* Copyright (C) 2005 Jens Axboe <axboe@suse.de>
* Copyright (C) 2006 Jens Axboe <axboe@kernel.dk>
*
+ * The license below covers all files distributed with fio unless otherwise
+ * noted in the file itself.
+ *
* This program is free software; you can redistribute it and/or modify
- * it under the terms of the GNU General Public License as published by
- * the Free Software Foundation; either version 2 of the License, or
- * (at your option) any later version.
+ * it under the terms of the GNU General Public License version 2 as
+ * published by the Free Software Foundation.
*
* This program is distributed in the hope that it will be useful,
* but WITHOUT ANY WARRANTY; without even the implied warranty of
#include <string.h>
#include <signal.h>
#include <time.h>
+#include <locale.h>
#include <assert.h>
#include <sys/stat.h>
#include <sys/wait.h>
#include <sys/ipc.h>
#include <sys/shm.h>
-#include <sys/ioctl.h>
#include <sys/mman.h>
#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;
int shm_id = 0;
int temp_stall_ts;
-char *fio_inst_prefix = _INST_PREFIX;
-
-extern unsigned long long mlock_size;
-
-#define should_fsync(td) ((td_write(td) || td_rw(td)) && (!(td)->odirect || (td)->override_sync))
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)
{
- int i;
+ td->runstate = runstate;
+}
- for (i = 0; i < thread_number; i++) {
- struct thread_data *td = &threads[i];
+static void terminate_threads(int group_id, int forced_kill)
+{
+ struct thread_data *td;
+ int i;
+ for_each_td(td, i) {
if (group_id == TERMINATE_ALL || groupid == td->groupid) {
td->terminate = 1;
td->start_delay = 0;
+ if (forced_kill)
+ td_set_runstate(td, TD_EXITED);
}
}
}
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;
}
}
-/*
- * The ->file_map[] contains a map of blocks we have or have not done io
- * to yet. Used to make sure we cover the entire range in a fair fashion.
- */
-static int random_map_free(struct thread_data *td, struct fio_file *f,
- unsigned long long block)
-{
- unsigned int idx = RAND_MAP_IDX(td, f, block);
- unsigned int bit = RAND_MAP_BIT(td, f, block);
-
- return (f->file_map[idx] & (1UL << bit)) == 0;
-}
-
-/*
- * Return the next free block in the map.
- */
-static int get_next_free_block(struct thread_data *td, struct fio_file *f,
- unsigned long long *b)
-{
- int i;
-
- *b = 0;
- i = 0;
- while ((*b) * td->min_bs < f->file_size) {
- if (f->file_map[i] != -1UL) {
- *b += ffz(f->file_map[i]);
- return 0;
- }
-
- *b += BLOCKS_PER_MAP;
- i++;
- }
-
- return 1;
-}
-
-/*
- * Mark a given offset as used in the map.
- */
-static void mark_random_map(struct thread_data *td, struct fio_file *f,
- struct io_u *io_u)
-{
- unsigned long long block = io_u->offset / (unsigned long long) td->min_bs;
- unsigned int blocks = 0;
-
- while (blocks < (io_u->buflen / td->min_bs)) {
- unsigned int idx, bit;
-
- if (!random_map_free(td, f, block))
- break;
-
- idx = RAND_MAP_IDX(td, f, block);
- bit = RAND_MAP_BIT(td, f, block);
-
- assert(idx < f->num_maps);
-
- f->file_map[idx] |= (1UL << bit);
- block++;
- blocks++;
- }
-
- if ((blocks * td->min_bs) < io_u->buflen)
- io_u->buflen = blocks * td->min_bs;
-}
-
-/*
- * For random io, generate a random new block and see if it's used. Repeat
- * until we find a free one. For sequential io, just return the end of
- * the last io issued.
- */
-static int get_next_offset(struct thread_data *td, struct fio_file *f,
- unsigned long long *offset)
-{
- unsigned long long b, rb;
- long r;
-
- if (!td->sequential) {
- unsigned long long max_blocks = td->io_size / td->min_bs;
- int loops = 50;
-
- do {
- r = os_random_long(&td->random_state);
- b = ((max_blocks - 1) * r / (unsigned long long) (RAND_MAX+1.0));
- rb = b + (f->file_offset / td->min_bs);
- loops--;
- } while (!random_map_free(td, f, rb) && loops);
-
- if (!loops) {
- if (get_next_free_block(td, f, &b))
- return 1;
- }
- } else
- b = f->last_pos / td->min_bs;
-
- *offset = (b * td->min_bs) + f->file_offset;
- if (*offset > f->file_size)
- return 1;
-
- return 0;
-}
-
-static unsigned int get_next_buflen(struct thread_data *td)
-{
- unsigned int buflen;
- long r;
-
- if (td->min_bs == td->max_bs)
- buflen = td->min_bs;
- else {
- r = os_random_long(&td->bsrange_state);
- buflen = (1 + (double) (td->max_bs - 1) * r / (RAND_MAX + 1.0));
- buflen = (buflen + td->min_bs - 1) & ~(td->min_bs - 1);
- }
-
- if (buflen > td->io_size - td->this_io_bytes[td->ddir]) {
- /*
- * if using direct/raw io, we may not be able to
- * shrink the size. so just fail it.
- */
- if (td->io_ops->flags & FIO_RAWIO)
- return 0;
-
- buflen = td->io_size - td->this_io_bytes[td->ddir];
- }
-
- return buflen;
-}
-
/*
* Check if we are above the minimum rate given.
*/
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);
- if (rate_quit)
- terminate_threads(td->groupid);
+ fprintf(f_out, "%s: min rate %u not met, got %luKiB/sec\n", td->name, td->ratemin, rate);
return 1;
}
}
}
/*
- * Return the data direction for the next io_u. If the job is a
- * mixed read/write workload, check the rwmix cycle and switch if
- * necessary.
+ * When job exits, we can cancel the in-flight IO if we are using async
+ * io. Attempt to do so.
*/
-static int get_rw_ddir(struct thread_data *td)
-{
- if (td_rw(td)) {
- struct timeval now;
- unsigned long elapsed;
-
- gettimeofday(&now, NULL);
- elapsed = mtime_since_now(&td->rwmix_switch);
-
- /*
- * Check if it's time to seed a new data direction.
- */
- if (elapsed >= td->rwmixcycle) {
- unsigned int v;
- long r;
-
- r = os_random_long(&td->rwmix_state);
- v = 1 + (int) (100.0 * (r / (RAND_MAX + 1.0)));
- if (v < td->rwmixread)
- td->rwmix_ddir = DDIR_READ;
- else
- td->rwmix_ddir = DDIR_WRITE;
- memcpy(&td->rwmix_switch, &now, sizeof(now));
- }
- return td->rwmix_ddir;
- } else if (td_read(td))
- return DDIR_READ;
- else
- return DDIR_WRITE;
-}
-
-static int td_io_prep(struct thread_data *td, struct io_u *io_u)
-{
- if (td->io_ops->prep && td->io_ops->prep(td, io_u))
- return 1;
-
- return 0;
-}
-
-void put_io_u(struct thread_data *td, struct io_u *io_u)
+static void cleanup_pending_aio(struct thread_data *td)
{
- io_u->file = NULL;
- list_del(&io_u->list);
- list_add(&io_u->list, &td->io_u_freelist);
- td->cur_depth--;
-}
+ struct list_head *entry, *n;
+ struct io_u *io_u;
+ int r;
-static int fill_io_u(struct thread_data *td, struct fio_file *f,
- struct io_u *io_u)
-{
/*
- * If using an iolog, grab next piece if any available.
+ * get immediately available events, if any
*/
- if (td->read_iolog)
- return read_iolog_get(td, io_u);
+ r = io_u_queued_complete(td, 0, NULL);
+ if (r < 0)
+ return;
/*
- * No log, let the seq/rand engine retrieve the next position.
+ * now cancel remaining active events
*/
- if (!get_next_offset(td, f, &io_u->offset)) {
- io_u->buflen = get_next_buflen(td);
-
- if (io_u->buflen) {
- io_u->ddir = get_rw_ddir(td);
+ if (td->io_ops->cancel) {
+ list_for_each_safe(entry, n, &td->io_u_busylist) {
+ io_u = list_entry(entry, struct io_u, list);
/*
- * If using a write iolog, store this entry.
+ * 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 (td->write_iolog)
- write_iolog_put(td, io_u);
-
- io_u->file = f;
- return 0;
+ 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);
+ }
}
}
- return 1;
-}
-
-#define queue_full(td) list_empty(&(td)->io_u_freelist)
-
-struct io_u *__get_io_u(struct thread_data *td)
-{
- struct io_u *io_u = NULL;
-
- if (!queue_full(td)) {
- io_u = list_entry(td->io_u_freelist.next, struct io_u, list);
-
- io_u->error = 0;
- io_u->resid = 0;
- list_del(&io_u->list);
- list_add(&io_u->list, &td->io_u_busylist);
- td->cur_depth++;
- }
-
- return io_u;
+ if (td->cur_depth)
+ r = io_u_queued_complete(td, td->cur_depth, NULL);
}
/*
- * Return an io_u to be processed. Gets a buflen and offset, sets direction,
- * etc. The returned io_u is fully ready to be prepped and submitted.
+ * Helper to handle the final sync of a file. Works just like the normal
+ * io path, just does everything sync.
*/
-static struct io_u *get_io_u(struct thread_data *td, struct fio_file *f)
+static int fio_io_sync(struct thread_data *td, struct fio_file *f)
{
- struct io_u *io_u;
+ struct io_u *io_u = __get_io_u(td);
+ int ret;
- io_u = __get_io_u(td);
if (!io_u)
- return NULL;
-
- if (td->zone_bytes >= td->zone_size) {
- td->zone_bytes = 0;
- f->last_pos += td->zone_skip;
- }
-
- if (fill_io_u(td, f, io_u)) {
- put_io_u(td, io_u);
- return NULL;
- }
-
- if (io_u->buflen + io_u->offset > f->file_size) {
- if (td->io_ops->flags & FIO_RAWIO) {
- put_io_u(td, io_u);
- return NULL;
- }
-
- io_u->buflen = f->file_size - io_u->offset;
- }
-
- if (!io_u->buflen) {
- put_io_u(td, io_u);
- return NULL;
- }
-
- if (!td->read_iolog && !td->sequential)
- mark_random_map(td, f, io_u);
-
- f->last_pos += io_u->buflen;
+ return 1;
- if (td->verify != VERIFY_NONE)
- populate_verify_io_u(td, io_u);
+ io_u->ddir = DDIR_SYNC;
+ io_u->file = f;
if (td_io_prep(td, io_u)) {
put_io_u(td, io_u);
- return NULL;
+ return 1;
}
- gettimeofday(&io_u->start_time, NULL);
- return io_u;
-}
-
-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;
- 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;
-}
-
-static int td_io_sync(struct thread_data *td, struct fio_file *f)
-{
- if (td->io_ops->sync)
- return td->io_ops->sync(td, f);
-
- return 0;
-}
-
-static int td_io_getevents(struct thread_data *td, int min, int max,
- struct timespec *t)
-{
- return td->io_ops->getevents(td, min, max, t);
-}
-
-static int td_io_queue(struct thread_data *td, struct io_u *io_u)
-{
- gettimeofday(&io_u->issue_time, NULL);
-
- return td->io_ops->queue(td, io_u);
-}
-
-#define iocb_time(iocb) ((unsigned long) (iocb)->data)
-
-static void io_completed(struct thread_data *td, struct io_u *io_u,
- struct io_completion_data *icd)
-{
- struct timeval e;
- unsigned long msec;
-
- gettimeofday(&e, NULL);
-
- if (!io_u->error) {
- unsigned int bytes = io_u->buflen - io_u->resid;
- const int idx = io_u->ddir;
-
- td->io_blocks[idx]++;
- td->io_bytes[idx] += bytes;
- td->zone_bytes += bytes;
- td->this_io_bytes[idx] += bytes;
-
- msec = mtime_since(&io_u->issue_time, &e);
-
- add_clat_sample(td, idx, msec);
- add_bw_sample(td, idx);
-
- if ((td_rw(td) || td_write(td)) && idx == DDIR_WRITE)
- log_io_piece(td, io_u);
-
- icd->bytes_done[idx] += bytes;
- } else
- icd->error = io_u->error;
-}
-
-static void ios_completed(struct thread_data *td,struct io_completion_data *icd)
-{
- struct io_u *io_u;
- int i;
-
- icd->error = 0;
- icd->bytes_done[0] = icd->bytes_done[1] = 0;
-
- for (i = 0; i < icd->nr; i++) {
- io_u = td->io_ops->event(td, i);
-
- io_completed(td, io_u, icd);
+requeue:
+ ret = td_io_queue(td, io_u);
+ if (ret < 0) {
+ td_verror(td, io_u->error);
put_io_u(td, io_u);
- }
-}
-
-/*
- * 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 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);
- }
-
- /*
- * 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);
-
- r = td->io_ops->cancel(td, io_u);
- if (!r)
- put_io_u(td, io_u);
+ return 1;
+ } else if (ret == FIO_Q_QUEUED) {
+ if (io_u_queued_complete(td, 1, NULL) < 0)
+ return 1;
+ } else if (ret == FIO_Q_COMPLETED) {
+ if (io_u->error) {
+ td_verror(td, io_u->error);
+ return 1;
}
- }
- 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 (io_u_sync_complete(td, io_u, NULL) < 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,
* 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 timeval t;
- 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
* read from disk.
*/
for_each_file(td, f, i) {
- td_io_sync(td, f);
- file_invalidate_cache(td, f);
+ if (fio_io_sync(td, f))
+ break;
+ if (file_invalidate_cache(td, f))
+ break;
}
- td_set_runstate(td, TD_VERIFYING);
-
- do {
- if (td->terminate)
- break;
+ if (td->error)
+ return;
- gettimeofday(&t, NULL);
- if (runtime_exceeded(td, &t))
- break;
+ td_set_runstate(td, TD_VERIFYING);
+ io_u = NULL;
+ while (!td->terminate) {
io_u = __get_io_u(td);
if (!io_u)
break;
- if (get_next_verify(td, io_u)) {
+ if (runtime_exceeded(td, &io_u->start_time)) {
put_io_u(td, io_u);
break;
}
- f = get_next_file(td);
- if (!f)
- break;
-
- io_u->file = f;
-
- if (td_io_prep(td, io_u)) {
+ if (get_next_verify(td, io_u)) {
put_io_u(td, io_u);
break;
}
- ret = td_io_queue(td, io_u);
- if (ret) {
+ if (td_io_prep(td, io_u)) {
put_io_u(td, io_u);
- td_verror(td, ret);
break;
}
+requeue:
+ ret = td_io_queue(td, io_u);
- /*
- * we have one pending to verify, do that while
- * we are doing io on the next one
- */
- if (do_io_u_verify(td, &v_io_u))
- break;
+ 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;
- ret = td_io_getevents(td, 1, 1, NULL);
- if (ret != 1) {
+ io_u->xfer_buflen = io_u->resid;
+ io_u->xfer_buf += bytes;
+ goto requeue;
+ }
+ ret = io_u_sync_complete(td, io_u, verify_io_u);
if (ret < 0)
- td_verror(td, ret);
+ break;
+ continue;
+ case FIO_Q_QUEUED:
+ break;
+ case FIO_Q_BUSY:
+ requeue_io_u(td, &io_u);
+ ret = td_io_commit(td);
+ break;
+ default:
+ assert(ret < 0);
+ td_verror(td, -ret);
break;
}
- v_io_u = td->io_ops->event(td, 0);
- icd.nr = 1;
- icd.error = 0;
- 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;
+ if (ret < 0 || td->error)
break;
- }
/*
- * if we can't submit more io, we need to verify now
+ * if we can queue more, do so. but check if there are
+ * completed io_u's first.
*/
- if (queue_full(td) && do_io_u_verify(td, &v_io_u))
- break;
+ min_events = 0;
+ if (queue_full(td) || ret == FIO_Q_BUSY) {
+ min_events = 1;
- } while (1);
+ if (td->cur_depth > td->iodepth_low)
+ min_events = td->cur_depth - td->iodepth_low;
+ }
- do_io_u_verify(td, &v_io_u);
+ /*
+ * 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, verify_io_u) < 0)
+ break;
+ }
if (td->cur_depth)
cleanup_pending_aio(td);
int i = 0;
while (!td->terminate) {
- gettimeofday(&e, NULL);
+ fio_gettime(&e, NULL);
if (runtime_exceeded(td, &e))
break;
*/
static void do_io(struct thread_data *td)
{
- struct io_completion_data icd;
- struct timeval s, e;
+ struct timeval s;
unsigned long usec;
- struct fio_file *f;
- int i;
+ int i, ret = 0;
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;
- int ret, min_evts = 0;
+ 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;
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));
- ret = td_io_queue(td, io_u);
- if (ret) {
+ if (runtime_exceeded(td, &s)) {
put_io_u(td, io_u);
- td_verror(td, ret);
break;
}
+requeue:
+ ret = td_io_queue(td, io_u);
- add_slat_sample(td, io_u->ddir, mtime_since(&io_u->start_time, &io_u->issue_time));
+ 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;
- if (td->cur_depth < td->iodepth) {
- timeout = &ts;
- min_evts = 0;
- } else {
- timeout = NULL;
- min_evts = 1;
+ 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);
+ if (bytes_done < 0)
+ ret = bytes_done;
+ 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);
+ ret = td_io_commit(td);
+ break;
+ default:
+ assert(ret < 0);
+ put_io_u(td, io_u);
+ break;
}
- ret = td_io_getevents(td, min_evts, td->cur_depth, timeout);
- if (ret < 0) {
- td_verror(td, ret);
+ if (ret < 0 || td->error)
break;
- } else if (!ret)
- continue;
- icd.nr = ret;
- ios_completed(td, &icd);
- if (icd.error) {
- td_verror(td, icd.error);
- break;
+ /*
+ * See if we need to complete some commands
+ */
+ if (ret == FIO_Q_QUEUED || ret == FIO_Q_BUSY) {
+ min_evts = 0;
+ if (queue_full(td) || ret == FIO_Q_BUSY) {
+ min_evts = 1;
+
+ if (td->cur_depth > td->iodepth_low)
+ min_evts = td->cur_depth - td->iodepth_low;
+ }
+
+ fio_gettime(&comp_time, NULL);
+ bytes_done = io_u_queued_complete(td, min_evts, NULL);
+ 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
*/
- gettimeofday(&e, NULL);
- usec = utime_since(&s, &e);
+ usec = utime_since(&s, &comp_time);
- rate_throttle(td, usec, icd.bytes_done[td->ddir]);
+ rate_throttle(td, usec, bytes_done, td->ddir);
- if (check_min_rate(td, &e)) {
- td_verror(td, ENOMEM);
+ if (check_min_rate(td, &comp_time)) {
+ if (exitall_on_terminate)
+ terminate_threads(td->groupid, 0);
+ td_verror(td, ENODATA);
break;
}
- if (runtime_exceeded(td, &e))
- break;
-
- if (td->thinktime)
- usec_sleep(td, td->thinktime);
+ if (td->thinktime) {
+ unsigned long long b;
- if (should_fsync(td) && td->fsync_blocks &&
- (td->io_blocks[DDIR_WRITE] % td->fsync_blocks) == 0)
- td_io_sync(td, f);
- }
+ b = td->io_blocks[0] + td->io_blocks[1];
+ if (!(b % td->thinktime_blocks)) {
+ int left;
- if (td->cur_depth)
- cleanup_pending_aio(td);
+ if (td->thinktime_spin)
+ __usec_sleep(td->thinktime_spin);
- if (should_fsync(td) && td->end_fsync) {
- td_set_runstate(td, TD_FSYNCING);
- for_each_file(td, f, i)
- td_io_sync(td, f);
+ left = td->thinktime - td->thinktime_spin;
+ if (left)
+ usec_sleep(td, left);
+ }
+ }
}
-}
-static int td_io_init(struct thread_data *td)
-{
- if (td->io_ops->init)
- return td->io_ops->init(td);
+ if (!td->error) {
+ struct fio_file *f;
- return 0;
+ if (td->cur_depth)
+ cleanup_pending_aio(td);
+
+ if (should_fsync(td) && td->end_fsync) {
+ td_set_runstate(td, TD_FSYNCING);
+ for_each_file(td, f, i)
+ fio_io_sync(td, f);
+ }
+ }
}
static void cleanup_io_u(struct thread_data *td)
free(io_u);
}
- if (td->mem_type == MEM_MALLOC)
- free(td->orig_buffer);
- else if (td->mem_type == MEM_SHM) {
- struct shmid_ds sbuf;
+ free_io_mem(td);
+}
- shmdt(td->orig_buffer);
- shmctl(td->shm_id, IPC_RMID, &sbuf);
- } else if (td->mem_type == MEM_MMAP)
- munmap(td->orig_buffer, td->orig_buffer_size);
- else
- log_err("Bad memory type %d\n", td->mem_type);
+/*
+ * "randomly" fill the buffer contents
+ */
+static void fill_rand_buf(struct io_u *io_u, int max_bs)
+{
+ int *ptr = io_u->buf;
- td->orig_buffer = NULL;
+ 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;
+ unsigned int max_bs;
int i, max_units;
char *p;
else
max_units = td->iodepth;
- td->orig_buffer_size = td->max_bs * max_units + MASK;
+ max_bs = max(td->max_bs[DDIR_READ], td->max_bs[DDIR_WRITE]);
+ td->orig_buffer_size = max_bs * max_units;
- if (td->mem_type == MEM_MALLOC)
- td->orig_buffer = malloc(td->orig_buffer_size);
- else if (td->mem_type == MEM_SHM) {
- td->shm_id = shmget(IPC_PRIVATE, td->orig_buffer_size, IPC_CREAT | 0600);
- if (td->shm_id < 0) {
- td_verror(td, errno);
- perror("shmget");
- return 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;
- td->orig_buffer = shmat(td->shm_id, NULL, 0);
- if (td->orig_buffer == (void *) -1) {
- td_verror(td, errno);
- perror("shmat");
- td->orig_buffer = NULL;
- return 1;
- }
- } else if (td->mem_type == MEM_MMAP) {
- td->orig_buffer = mmap(NULL, td->orig_buffer_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
- if (td->orig_buffer == MAP_FAILED) {
- td_verror(td, errno);
- perror("mmap");
- td->orig_buffer = NULL;
- return 1;
- }
- }
+ if (allocate_io_mem(td))
+ return 1;
p = ALIGN(td->orig_buffer);
for (i = 0; i < max_units; i++) {
memset(io_u, 0, sizeof(*io_u));
INIT_LIST_HEAD(&io_u->list);
- io_u->buf = p + td->max_bs * i;
+ io_u->buf = p + max_bs * i;
+ if (td_write(td) || td_rw(td))
+ fill_rand_buf(io_u, max_bs);
+
io_u->index = i;
+ io_u->flags = IO_U_F_FREE;
list_add(&io_u->list, &td->io_u_freelist);
}
+ io_u_init_timeout();
+
return 0;
}
FILE *f;
int ret;
+ if (td->io_ops->flags & FIO_CPUIO)
+ return 0;
+
sprintf(tmp, "%s/queue/scheduler", td->sysfs_root);
f = fopen(tmp, "r+");
struct fio_file *f;
int i;
- 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->last_was_sync = 0;
+
for_each_file(td, f, i) {
+ f->last_completed_pos = 0;
+
f->last_pos = 0;
if (td->io_ops->flags & FIO_SYNCIO)
lseek(f->fd, SEEK_SET, 0);
*/
static void *thread_main(void *data)
{
+ unsigned long long runtime[2];
struct thread_data *td = data;
if (!td->use_thread)
INIT_LIST_HEAD(&td->io_u_freelist);
INIT_LIST_HEAD(&td->io_u_busylist);
+ INIT_LIST_HEAD(&td->io_u_requeues);
INIT_LIST_HEAD(&td->io_hist_list);
INIT_LIST_HEAD(&td->io_log_list);
goto err;
}
- if (td_io_init(td))
- goto err;
-
if (init_iolog(td))
goto err;
if (!td->create_serialize && setup_files(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))
+ goto err;
- gettimeofday(&td->epoch, NULL);
+ if (td->exec_prerun) {
+ if (system(td->exec_prerun) < 0)
+ goto err;
+ }
- if (td->exec_prerun)
- system(td->exec_prerun);
+ fio_gettime(&td->epoch, NULL);
+ memcpy(&td->timeout_end, &td->epoch, sizeof(td->epoch));
+ getrusage(RUSAGE_SELF, &td->ts.ru_start);
+ runtime[0] = runtime[1] = 0;
while (td->loops--) {
- getrusage(RUSAGE_SELF, &td->ru_start);
- gettimeofday(&td->start, NULL);
- memcpy(&td->stat_sample_time, &td->start, sizeof(td->start));
+ fio_gettime(&td->start, NULL);
+ memcpy(&td->ts.stat_sample_time, &td->start, sizeof(td->start));
if (td->ratemin)
- memcpy(&td->lastrate, &td->stat_sample_time, sizeof(td->lastrate));
+ memcpy(&td->lastrate, &td->ts.stat_sample_time, sizeof(td->lastrate));
clear_io_state(td);
prune_io_piece_log(td);
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;
continue;
clear_io_state(td);
- gettimeofday(&td->start, NULL);
+ fio_gettime(&td->start, NULL);
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;
}
- 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)
+ update_rusage_stat(td);
+ fio_gettime(&td->end_time, NULL);
+ td->runtime[0] = runtime[0] / 1000;
+ td->runtime[1] = runtime[1] / 1000;
+
+ 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->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 *) (unsigned long) 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) (unsigned long) ret;
}
/*
*/
static void reap_threads(int *nr_running, int *t_rate, int *m_rate)
{
- int i, cputhreads;
+ struct thread_data *td;
+ int i, cputhreads, pending, status, ret;
/*
* reap exited threads (TD_EXITED -> TD_REAPED)
*/
- for (i = 0, cputhreads = 0; i < thread_number; i++) {
- struct thread_data *td = &threads[i];
+ pending = cputhreads = 0;
+ for_each_td(td, i) {
+ int flags = 0;
- if (td->io_ops->flags & FIO_CPUIO)
+ /*
+ * ->io_ops is NULL for a thread that has closed its
+ * io engine
+ */
+ if (td->io_ops && td->io_ops->flags & FIO_CPUIO)
cputhreads++;
- if (td->runstate != TD_EXITED)
+ if (!td->pid || td->runstate == TD_REAPED)
+ continue;
+ if (td->use_thread) {
+ if (td->runstate == TD_EXITED) {
+ td_set_runstate(td, TD_REAPED);
+ goto reaped;
+ }
continue;
+ }
- td_set_runstate(td, TD_REAPED);
+ flags = WNOHANG;
+ if (td->runstate == TD_EXITED)
+ flags = 0;
+ /*
+ * 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", 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);
+
+ log_err("fio: pid=%d, got signal=%d\n", td->pid, sig);
+ td_set_runstate(td, TD_REAPED);
+ goto reaped;
+ }
+ if (WIFEXITED(status)) {
+ if (WEXITSTATUS(status) && !td->error)
+ td->error = WEXITSTATUS(status);
+
+ td_set_runstate(td, TD_REAPED);
+ goto reaped;
+ }
+ }
+
+ /*
+ * thread is not dead, continue
+ */
+ continue;
+reaped:
if (td->use_thread) {
long ret;
if (pthread_join(td->thread, (void *) &ret))
- perror("thread_join");
- } else
- waitpid(td->pid, NULL, 0);
+ perror("pthread_join");
+ }
(*nr_running)--;
(*m_rate) -= td->ratemin;
(*t_rate) -= td->rate;
- }
-
- if (*nr_running == cputhreads)
- terminate_threads(TERMINATE_ALL);
-}
-
-static void fio_unpin_memory(void *pinned)
-{
- if (pinned) {
- if (munlock(pinned, mlock_size) < 0)
- perror("munlock");
- munmap(pinned, mlock_size);
- }
-}
-
-static void *fio_pin_memory(void)
-{
- unsigned long long phys_mem;
- void *ptr;
-
- if (!mlock_size)
- return NULL;
- /*
- * Don't allow mlock of more than real_mem-128MB
- */
- phys_mem = os_phys_mem();
- if (phys_mem) {
- if ((mlock_size + 128 * 1024 * 1024) > phys_mem) {
- mlock_size = phys_mem - 128 * 1024 * 1024;
- fprintf(f_out, "fio: limiting mlocked memory to %lluMiB\n", mlock_size >> 20);
- }
- }
-
- ptr = mmap(NULL, mlock_size, PROT_READ | PROT_WRITE, MAP_PRIVATE | OS_MAP_ANON, 0, 0);
- if (!ptr) {
- perror("malloc locked mem");
- return NULL;
- }
- if (mlock(ptr, mlock_size) < 0) {
- munmap(ptr, mlock_size);
- perror("mlock");
- return NULL;
+ if (td->error)
+ exit_value++;
}
- return ptr;
+ if (*nr_running == cputhreads && !pending)
+ terminate_threads(TERMINATE_ALL, 0);
}
/*
struct thread_data *td;
unsigned long spent;
int i, todo, nr_running, m_rate, t_rate, nr_started;
- void *mlocked_mem;
- mlocked_mem = fio_pin_memory();
+ if (fio_pin_memory())
+ return;
if (!terse_output) {
printf("Starting %d thread%s\n", thread_number, thread_number > 1 ? "s" : "");
nr_started = 0;
m_rate = t_rate = 0;
- for (i = 0; i < thread_number; i++) {
- td = &threads[i];
-
+ 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,
* 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);
}
- time_init();
+ set_genesis_time();
while (todo) {
struct thread_data *map[MAX_JOBS];
/*
* create threads (TD_NOT_CREATED -> TD_CREATED)
*/
- for (i = 0; i < thread_number; i++) {
- td = &threads[i];
-
+ for_each_td(td, i) {
if (td->runstate != TD_NOT_CREATED)
continue;
if (fork())
fio_sem_down(&startup_sem);
else {
- fork_main(shm_id, i);
- exit(0);
+ int ret = fork_main(shm_id, i);
+
+ exit(ret);
}
}
}
* Wait for the started threads to transition to
* TD_INITIALIZED.
*/
- gettimeofday(&this_start, NULL);
+ fio_gettime(&this_start, NULL);
left = this_jobs;
- while (left) {
+ while (left && !fio_abort) {
if (mtime_since_now(&this_start) > JOB_START_TIMEOUT)
break;
/*
* start created threads (TD_INITIALIZED -> TD_RUNNING).
*/
- for (i = 0; i < thread_number; i++) {
- td = &threads[i];
-
+ for_each_td(td, i) {
if (td->runstate != TD_INITIALIZED)
continue;
}
update_io_ticks();
- fio_unpin_memory(mlocked_mem);
+ fio_unpin_memory();
}
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;
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]);
+ }
+
+ set_genesis_time();
+
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;
}