X-Git-Url: https://git.kernel.dk/?p=fio.git;a=blobdiff_plain;f=io_u.c;h=64ff73cd5555b8d4efa0303085f7944cae07544c;hp=d03049e52332e2b01510364863af812ecfd8e496;hb=f754086950c511d48af21aee5c49e9d7985d3d32;hpb=211c9b8948b43ed47aedf1227e0444a58db015e4

diff --git a/io_u.c b/io_u.c
index d03049e5..64ff73cd 100644
--- a/io_u.c
+++ b/io_u.c
@@ -191,23 +191,23 @@ static inline int should_sort_io(struct thread_data *td)
 	return 1;
 }
 
-static int should_do_random(struct thread_data *td)
+static int should_do_random(struct thread_data *td, enum fio_ddir ddir)
 {
 	unsigned int v;
 	unsigned long r;
 
-	if (td->o.perc_rand == 100)
+	if (td->o.perc_rand[ddir] == 100)
 		return 1;
 
 	if (td->o.use_os_rand) {
-		r = os_random_long(&td->seq_rand_state);
+		r = os_random_long(&td->seq_rand_state[ddir]);
 		v = 1 + (int) (100.0 * (r / (OS_RAND_MAX + 1.0)));
 	} else {
-		r = __rand(&td->__seq_rand_state);
+		r = __rand(&td->__seq_rand_state[ddir]);
 		v = 1 + (int) (100.0 * (r / (FRAND_MAX + 1.0)));
 	}
 
-	return v <= td->o.perc_rand;
+	return v <= td->o.perc_rand[ddir];
 }
 
 static int get_next_rand_offset(struct thread_data *td, struct fio_file *f,
@@ -293,7 +293,8 @@ static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
 }
 
 static int get_next_block(struct thread_data *td, struct io_u *io_u,
-			  enum fio_ddir ddir, int rw_seq)
+			  enum fio_ddir ddir, int rw_seq,
+			  unsigned int *is_random)
 {
 	struct fio_file *f = io_u->file;
 	uint64_t b, offset;
@@ -305,23 +306,30 @@ static int get_next_block(struct thread_data *td, struct io_u *io_u,
 
 	if (rw_seq) {
 		if (td_random(td)) {
-			if (should_do_random(td))
+			if (should_do_random(td, ddir)) {
 				ret = get_next_rand_block(td, f, ddir, &b);
-			else {
+				*is_random = 1;
+			} else {
+				*is_random = 0;
 				io_u->flags |= IO_U_F_BUSY_OK;
 				ret = get_next_seq_offset(td, f, ddir, &offset);
 				if (ret)
 					ret = get_next_rand_block(td, f, ddir, &b);
 			}
-		} else
+		} else {
+			*is_random = 0;
 			ret = get_next_seq_offset(td, f, ddir, &offset);
+		}
 	} else {
 		io_u->flags |= IO_U_F_BUSY_OK;
+		*is_random = 0;
 
 		if (td->o.rw_seq == RW_SEQ_SEQ) {
 			ret = get_next_seq_offset(td, f, ddir, &offset);
-			if (ret)
+			if (ret) {
 				ret = get_next_rand_block(td, f, ddir, &b);
+				*is_random = 0;
+			}
 		} else if (td->o.rw_seq == RW_SEQ_IDENT) {
 			if (f->last_start != -1ULL)
 				offset = f->last_start - f->file_offset;
@@ -353,7 +361,8 @@ static int get_next_block(struct thread_data *td, struct io_u *io_u,
  * 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 io_u *io_u)
+static int __get_next_offset(struct thread_data *td, struct io_u *io_u,
+			     unsigned int *is_random)
 {
 	struct fio_file *f = io_u->file;
 	enum fio_ddir ddir = io_u->ddir;
@@ -366,7 +375,7 @@ static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
 		td->ddir_seq_nr = td->o.ddir_seq_nr;
 	}
 
-	if (get_next_block(td, io_u, ddir, rw_seq_hit))
+	if (get_next_block(td, io_u, ddir, rw_seq_hit, is_random))
 		return 1;
 
 	if (io_u->offset >= f->io_size) {
@@ -387,16 +396,17 @@ static int __get_next_offset(struct thread_data *td, struct io_u *io_u)
 	return 0;
 }
 
-static int get_next_offset(struct thread_data *td, struct io_u *io_u)
+static int get_next_offset(struct thread_data *td, struct io_u *io_u,
+			   unsigned int *is_random)
 {
 	if (td->flags & TD_F_PROFILE_OPS) {
 		struct prof_io_ops *ops = &td->prof_io_ops;
 
 		if (ops->fill_io_u_off)
-			return ops->fill_io_u_off(td, io_u);
+			return ops->fill_io_u_off(td, io_u, is_random);
 	}
 
-	return __get_next_offset(td, io_u);
+	return __get_next_offset(td, io_u, is_random);
 }
 
 static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
@@ -407,14 +417,20 @@ static inline int io_u_fits(struct thread_data *td, struct io_u *io_u,
 	return io_u->offset + buflen <= f->io_size + get_start_offset(td);
 }
 
-static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
+static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u,
+				      unsigned int is_random)
 {
-	const int ddir = io_u->ddir;
+	int ddir = io_u->ddir;
 	unsigned int buflen = 0;
 	unsigned int minbs, maxbs;
 	unsigned long r, rand_max;
 
-	assert(ddir_rw(ddir));
+	assert(ddir_rw(io_u->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];
@@ -459,6 +475,10 @@ static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
 			}
 		}
 
+		if (td->o.do_verify && td->o.verify != VERIFY_NONE)
+			buflen = (buflen + td->o.verify_interval - 1) &
+				~(td->o.verify_interval - 1);
+
 		if (!td->o.bs_unaligned && is_power_of_2(minbs))
 			buflen = (buflen + minbs - 1) & ~(minbs - 1);
 
@@ -467,16 +487,17 @@ static unsigned int __get_next_buflen(struct thread_data *td, struct io_u *io_u)
 	return buflen;
 }
 
-static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u)
+static unsigned int get_next_buflen(struct thread_data *td, struct io_u *io_u,
+				    unsigned int is_random)
 {
 	if (td->flags & TD_F_PROFILE_OPS) {
 		struct prof_io_ops *ops = &td->prof_io_ops;
 
 		if (ops->fill_io_u_size)
-			return ops->fill_io_u_size(td, io_u);
+			return ops->fill_io_u_size(td, io_u, is_random);
 	}
 
-	return __get_next_buflen(td, io_u);
+	return __get_next_buflen(td, io_u, is_random);
 }
 
 static void set_rwmix_bytes(struct thread_data *td)
@@ -511,6 +532,24 @@ static inline enum fio_ddir get_rand_ddir(struct thread_data *td)
 	return DDIR_WRITE;
 }
 
+void io_u_quiesce(struct thread_data *td)
+{
+	/*
+	 * We are going to sleep, ensure that we flush anything pending as
+	 * not to skew our latency numbers.
+	 *
+	 * Changed to only monitor 'in flight' requests here instead of the
+	 * td->cur_depth, b/c td->cur_depth does not accurately represent
+	 * io's that have been actually submitted to an async engine,
+	 * and cur_depth is meaningless for sync engines.
+	 */
+	while (td->io_u_in_flight) {
+		int fio_unused ret;
+
+		ret = io_u_queued_complete(td, 1, NULL);
+	}
+}
+
 static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
 {
 	enum fio_ddir odir = ddir ^ 1;
@@ -547,20 +586,7 @@ static enum fio_ddir rate_ddir(struct thread_data *td, enum fio_ddir ddir)
 	} else
 		usec = td->rate_pending_usleep[ddir];
 
-	/*
-	 * We are going to sleep, ensure that we flush anything pending as
-	 * not to skew our latency numbers.
-	 *
-	 * Changed to only monitor 'in flight' requests here instead of the
-	 * td->cur_depth, b/c td->cur_depth does not accurately represent
-	 * io's that have been actually submitted to an async engine,
-	 * and cur_depth is meaningless for sync engines.
-	 */
-	while (td->io_u_in_flight) {
-		int fio_unused ret;
-
-		ret = io_u_queued_complete(td, 1, NULL);
-	}
+	io_u_quiesce(td);
 
 	fio_gettime(&t, NULL);
 	usec_sleep(td, usec);
@@ -671,8 +697,7 @@ void put_io_u(struct thread_data *td, struct io_u *io_u)
 
 	if (io_u->flags & IO_U_F_IN_CUR_DEPTH)
 		td->cur_depth--;
-	flist_del_init(&io_u->list);
-	flist_add(&io_u->list, &td->io_u_freelist);
+	io_u_qpush(&td->io_u_freelist, io_u);
 	td_io_u_unlock(td);
 	td_io_u_free_notify(td);
 }
@@ -699,14 +724,16 @@ void requeue_io_u(struct thread_data *td, struct io_u **io_u)
 	__io_u->flags &= ~IO_U_F_FLIGHT;
 	if (__io_u->flags & IO_U_F_IN_CUR_DEPTH)
 		td->cur_depth--;
-	flist_del(&__io_u->list);
-	flist_add_tail(&__io_u->list, &td->io_u_requeues);
+
+	io_u_rpush(&td->io_u_requeues, __io_u);
 	td_io_u_unlock(td);
 	*io_u = NULL;
 }
 
 static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 {
+	unsigned int is_random;
+
 	if (td->io_ops->flags & FIO_NOIO)
 		goto out;
 
@@ -732,12 +759,12 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 	 * No log, let the seq/rand engine retrieve the next buflen and
 	 * position.
 	 */
-	if (get_next_offset(td, io_u)) {
+	if (get_next_offset(td, io_u, &is_random)) {
 		dprint(FD_IO, "io_u %p, failed getting offset\n", io_u);
 		return 1;
 	}
 
-	io_u->buflen = get_next_buflen(td, io_u);
+	io_u->buflen = get_next_buflen(td, io_u, is_random);
 	if (!io_u->buflen) {
 		dprint(FD_IO, "io_u %p, failed getting buflen\n", io_u);
 		return 1;
@@ -1099,18 +1126,148 @@ static int set_io_u_file(struct thread_data *td, struct io_u *io_u)
 	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;
+	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)
+{
+	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;
+}
+
+static void lat_target_success(struct thread_data *td)
+{
+	const unsigned int qd = td->latency_qd;
+
+	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 != td->o.iodepth)
+		td->latency_qd = (td->latency_qd + td->latency_qd_high) / 2;
+	else
+		td->latency_qd *= 2;
+
+	if (td->latency_qd > td->o.iodepth)
+		td->latency_qd = td->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
+	 */
+	if (td->latency_qd == qd)
+		td->done = 1;
+
+	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 = NULL;
+	struct io_u *io_u;
 
 	td_io_u_lock(td);
 
 again:
-	if (!flist_empty(&td->io_u_requeues))
-		io_u = flist_entry(td->io_u_requeues.next, struct io_u, list);
+	if (!io_u_rempty(&td->io_u_requeues))
+		io_u = io_u_rpop(&td->io_u_requeues);
 	else if (!queue_full(td)) {
-		io_u = flist_entry(td->io_u_freelist.next, struct io_u, list);
+		io_u = io_u_qpop(&td->io_u_freelist);
 
 		io_u->buflen = 0;
 		io_u->resid = 0;
@@ -1126,10 +1283,9 @@ again:
 
 		io_u->error = 0;
 		io_u->acct_ddir = -1;
-		flist_del(&io_u->list);
-		flist_add_tail(&io_u->list, &td->io_u_busylist);
 		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
@@ -1363,11 +1519,18 @@ static void account_io_completion(struct thread_data *td, struct io_u *io_u,
 		tusec = utime_since(&io_u->start_time, &icd->time);
 		add_lat_sample(td, idx, tusec, bytes);
 
-		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->flags & TD_F_PROFILE_OPS) {
+			struct prof_io_ops *ops = &td->prof_io_ops;
+
+			if (ops->io_u_lat)
+				icd->error = ops->io_u_lat(td, tusec);
+		}
+
+		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);
 		}
 	}
 
@@ -1379,7 +1542,10 @@ static void account_io_completion(struct thread_data *td, struct io_u *io_u,
 	if (!td->o.disable_bw)
 		add_bw_sample(td, idx, bytes, &icd->time);
 
-	add_iops_sample(td, idx, &icd->time);
+	add_iops_sample(td, idx, bytes, &icd->time);
+
+	if (td->o.number_ios && !--td->o.number_ios)
+		td->done = 1;
 }
 
 static long long usec_for_io(struct thread_data *td, enum fio_ddir ddir)
@@ -1403,6 +1569,15 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 	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)) {
@@ -1458,12 +1633,6 @@ static void io_completed(struct thread_data *td, struct io_u *io_u,
 					 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) {
@@ -1602,25 +1771,35 @@ void io_u_queued(struct thread_data *td, struct io_u *io_u)
 	}
 }
 
-/*
- * "randomly" fill the buffer contents
- */
-void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
-		      unsigned int min_write, unsigned int max_bs)
+void fill_io_buffer(struct thread_data *td, void *buf, unsigned int min_write,
+		    unsigned int max_bs)
 {
-	io_u->buf_filled_len = 0;
-
-	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) {
 			unsigned int seg = min_write;
 
 			seg = min(min_write, td->o.compress_chunk);
-			fill_random_buf_percentage(&td->buf_state, io_u->buf,
+			if (!seg)
+				seg = min_write;
+
+			fill_random_buf_percentage(&td->buf_state, buf,
 						perc, seg, max_bs);
 		} else
-			fill_random_buf(&td->buf_state, io_u->buf, max_bs);
+			fill_random_buf(&td->buf_state, buf, max_bs);
 	} else
-		memset(io_u->buf, 0, max_bs);
+		memset(buf, 0, max_bs);
+}
+
+/*
+ * "randomly" fill the buffer contents
+ */
+void io_u_fill_buffer(struct thread_data *td, struct io_u *io_u,
+		      unsigned int min_write, unsigned int max_bs)
+{
+	io_u->buf_filled_len = 0;
+	fill_io_buffer(td, io_u->buf, min_write, max_bs);
 }