#include "err.h"
#include "lib/pow2.h"
#include "minmax.h"
+#include "zbd.h"
+#include "target.h"
struct io_completion_data {
int nr; /* input */
/*
* Mark a given offset as used in the map.
*/
-static void mark_random_map(struct thread_data *td, struct io_u *io_u)
+static uint64_t mark_random_map(struct thread_data *td, struct io_u *io_u,
+ uint64_t offset, uint64_t buflen)
{
unsigned long long min_bs = td->o.min_bs[io_u->ddir];
struct fio_file *f = io_u->file;
unsigned long long nr_blocks;
uint64_t block;
- block = (io_u->offset - f->file_offset) / (uint64_t) min_bs;
- nr_blocks = (io_u->buflen + min_bs - 1) / min_bs;
+ block = (offset - f->file_offset) / (uint64_t) min_bs;
+ nr_blocks = (buflen + min_bs - 1) / min_bs;
+ assert(nr_blocks > 0);
- if (!(io_u->flags & IO_U_F_BUSY_OK))
+ if (!(io_u->flags & IO_U_F_BUSY_OK)) {
nr_blocks = axmap_set_nr(f->io_axmap, block, nr_blocks);
+ assert(nr_blocks > 0);
+ }
+
+ if ((nr_blocks * min_bs) < buflen)
+ buflen = nr_blocks * min_bs;
- if ((nr_blocks * min_bs) < io_u->buflen)
- io_u->buflen = nr_blocks * min_bs;
+ return buflen;
}
static uint64_t last_block(struct thread_data *td, struct fio_file *f,
if (max_size > f->real_file_size)
max_size = f->real_file_size;
- if (td->o.zone_range)
+ if (td->o.zone_mode == ZONE_MODE_STRIDED && td->o.zone_range)
max_size = td->o.zone_range;
if (td->o.min_bs[ddir] > td->o.ba[ddir])
void put_io_u(struct thread_data *td, struct io_u *io_u)
{
+ const bool needs_lock = td_async_processing(td);
+
+ if (io_u->post_submit) {
+ io_u->post_submit(io_u, io_u->error == 0);
+ io_u->post_submit = NULL;
+ }
+
if (td->parent)
td = td->parent;
- td_io_u_lock(td);
+ if (needs_lock)
+ __td_io_u_lock(td);
if (io_u->file && !(io_u->flags & IO_U_F_NO_FILE_PUT))
put_file_log(td, io_u->file);
}
io_u_qpush(&td->io_u_freelist, io_u);
td_io_u_free_notify(td);
- td_io_u_unlock(td);
+
+ if (needs_lock)
+ __td_io_u_unlock(td);
}
void clear_io_u(struct thread_data *td, struct io_u *io_u)
void requeue_io_u(struct thread_data *td, struct io_u **io_u)
{
+ const bool needs_lock = td_async_processing(td);
struct io_u *__io_u = *io_u;
enum fio_ddir ddir = acct_ddir(__io_u);
if (td->parent)
td = td->parent;
- td_io_u_lock(td);
+ if (needs_lock)
+ __td_io_u_lock(td);
io_u_set(td, __io_u, IO_U_F_FREE);
if ((__io_u->flags & IO_U_F_FLIGHT) && ddir_rw(ddir))
io_u_rpush(&td->io_u_requeues, __io_u);
td_io_u_free_notify(td);
- td_io_u_unlock(td);
+
+ if (needs_lock)
+ __td_io_u_unlock(td);
+
*io_u = NULL;
}
-static void __fill_io_u_zone(struct thread_data *td, struct io_u *io_u)
+static void setup_strided_zone_mode(struct thread_data *td, struct io_u *io_u)
{
struct fio_file *f = io_u->file;
+ assert(td->o.zone_mode == ZONE_MODE_STRIDED);
+ assert(td->o.zone_size);
+ assert(td->o.zone_range);
+
/*
* See if it's time to switch to a new zone
*/
static int fill_io_u(struct thread_data *td, struct io_u *io_u)
{
bool is_random;
+ uint64_t offset;
+ enum io_u_action ret;
if (td_ioengine_flagged(td, FIO_NOIO))
goto out;
if (!ddir_rw(io_u->ddir))
goto out;
- /*
- * When file is zoned zone_range is always positive
- */
- if (td->o.zone_range)
- __fill_io_u_zone(td, io_u);
+ if (td->o.zone_mode == ZONE_MODE_STRIDED)
+ setup_strided_zone_mode(td, io_u);
/*
* No log, let the seq/rand engine retrieve the next buflen and
return 1;
}
+ offset = io_u->offset;
+ if (td->o.zone_mode == ZONE_MODE_ZBD) {
+ ret = zbd_adjust_block(td, io_u);
+ if (ret == io_u_eof)
+ return 1;
+ }
+
if (io_u->offset + io_u->buflen > io_u->file->real_file_size) {
dprint(FD_IO, "io_u %p, off=0x%llx + len=0x%llx exceeds file size=0x%llx\n",
io_u,
* mark entry before potentially trimming io_u
*/
if (td_random(td) && file_randommap(td, io_u->file))
- mark_random_map(td, io_u);
+ io_u->buflen = mark_random_map(td, io_u, offset, io_u->buflen);
out:
dprint_io_u(io_u, "fill");
if (!fill_io_u(td, io_u))
break;
+ if (io_u->post_submit) {
+ io_u->post_submit(io_u, false);
+ io_u->post_submit = NULL;
+ }
+
put_file_log(td, f);
td_io_close_file(td, f);
io_u->file = NULL;
return 0;
}
-static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
- unsigned long long tnsec, unsigned long long max_nsec)
-{
- if (!td->error)
- log_err("fio: latency of %llu nsec exceeds specified max (%llu nsec)\n", tnsec, max_nsec);
- td_verror(td, ETIMEDOUT, "max latency exceeded");
- icd->error = ETIMEDOUT;
-}
-
-static void lat_new_cycle(struct thread_data *td)
-{
- fio_gettime(&td->latency_ts, NULL);
- td->latency_ios = ddir_rw_sum(td->io_blocks);
- td->latency_failed = 0;
-}
-
-/*
- * We had an IO outside the latency target. Reduce the queue depth. If we
- * are at QD=1, then it's time to give up.
- */
-static bool __lat_target_failed(struct thread_data *td)
-{
- if (td->latency_qd == 1)
- return true;
-
- td->latency_qd_high = td->latency_qd;
-
- if (td->latency_qd == td->latency_qd_low)
- td->latency_qd_low--;
-
- td->latency_qd = (td->latency_qd + td->latency_qd_low) / 2;
-
- dprint(FD_RATE, "Ramped down: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
-
- /*
- * When we ramp QD down, quiesce existing IO to prevent
- * a storm of ramp downs due to pending higher depth.
- */
- io_u_quiesce(td);
- lat_new_cycle(td);
- return false;
-}
-
-static bool lat_target_failed(struct thread_data *td)
-{
- if (td->o.latency_percentile.u.f == 100.0)
- return __lat_target_failed(td);
-
- td->latency_failed++;
- return false;
-}
-
-void lat_target_init(struct thread_data *td)
-{
- td->latency_end_run = 0;
-
- if (td->o.latency_target) {
- dprint(FD_RATE, "Latency target=%llu\n", td->o.latency_target);
- fio_gettime(&td->latency_ts, NULL);
- td->latency_qd = 1;
- td->latency_qd_high = td->o.iodepth;
- td->latency_qd_low = 1;
- td->latency_ios = ddir_rw_sum(td->io_blocks);
- } else
- td->latency_qd = td->o.iodepth;
-}
-
-void lat_target_reset(struct thread_data *td)
-{
- if (!td->latency_end_run)
- lat_target_init(td);
-}
-
-static void lat_target_success(struct thread_data *td)
-{
- const unsigned int qd = td->latency_qd;
- struct thread_options *o = &td->o;
-
- td->latency_qd_low = td->latency_qd;
-
- /*
- * If we haven't failed yet, we double up to a failing value instead
- * of bisecting from highest possible queue depth. If we have set
- * a limit other than td->o.iodepth, bisect between that.
- */
- if (td->latency_qd_high != o->iodepth)
- td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
- else
- td->latency_qd *= 2;
-
- if (td->latency_qd > o->iodepth)
- td->latency_qd = o->iodepth;
-
- dprint(FD_RATE, "Ramped up: %d %d %d\n", td->latency_qd_low, td->latency_qd, td->latency_qd_high);
-
- /*
- * Same as last one, we are done. Let it run a latency cycle, so
- * we get only the results from the targeted depth.
- */
- if (td->latency_qd == qd) {
- if (td->latency_end_run) {
- dprint(FD_RATE, "We are done\n");
- td->done = 1;
- } else {
- dprint(FD_RATE, "Quiesce and final run\n");
- io_u_quiesce(td);
- td->latency_end_run = 1;
- reset_all_stats(td);
- reset_io_stats(td);
- }
- }
-
- lat_new_cycle(td);
-}
-
-/*
- * Check if we can bump the queue depth
- */
-void lat_target_check(struct thread_data *td)
-{
- uint64_t usec_window;
- uint64_t ios;
- double success_ios;
-
- usec_window = utime_since_now(&td->latency_ts);
- if (usec_window < td->o.latency_window)
- return;
-
- ios = ddir_rw_sum(td->io_blocks) - td->latency_ios;
- success_ios = (double) (ios - td->latency_failed) / (double) ios;
- success_ios *= 100.0;
-
- dprint(FD_RATE, "Success rate: %.2f%% (target %.2f%%)\n", success_ios, td->o.latency_percentile.u.f);
-
- if (success_ios >= td->o.latency_percentile.u.f)
- lat_target_success(td);
- else
- __lat_target_failed(td);
-}
-
/*
* If latency target is enabled, we might be ramping up or down and not
* using the full queue depth available.
if (qempty)
return true;
- if (!td->o.latency_target)
+ if (!td->o.latency_target || td->o.iodepth_mode != IOD_STEPPED)
return false;
return td->cur_depth >= td->latency_qd;
struct io_u *__get_io_u(struct thread_data *td)
{
+ const bool needs_lock = td_async_processing(td);
struct io_u *io_u = NULL;
int ret;
if (td->stop_io)
return NULL;
- td_io_u_lock(td);
+ if (needs_lock)
+ __td_io_u_lock(td);
again:
if (!io_u_rempty(&td->io_u_requeues))
goto again;
}
- td_io_u_unlock(td);
+ if (needs_lock)
+ __td_io_u_unlock(td);
+
return io_u;
}
|| td->o.gtod_reduce;
}
+static void trim_block_info(struct thread_data *td, struct io_u *io_u)
+{
+ uint32_t *info = io_u_block_info(td, io_u);
+
+ if (BLOCK_INFO_STATE(*info) >= BLOCK_STATE_TRIM_FAILURE)
+ return;
+
+ *info = BLOCK_INFO(BLOCK_STATE_TRIMMED, BLOCK_INFO_TRIMS(*info) + 1);
+}
+
static void account_io_completion(struct thread_data *td, struct io_u *io_u,
struct io_completion_data *icd,
const enum fio_ddir idx, unsigned int bytes)
icd->error = ops->io_u_lat(td, tnsec);
}
- if (td->o.max_latency && tnsec > td->o.max_latency)
- lat_fatal(td, icd, tnsec, td->o.max_latency);
+ if (td->o.max_latency && tnsec > td->o.max_latency) {
+ icd->error = ETIMEDOUT;
+ lat_fatal(td, tnsec, td->o.max_latency);
+ }
if (td->o.latency_target && tnsec > td->o.latency_target) {
- if (lat_target_failed(td))
- lat_fatal(td, icd, tnsec, td->o.latency_target);
+ if (lat_target_failed(td)) {
+ icd->error = ETIMEDOUT;
+ lat_fatal(td, tnsec, td->o.latency_target);
+ }
}
}
} else if (ddir_sync(idx) && !td->o.disable_clat)
add_sync_clat_sample(&td->ts, llnsec);
- if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM) {
- uint32_t *info = io_u_block_info(td, io_u);
- if (BLOCK_INFO_STATE(*info) < BLOCK_STATE_TRIM_FAILURE) {
- if (io_u->ddir == DDIR_TRIM) {
- *info = BLOCK_INFO(BLOCK_STATE_TRIMMED,
- BLOCK_INFO_TRIMS(*info) + 1);
- } else if (io_u->ddir == DDIR_WRITE) {
- *info = BLOCK_INFO_SET_STATE(BLOCK_STATE_WRITTEN,
- *info);
- }
- }
- }
+ if (td->ts.nr_block_infos && io_u->ddir == DDIR_TRIM)
+ trim_block_info(td, io_u);
}
static void file_log_write_comp(const struct thread_data *td, struct fio_file *f,
static bool should_account(struct thread_data *td)
{
- return ramp_time_over(td) && (td->runstate == TD_RUNNING ||
- td->runstate == TD_VERIFYING);
+ return lat_step_account(td) && ramp_time_over(td) &&
+ (td->runstate == TD_RUNNING || td->runstate == TD_VERIFYING);
}
static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,