#include <sys/wait.h>
#include <sys/ipc.h>
#include <sys/mman.h>
+#include <math.h>
#include "fio.h"
#ifndef FIO_NO_HAVE_SHM_H
#include "cgroup.h"
#include "profile.h"
#include "lib/rand.h"
-#include "memalign.h"
+#include "lib/memalign.h"
#include "server.h"
#include "lib/getrusage.h"
#include "idletime.h"
#include "err.h"
#include "lib/tp.h"
+#include "workqueue.h"
+#include "lib/mountcheck.h"
static pthread_t helper_thread;
static pthread_mutex_t helper_lock;
}
}
-static void sig_show_status(int sig)
+void sig_show_status(int sig)
{
show_running_run_stats();
}
/*
* 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;
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];
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 {
if (spent)
rate = ((bytes - td->rate_bytes[ddir]) * 1000) / spent;
log_err("%s: min rate %u not met, got"
" %luKB/sec\n", td->o.name,
ratemin, rate);
- return 1;
+ return true;
}
}
} else {
if (iops < rate_iops) {
log_err("%s: min iops rate %u not met\n",
td->o.name, rate_iops);
- return 1;
+ return true;
} else {
if (spent)
rate = ((iops - td->rate_blocks[ddir]) * 1000) / spent;
log_err("%s: min iops rate %u not met,"
" got %lu\n", td->o.name,
rate_iops_min, rate);
+ return true;
}
}
}
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;
/*
* 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;
}
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:
if (ret < 0) {
td_verror(td, io_u->error, "td_io_queue");
put_io_u(td, io_u);
- return 1;
+ return true;
} else if (ret == FIO_Q_QUEUED) {
- if (io_u_queued_complete(td, 1, NULL) < 0)
- return 1;
+ if (io_u_queued_complete(td, 1) < 0)
+ return true;
} else if (ret == FIO_Q_COMPLETED) {
if (io_u->error) {
td_verror(td, io_u->error, "td_io_queue");
- return 1;
+ return true;
}
- if (io_u_sync_complete(td, io_u, NULL) < 0)
- return 1;
+ if (io_u_sync_complete(td, io_u) < 0)
+ return true;
} else if (ret == FIO_Q_BUSY) {
if (td_io_commit(td))
- return 1;
+ return true;
goto requeue;
}
- return 0;
+ return false;
}
static int fio_file_fsync(struct thread_data *td, struct fio_file *f)
__update_tv_cache(td);
}
-static inline int runtime_exceeded(struct thread_data *td, struct timeval *t)
+static inline bool runtime_exceeded(struct thread_data *td, struct timeval *t)
{
if (in_ramp_time(td))
- return 0;
+ return false;
if (!td->o.timeout)
- return 0;
+ return false;
if (utime_since(&td->epoch, t) >= td->o.timeout)
- return 1;
+ return true;
- return 0;
+ return false;
+}
+
+/*
+ * 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 int break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
- int *retptr)
+static bool break_on_this_error(struct thread_data *td, enum fio_ddir ddir,
+ int *retptr)
{
int ret = *retptr;
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)) {
/*
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
*/
td_clear_error(td);
fio_mark_td_terminate(td);
- return 1;
+ 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)
}
}
+static int wait_for_completions(struct thread_data *td, struct timeval *time)
+{
+ const int full = queue_full(td);
+ int min_evts = 0;
+ int ret;
+
+ /*
+ * if the queue is full, we MUST reap at least 1 event
+ */
+ min_evts = min(td->o.iodepth_batch_complete_min, td->cur_depth);
+ if ((full && !min_evts) || !td->o.iodepth_batch_complete_min)
+ min_evts = 1;
+
+ if (time && (__should_check_rate(td, DDIR_READ) ||
+ __should_check_rate(td, DDIR_WRITE) ||
+ __should_check_rate(td, DDIR_TRIM)))
+ fio_gettime(time, NULL);
+
+ do {
+ ret = io_u_queued_complete(td, min_evts);
+ if (ret < 0)
+ break;
+ } while (full && (td->cur_depth > td->o.iodepth_low));
+
+ return ret;
+}
+
+int io_queue_event(struct thread_data *td, struct io_u *io_u, int *ret,
+ enum fio_ddir ddir, uint64_t *bytes_issued, int from_verify,
+ struct timeval *comp_time)
+{
+ int ret2;
+
+ switch (*ret) {
+ case FIO_Q_COMPLETED:
+ if (io_u->error) {
+ *ret = -io_u->error;
+ clear_io_u(td, io_u);
+ } else if (io_u->resid) {
+ int bytes = io_u->xfer_buflen - io_u->resid;
+ struct fio_file *f = io_u->file;
+
+ if (bytes_issued)
+ *bytes_issued += bytes;
+
+ if (!from_verify)
+ trim_io_piece(td, io_u);
+
+ /*
+ * zero read, fail
+ */
+ if (!bytes) {
+ if (!from_verify)
+ unlog_io_piece(td, io_u);
+ td_verror(td, EIO, "full resid");
+ put_io_u(td, io_u);
+ break;
+ }
+
+ io_u->xfer_buflen = io_u->resid;
+ io_u->xfer_buf += bytes;
+ io_u->offset += bytes;
+
+ if (ddir_rw(io_u->ddir))
+ td->ts.short_io_u[io_u->ddir]++;
+
+ f = io_u->file;
+ if (io_u->offset == f->real_file_size)
+ goto sync_done;
+
+ requeue_io_u(td, &io_u);
+ } else {
+sync_done:
+ if (comp_time && (__should_check_rate(td, DDIR_READ) ||
+ __should_check_rate(td, DDIR_WRITE) ||
+ __should_check_rate(td, DDIR_TRIM)))
+ fio_gettime(comp_time, NULL);
+
+ *ret = io_u_sync_complete(td, io_u);
+ if (*ret < 0)
+ break;
+ }
+ return 0;
+ case FIO_Q_QUEUED:
+ /*
+ * if the engine doesn't have a commit hook,
+ * the io_u is really queued. if it does have such
+ * a hook, it has to call io_u_queued() itself.
+ */
+ if (td->io_ops->commit == NULL)
+ io_u_queued(td, io_u);
+ if (bytes_issued)
+ *bytes_issued += io_u->xfer_buflen;
+ break;
+ case FIO_Q_BUSY:
+ if (!from_verify)
+ unlog_io_piece(td, io_u);
+ requeue_io_u(td, &io_u);
+ ret2 = td_io_commit(td);
+ if (ret2 < 0)
+ *ret = ret2;
+ break;
+ default:
+ assert(*ret < 0);
+ td_verror(td, -(*ret), "td_io_queue");
+ break;
+ }
+
+ if (break_on_this_error(td, ddir, ret))
+ return 1;
+
+ return 0;
+}
+
+static inline bool io_in_polling(struct thread_data *td)
+{
+ return !td->o.iodepth_batch_complete_min &&
+ !td->o.iodepth_batch_complete_max;
+}
+
/*
* The main verify engine. Runs over the writes we previously submitted,
* reads the blocks back in, and checks the crc/md5 of the data.
*/
static void do_verify(struct thread_data *td, uint64_t verify_bytes)
{
- uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
struct fio_file *f;
struct io_u *io_u;
int ret, min_events;
io_u = NULL;
while (!td->terminate) {
enum fio_ddir ddir;
- int ret2, full;
+ int full;
update_tv_cache(td);
check_update_rusage(td);
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) {
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;
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;
/*
*/
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;
}
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);
dprint(FD_VERIFY, "exiting loop\n");
}
-static unsigned int exceeds_number_ios(struct thread_data *td)
+static bool exceeds_number_ios(struct thread_data *td)
{
unsigned long long number_ios;
if (!td->o.number_ios)
- return 0;
+ return false;
- number_ios = ddir_rw_sum(td->this_io_blocks);
+ number_ios = ddir_rw_sum(td->io_blocks);
number_ios += td->io_u_queued + td->io_u_in_flight;
- return number_ios >= td->o.number_ios;
+ return number_ios >= (td->o.number_ios * td->loops);
+}
+
+static bool io_issue_bytes_exceeded(struct thread_data *td)
+{
+ unsigned long long bytes, limit;
+
+ if (td_rw(td))
+ bytes = td->io_issue_bytes[DDIR_READ] + td->io_issue_bytes[DDIR_WRITE];
+ else if (td_write(td))
+ bytes = td->io_issue_bytes[DDIR_WRITE];
+ else if (td_read(td))
+ bytes = td->io_issue_bytes[DDIR_READ];
+ else
+ bytes = td->io_issue_bytes[DDIR_TRIM];
+
+ if (td->o.io_limit)
+ limit = td->o.io_limit;
+ else
+ limit = td->o.size;
+
+ limit *= td->loops;
+ return bytes >= limit || exceeds_number_ios(td);
}
-static int io_bytes_exceeded(struct thread_data *td)
+static bool io_complete_bytes_exceeded(struct thread_data *td)
{
unsigned long long bytes, 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;
+}
+
/*
* 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 uint64_t do_io(struct thread_data *td)
{
- uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
unsigned int i;
int ret = 0;
uint64_t total_bytes, bytes_issued = 0;
lat_target_init(td);
+ total_bytes = td->o.size;
+ /*
+ * Allow random overwrite workloads to write up to io_limit
+ * before starting verification phase as 'size' doesn't apply.
+ */
+ if (td_write(td) && td_random(td) && td->o.norandommap)
+ total_bytes = max(total_bytes, (uint64_t) td->o.io_limit);
/*
* If verify_backlog is enabled, we'll run the verify in this
* handler as well. For that case, we may need up to twice the
* amount of bytes.
*/
- total_bytes = td->o.size;
if (td->o.verify != VERIFY_NONE &&
(td_write(td) && td->o.verify_backlog))
total_bytes += td->o.size;
+ /* In trimwrite mode, each byte is trimmed and then written, so
+ * allow total_bytes to be twice as big */
+ if (td_trimwrite(td))
+ total_bytes += td->total_io_size;
+
while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
- (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td) ||
+ (!flist_empty(&td->trim_list)) || !io_issue_bytes_exceeded(td) ||
td->o.time_based) {
struct timeval comp_time;
- int min_evts = 0;
struct io_u *io_u;
- int ret2, full;
+ int full;
enum fio_ddir ddir;
check_update_rusage(td);
if (flow_threshold_exceeded(td))
continue;
- if (bytes_issued >= total_bytes)
+ if (!td->o.time_based && bytes_issued >= total_bytes)
break;
io_u = get_io_u(td);
!td->o.experimental_verify)
log_io_piece(td, io_u);
- ret = td_io_queue(td, io_u);
- switch (ret) {
- case FIO_Q_COMPLETED:
- if (io_u->error) {
- ret = -io_u->error;
- unlog_io_piece(td, io_u);
- 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;
+ 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);
- trim_io_piece(td, io_u);
+ if (td->error)
+ break;
- /*
- * zero read, fail
- */
- if (!bytes) {
- unlog_io_piece(td, io_u);
- td_verror(td, EIO, "full resid");
- put_io_u(td, io_u);
- break;
- }
+ ret = workqueue_enqueue(&td->io_wq, &io_u->work);
+ if (ret)
+ ret = FIO_Q_QUEUED;
+ else
+ ret = FIO_Q_BUSY;
- io_u->xfer_buflen = io_u->resid;
- io_u->xfer_buf += bytes;
- io_u->offset += bytes;
+ if (ret == FIO_Q_QUEUED && ddir_rw(ddir)) {
+ td->io_issues[ddir]++;
+ td->io_issue_bytes[ddir] += blen;
+ td->rate_io_issue_bytes[ddir] += blen;
+ }
- if (ddir_rw(io_u->ddir))
- td->ts.short_io_u[io_u->ddir]++;
+ if (should_check_rate(td))
+ td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
- if (io_u->offset == f->real_file_size)
- goto sync_done;
+ } else {
+ ret = td_io_queue(td, io_u);
- 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);
+ if (should_check_rate(td))
+ td->rate_next_io_time[ddir] = usec_for_io(td, ddir);
- ret = io_u_sync_complete(td, io_u, bytes_done);
- if (ret < 0)
- break;
- bytes_issued += io_u->xfer_buflen;
- }
- 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:
- 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);
- put_io_u(td, io_u);
- break;
- }
-
- if (break_on_this_error(td, ddir, &ret))
- break;
+ if (io_queue_event(td, io_u, &ret, ddir, &bytes_issued, 0, &comp_time))
+ break;
- /*
- * See if we need to complete some commands. Note that we
- * can get BUSY even without IO queued, if the system is
- * resource starved.
- */
-reap:
- full = queue_full(td) || (ret == FIO_Q_BUSY && td->cur_depth);
- if (full || !td->o.iodepth_batch_complete) {
- min_evts = min(td->o.iodepth_batch_complete,
- td->cur_depth);
/*
- * if the queue is full, we MUST reap at least 1 event
+ * See if we need to complete some commands. Note that
+ * we can get BUSY even without IO queued, if the
+ * system is resource starved.
*/
- if (full && !min_evts)
- min_evts = 1;
-
- if (__should_check_rate(td, DDIR_READ) ||
- __should_check_rate(td, DDIR_WRITE) ||
- __should_check_rate(td, DDIR_TRIM))
- fio_gettime(&comp_time, NULL);
-
- do {
- ret = io_u_queued_complete(td, min_evts, bytes_done);
- if (ret < 0)
- break;
-
- } while (full && (td->cur_depth > td->o.iodepth_low));
+reap:
+ full = queue_full(td) ||
+ (ret == FIO_Q_BUSY && td->cur_depth);
+ if (full || io_in_polling(td))
+ ret = wait_for_completions(td, &comp_time);
}
-
if (ret < 0)
break;
- if (!ddir_rw_sum(bytes_done) && !(td->io_ops->flags & FIO_NOIO))
+ if (!ddir_rw_sum(td->bytes_done) &&
+ !(td->io_ops->flags & FIO_NOIO))
continue;
- if (!in_ramp_time(td) && should_check_rate(td, bytes_done)) {
- if (check_min_rate(td, &comp_time, bytes_done)) {
+ if (!in_ramp_time(td) && should_check_rate(td)) {
+ if (check_min_rate(td, &comp_time)) {
if (exitall_on_terminate)
fio_terminate_threads(td->groupid);
td_verror(td, EIO, "check_min_rate");
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;
}
if (!ddir_rw_sum(td->this_io_bytes))
td->done = 1;
- return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
+ return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
}
static void cleanup_io_u(struct thread_data *td)
/*
* Read back and check that the selected scheduler is now the default.
*/
+ 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;
}
- tmp[sizeof(tmp) - 1] = '\0';
+ /*
+ * either a list of io schedulers or "none\n" is expected.
+ */
+ tmp[strlen(tmp) - 1] = '\0';
sprintf(tmp2, "[%s]", td->o.ioscheduler);
return 0;
}
-static int keep_running(struct thread_data *td)
+static bool keep_running(struct thread_data *td)
{
unsigned long long limit;
if (td->done)
- return 0;
+ return false;
if (td->o.time_based)
- return 1;
+ return true;
if (td->o.loops) {
td->o.loops--;
- return 1;
+ return true;
}
if (exceeds_number_ios(td))
- return 0;
+ return false;
if (td->o.io_limit)
limit = td->o.io_limit;
*/
diff = limit - ddir_rw_sum(td->io_bytes);
if (diff < td_max_bs(td))
- return 0;
+ return false;
if (fio_files_done(td))
- return 0;
+ return false;
- return 1;
+ return true;
}
- return 0;
+ return false;
}
static int exec_string(struct thread_options *o, const char *string, const char *mode)
{
- int ret, newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
+ size_t newlen = strlen(string) + strlen(o->name) + strlen(mode) + 9 + 1;
+ int ret;
char *str;
str = malloc(newlen);
*/
static uint64_t do_dry_run(struct thread_data *td)
{
- uint64_t bytes_done[DDIR_RWDIR_CNT] = { 0, 0, 0 };
-
td_set_runstate(td, TD_RUNNING);
while ((td->o.read_iolog_file && !flist_empty(&td->io_log_list)) ||
- (!flist_empty(&td->trim_list)) || !io_bytes_exceeded(td)) {
+ (!flist_empty(&td->trim_list)) || !io_complete_bytes_exceeded(td)) {
struct io_u *io_u;
int ret;
if (!io_u)
break;
- io_u->flags |= IO_U_F_FLIGHT;
+ io_u_set(io_u, IO_U_F_FLIGHT);
io_u->error = 0;
io_u->resid = 0;
if (ddir_rw(acct_ddir(io_u)))
!td->o.experimental_verify)
log_io_piece(td, io_u);
- ret = io_u_sync_complete(td, io_u, bytes_done);
+ ret = io_u_sync_complete(td, io_u);
(void) ret;
}
- return bytes_done[DDIR_WRITE] + bytes_done[DDIR_TRIM];
+ return td->bytes_done[DDIR_WRITE] + td->bytes_done[DDIR_TRIM];
+}
+
+static void io_workqueue_fn(struct submit_worker *sw,
+ struct workqueue_work *work)
+{
+ struct io_u *io_u = container_of(work, struct io_u, work);
+ const enum fio_ddir ddir = io_u->ddir;
+ struct thread_data *td = sw->private;
+ int ret;
+
+ dprint(FD_RATE, "io_u %p queued by %u\n", io_u, gettid());
+
+ io_u_set(io_u, IO_U_F_NO_FILE_PUT);
+
+ td->cur_depth++;
+
+ do {
+ ret = td_io_queue(td, io_u);
+ if (ret != FIO_Q_BUSY)
+ break;
+ ret = io_u_queued_complete(td, 1);
+ if (ret > 0)
+ td->cur_depth -= ret;
+ io_u_clear(io_u, IO_U_F_FLIGHT);
+ } while (1);
+
+ dprint(FD_RATE, "io_u %p ret %d by %u\n", io_u, ret, gettid());
+
+ io_queue_event(td, io_u, &ret, ddir, NULL, 0, NULL);
+
+ if (ret == FIO_Q_COMPLETED)
+ td->cur_depth--;
+ else if (ret == FIO_Q_QUEUED) {
+ unsigned int min_evts;
+
+ if (td->o.iodepth == 1)
+ min_evts = 1;
+ else
+ min_evts = 0;
+
+ ret = io_u_queued_complete(td, min_evts);
+ if (ret > 0)
+ td->cur_depth -= ret;
+ } else if (ret == FIO_Q_BUSY) {
+ ret = io_u_queued_complete(td, td->cur_depth);
+ if (ret > 0)
+ td->cur_depth -= ret;
+ }
+}
+
+static bool io_workqueue_pre_sleep_flush_fn(struct submit_worker *sw)
+{
+ struct thread_data *td = sw->private;
+
+ if (td->io_u_queued || td->cur_depth || td->io_u_in_flight)
+ return true;
+
+ return false;
+}
+
+static void io_workqueue_pre_sleep_fn(struct submit_worker *sw)
+{
+ struct thread_data *td = sw->private;
+ int ret;
+
+ ret = io_u_quiesce(td);
+ if (ret > 0)
+ td->cur_depth -= ret;
+}
+
+static int io_workqueue_alloc_fn(struct submit_worker *sw)
+{
+ struct thread_data *td;
+
+ td = calloc(1, sizeof(*td));
+ sw->private = td;
+ return 0;
+}
+
+static void io_workqueue_free_fn(struct submit_worker *sw)
+{
+ free(sw->private);
+ sw->private = NULL;
+}
+
+static int io_workqueue_init_worker_fn(struct submit_worker *sw)
+{
+ struct thread_data *parent = sw->wq->td;
+ struct thread_data *td = sw->private;
+ int fio_unused ret;
+
+ memcpy(&td->o, &parent->o, sizeof(td->o));
+ memcpy(&td->ts, &parent->ts, sizeof(td->ts));
+ td->o.uid = td->o.gid = -1U;
+ dup_files(td, parent);
+ td->eo = parent->eo;
+ fio_options_mem_dupe(td);
+
+ if (ioengine_load(td))
+ goto err;
+
+ if (td->o.odirect)
+ td->io_ops->flags |= FIO_RAWIO;
+
+ td->pid = gettid();
+
+ INIT_FLIST_HEAD(&td->io_log_list);
+ INIT_FLIST_HEAD(&td->io_hist_list);
+ INIT_FLIST_HEAD(&td->verify_list);
+ INIT_FLIST_HEAD(&td->trim_list);
+ INIT_FLIST_HEAD(&td->next_rand_list);
+ td->io_hist_tree = RB_ROOT;
+
+ td->o.iodepth = 1;
+ if (td_io_init(td))
+ goto err_io_init;
+
+ fio_gettime(&td->epoch, NULL);
+ fio_getrusage(&td->ru_start);
+ clear_io_state(td, 1);
+
+ td_set_runstate(td, TD_RUNNING);
+ td->flags |= TD_F_CHILD;
+ td->parent = parent;
+ return 0;
+
+err_io_init:
+ close_ioengine(td);
+err:
+ return 1;
+
+}
+
+static void io_workqueue_exit_worker_fn(struct submit_worker *sw)
+{
+ struct thread_data *td = sw->private;
+
+ fio_options_free(td);
+ close_and_free_files(td);
+ if (td->io_ops)
+ close_ioengine(td);
+ td_set_runstate(td, TD_EXITED);
+}
+
+#ifdef CONFIG_SFAA
+static void sum_val(uint64_t *dst, uint64_t *src)
+{
+ if (*src) {
+ __sync_fetch_and_add(dst, *src);
+ *src = 0;
+ }
+}
+#else
+static void sum_val(uint64_t *dst, uint64_t *src)
+{
+ if (*src) {
+ *dst += *src;
+ *src = 0;
+ }
+}
+#endif
+
+static void pthread_double_unlock(pthread_mutex_t *lock1,
+ pthread_mutex_t *lock2)
+{
+#ifndef CONFIG_SFAA
+ pthread_mutex_unlock(lock1);
+ pthread_mutex_unlock(lock2);
+#endif
+}
+
+static void pthread_double_lock(pthread_mutex_t *lock1, pthread_mutex_t *lock2)
+{
+#ifndef CONFIG_SFAA
+ if (lock1 < lock2) {
+ pthread_mutex_lock(lock1);
+ pthread_mutex_lock(lock2);
+ } else {
+ pthread_mutex_lock(lock2);
+ pthread_mutex_lock(lock1);
+ }
+#endif
+}
+
+static void sum_ddir(struct thread_data *dst, struct thread_data *src,
+ enum fio_ddir ddir)
+{
+ pthread_double_lock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
+
+ sum_val(&dst->io_bytes[ddir], &src->io_bytes[ddir]);
+ sum_val(&dst->io_blocks[ddir], &src->io_blocks[ddir]);
+ sum_val(&dst->this_io_blocks[ddir], &src->this_io_blocks[ddir]);
+ sum_val(&dst->this_io_bytes[ddir], &src->this_io_bytes[ddir]);
+ sum_val(&dst->bytes_done[ddir], &src->bytes_done[ddir]);
+
+ pthread_double_unlock(&dst->io_wq.stat_lock, &src->io_wq.stat_lock);
+}
+
+static void io_workqueue_update_acct_fn(struct submit_worker *sw)
+{
+ struct thread_data *src = sw->private;
+ struct thread_data *dst = sw->wq->td;
+
+ if (td_read(src))
+ sum_ddir(dst, src, DDIR_READ);
+ if (td_write(src))
+ sum_ddir(dst, src, DDIR_WRITE);
+ if (td_trim(src))
+ sum_ddir(dst, src, DDIR_TRIM);
+
}
+struct workqueue_ops rated_wq_ops = {
+ .fn = io_workqueue_fn,
+ .pre_sleep_flush_fn = io_workqueue_pre_sleep_flush_fn,
+ .pre_sleep_fn = io_workqueue_pre_sleep_fn,
+ .update_acct_fn = io_workqueue_update_acct_fn,
+ .alloc_worker_fn = io_workqueue_alloc_fn,
+ .free_worker_fn = io_workqueue_free_fn,
+ .init_worker_fn = io_workqueue_init_worker_fn,
+ .exit_worker_fn = io_workqueue_exit_worker_fn,
+};
+
/*
* Entry point for the thread based jobs. The process based jobs end up
* here as well, after a little setup.
*/
static void *thread_main(void *data)
{
- unsigned long long elapsed;
+ unsigned long long elapsed_us[DDIR_RWDIR_CNT] = { 0, };
struct thread_data *td = data;
struct thread_options *o = &td->o;
pthread_condattr_t attr;
fio_verify_init(td);
+ if ((o->io_submit_mode == IO_MODE_OFFLOAD) &&
+ workqueue_init(td, &td->io_wq, &rated_wq_ops, td->o.iodepth))
+ goto err;
+
fio_gettime(&td->epoch, NULL);
fio_getrusage(&td->ru_start);
+ memcpy(&td->bw_sample_time, &td->epoch, sizeof(td->epoch));
+ memcpy(&td->iops_sample_time, &td->epoch, sizeof(td->epoch));
+
+ if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
+ o->ratemin[DDIR_TRIM]) {
+ memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
+ sizeof(td->bw_sample_time));
+ memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
+ sizeof(td->bw_sample_time));
+ memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
+ sizeof(td->bw_sample_time));
+ }
+
clear_state = 0;
while (keep_running(td)) {
uint64_t verify_bytes;
fio_gettime(&td->start, NULL);
- memcpy(&td->bw_sample_time, &td->start, sizeof(td->start));
- memcpy(&td->iops_sample_time, &td->start, sizeof(td->start));
memcpy(&td->tv_cache, &td->start, sizeof(td->start));
- if (o->ratemin[DDIR_READ] || o->ratemin[DDIR_WRITE] ||
- o->ratemin[DDIR_TRIM]) {
- memcpy(&td->lastrate[DDIR_READ], &td->bw_sample_time,
- sizeof(td->bw_sample_time));
- memcpy(&td->lastrate[DDIR_WRITE], &td->bw_sample_time,
- sizeof(td->bw_sample_time));
- memcpy(&td->lastrate[DDIR_TRIM], &td->bw_sample_time,
- sizeof(td->bw_sample_time));
- }
-
if (clear_state)
- clear_io_state(td);
+ clear_io_state(td, 0);
prune_io_piece_log(td);
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]) {
- elapsed = mtime_since_now(&td->start);
- td->ts.runtime[DDIR_READ] += elapsed;
- }
- if (td_write(td) && td->io_bytes[DDIR_WRITE]) {
- elapsed = mtime_since_now(&td->start);
- td->ts.runtime[DDIR_WRITE] += elapsed;
- }
- if (td_trim(td) && td->io_bytes[DDIR_TRIM]) {
- elapsed = mtime_since_now(&td->start);
- td->ts.runtime[DDIR_TRIM] += elapsed;
- }
+ 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);
(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);
+ /*
+ * See comment further up for why this is done here.
+ */
+ check_update_rusage(td);
+
fio_mutex_down(stat_mutex);
- td->ts.runtime[DDIR_READ] += mtime_since_now(&td->start);
+ update_runtime(td, elapsed_us, DDIR_READ);
fio_gettime(&td->start, NULL);
fio_mutex_up(stat_mutex);
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))) {
- struct all_io_list *state;
- size_t sz;
-
- state = get_all_io_list(td->thread_number, &sz);
- if (state) {
- __verify_save_state(state, "local");
- free(state);
- }
- }
+ (td->o.verify != VERIFY_NONE && td_write(td)))
+ verify_save_state(td->thread_number);
fio_unpin_memory(td);
fio_writeout_logs(td);
+ if (o->io_submit_mode == IO_MODE_OFFLOAD)
+ workqueue_exit(&td->io_wq);
+
if (td->flags & TD_F_COMPRESS_LOG)
tp_exit(&td->tp_data);
fio_terminate_threads(TERMINATE_ALL);
}
-static int __check_trigger_file(void)
+static bool __check_trigger_file(void)
{
struct stat sb;
if (!trigger_file)
- return 0;
+ return false;
if (stat(trigger_file, &sb))
- return 0;
+ return false;
if (unlink(trigger_file) < 0)
log_err("fio: failed to unlink %s: %s\n", trigger_file,
strerror(errno));
- return 1;
+ return true;
}
-static int trigger_timedout(void)
+static bool trigger_timedout(void)
{
if (trigger_timeout)
return time_since_genesis() >= trigger_timeout;
- return 0;
+ return false;
}
void exec_trigger(const char *cmd)
if (nr_clients)
fio_clients_send_trigger(trigger_remote_cmd);
else {
- verify_save_state();
+ verify_save_state(IO_LIST_ALL);
fio_terminate_threads(TERMINATE_ALL);
exec_trigger(trigger_cmd);
}
if (is_backend) {
void *data;
+ int ver;
ret = fio_server_get_verify_state(td->o.name,
- td->thread_number - 1, &data);
+ td->thread_number - 1, &data, &ver);
if (!ret)
- verify_convert_assign_state(td, data);
+ verify_convert_assign_state(td, data, ver);
} else
ret = verify_load_state(td, "local");
usleep(usecs);
}
+static bool check_mount_writes(struct thread_data *td)
+{
+ struct fio_file *f;
+ unsigned int i;
+
+ if (!td_write(td) || td->o.allow_mounted_write)
+ return false;
+
+ for_each_file(td, f, i) {
+ if (f->filetype != FIO_TYPE_BD)
+ continue;
+ if (device_is_mounted(f->file_name))
+ goto mounted;
+ }
+
+ return false;
+mounted:
+ log_err("fio: %s appears mounted, and 'allow_mounted_write' isn't set. Aborting.", f->file_name);
+ return true;
+}
+
/*
* Main function for kicking off and reaping jobs, as needed.
*/
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,
gettimeofday(&tv, NULL);
ts.tv_sec = tv.tv_sec + sec;
ts.tv_nsec = (tv.tv_usec * 1000) + nsec;
- if (ts.tv_nsec > 1000000000ULL) {
+
+ if (ts.tv_nsec >= 1000000000ULL) {
ts.tv_nsec -= 1000000000ULL;
ts.tv_sec++;
}
for (i = 0; i < DDIR_RWDIR_CNT; i++) {
struct io_log *log = agg_io_log[i];
- flush_log(log);
+ flush_log(log, 0);
free_log(log);
}
}
projects / fio.git / blobdiff