#include "trim.h"
#include "lib/rand.h"
#include "lib/axmap.h"
+#include "err.h"
struct io_completion_data {
int nr; /* input */
dprint(FD_RANDOM, "off rand %llu\n", (unsigned long long) r);
- *b = (lastb - 1) * (r / ((uint64_t) rmax + 1.0));
+ *b = lastb * (r / ((uint64_t) rmax + 1.0));
} else {
uint64_t off = 0;
{
assert(ddir_rw(ddir));
- if (f->last_pos >= f->io_size + get_start_offset(td) && td->o.time_based)
+ if (f->last_pos >= f->io_size + get_start_offset(td, f) && td->o.time_based)
f->last_pos = f->last_pos - f->io_size;
if (f->last_pos < f->real_file_size) {
{
struct fio_file *f = io_u->file;
- return io_u->offset + buflen <= f->io_size + get_start_offset(td);
+ return io_u->offset + buflen <= f->io_size + get_start_offset(td, f);
}
static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
unsigned int minbs, maxbs;
unsigned long r, rand_max;
- assert(ddir_rw(io_u->ddir));
+ assert(ddir_rw(ddir));
if (td->o.bs_is_seq_rand)
ddir = is_random ? DDIR_WRITE: DDIR_READ;
- else
- ddir = io_u->ddir;
minbs = td->o.min_bs[ddir];
maxbs = td->o.max_bs[ddir];
if (!fio_file_open(f)) {
int err;
+ if (td->nr_open_files >= td->o.open_files)
+ return ERR_PTR(-EBUSY);
+
err = td_io_open_file(td, f);
if (err)
continue;
if (!fio_file_open(f)) {
int err;
+ if (td->nr_open_files >= td->o.open_files)
+ return ERR_PTR(-EBUSY);
+
err = td_io_open_file(td, f);
if (err) {
dprint(FD_FILE, "error %d on open of %s\n",
else
f = get_next_file_rand(td, FIO_FILE_open, FIO_FILE_closing);
+ if (IS_ERR(f))
+ return f;
+
td->file_service_file = f;
td->file_service_left = td->file_service_nr - 1;
out:
- dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
+ if (f)
+ dprint(FD_FILE, "get_next_file: %p [%s]\n", f, f->file_name);
+ else
+ dprint(FD_FILE, "get_next_file: NULL\n");
return f;
}
return __get_next_file(td);
}
-static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
+static long set_io_u_file(struct thread_data *td, struct io_u *io_u)
{
struct fio_file *f;
do {
f = get_next_file(td);
- if (!f)
- return 1;
+ if (IS_ERR_OR_NULL(f))
+ return PTR_ERR(f);
io_u->file = f;
get_file(f);
return 0;
}
+static void lat_fatal(struct thread_data *td, struct io_completion_data *icd,
+ unsigned long tusec, unsigned long max_usec)
+{
+ if (!td->error)
+ log_err("fio: latency of %lu usec exceeds specified max (%lu usec)\n", tusec, max_usec);
+ 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 int __lat_target_failed(struct thread_data *td)
+{
+ if (td->latency_qd == 1)
+ return 1;
+
+ 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 0;
+}
+
+static int 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 0;
+}
+
+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.
+ */
+int queue_full(struct thread_data *td)
+{
+ const int qempty = io_u_qempty(&td->io_u_freelist);
+
+ if (qempty)
+ return 1;
+ if (!td->o.latency_target)
+ return 0;
+
+ return td->cur_depth >= td->latency_qd;
+}
struct io_u *__get_io_u(struct thread_data *td)
{
- struct io_u *io_u;
+ struct io_u *io_u = NULL;
td_io_u_lock(td);
again:
if (!io_u_rempty(&td->io_u_requeues))
io_u = io_u_rpop(&td->io_u_requeues);
- else if (!io_u_qempty(&td->io_u_freelist))
+ else if (!queue_full(td)) {
io_u = io_u_qpop(&td->io_u_freelist);
- if (io_u) {
+ io_u->file = NULL;
io_u->buflen = 0;
io_u->resid = 0;
- io_u->file = NULL;
io_u->end_io = NULL;
}
io_u->acct_ddir = -1;
td->cur_depth++;
io_u->flags |= IO_U_F_IN_CUR_DEPTH;
+ io_u->ipo = NULL;
} else if (td->o.verify_async) {
/*
* We ran out, wait for async verify threads to finish and
struct fio_file *f;
struct io_u *io_u;
int do_scramble = 0;
+ long ret = 0;
io_u = __get_io_u(td);
if (!io_u) {
if (read_iolog_get(td, io_u))
goto err_put;
} else if (set_io_u_file(td, io_u)) {
+ ret = -EBUSY;
dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
goto err_put;
}
f = io_u->file;
+ if (!f) {
+ dprint(FD_IO, "io_u %p, setting file failed\n", io_u);
+ goto err_put;
+ }
+
assert(fio_file_open(f));
if (ddir_rw(io_u->ddir)) {
if (td->flags & TD_F_REFILL_BUFFERS) {
io_u_fill_buffer(td, io_u,
io_u->xfer_buflen, io_u->xfer_buflen);
- } else if (td->flags & TD_F_SCRAMBLE_BUFFERS)
+ } else if ((td->flags & TD_F_SCRAMBLE_BUFFERS) &&
+ !(td->flags & TD_F_COMPRESS))
do_scramble = 1;
if (td->flags & TD_F_VER_NONE) {
populate_verify_io_u(td, io_u);
err_put:
dprint(FD_IO, "get_io_u failed\n");
put_io_u(td, io_u);
- return NULL;
+ return ERR_PTR(ret);
}
void io_u_log_error(struct thread_data *td, struct io_u *io_u)
td_verror(td, io_u->error, "io_u error");
}
+static inline int gtod_reduce(struct thread_data *td)
+{
+ return td->o.disable_clat && td->o.disable_lat && td->o.disable_slat
+ && td->o.disable_bw;
+}
+
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)
{
unsigned long lusec = 0;
- if (!td->o.disable_clat || !td->o.disable_bw)
+ if (!gtod_reduce(td))
lusec = utime_since(&io_u->issue_time, &icd->time);
if (!td->o.disable_lat) {
icd->error = ops->io_u_lat(td, tusec);
}
- if (td->o.max_latency && tusec > td->o.max_latency) {
- if (!td->error)
- log_err("fio: latency of %lu usec exceeds specified max (%u usec)\n", tusec, td->o.max_latency);
- td_verror(td, ETIMEDOUT, "max latency exceeded");
- icd->error = ETIMEDOUT;
+ if (td->o.max_latency && tusec > td->o.max_latency)
+ lat_fatal(td, icd, tusec, td->o.max_latency);
+ if (td->o.latency_target && tusec > td->o.latency_target) {
+ if (lat_target_failed(td))
+ lat_fatal(td, icd, tusec, td->o.latency_target);
}
}
if (!td->o.disable_bw)
add_bw_sample(td, idx, bytes, &icd->time);
- add_iops_sample(td, idx, bytes, &icd->time);
+ if (!gtod_reduce(td))
+ add_iops_sample(td, idx, bytes, &icd->time);
}
static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
td_io_u_lock(td);
assert(io_u->flags & IO_U_F_FLIGHT);
io_u->flags &= ~(IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
+
+ /*
+ * Mark IO ok to verify
+ */
+ if (io_u->ipo) {
+ io_u->ipo->flags &= ~IP_F_IN_FLIGHT;
+ write_barrier();
+ }
+
td_io_u_unlock(td);
if (ddir_sync(io_u->ddir)) {
utime_since_now(&td->start));
}
- if (td_write(td) && idx == DDIR_WRITE &&
- td->o.do_verify &&
- td->o.verify != VERIFY_NONE &&
- !td->o.experimental_verify)
- log_io_piece(td, io_u);
-
icd->bytes_done[idx] += bytes;
if (io_u->end_io) {
int nr)
{
int ddir;
- if (!td->o.disable_clat || !td->o.disable_bw)
+
+ if (!gtod_reduce(td))
fio_gettime(&icd->time, NULL);
icd->nr = nr;
void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
unsigned int max_bs)
{
- if (!td->o.zero_buffers) {
+ if (td->o.buffer_pattern_bytes)
+ fill_buffer_pattern(td, buf, max_bs);
+ else if (!td->o.zero_buffers) {
unsigned int perc = td->o.compress_percentage;
if (perc) {