X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=fio.c;h=78deaf66a732b6e92434cf0ccef52b66b62d5a71;hp=a175be7a1f6bc07c13107d5f8be9152e14c67aea;hb=97601024a53586d77a368763f08be9ee483fdc9c;hpb=02bcaa8c31feb93c61b701d143a7eea3efd2124d diff --git a/fio.c b/fio.c index a175be7a..78deaf66 100644 --- a/fio.c +++ b/fio.c @@ -26,20 +26,20 @@ #include #include #include +#include #include #include #include #include #include -#include #include #include "fio.h" #include "os.h" -#define MASK (4095) - -#define ALIGN(buf) (char *) (((unsigned long) (buf) + MASK) & ~(MASK)) +static unsigned long page_mask; +#define ALIGN(buf) \ + (char *) (((unsigned long) (buf) + page_mask) & ~page_mask) int groupid = 0; int thread_number = 0; @@ -47,11 +47,20 @@ int shm_id = 0; int temp_stall_ts; static volatile int startup_sem; +static volatile int fio_abort; +static int exit_value; + +struct io_log *agg_io_log[2]; #define TERMINATE_ALL (-1) #define JOB_START_TIMEOUT (5 * 1000) -static void terminate_threads(int group_id) +static inline void td_set_runstate(struct thread_data *td, int runstate) +{ + td->runstate = runstate; +} + +static void terminate_threads(int group_id, int forced_kill) { struct thread_data *td; int i; @@ -60,6 +69,8 @@ static void terminate_threads(int group_id) if (group_id == TERMINATE_ALL || groupid == td->groupid) { td->terminate = 1; td->start_delay = 0; + if (forced_kill) + td_set_runstate(td, TD_EXITED); } } } @@ -73,9 +84,9 @@ static void sig_handler(int sig) print_thread_status(); break; default: - printf("\nfio: terminating on signal\n"); + printf("\nfio: terminating on signal %d\n", sig); fflush(stdout); - terminate_threads(TERMINATE_ALL); + terminate_threads(TERMINATE_ALL, 0); break; } } @@ -105,7 +116,7 @@ static int check_min_rate(struct thread_data *td, struct timeval *now) rate = (td->this_io_bytes[ddir] - td->rate_bytes) / spent; if (rate < td->ratemin) { - fprintf(f_out, "%s: min rate %d not met, got %ldKiB/sec\n", td->name, td->ratemin, rate); + fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate); return 1; } } @@ -125,11 +136,6 @@ static inline int runtime_exceeded(struct thread_data *td, struct timeval *t) return 0; } -static inline void td_set_runstate(struct thread_data *td, int runstate) -{ - td->runstate = runstate; -} - static struct fio_file *get_next_file(struct thread_data *td) { unsigned int old_next_file = td->next_file; @@ -157,20 +163,14 @@ static struct fio_file *get_next_file(struct thread_data *td) */ 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 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) { - icd.nr = r; - ios_completed(td, &icd); - } + io_u_queued_complete(td, 0, NULL); /* * now cancel remaining active events @@ -185,13 +185,8 @@ static void cleanup_pending_aio(struct thread_data *td) } } - if (td->cur_depth) { - r = td_io_getevents(td, td->cur_depth, td->cur_depth, NULL); - if (r > 0) { - icd.nr = r; - ios_completed(td, &icd); - } - } + if (td->cur_depth) + io_u_queued_complete(td, td->cur_depth, NULL); } /* @@ -201,7 +196,6 @@ 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) @@ -216,23 +210,20 @@ static int fio_io_sync(struct thread_data *td, struct fio_file *f) } ret = td_io_queue(td, io_u); - if (ret) { + if (ret < 0) { td_verror(td, io_u->error); put_io_u(td, io_u); return 1; - } - - ret = td_io_getevents(td, 1, td->cur_depth, NULL); - if (ret < 0) { - td_verror(td, ret); - return 1; - } + } else if (ret == FIO_Q_QUEUED) { + if (io_u_queued_complete(td, 1, NULL)) + return 1; + } else if (ret == FIO_Q_COMPLETED) { + if (io_u->error) { + td_verror(td, io_u->error); + return 1; + } - icd.nr = ret; - ios_completed(td, &icd); - if (icd.error) { - td_verror(td, icd.error); - return 1; + io_u_sync_complete(td, io_u, NULL); } return 0; @@ -242,12 +233,11 @@ static int fio_io_sync(struct thread_data *td, struct fio_file *f) * The main verify engine. Runs over the writes we previusly submitted, * reads the blocks back in, and checks the crc/md5 of the data. */ -void do_verify(struct thread_data *td) +static void do_verify(struct thread_data *td) { - struct io_u *io_u, *v_io_u = NULL; - struct io_completion_data icd; struct fio_file *f; - int ret, i; + struct io_u *io_u; + int ret, i, min_events; /* * sync io first and invalidate cache, to make sure we really @@ -260,78 +250,67 @@ void do_verify(struct thread_data *td) td_set_runstate(td, TD_VERIFYING); - do { - if (td->terminate) - break; - + io_u = NULL; + while (!td->terminate) { io_u = __get_io_u(td); if (!io_u) break; - if (runtime_exceeded(td, &io_u->start_time)) { - put_io_u(td, io_u); + if (runtime_exceeded(td, &io_u->start_time)) break; - } - if (get_next_verify(td, io_u)) { - put_io_u(td, io_u); + if (get_next_verify(td, io_u)) break; - } - f = get_next_file(td); - if (!f) + if (td_io_prep(td, io_u)) break; - io_u->file = f; +requeue: + ret = td_io_queue(td, io_u); - if (td_io_prep(td, io_u)) { - put_io_u(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) { + int bytes = io_u->xfer_buflen - io_u->resid; + + io_u->xfer_buflen = io_u->resid; + io_u->xfer_buf += bytes; + goto requeue; + } + if (io_u_sync_complete(td, io_u, verify_io_u)) + break; + continue; + case FIO_Q_QUEUED: + break; + default: + assert(ret < 0); + td_verror(td, -ret); break; } - ret = td_io_queue(td, io_u); - if (ret) { - td_verror(td, io_u->error); - put_io_u(td, io_u); + if (ret < 0) break; - } /* - * we have one pending to verify, do that while - * we are doing io on the next one + * if we can queue more, do so. but check if there are + * completed io_u's first. */ - if (do_io_u_verify(td, &v_io_u)) - break; - - ret = td_io_getevents(td, 1, 1, NULL); - if (ret != 1) { - if (ret < 0) - td_verror(td, ret); - break; - } - - v_io_u = td->io_ops->event(td, 0); - icd.nr = 1; - icd.error = 0; - fio_gettime(&icd.time, NULL); - io_completed(td, v_io_u, &icd); - - if (icd.error) { - td_verror(td, icd.error); - put_io_u(td, v_io_u); - v_io_u = NULL; - break; - } + min_events = 0; + if (queue_full(td)) + min_events = 1; /* - * if we can't submit more io, we need to verify now + * Reap required number of io units, if any, and do the + * verification on them through the callback handler */ - if (queue_full(td) && do_io_u_verify(td, &v_io_u)) + if (io_u_queued_complete(td, min_events, verify_io_u)) break; + } - } while (1); - - do_io_u_verify(td, &v_io_u); + if (io_u) + put_io_u(td, io_u); if (td->cur_depth) cleanup_pending_aio(td); @@ -370,7 +349,6 @@ static void do_cpuio(struct thread_data *td) */ static void do_io(struct thread_data *td) { - struct io_completion_data icd; struct timeval s; unsigned long usec; struct fio_file *f; @@ -378,9 +356,9 @@ static void do_io(struct thread_data *td) td_set_runstate(td, TD_RUNNING); - while (td->this_io_bytes[td->ddir] < td->io_size) { - struct timespec ts = { .tv_sec = 0, .tv_nsec = 0}; - struct timespec *timeout; + while ((td->this_io_bytes[0] + td->this_io_bytes[1]) < td->io_size) { + struct timeval comp_time; + long bytes_done = 0; int min_evts = 0; struct io_u *io_u; @@ -397,58 +375,90 @@ static void do_io(struct thread_data *td) memcpy(&s, &io_u->start_time, sizeof(s)); + if (runtime_exceeded(td, &s)) { + put_io_u(td, io_u); + break; + } +requeue: ret = td_io_queue(td, io_u); - if (ret) { - td_verror(td, io_u->error); + + 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) { + int bytes = io_u->xfer_buflen - io_u->resid; + + io_u->xfer_buflen = io_u->resid; + io_u->xfer_buf += bytes; + goto requeue; + } + fio_gettime(&comp_time, NULL); + bytes_done = io_u_sync_complete(td, io_u, NULL); + break; + case FIO_Q_QUEUED: + break; + default: + assert(ret < 0); put_io_u(td, io_u); break; } + if (ret < 0) + break; + add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time)); - if (td->cur_depth < td->iodepth) { - timeout = &ts; + /* + * See if we need to complete some commands + */ + if (ret == FIO_Q_QUEUED) { min_evts = 0; - } else { - timeout = NULL; - min_evts = 1; + if (queue_full(td)) + min_evts = 1; + + fio_gettime(&comp_time, NULL); + bytes_done = io_u_queued_complete(td, min_evts, NULL); + if (bytes_done < 0) + break; } - ret = td_io_getevents(td, min_evts, td->cur_depth, timeout); - if (ret < 0) { - td_verror(td, ret); - break; - } else if (!ret) + if (!bytes_done) continue; - icd.nr = ret; - ios_completed(td, &icd); - if (icd.error) { - td_verror(td, icd.error); - break; - } - /* * 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, td->ddir); - if (check_min_rate(td, &icd.time)) { + if (check_min_rate(td, &comp_time)) { if (exitall_on_terminate) - terminate_threads(td->groupid); - td_verror(td, ENOMEM); + terminate_threads(td->groupid, 0); + td_verror(td, ENODATA); break; } - if (runtime_exceeded(td, &icd.time)) - break; + if (td->thinktime) { + unsigned long long b; + + b = td->io_blocks[0] + td->io_blocks[1]; + if (!(b % td->thinktime_blocks)) { + int left; - if (td->thinktime) - usec_sleep(td, td->thinktime); + if (td->thinktime_spin) + __usec_sleep(td->thinktime_spin); + + left = td->thinktime - td->thinktime_spin; + if (left) + usec_sleep(td, left); + } + } } if (!td->error) { @@ -507,7 +517,12 @@ static int init_io_u(struct thread_data *td) max_units = td->iodepth; max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]); - td->orig_buffer_size = max_bs * max_units + MASK; + td->orig_buffer_size = max_bs * max_units; + + 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 (allocate_io_mem(td)) return 1; @@ -605,6 +620,7 @@ static void clear_io_state(struct thread_data *td) */ static void *thread_main(void *data) { + unsigned long long runtime[2]; struct thread_data *td = data; if (!td->use_thread) @@ -625,9 +641,6 @@ static void *thread_main(void *data) goto err; } - if (td_io_init(td)) - goto err; - if (init_iolog(td)) goto err; @@ -658,13 +671,23 @@ static void *thread_main(void *data) if (open_files(td)) goto err; - fio_gettime(&td->epoch, NULL); + /* + * Do this late, as some IO engines would like to have the + * files setup prior to initializing structures. + */ + if (td_io_init(td)) + goto err; - if (td->exec_prerun) - system(td->exec_prerun); + if (td->exec_prerun) { + if (system(td->exec_prerun) < 0) + goto err; + } + fio_gettime(&td->epoch, NULL); + getrusage(RUSAGE_SELF, &td->ru_start); + + runtime[0] = runtime[1] = 0; while (td->loops--) { - getrusage(RUSAGE_SELF, &td->ru_start); fio_gettime(&td->start, NULL); memcpy(&td->stat_sample_time, &td->start, sizeof(td->start)); @@ -679,11 +702,9 @@ static void *thread_main(void *data) else do_io(td); - td->runtime[td->ddir] += mtime_since_now(&td->start); + runtime[td->ddir] += utime_since_now(&td->start); if (td_rw(td) && td->io_bytes[td->ddir ^ 1]) - td->runtime[td->ddir ^ 1] = td->runtime[td->ddir]; - - update_rusage_stat(td); + runtime[td->ddir ^ 1] = runtime[td->ddir]; if (td->error || td->terminate) break; @@ -696,12 +717,17 @@ static void *thread_main(void *data) do_verify(td); - td->runtime[DDIR_READ] += mtime_since_now(&td->start); + runtime[DDIR_READ] += utime_since_now(&td->start); if (td->error || td->terminate) break; } + 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) @@ -710,40 +736,45 @@ static void *thread_main(void *data) finish_log(td, td->clat_log, "clat"); if (td->write_iolog_file) write_iolog_close(td); - if (td->exec_postrun) - system(td->exec_postrun); + if (td->exec_postrun) { + if (system(td->exec_postrun) < 0) + log_err("fio: postrun %s failed\n", td->exec_postrun); + } if (exitall_on_terminate) - terminate_threads(td->groupid); + terminate_threads(td->groupid, 0); err: + if (td->error) + printf("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror); close_files(td); close_ioengine(td); cleanup_io_u(td); td_set_runstate(td, TD_EXITED); - return NULL; - + return (void *) td->error; } /* * We cannot pass the td data into a forked process, so attach the td and * pass it to the thread worker. */ -static void *fork_main(int shmid, int offset) +static int fork_main(int shmid, int offset) { struct thread_data *td; - void *data; + void *data, *ret; data = shmat(shmid, NULL, 0); if (data == (void *) -1) { + int __err = errno; + perror("shmat"); - return NULL; + return __err; } td = data + offset * sizeof(struct thread_data); - thread_main(td); + ret = thread_main(td); shmdt(data); - return NULL; + return (int) ret; } /* @@ -752,7 +783,7 @@ static void *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; + int i, cputhreads, pending, status, ret; /* * reap exited threads (TD_EXITED -> TD_REAPED) @@ -766,6 +797,22 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate) if (td->io_ops && td->io_ops->flags & FIO_CPUIO) cputhreads++; + 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); + td_set_runstate(td, TD_REAPED); + goto reaped; + } + } + if (td->runstate != TD_EXITED) { if (td->runstate < TD_RUNNING) pending++; @@ -773,6 +820,9 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate) continue; } + if (td->error) + exit_value++; + td_set_runstate(td, TD_REAPED); if (td->use_thread) { @@ -780,16 +830,26 @@ static void reap_threads(int *nr_running, int *t_rate, int *m_rate) if (pthread_join(td->thread, (void *) &ret)) perror("thread_join"); - } else - waitpid(td->pid, NULL, 0); + } 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++; + } + } + +reaped: (*nr_running)--; (*m_rate) -= td->ratemin; (*t_rate) -= td->rate; } if (*nr_running == cputhreads && !pending) - terminate_threads(TERMINATE_ALL); + terminate_threads(TERMINATE_ALL, 0); } /* @@ -820,10 +880,10 @@ static void run_threads(void) for_each_td(td, i) { print_status_init(td->thread_number - 1); - init_disk_util(td); - - if (!td->create_serialize) + if (!td->create_serialize) { + init_disk_util(td); continue; + } /* * do file setup here so it happens sequentially, @@ -831,9 +891,14 @@ static void run_threads(void) * client data interspersed on disk */ if (setup_files(td)) { + exit_value++; + if (td->error) + log_err("fio: pid=%d, err=%d/%s\n", td->pid, td->error, td->verror); td_set_runstate(td, TD_REAPED); todo--; } + + init_disk_util(td); } while (todo) { @@ -885,8 +950,9 @@ static void run_threads(void) if (fork()) fio_sem_down(&startup_sem); else { - fork_main(shm_id, i); - exit(0); + int ret = fork_main(shm_id, i); + + exit(ret); } } } @@ -897,7 +963,7 @@ static void run_threads(void) */ fio_gettime(&this_start, NULL); left = this_jobs; - while (left) { + while (left && !fio_abort) { if (mtime_since_now(&this_start) > JOB_START_TIMEOUT) break; @@ -963,6 +1029,15 @@ static void run_threads(void) int main(int argc, char *argv[]) { + long ps; + + /* + * We need locale for number printing, if it isn't set then just + * go with the US format. + */ + if (!getenv("LC_NUMERIC")) + setlocale(LC_NUMERIC, "en_US"); + if (parse_options(argc, argv)) return 1; @@ -971,10 +1046,30 @@ int main(int argc, char *argv[]) return 1; } + ps = sysconf(_SC_PAGESIZE); + if (ps < 0) { + log_err("Failed to get page size\n"); + return 1; + } + + page_mask = ps - 1; + + if (write_bw_log) { + setup_log(&agg_io_log[DDIR_READ]); + setup_log(&agg_io_log[DDIR_WRITE]); + } + disk_util_timer_arm(); run_threads(); - show_run_stats(); - return 0; + 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"); + } + } + + return exit_value; }