X-Git-Url: https://git.kernel.dk/?a=blobdiff_plain;f=fio.c;h=43197b76a7649c54279f2dc82da036f9e023de09;hb=01743ee1718e7ec4b16ae3e53c8f64900c6052cc;hp=3e55f92b31f49199f444ac62c5ce87890c7d85f6;hpb=22819ec237297fc39435ed566bee01a4225bfb39;p=fio.git diff --git a/fio.c b/fio.c index 3e55f92b..43197b76 100644 --- a/fio.c +++ b/fio.c @@ -35,20 +35,26 @@ #include #include "fio.h" -#include "os.h" +#include "hash.h" +#include "smalloc.h" -static unsigned long page_mask; +unsigned long page_mask; +unsigned long page_size; #define ALIGN(buf) \ (char *) (((unsigned long) (buf) + page_mask) & ~page_mask) int groupid = 0; int thread_number = 0; +int nr_process = 0; +int nr_thread = 0; int shm_id = 0; int temp_stall_ts; +unsigned long done_secs = 0; -static volatile int startup_sem; +static struct fio_mutex *startup_mutex; static volatile int fio_abort; static int exit_value; +static struct itimerval itimer; struct io_log *agg_io_log[2]; @@ -57,48 +63,101 @@ struct io_log *agg_io_log[2]; static inline void td_set_runstate(struct thread_data *td, int runstate) { + if (td->runstate == runstate) + return; + + dprint(FD_PROCESS, "pid=%d: runstate %d -> %d\n", (int) td->pid, + td->runstate, runstate); td->runstate = runstate; } -static void terminate_threads(int group_id, int forced_kill) +static void terminate_threads(int group_id) { struct thread_data *td; int i; + dprint(FD_PROCESS, "terminate group_id=%d\n", group_id); + for_each_td(td, i) { if (group_id == TERMINATE_ALL || groupid == td->groupid) { + dprint(FD_PROCESS, "setting terminate on %s/%d\n", + td->o.name, (int) td->pid); td->terminate = 1; - td->start_delay = 0; - if (forced_kill) - td_set_runstate(td, TD_EXITED); + td->o.start_delay = 0; + + /* + * if the thread is running, just let it exit + */ + if (td->runstate < TD_RUNNING) + kill(td->pid, SIGQUIT); + else { + struct ioengine_ops *ops = td->io_ops; + + if (ops && (ops->flags & FIO_SIGQUIT)) + kill(td->pid, SIGQUIT); + } } } } -static void sig_handler(int sig) +static void status_timer_arm(void) { - switch (sig) { - case SIGALRM: - update_io_ticks(); - disk_util_timer_arm(); - print_thread_status(); - break; - default: - printf("\nfio: terminating on signal %d\n", sig); - fflush(stdout); - terminate_threads(TERMINATE_ALL, 0); - break; + itimer.it_value.tv_sec = 0; + itimer.it_value.tv_usec = DISK_UTIL_MSEC * 1000; + setitimer(ITIMER_REAL, &itimer, NULL); +} + +static void sig_alrm(int sig) +{ + if (threads) { + update_io_ticks(); + print_thread_status(); + status_timer_arm(); + } +} + +static void sig_int(int sig) +{ + if (threads) { + printf("\nfio: terminating on signal %d\n", sig); + fflush(stdout); + terminate_threads(TERMINATE_ALL); } } +/* + * We need to rearm on BSD/solaris. Switch this to sigaction in the future... + */ +static void set_sig_handlers(void) +{ + struct sigaction act; + + memset(&act, 0, sizeof(act)); + act.sa_handler = sig_alrm; + act.sa_flags = SA_RESTART; + sigaction(SIGALRM, &act, NULL); + + memset(&act, 0, sizeof(act)); + act.sa_handler = sig_int; + act.sa_flags = SA_RESTART; + sigaction(SIGINT, &act, NULL); +} + /* * Check if we are above the minimum rate given. */ static int check_min_rate(struct thread_data *td, struct timeval *now) { + unsigned long long bytes = 0; + unsigned long iops = 0; unsigned long spent; unsigned long rate; - int ddir = td->ddir; + + /* + * No minimum rate set, always ok + */ + if (!td->o.ratemin && !td->o.rate_iops_min) + return 0; /* * allow a 2 second settle period in the beginning @@ -106,98 +165,120 @@ static int check_min_rate(struct thread_data *td, struct timeval *now) if (mtime_since(&td->start, now) < 2000) return 0; + if (td_read(td)) { + iops += td->io_blocks[DDIR_READ]; + bytes += td->this_io_bytes[DDIR_READ]; + } + if (td_write(td)) { + iops += td->io_blocks[DDIR_WRITE]; + bytes += td->this_io_bytes[DDIR_WRITE]; + } + /* * if rate blocks is set, sample is running */ - if (td->rate_bytes) { + if (td->rate_bytes || td->rate_blocks) { spent = mtime_since(&td->lastrate, now); - if (spent < td->ratecycle) + if (spent < td->o.ratecycle) return 0; - rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent; - if (rate < td->ratemin) { - fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate); - return 1; + if (td->o.rate) { + /* + * check bandwidth specified rate + */ + if (bytes < td->rate_bytes) { + log_err("%s: min rate %u not met\n", td->o.name, + td->o.ratemin); + return 1; + } else { + rate = (bytes - td->rate_bytes) / spent; + if (rate < td->o.ratemin || + bytes < td->rate_bytes) { + log_err("%s: min rate %u not met, got" + " %luKiB/sec\n", td->o.name, + td->o.ratemin, rate); + return 1; + } + } + } else { + /* + * checks iops specified rate + */ + if (iops < td->o.rate_iops) { + log_err("%s: min iops rate %u not met\n", + td->o.name, td->o.rate_iops); + return 1; + } else { + rate = (iops - td->rate_blocks) / spent; + if (rate < td->o.rate_iops_min || + iops < td->rate_blocks) { + log_err("%s: min iops rate %u not met," + " got %lu\n", td->o.name, + td->o.rate_iops_min, + rate); + } + } } } - td->rate_bytes = td->this_io_bytes[ddir]; + td->rate_bytes = bytes; + td->rate_blocks = iops; memcpy(&td->lastrate, now, sizeof(*now)); return 0; } static inline int runtime_exceeded(struct thread_data *td, struct timeval *t) { - if (!td->timeout) + if (!td->o.timeout) return 0; - if (mtime_since(&td->epoch, t) >= td->timeout * 1000) + if (mtime_since(&td->epoch, t) >= td->o.timeout * 1000) return 1; return 0; } -static struct fio_file *get_next_file(struct thread_data *td) -{ - unsigned int old_next_file = td->next_file; - struct fio_file *f; - - do { - f = &td->files[td->next_file]; - - td->next_file++; - if (td->next_file >= td->nr_files) - td->next_file = 0; - - if (f->fd != -1) - break; - - f = NULL; - } while (td->next_file != old_next_file); - - return f; -} - /* * When job exits, we can cancel the in-flight IO if we are using async * io. Attempt to do so. */ static void cleanup_pending_aio(struct thread_data *td) { - struct timespec ts = { .tv_sec = 0, .tv_nsec = 0}; - struct list_head *entry, *n; - struct io_completion_data icd; + struct flist_head *entry, *n; struct io_u *io_u; int r; /* * get immediately available events, if any */ - r = td_io_getevents(td, 0, td->cur_depth, &ts); - if (r > 0) { - init_icd(&icd, NULL, r); - ios_completed(td, &icd); - } + r = io_u_queued_complete(td, 0); + if (r < 0) + return; /* * now cancel remaining active events */ if (td->io_ops->cancel) { - list_for_each_safe(entry, n, &td->io_u_busylist) { - io_u = list_entry(entry, struct io_u, list); + flist_for_each_safe(entry, n, &td->io_u_busylist) { + io_u = flist_entry(entry, struct io_u, list); - r = td->io_ops->cancel(td, io_u); - if (!r) + /* + * if the io_u isn't in flight, then that generally + * means someone leaked an io_u. complain but fix + * it up, so we don't stall here. + */ + if ((io_u->flags & IO_U_F_FLIGHT) == 0) { + log_err("fio: non-busy IO on busy list\n"); put_io_u(td, io_u); + } else { + r = td->io_ops->cancel(td, io_u); + if (!r) + put_io_u(td, io_u); + } } } - if (td->cur_depth) { - r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL); - if (r > 0) { - init_icd(&icd, NULL, r); - ios_completed(td, &icd); - } - } + if (td->cur_depth) + r = io_u_queued_complete(td, td->cur_depth); } /* @@ -207,7 +288,6 @@ static void cleanup_pending_aio(struct thread_data *td) static int fio_io_sync(struct thread_data *td, struct fio_file *f) { struct io_u *io_u = __get_io_u(td); - struct io_completion_data icd; int ret; if (!io_u) @@ -221,232 +301,265 @@ static int fio_io_sync(struct thread_data *td, struct fio_file *f) return 1; } +requeue: ret = td_io_queue(td, io_u); if (ret < 0) { - td_verror(td, io_u->error); + td_verror(td, io_u->error, "td_io_queue"); put_io_u(td, io_u); return 1; } else if (ret == FIO_Q_QUEUED) { - ret = td_io_getevents(td, 1, td->cur_depth, NULL); - if (ret < 0) { - td_verror(td, -ret); + if (io_u_queued_complete(td, 1) < 0) return 1; - } - - init_icd(&icd, NULL, ret); - ios_completed(td, &icd); - if (icd.error) { - td_verror(td, icd.error); - return 1; - } } else if (ret == FIO_Q_COMPLETED) { if (io_u->error) { - td_verror(td, io_u->error); + td_verror(td, io_u->error, "td_io_queue"); return 1; } - init_icd(&icd, NULL, 1); - io_completed(td, io_u, &icd); - put_io_u(td, io_u); + if (io_u_sync_complete(td, io_u) < 0) + return 1; + } else if (ret == FIO_Q_BUSY) { + if (td_io_commit(td)) + return 1; + goto requeue; } return 0; } /* - * The main verify engine. Runs over the writes we previusly submitted, + * 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) { struct fio_file *f; struct io_u *io_u; - int ret, i, min_events; + int ret, min_events; + unsigned int i; /* * sync io first and invalidate cache, to make sure we really * read from disk. */ for_each_file(td, f, i) { - fio_io_sync(td, f); - file_invalidate_cache(td, f); + if (!(f->flags & FIO_FILE_OPEN)) + continue; + if (fio_io_sync(td, f)) + break; + if (file_invalidate_cache(td, f)) + break; } + if (td->error) + return; + td_set_runstate(td, TD_VERIFYING); io_u = NULL; while (!td->terminate) { - struct io_completion_data icd; - struct timespec *timeout; + int ret2; io_u = __get_io_u(td); if (!io_u) break; - if (runtime_exceeded(td, &io_u->start_time)) + if (runtime_exceeded(td, &io_u->start_time)) { + put_io_u(td, io_u); + td->terminate = 1; break; + } - if (get_next_verify(td, io_u)) + if (get_next_verify(td, io_u)) { + put_io_u(td, io_u); break; + } - if (td_io_prep(td, io_u)) + if (td_io_prep(td, io_u)) { + put_io_u(td, io_u); break; + } -requeue: - ret = td_io_queue(td, io_u); + io_u->end_io = verify_io_u; + ret = td_io_queue(td, io_u); switch (ret) { case FIO_Q_COMPLETED: if (io_u->error) ret = -io_u->error; - if (io_u->xfer_buflen != io_u->resid && io_u->resid) { + else if (io_u->resid) { int bytes = io_u->xfer_buflen - io_u->resid; + struct fio_file *f = io_u->file; + + /* + * zero read, fail + */ + if (!bytes) { + td_verror(td, ENODATA, "full resid"); + put_io_u(td, io_u); + break; + } io_u->xfer_buflen = io_u->resid; io_u->xfer_buf += bytes; - goto requeue; - } - init_icd(&icd, verify_io_u, 1); - io_completed(td, io_u, &icd); - if (icd.error) { - ret = icd.error; - break; + io_u->offset += bytes; + + td->ts.short_io_u[io_u->ddir]++; + + 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); + if (ret < 0) + break; } - put_io_u(td, io_u); 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_verror(td, -ret, "td_io_queue"); break; } - if (ret < 0) + if (ret < 0 || td->error) break; /* * if we can queue more, do so. but check if there are * completed io_u's first. */ - if (queue_full(td)) { - timeout = NULL; - min_events = 1; - } else { - struct timespec ts; - - ts.tv_sec = 0; - ts.tv_nsec = 0; - timeout = &ts; - min_events = 0; + min_events = 0; + if (queue_full(td) || ret == FIO_Q_BUSY) { + if (td->cur_depth >= td->o.iodepth_low) + min_events = td->cur_depth - td->o.iodepth_low; + if (!min_events) + min_events = 1; } /* * Reap required number of io units, if any, and do the * verification on them through the callback handler */ - ret = td_io_getevents(td, min_events, td->cur_depth, timeout); - if (ret < 0) { - td_verror(td, -ret); + if (io_u_queued_complete(td, min_events) < 0) break; - } else if (!ret) - continue; - - init_icd(&icd, verify_io_u, ret); - ios_completed(td, &icd); - - if (icd.error) { - td_verror(td, icd.error); - break; - } } - if (io_u) - put_io_u(td, io_u); + if (!td->error) { + min_events = td->cur_depth; - if (td->cur_depth) + if (min_events) + ret = io_u_queued_complete(td, min_events); + } else cleanup_pending_aio(td); td_set_runstate(td, TD_RUNNING); } -/* - * Not really an io thread, all it does is burn CPU cycles in the specified - * manner. - */ -static void do_cpuio(struct thread_data *td) -{ - struct timeval e; - int split = 100 / td->cpuload; - int i = 0; - - while (!td->terminate) { - fio_gettime(&e, NULL); - - if (runtime_exceeded(td, &e)) - break; - - if (!(i % split)) - __usec_sleep(10000); - else - usec_sleep(td, 10000); - - i++; - } -} - /* * Main IO worker function. It retrieves io_u's to process and queues * and reaps them, checking for rate and errors along the way. */ static void do_io(struct thread_data *td) { - struct io_completion_data icd; struct timeval s; unsigned long usec; - struct fio_file *f; - int i, ret = 0; + unsigned int i; + int ret = 0; td_set_runstate(td, TD_RUNNING); - while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) { - struct timespec *timeout; + while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->o.size) { + struct timeval comp_time; + long bytes_done = 0; int min_evts = 0; struct io_u *io_u; + int ret2; if (td->terminate) break; - f = get_next_file(td); - if (!f) - break; - - io_u = get_io_u(td, f); + io_u = get_io_u(td); if (!io_u) break; memcpy(&s, &io_u->start_time, sizeof(s)); -requeue: - ret = td_io_queue(td, io_u); + if (runtime_exceeded(td, &s)) { + put_io_u(td, io_u); + td->terminate = 1; + break; + } + + /* + * Add verification end_io handler, if asked to verify + * a previously written file. + */ + if (td->o.verify != VERIFY_NONE && io_u->ddir == DDIR_READ) { + io_u->end_io = verify_io_u; + td_set_runstate(td, TD_VERIFYING); + } 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; - break; - } - if (io_u->xfer_buflen != io_u->resid && io_u->resid) { + if (io_u->error) + ret = -io_u->error; + else if (io_u->resid) { int bytes = io_u->xfer_buflen - io_u->resid; + struct fio_file *f = io_u->file; + + /* + * zero read, fail + */ + if (!bytes) { + td_verror(td, ENODATA, "full resid"); + put_io_u(td, io_u); + break; + } io_u->xfer_buflen = io_u->resid; io_u->xfer_buf += bytes; - goto requeue; + io_u->offset += bytes; + + td->ts.short_io_u[io_u->ddir]++; + + if (io_u->offset == f->real_file_size) + goto sync_done; + + requeue_io_u(td, &io_u); + } else { +sync_done: + fio_gettime(&comp_time, NULL); + bytes_done = io_u_sync_complete(td, io_u); + if (bytes_done < 0) + ret = bytes_done; } - init_icd(&icd, NULL, 1); - io_completed(td, io_u, &icd); - put_io_u(td, io_u); 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); + 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); @@ -454,115 +567,105 @@ requeue: break; } - if (ret < 0) + if (ret < 0 || td->error) break; - add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time)); - - if (ret == FIO_Q_QUEUED) { - if (td->cur_depth < td->iodepth) { - struct timespec ts; - - ts.tv_sec = 0; - ts.tv_nsec = 0; - timeout = &ts; - min_evts = 0; - } else { - timeout = NULL; + /* + * See if we need to complete some commands + */ + if (queue_full(td) || ret == FIO_Q_BUSY) { + min_evts = 0; + if (td->cur_depth >= td->o.iodepth_low) + min_evts = td->cur_depth - td->o.iodepth_low; + if (!min_evts) min_evts = 1; - } - - ret = td_io_getevents(td, min_evts, td->cur_depth, timeout); - if (ret < 0) { - td_verror(td, -ret); - break; - } else if (!ret) - continue; - - init_icd(&icd, NULL, ret); - ios_completed(td, &icd); - if (icd.error) { - td_verror(td, icd.error); + fio_gettime(&comp_time, NULL); + bytes_done = io_u_queued_complete(td, min_evts); + if (bytes_done < 0) break; - } } + if (!bytes_done) + continue; + /* * the rate is batched for now, it should work for batches * of completions except the very first one which may look * a little bursty */ - usec = utime_since(&s, &icd.time); + usec = utime_since(&s, &comp_time); - rate_throttle(td, usec, icd.bytes_done[td->ddir], td->ddir); + rate_throttle(td, usec, bytes_done); - if (check_min_rate(td, &icd.time)) { + if (check_min_rate(td, &comp_time)) { if (exitall_on_terminate) - terminate_threads(td->groupid, 0); - td_verror(td, ENODATA); + terminate_threads(td->groupid); + td_verror(td, ENODATA, "check_min_rate"); break; } - if (runtime_exceeded(td, &icd.time)) - break; - - if (td->thinktime) { + if (td->o.thinktime) { unsigned long long b; b = td->io_blocks[0] + td->io_blocks[1]; - if (!(b % td->thinktime_blocks)) { + if (!(b % td->o.thinktime_blocks)) { int left; - if (td->thinktime_spin) - __usec_sleep(td->thinktime_spin); + if (td->o.thinktime_spin) + __usec_sleep(td->o.thinktime_spin); - left = td->thinktime - td->thinktime_spin; + left = td->o.thinktime - td->o.thinktime_spin; if (left) usec_sleep(td, left); } } } + if (td->o.fill_device && td->error == ENOSPC) { + td->error = 0; + td->terminate = 1; + } if (!td->error) { - if (td->cur_depth) - cleanup_pending_aio(td); + struct fio_file *f; - if (should_fsync(td) && td->end_fsync) { + i = td->cur_depth; + if (i) + ret = io_u_queued_complete(td, i); + + if (should_fsync(td) && td->o.end_fsync) { td_set_runstate(td, TD_FSYNCING); - for_each_file(td, f, i) + + for_each_file(td, f, i) { + if (!(f->flags & FIO_FILE_OPEN)) + continue; fio_io_sync(td, f); + } } - } + } else + cleanup_pending_aio(td); + + /* + * stop job if we failed doing any IO + */ + if ((td->this_io_bytes[0] + td->this_io_bytes[1]) == 0) + td->done = 1; } static void cleanup_io_u(struct thread_data *td) { - struct list_head *entry, *n; + struct flist_head *entry, *n; struct io_u *io_u; - list_for_each_safe(entry, n, &td->io_u_freelist) { - io_u = list_entry(entry, struct io_u, list); + flist_for_each_safe(entry, n, &td->io_u_freelist) { + io_u = flist_entry(entry, struct io_u, list); - list_del(&io_u->list); + flist_del(&io_u->list); free(io_u); } free_io_mem(td); } -/* - * "randomly" fill the buffer contents - */ -static void fill_rand_buf(struct io_u *io_u, int max_bs) -{ - int *ptr = io_u->buf; - - while ((void *) ptr - io_u->buf < max_bs) { - *ptr = rand() * 0x9e370001; - ptr++; - } -} - static int init_io_u(struct thread_data *td) { struct io_u *io_u; @@ -570,39 +673,52 @@ static int init_io_u(struct thread_data *td) int i, max_units; char *p; - if (td->io_ops->flags & FIO_CPUIO) - return 0; + max_units = td->o.iodepth; + max_bs = max(td->o.max_bs[DDIR_READ], td->o.max_bs[DDIR_WRITE]); + td->orig_buffer_size = (unsigned long long) max_bs + * (unsigned long long) max_units; - if (td->io_ops->flags & FIO_SYNCIO) - max_units = 1; - else - max_units = td->iodepth; + if (td->o.mem_type == MEM_SHMHUGE || td->o.mem_type == MEM_MMAPHUGE) { + unsigned long bs; - max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]); - td->orig_buffer_size = max_bs * max_units; + bs = td->orig_buffer_size + td->o.hugepage_size - 1; + td->orig_buffer_size = bs & ~(td->o.hugepage_size - 1); + } - if (td->mem_type == MEM_SHMHUGE || td->mem_type == MEM_MMAPHUGE) - td->orig_buffer_size = (td->orig_buffer_size + td->hugepage_size - 1) & ~(td->hugepage_size - 1); - else - td->orig_buffer_size += page_mask; + if (td->orig_buffer_size != (size_t) td->orig_buffer_size) { + log_err("fio: IO memory too large. Reduce max_bs or iodepth\n"); + return 1; + } if (allocate_io_mem(td)) return 1; - p = ALIGN(td->orig_buffer); + if (td->o.odirect) + p = ALIGN(td->orig_buffer); + else + p = td->orig_buffer; + for (i = 0; i < max_units; i++) { + if (td->terminate) + return 1; io_u = malloc(sizeof(*io_u)); memset(io_u, 0, sizeof(*io_u)); - INIT_LIST_HEAD(&io_u->list); + INIT_FLIST_HEAD(&io_u->list); + + if (!(td->io_ops->flags & FIO_NOIO)) { + io_u->buf = p + max_bs * i; - io_u->buf = p + max_bs * i; - if (td_write(td) || td_rw(td)) - fill_rand_buf(io_u, max_bs); + if (td_write(td) && !td->o.refill_buffers) + io_u_fill_buffer(td, io_u, max_bs); + } io_u->index = i; - list_add(&io_u->list, &td->io_u_freelist); + io_u->flags = IO_U_F_FREE; + flist_add(&io_u->list, &td->io_u_freelist); } + io_u_init_timeout(); + return 0; } @@ -612,23 +728,28 @@ static int switch_ioscheduler(struct thread_data *td) FILE *f; int ret; - if (td->io_ops->flags & FIO_CPUIO) + if (td->io_ops->flags & FIO_DISKLESSIO) return 0; sprintf(tmp, "%s/queue/scheduler", td->sysfs_root); f = fopen(tmp, "r+"); if (!f) { - td_verror(td, errno); + if (errno == ENOENT) { + log_err("fio: os or kernel doesn't support IO scheduler" + " switching\n"); + return 0; + } + td_verror(td, errno, "fopen iosched"); return 1; } /* * Set io scheduler. */ - ret = fwrite(td->ioscheduler, strlen(td->ioscheduler), 1, f); + ret = fwrite(td->o.ioscheduler, strlen(td->o.ioscheduler), 1, f); if (ferror(f) || ret != 1) { - td_verror(td, errno); + td_verror(td, errno, "fwrite"); fclose(f); return 1; } @@ -640,15 +761,15 @@ static int switch_ioscheduler(struct thread_data *td) */ ret = fread(tmp, 1, sizeof(tmp), f); if (ferror(f) || ret < 0) { - td_verror(td, errno); + td_verror(td, errno, "fread"); fclose(f); return 1; } - sprintf(tmp2, "[%s]", td->ioscheduler); + sprintf(tmp2, "[%s]", td->o.ioscheduler); if (!strstr(tmp, tmp2)) { - log_err("fio: io scheduler %s not found\n", td->ioscheduler); - td_verror(td, EINVAL); + log_err("fio: io scheduler %s not found\n", td->o.ioscheduler); + td_verror(td, EINVAL, "iosched_switch"); fclose(f); return 1; } @@ -657,23 +778,59 @@ static int switch_ioscheduler(struct thread_data *td) return 0; } -static void clear_io_state(struct thread_data *td) +static int keep_running(struct thread_data *td) +{ + unsigned long long io_done; + + if (td->done) + return 0; + if (td->o.time_based) + return 1; + if (td->o.loops) { + td->o.loops--; + return 1; + } + + io_done = td->io_bytes[DDIR_READ] + td->io_bytes[DDIR_WRITE] + + td->io_skip_bytes; + if (io_done < td->o.size) + return 1; + + return 0; +} + +static int clear_io_state(struct thread_data *td) { struct fio_file *f; - int i; + unsigned int i; + int ret; - td->stat_io_bytes[0] = td->stat_io_bytes[1] = 0; + td->ts.stat_io_bytes[0] = td->ts.stat_io_bytes[1] = 0; td->this_io_bytes[0] = td->this_io_bytes[1] = 0; td->zone_bytes = 0; + td->rate_bytes = 0; + td->rate_blocks = 0; + td->rw_end_set[0] = td->rw_end_set[1] = 0; - for_each_file(td, f, i) { - f->last_pos = 0; - if (td->io_ops->flags & FIO_SYNCIO) - lseek(f->fd, SEEK_SET, 0); + td->last_was_sync = 0; + + /* + * reset file done count if we are to start over + */ + if (td->o.time_based || td->o.loops) + td->nr_done_files = 0; - if (f->file_map) - memset(f->file_map, 0, f->num_maps * sizeof(long)); + close_files(td); + + ret = 0; + for_each_file(td, f, i) { + f->flags &= ~FIO_FILE_DONE; + ret = td_io_open_file(td, f); + if (ret) + break; } + + return ret; } /* @@ -682,99 +839,134 @@ static void clear_io_state(struct thread_data *td) */ static void *thread_main(void *data) { - unsigned long long runtime[2]; + unsigned long long runtime[2], elapsed; struct thread_data *td = data; + int clear_state; - if (!td->use_thread) + if (!td->o.use_thread) setsid(); td->pid = getpid(); - INIT_LIST_HEAD(&td->io_u_freelist); - INIT_LIST_HEAD(&td->io_u_busylist); - INIT_LIST_HEAD(&td->io_hist_list); - INIT_LIST_HEAD(&td->io_log_list); + dprint(FD_PROCESS, "jobs pid=%d started\n", (int) td->pid); - if (init_io_u(td)) + INIT_FLIST_HEAD(&td->io_u_freelist); + INIT_FLIST_HEAD(&td->io_u_busylist); + INIT_FLIST_HEAD(&td->io_u_requeues); + INIT_FLIST_HEAD(&td->io_log_list); + INIT_FLIST_HEAD(&td->io_hist_list); + td->io_hist_tree = RB_ROOT; + + td_set_runstate(td, TD_INITIALIZED); + fio_mutex_up(startup_mutex); + fio_mutex_down(td->mutex); + + /* + * the ->mutex mutex is now no longer used, close it to avoid + * eating a file descriptor + */ + fio_mutex_remove(td->mutex); + + /* + * May alter parameters that init_io_u() will use, so we need to + * do this first. + */ + if (init_iolog(td)) goto err; - if (fio_setaffinity(td) == -1) { - td_verror(td, errno); + if (init_io_u(td)) goto err; - } - if (init_iolog(td)) + if (td->o.cpumask_set && fio_setaffinity(td) == -1) { + td_verror(td, errno, "cpu_set_affinity"); goto err; + } - if (td->ioprio) { + if (td->ioprio_set) { if (ioprio_set(IOPRIO_WHO_PROCESS, 0, td->ioprio) == -1) { - td_verror(td, errno); + td_verror(td, errno, "ioprio_set"); goto err; } } - if (nice(td->nice) == -1) { - td_verror(td, errno); + if (nice(td->o.nice) == -1) { + td_verror(td, errno, "nice"); goto err; } - if (init_random_state(td)) + if (td->o.ioscheduler && switch_ioscheduler(td)) goto err; - if (td->ioscheduler && switch_ioscheduler(td)) + if (!td->o.create_serialize && setup_files(td)) goto err; - td_set_runstate(td, TD_INITIALIZED); - fio_sem_up(&startup_sem); - fio_sem_down(&td->mutex); - - if (!td->create_serialize && setup_files(td)) + if (td_io_init(td)) goto err; + if (open_files(td)) goto err; - /* - * Do this late, as some IO engines would like to have the - * files setup prior to initializing structures. - */ - if (td_io_init(td)) + if (init_random_map(td)) goto err; - if (td->exec_prerun) { - if (system(td->exec_prerun) < 0) + if (td->o.exec_prerun) { + if (system(td->o.exec_prerun) < 0) goto err; } fio_gettime(&td->epoch, NULL); - getrusage(RUSAGE_SELF, &td->ru_start); + memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch)); + getrusage(RUSAGE_SELF, &td->ts.ru_start); runtime[0] = runtime[1] = 0; - while (td->loops--) { + clear_state = 0; + while (keep_running(td)) { fio_gettime(&td->start, NULL); - memcpy(&td->stat_sample_time, &td->start, sizeof(td->start)); + memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start)); - if (td->ratemin) - memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate)); + if (td->o.ratemin) + memcpy(&td->lastrate, &td->ts.stat_sample_time, + sizeof(td->lastrate)); + + if (clear_state && clear_io_state(td)) + break; - clear_io_state(td); prune_io_piece_log(td); - if (td->io_ops->flags & FIO_CPUIO) - do_cpuio(td); - else - do_io(td); + do_io(td); - runtime[td->ddir] += utime_since_now(&td->start); - if (td_rw(td) && td->io_bytes[td->ddir ^ 1]) - runtime[td->ddir ^ 1] = runtime[td->ddir]; + clear_state = 1; + + if (td_read(td) && td->io_bytes[DDIR_READ]) { + if (td->rw_end_set[DDIR_READ]) + elapsed = utime_since(&td->start, + &td->rw_end[DDIR_READ]); + else + elapsed = utime_since_now(&td->start); + + runtime[DDIR_READ] += elapsed; + } + if (td_write(td) && td->io_bytes[DDIR_WRITE]) { + if (td->rw_end_set[DDIR_WRITE]) + elapsed = utime_since(&td->start, + &td->rw_end[DDIR_WRITE]); + else + elapsed = utime_since_now(&td->start); + + runtime[DDIR_WRITE] += elapsed; + } if (td->error || td->terminate) break; - if (td->verify == VERIFY_NONE) + if (!td->o.do_verify || + td->o.verify == VERIFY_NONE || + (td->io_ops->flags & FIO_UNIDIR)) continue; - clear_io_state(td); + if (clear_io_state(td)) + break; + fio_gettime(&td->start, NULL); do_verify(td); @@ -786,34 +978,43 @@ static void *thread_main(void *data) } update_rusage_stat(td); - fio_gettime(&td->end_time, NULL); - td->runtime[0] = runtime[0] / 1000; - td->runtime[1] = runtime[1] / 1000; - - if (td->bw_log) - finish_log(td, td->bw_log, "bw"); - if (td->slat_log) - finish_log(td, td->slat_log, "slat"); - if (td->clat_log) - finish_log(td, td->clat_log, "clat"); - if (td->write_iolog_file) - write_iolog_close(td); - if (td->exec_postrun) { - if (system(td->exec_postrun) < 0) - log_err("fio: postrun %s failed\n", td->exec_postrun); + td->ts.runtime[0] = (runtime[0] + 999) / 1000; + td->ts.runtime[1] = (runtime[1] + 999) / 1000; + td->ts.total_run_time = mtime_since_now(&td->epoch); + td->ts.io_bytes[0] = td->io_bytes[0]; + td->ts.io_bytes[1] = td->io_bytes[1]; + + if (td->ts.bw_log) + finish_log(td, td->ts.bw_log, "bw"); + if (td->ts.slat_log) + finish_log(td, td->ts.slat_log, "slat"); + if (td->ts.clat_log) + finish_log(td, td->ts.clat_log, "clat"); + if (td->o.exec_postrun) { + if (system(td->o.exec_postrun) < 0) + log_err("fio: postrun %s failed\n", td->o.exec_postrun); } if (exitall_on_terminate) - terminate_threads(td->groupid, 0); + terminate_threads(td->groupid); err: if (td->error) - printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror); - close_files(td); + printf("fio: pid=%d, err=%d/%s\n", (int) td->pid, td->error, + td->verror); + close_and_free_files(td); close_ioengine(td); cleanup_io_u(td); + + /* + * do this very late, it will log file closing as well + */ + if (td->o.write_iolog_file) + write_iolog_close(td); + + options_mem_free(td); td_set_runstate(td, TD_EXITED); - return (void *) td->error; + return (void *) (unsigned long) td->error; } /* @@ -836,7 +1037,7 @@ static int fork_main(int shmid, int offset) td = data + offset * sizeof(struct thread_data); ret = thread_main(td); shmdt(data); - return (int) ret; + return (int) (unsigned long) ret; } /* @@ -845,73 +1046,93 @@ static int fork_main(int shmid, int offset) static void reap_threads(int *nr_running, int *t_rate, int *m_rate) { struct thread_data *td; - int i, cputhreads, pending, status, ret; + int i, cputhreads, realthreads, pending, status, ret; /* * reap exited threads (TD_EXITED -> TD_REAPED) */ - pending = cputhreads = 0; + realthreads = pending = cputhreads = 0; for_each_td(td, i) { + int flags = 0; + /* * ->io_ops is NULL for a thread that has closed its * io engine */ - if (td->io_ops && td->io_ops->flags & FIO_CPUIO) + if (td->io_ops && !strcmp(td->io_ops->name, "cpuio")) cputhreads++; + else + realthreads++; - if (td->runstate < TD_EXITED) { - /* - * check if someone quit or got killed in an unusual way - */ - ret = waitpid(td->pid, &status, WNOHANG); - if (ret < 0) - perror("waitpid"); - else if ((ret == td->pid) && WIFSIGNALED(status)) { - int sig = WTERMSIG(status); - - log_err("fio: pid=%d, got signal=%d\n", td->pid, sig); + if (!td->pid) { + pending++; + continue; + } + if (td->runstate == TD_REAPED) + continue; + if (td->o.use_thread) { + if (td->runstate == TD_EXITED) { td_set_runstate(td, TD_REAPED); goto reaped; } - } - - if (td->runstate != TD_EXITED) { - if (td->runstate < TD_RUNNING) - pending++; - continue; } - if (td->error) - exit_value++; + flags = WNOHANG; + if (td->runstate == TD_EXITED) + flags = 0; - td_set_runstate(td, TD_REAPED); + /* + * check if someone quit or got killed in an unusual way + */ + ret = waitpid(td->pid, &status, flags); + if (ret < 0) { + if (errno == ECHILD) { + log_err("fio: pid=%d disappeared %d\n", + (int) td->pid, td->runstate); + td_set_runstate(td, TD_REAPED); + goto reaped; + } + perror("waitpid"); + } else if (ret == td->pid) { + if (WIFSIGNALED(status)) { + int sig = WTERMSIG(status); - if (td->use_thread) { - long ret; + if (sig != SIGQUIT) + log_err("fio: pid=%d, got signal=%d\n", + (int) td->pid, sig); + td_set_runstate(td, TD_REAPED); + goto reaped; + } + if (WIFEXITED(status)) { + if (WEXITSTATUS(status) && !td->error) + td->error = WEXITSTATUS(status); - if (pthread_join(td->thread, (void *) &ret)) - perror("thread_join"); - } else { - int status; - - ret = waitpid(td->pid, &status, 0); - if (ret < 0) - perror("waitpid"); - else if (WIFEXITED(status) && WEXITSTATUS(status)) { - if (!exit_value) - exit_value++; + td_set_runstate(td, TD_REAPED); + goto reaped; } } + /* + * thread is not dead, continue + */ + pending++; + continue; reaped: (*nr_running)--; - (*m_rate) -= td->ratemin; - (*t_rate) -= td->rate; + (*m_rate) -= td->o.ratemin; + (*t_rate) -= td->o.rate; + if (!td->pid) + pending--; + + if (td->error) + exit_value++; + + done_secs += mtime_since_now(&td->epoch) / 1000; } - if (*nr_running == cputhreads && !pending) - terminate_threads(TERMINATE_ALL, 0); + if (*nr_running == cputhreads && !pending && realthreads) + terminate_threads(TERMINATE_ALL); } /* @@ -927,12 +1148,21 @@ static void run_threads(void) return; if (!terse_output) { - printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : ""); + printf("Starting "); + if (nr_thread) + printf("%d thread%s", nr_thread, + nr_thread > 1 ? "s" : ""); + if (nr_process) { + if (nr_thread) + printf(" and "); + printf("%d process%s", nr_process, + nr_process > 1 ? "es" : ""); + } + printf("\n"); fflush(stdout); } - signal(SIGINT, sig_handler); - signal(SIGALRM, sig_handler); + set_sig_handlers(); todo = thread_number; nr_running = 0; @@ -942,7 +1172,7 @@ static void run_threads(void) for_each_td(td, i) { print_status_init(td->thread_number - 1); - if (!td->create_serialize) { + if (!td->o.create_serialize) { init_disk_util(td); continue; } @@ -955,14 +1185,28 @@ static void run_threads(void) if (setup_files(td)) { exit_value++; if (td->error) - log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror); + log_err("fio: pid=%d, err=%d/%s\n", + (int) td->pid, td->error, td->verror); td_set_runstate(td, TD_REAPED); todo--; + } else { + struct fio_file *f; + unsigned int i; + + /* + * for sharing to work, each job must always open + * its own files. so close them, if we opened them + * for creation + */ + for_each_file(td, f, i) + td_io_close_file(td, f); } init_disk_util(td); } + set_genesis_time(); + while (todo) { struct thread_data *map[MAX_JOBS]; struct timeval this_start; @@ -984,15 +1228,18 @@ static void run_threads(void) continue; } - if (td->start_delay) { + if (td->o.start_delay) { spent = mtime_since_genesis(); - if (td->start_delay * 1000 > spent) + if (td->o.start_delay * 1000 > spent) continue; } - if (td->stonewall && (nr_started || nr_running)) + if (td->o.stonewall && (nr_started || nr_running)) { + dprint(FD_PROCESS, "%s: stonewall wait\n", + td->o.name); break; + } /* * Set state to created. Thread will transition @@ -1000,23 +1247,30 @@ static void run_threads(void) */ td_set_runstate(td, TD_CREATED); map[this_jobs++] = td; - fio_sem_init(&startup_sem, 1); nr_started++; - if (td->use_thread) { - if (pthread_create(&td->thread, NULL, thread_main, td)) { - perror("thread_create"); + if (td->o.use_thread) { + dprint(FD_PROCESS, "will pthread_create\n"); + if (pthread_create(&td->thread, NULL, + thread_main, td)) { + perror("pthread_create"); nr_started--; + break; } + if (pthread_detach(td->thread) < 0) + perror("pthread_detach"); } else { - if (fork()) - fio_sem_down(&startup_sem); - else { + pid_t pid; + dprint(FD_PROCESS, "will fork\n"); + pid = fork(); + if (!pid) { int ret = fork_main(shm_id, i); - exit(ret); - } + _exit(ret); + } else if (i == fio_debug_jobno) + *fio_debug_jobp = pid; } + fio_mutex_down(startup_mutex); } /* @@ -1068,10 +1322,10 @@ static void run_threads(void) td_set_runstate(td, TD_RUNNING); nr_running++; nr_started--; - m_rate += td->ratemin; - t_rate += td->rate; + m_rate += td->o.ratemin; + t_rate += td->o.rate; todo--; - fio_sem_up(&td->mutex); + fio_mutex_up(td->mutex); } reap_threads(&nr_running, &t_rate, &m_rate); @@ -1093,6 +1347,8 @@ int main(int argc, char *argv[]) { long ps; + sinit(); + /* * We need locale for number printing, if it isn't set then just * go with the US format. @@ -1103,10 +1359,8 @@ int main(int argc, char *argv[]) if (parse_options(argc, argv)) return 1; - if (!thread_number) { - log_err("Nothing to do\n"); - return 1; - } + if (!thread_number) + return 0; ps = sysconf(_SC_PAGESIZE); if (ps < 0) { @@ -1114,6 +1368,7 @@ int main(int argc, char *argv[]) return 1; } + page_size = ps; page_mask = ps - 1; if (write_bw_log) { @@ -1121,17 +1376,23 @@ int main(int argc, char *argv[]) setup_log(&agg_io_log[DDIR_WRITE]); } - disk_util_timer_arm(); + startup_mutex = fio_mutex_init(0); + + set_genesis_time(); + + status_timer_arm(); run_threads(); if (!fio_abort) { 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_READ], "agg-read_bw.log"); + __finish_log(agg_io_log[DDIR_WRITE], + "agg-write_bw.log"); } } + fio_mutex_remove(startup_mutex); return exit_value; }