td->io_u_lock, NULL); td->io_hist_tree = RB_ROOT; - pthread_condattr_init(&attr); - pthread_cond_init(&td->verify_cond, &attr); - pthread_cond_init(&td->free_cond, &attr); + ret = mutex_cond_init_pshared(&td->io_u_lock, &td->free_cond); + if (ret) { + td_verror(td, ret, "mutex_cond_init_pshared"); + goto err; + } + ret = cond_init_pshared(&td->verify_cond); + if (ret) { + td_verror(td, ret, "mutex_cond_pshared"); + goto err; + } td_set_runstate(td, TD_INITIALIZED); dprint(FD_MUTEX, "up startup_mutex\n"); @@ -1155,13 +1515,6 @@ static void *thread_main(void *data) fio_mutex_down(td->mutex); dprint(FD_MUTEX, "done waiting on td->mutex\n"); - /* - * the ->mutex mutex is now no longer used, close it to avoid - * eating a file descriptor - */ - fio_mutex_remove(td->mutex); - td->mutex = NULL; - /* * A new gid requires privilege, so we need to do this before setting * the uid. @@ -1175,6 +1528,14 @@ static void *thread_main(void *data) goto err; } + /* + * Do this early, we don't want the compress threads to be limited + * to the same CPUs as the IO workers. So do this before we set + * any potential CPU affinity + */ + if (iolog_compress_init(td, sk_out)) + goto err; + /* * If we have a gettimeofday() thread, make sure we exclude that * thread from this job @@ -1186,7 +1547,16 @@ 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) { + log_err("fio: no CPUs set\n"); + log_err("fio: Try increasing number of available CPUs\n"); + td_verror(td, EINVAL, "cpus_split"); + goto err; + } + } ret = fio_setaffinity(td->pid, o->cpumask); if (ret == -1) { td_verror(td, errno, "cpu_set_affinity"); @@ -1196,16 +1566,19 @@ 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"); @@ -1213,14 +1586,17 @@ 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(); @@ -1233,6 +1609,9 @@ static void *thread_main(void *data) break; } + if (mask) + numa_free_nodemask(mask); + } } #endif @@ -1253,7 +1632,8 @@ 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"); @@ -1292,49 +1672,77 @@ static void *thread_main(void *data) fio_verify_init(td); + if (rate_submit_init(td, sk_out)) + 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, 0); - if (clear_state) - clear_io_state(td); + if (o->unlink_each_loop && unlink_all_files(td)) + break; + } prune_io_piece_log(td); - verify_bytes = do_io(td); + if (td->o.verify_only && (td_write(td) || td_rw(td))) + verify_bytes = do_dry_run(td); + else { + uint64_t bytes_done[DDIR_RWDIR_CNT]; - clear_state = 1; + do_io(td, bytes_done); - 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; + if (!ddir_rw_sum(bytes_done)) { + fio_mark_td_terminate(td); + verify_bytes = 0; + } else { + verify_bytes = bytes_done[DDIR_WRITE] + + bytes_done[DDIR_TRIM]; + } } + clear_state = 1; + + /* + * 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; @@ -1343,71 +1751,49 @@ 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; } + td_set_runstate(td, TD_FINISHING); + 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); - 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"); - } + td_writeout_logs(td, true); + + iolog_compress_exit(td); + rate_submit_exit(td); - fio_mutex_up(writeout_mutex); if (o->exec_postrun) exec_string(o, o->exec_postrun, (const char *)"postrun"); - if (exitall_on_terminate) + if (exitall_on_terminate || (o->exitall_error && td->error)) fio_terminate_threads(td->groupid); err: @@ -1422,11 +1808,21 @@ err: cleanup_io_u(td); close_ioengine(td); cgroup_shutdown(td, &cgroup_mnt); + verify_free_state(td); + + if (td->zone_state_index) { + int i; - if (o->cpumask_set) { - int ret = fio_cpuset_exit(&o->cpumask); + for (i = 0; i < DDIR_RWDIR_CNT; i++) + free(td->zone_state_index[i]); + free(td->zone_state_index); + td->zone_state_index = NULL; + } - 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"); } /* @@ -1435,42 +1831,27 @@ 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); - return (void *) (uintptr_t) td->error; -} - - -/* - * We cannot pass the td data into a forked process, so attach the td and - * pass it to the thread worker. - */ -static int fork_main(int shmid, int offset) -{ - struct thread_data *td; - void *data, *ret; - -#ifndef __hpux - data = shmat(shmid, NULL, 0); - if (data == (void *) -1) { - int __err = errno; - perror("shmat"); - return __err; - } -#else /* - * HP-UX inherits shm mappings? + * Do this last after setting our runstate to exited, so we + * know that the stat thread is signaled. */ - data = threads; -#endif + check_update_rusage(td); - td = data + offset * sizeof(struct thread_data); - ret = thread_main(td); - shmdt(data); - return (int) (uintptr_t) ret; + sk_out_drop(); + return (void *) (uintptr_t) td->error; +} + +static void dump_td_info(struct thread_data *td) +{ + log_err("fio: job '%s' (state=%d) hasn't exited in %lu seconds, it " + "appears to be stuck. Doing forceful exit of this job.\n", + td->o.name, td->runstate, + (unsigned long) time_since_now(&td->terminate_time)); } /* @@ -1550,6 +1931,18 @@ 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 && + td->runstate < TD_FSYNCING && + 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 */ @@ -1573,37 +1966,157 @@ 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; + + ret = fio_server_get_verify_state(td->o.name, + td->thread_number - 1, &data); + if (!ret) + verify_assign_state(td, data); + } 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; +} + +static bool waitee_running(struct thread_data *me) +{ + const char *waitee = me->o.wait_for; + const char *self = me->o.name; + struct thread_data *td; + int i; + + if (!waitee) + return false; + + for_each_td(td, i) { + if (!strcmp(td->o.name, self) || strcmp(td->o.name, waitee)) + continue; + + if (td->runstate < TD_EXITED) { + dprint(FD_PROCESS, "%s fenced by %s(%s)\n", + self, td->o.name, + runstate_to_name(td->runstate)); + return true; + } + } + + dprint(FD_PROCESS, "%s: %s completed, can run\n", self, waitee); + return false; +} + /* * Main function for kicking off and reaping jobs, as needed. */ -static void run_threads(void) +static void run_threads(struct sk_out *sk_out) { struct thread_data *td; - unsigned long spent; unsigned int i, todo, nr_running, m_rate, t_rate, nr_started; + uint64_t spent; if (fio_gtod_offload && fio_start_gtod_thread()) return; - + fio_idle_prof_init(); set_sig_handlers(); 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, @@ -1615,7 +2128,7 @@ static void run_threads(void) nr_process > 1 ? "es" : ""); } log_info("\n"); - fflush(stdout); + log_info_flush(); } todo = thread_number; @@ -1629,12 +2142,16 @@ 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", @@ -1666,6 +2183,7 @@ static void run_threads(void) struct thread_data *map[REAL_MAX_JOBS]; struct timeval this_start; int this_jobs = 0, left; + struct fork_data *fd; /* * create threads (TD_NOT_CREATED -> TD_CREATED) @@ -1684,9 +2202,9 @@ static void run_threads(void) } if (td->o.start_delay) { - spent = mtime_since_genesis(); + spent = utime_since_genesis(); - if (td->o.start_delay * 1000 > spent) + if (td->o.start_delay > spent) continue; } @@ -1696,6 +2214,12 @@ static void run_threads(void) break; } + if (waitee_running(td)) { + dprint(FD_PROCESS, "%s: waiting for %s\n", + td->o.name, td->o.wait_for); + continue; + } + init_disk_util(td); td->rusage_sem = fio_mutex_init(FIO_MUTEX_LOCKED); @@ -1709,15 +2233,20 @@ static void run_threads(void) map[this_jobs++] = td; nr_started++; + fd = calloc(1, sizeof(*fd)); + fd->td = td; + fd->sk_out = sk_out; + if (td->o.use_thread) { int ret; dprint(FD_PROCESS, "will pthread_create\n"); ret = pthread_create(&td->thread, NULL, - thread_main, td); + thread_main, fd); if (ret) { log_err("pthread_create: %s\n", strerror(ret)); + free(fd); nr_started--; break; } @@ -1730,14 +2259,15 @@ static void run_threads(void) dprint(FD_PROCESS, "will fork\n"); pid = fork(); if (!pid) { - int ret = fork_main(shm_id, i); + int ret; + ret = (int)(uintptr_t)thread_main(fd); _exit(ret); } else if (i == fio_debug_jobno) *fio_debug_jobp = pid; } dprint(FD_MUTEX, "wait on startup_mutex\n"); - if (fio_mutex_down_timeout(startup_mutex, 10)) { + if (fio_mutex_down_timeout(startup_mutex, 10000)) { log_err("fio: job startup hung? exiting.\n"); fio_terminate_threads(TERMINATE_ALL); fio_abort = 1; @@ -1822,67 +2352,13 @@ static void run_threads(void) update_io_ticks(); } -void wait_for_disk_thread_exit(void) -{ - fio_mutex_down(disk_thread_mutex); -} - static void free_disk_util(void) { - disk_util_start_exit(); - wait_for_disk_thread_exit(); disk_util_prune_entries(); + helper_thread_destroy(); } -static void *disk_thread_main(void *data) -{ - int ret = 0; - - fio_mutex_up(startup_mutex); - - while (threads && !ret) { - usleep(DISK_UTIL_MSEC * 1000); - if (!threads) - break; - ret = update_io_ticks(); - - if (!is_backend) - print_thread_status(); - } - - fio_mutex_up(disk_thread_mutex); - return NULL; -} - -static int create_disk_util_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; - } - - ret = pthread_detach(disk_util_thread); - if (ret) { - fio_mutex_remove(disk_thread_mutex); - log_err("Can't detatch disk util thread: %s\n", strerror(ret)); - return 1; - } - - dprint(FD_MUTEX, "wait on startup_mutex\n"); - fio_mutex_down(startup_mutex); - dprint(FD_MUTEX, "done waiting on startup_mutex\n"); - return 0; -} - -int fio_backend(void) +int fio_backend(struct sk_out *sk_out) { struct thread_data *td; int i; @@ -1897,40 +2373,49 @@ 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(); + helper_thread_create(startup_mutex, sk_out); cgroup_list = smalloc(sizeof(*cgroup_list)); INIT_FLIST_HEAD(cgroup_list); - run_threads(); + run_threads(sk_out); + + 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, false); + 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); @@ -1938,8 +2423,6 @@ 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; }