X-Git-Url: https://git.kernel.dk/?a=blobdiff_plain;f=io_u.c;h=ca7ee68fde393e8cbd440a5deef9b830a1dbff41;hb=af10f514b5e8713f0821c34fadfe22c91d137720;hp=586a4befdce0969e596f9243abc73044b4b75fda;hpb=c4cb947e8f92c10835164b67deed06828cfc01be;p=fio.git

diff --git a/io_u.c b/io_u.c
index 586a4bef..ca7ee68f 100644
--- a/io_u.c
+++ b/io_u.c
@@ -355,7 +355,7 @@ static int get_next_seq_offset(struct thread_data *td, struct fio_file *f,
 	 * and invalidate the cache, if we need to.
 	 */
 	if (f->last_pos[ddir] >= f->io_size + get_start_offset(td, f) &&
-	    o->time_based) {
+	    o->time_based && o->nr_files == 1) {
 		f->last_pos[ddir] = f->file_offset;
 		loop_cache_invalidate(td, f);
 	}
@@ -417,7 +417,13 @@ static int get_next_block(struct thread_data *td, struct io_u *io_u,
 
 	b = offset = -1ULL;
 
-	if (rw_seq) {
+	if (td_randtrimwrite(td) && ddir == DDIR_WRITE) {
+		/* don't mark randommap for these writes */
+		io_u_set(td, io_u, IO_U_F_BUSY_OK);
+		offset = f->last_start[DDIR_TRIM];
+		*is_random = true;
+		ret = 0;
+	} else if (rw_seq) {
 		if (td_random(td)) {
 			if (should_do_random(td, ddir)) {
 				ret = get_next_rand_block(td, f, ddir, &b);
@@ -507,6 +513,24 @@ static int get_next_offset(struct thread_data *td, struct io_u *io_u,
 		return 1;
 	}
 
+	/*
+	 * For randtrimwrite, we decide whether to issue a trim or a write
+	 * based on whether the offsets for the most recent trim and write
+	 * operations match. If they don't match that means we just issued a
+	 * new trim and the next operation should be a write. If they *do*
+	 * match that means we just completed a trim+write pair and the next
+	 * command should be a trim.
+	 *
+	 * This works fine for sequential workloads but for random workloads
+	 * it's possible to complete a trim+write pair and then have the next
+	 * randomly generated offset match the previous offset. If that happens
+	 * we need to alter the offset for the last write operation in order
+	 * to ensure that we issue a write operation the next time through.
+	 */
+	if (td_randtrimwrite(td) && ddir == DDIR_TRIM &&
+	    f->last_start[DDIR_TRIM] == io_u->offset)
+		f->last_start[DDIR_WRITE]--;
+
 	io_u->verify_offset = io_u->offset;
 	return 0;
 }
@@ -530,6 +554,12 @@ static unsigned long long get_next_buflen(struct thread_data *td, struct io_u *i
 
 	assert(ddir_rw(ddir));
 
+	if (td_randtrimwrite(td) && ddir == DDIR_WRITE) {
+		struct fio_file *f = io_u->file;
+
+		return f->last_pos[DDIR_TRIM] - f->last_start[DDIR_TRIM];
+	}
+
 	if (td->o.bs_is_seq_rand)
 		ddir = is_random ? DDIR_WRITE : DDIR_READ;
 
@@ -755,7 +785,15 @@ static enum fio_ddir get_rw_ddir(struct thread_data *td)
 	else
 		ddir = DDIR_INVAL;
 
-	td->rwmix_ddir = rate_ddir(td, ddir);
+	if (!should_check_rate(td)) {
+		/*
+		 * avoid time-consuming call to utime_since_now() if rate checking
+		 * isn't being used. this imrpoves IOPs 50%. See:
+		 * https://github.com/axboe/fio/issues/1501#issuecomment-1418327049
+		 */
+		td->rwmix_ddir = ddir;
+	} else
+		td->rwmix_ddir = rate_ddir(td, ddir);
 	return td->rwmix_ddir;
 }
 
@@ -768,7 +806,7 @@ static void set_rw_ddir(struct thread_data *td, struct io_u *io_u)
 
 	if (td_trimwrite(td)) {
 		struct fio_file *f = io_u->file;
-		if (f->last_pos[DDIR_WRITE] == f->last_pos[DDIR_TRIM])
+		if (f->last_start[DDIR_WRITE] == f->last_start[DDIR_TRIM])
 			ddir = DDIR_TRIM;
 		else
 			ddir = DDIR_WRITE;
@@ -946,10 +984,15 @@ static int fill_io_u(struct thread_data *td, struct io_u *io_u)
 	offset = io_u->offset;
 	if (td->o.zone_mode == ZONE_MODE_ZBD) {
 		ret = zbd_adjust_block(td, io_u);
-		if (ret == io_u_eof)
+		if (ret == io_u_eof) {
+			dprint(FD_IO, "zbd_adjust_block() returned io_u_eof\n");
 			return 1;
+		}
 	}
 
+	if (td->o.fdp)
+		fdp_fill_dspec_data(td, io_u);
+
 	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,
@@ -993,7 +1036,7 @@ static void __io_u_mark_map(uint64_t *map, unsigned int nr)
 		break;
 	case 1 ... 4:
 		idx = 1;
-		fallthrough;
+		fio_fallthrough;
 	case 0:
 		break;
 	}
@@ -1035,7 +1078,7 @@ void io_u_mark_depth(struct thread_data *td, unsigned int nr)
 		break;
 	case 2 ... 3:
 		idx = 1;
-		fallthrough;
+		fio_fallthrough;
 	case 1:
 		break;
 	}
@@ -1076,7 +1119,7 @@ static void io_u_mark_lat_nsec(struct thread_data *td, unsigned long long nsec)
 		break;
 	case 2 ... 3:
 		idx = 1;
-		fallthrough;
+		fio_fallthrough;
 	case 0 ... 1:
 		break;
 	}
@@ -1118,7 +1161,7 @@ static void io_u_mark_lat_usec(struct thread_data *td, unsigned long long usec)
 		break;
 	case 2 ... 3:
 		idx = 1;
-		fallthrough;
+		fio_fallthrough;
 	case 0 ... 1:
 		break;
 	}
@@ -1166,7 +1209,7 @@ static void io_u_mark_lat_msec(struct thread_data *td, unsigned long long msec)
 		break;
 	case 2 ... 3:
 		idx = 1;
-		fallthrough;
+		fio_fallthrough;
 	case 0 ... 1:
 		break;
 	}
@@ -1595,7 +1638,7 @@ again:
 		assert(io_u->flags & IO_U_F_FREE);
 		io_u_clear(td, io_u, IO_U_F_FREE | IO_U_F_NO_FILE_PUT |
 				 IO_U_F_TRIMMED | IO_U_F_BARRIER |
-				 IO_U_F_VER_LIST | IO_U_F_HIGH_PRIO);
+				 IO_U_F_VER_LIST);
 
 		io_u->error = 0;
 		io_u->acct_ddir = -1;
@@ -1803,6 +1846,7 @@ struct io_u *get_io_u(struct thread_data *td)
 	 * Remember the issuing context priority. The IO engine may change this.
 	 */
 	io_u->ioprio = td->ioprio;
+	io_u->clat_prio_index = 0;
 out:
 	assert(io_u->file);
 	if (!td_io_prep(td, io_u)) {
@@ -1889,7 +1933,7 @@ static void account_io_completion(struct thread_data *td, struct io_u *io_u,
 
 		tnsec = ntime_since(&io_u->start_time, &icd->time);
 		add_lat_sample(td, idx, tnsec, bytes, io_u->offset,
-			       io_u->ioprio, io_u_is_high_prio(io_u));
+			       io_u->ioprio, io_u->clat_prio_index);
 
 		if (td->flags & TD_F_PROFILE_OPS) {
 			struct prof_io_ops *ops = &td->prof_io_ops;
@@ -1911,7 +1955,7 @@ static void account_io_completion(struct thread_data *td, struct io_u *io_u,
 	if (ddir_rw(idx)) {
 		if (!td->o.disable_clat) {
 			add_clat_sample(td, idx, llnsec, bytes, io_u->offset,
-					io_u->ioprio, io_u_is_high_prio(io_u));
+					io_u->ioprio, io_u->clat_prio_index);
 			io_u_mark_latency(td, llnsec);
 		}
 
@@ -1965,7 +2009,7 @@ static void io_completed(struct thread_data *td, struct io_u **io_u_ptr,
 	dprint_io_u(io_u, "complete");
 
 	assert(io_u->flags & IO_U_F_FLIGHT);
-	io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK);
+	io_u_clear(td, io_u, IO_U_F_FLIGHT | IO_U_F_BUSY_OK | IO_U_F_PATTERN_DONE);
 
 	/*
 	 * Mark IO ok to verify
@@ -2090,13 +2134,26 @@ static void ios_completed(struct thread_data *td,
 	}
 }
 
+static void io_u_update_bytes_done(struct thread_data *td,
+				   struct io_completion_data *icd)
+{
+	int ddir;
+
+	if (td->runstate == TD_VERIFYING) {
+		td->bytes_verified += icd->bytes_done[DDIR_READ];
+		return;
+	}
+
+	for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
+		td->bytes_done[ddir] += icd->bytes_done[ddir];
+}
+
 /*
  * Complete a single io_u for the sync engines.
  */
 int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
 {
 	struct io_completion_data icd;
-	int ddir;
 
 	init_icd(td, &icd, 1);
 	io_completed(td, &io_u, &icd);
@@ -2109,8 +2166,7 @@ int io_u_sync_complete(struct thread_data *td, struct io_u *io_u)
 		return -1;
 	}
 
-	for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
-		td->bytes_done[ddir] += icd.bytes_done[ddir];
+	io_u_update_bytes_done(td, &icd);
 
 	return 0;
 }
@@ -2122,7 +2178,7 @@ int io_u_queued_complete(struct thread_data *td, int min_evts)
 {
 	struct io_completion_data icd;
 	struct timespec *tvp = NULL;
-	int ret, ddir;
+	int ret;
 	struct timespec ts = { .tv_sec = 0, .tv_nsec = 0, };
 
 	dprint(FD_IO, "io_u_queued_complete: min=%d\n", min_evts);
@@ -2148,8 +2204,7 @@ int io_u_queued_complete(struct thread_data *td, int min_evts)
 		return -1;
 	}
 
-	for (ddir = 0; ddir < DDIR_RWDIR_CNT; ddir++)
-		td->bytes_done[ddir] += icd.bytes_done[ddir];
+	io_u_update_bytes_done(td, &icd);
 
 	return ret;
 }
@@ -2230,27 +2285,30 @@ void fill_io_buffer(struct thread_data *td, void *buf, unsigned long long min_wr
 
 	if (o->compress_percentage || o->dedupe_percentage) {
 		unsigned int perc = td->o.compress_percentage;
-		struct frand_state *rs;
+		struct frand_state *rs = NULL;
 		unsigned long long left = max_bs;
 		unsigned long long this_write;
 
 		do {
-			rs = get_buf_state(td);
+			/*
+			 * Buffers are either entirely dedupe-able or not.
+			 * If we choose to dedup, the buffer should undergo
+			 * the same manipulation as the original write. Which
+			 * means we should retrack the steps we took for compression
+			 * as well.
+			 */
+			if (!rs)
+				rs = get_buf_state(td);
 
 			min_write = min(min_write, left);
 
-			if (perc) {
-				this_write = min_not_zero(min_write,
-							(unsigned long long) td->o.compress_chunk);
+			this_write = min_not_zero(min_write,
+						(unsigned long long) td->o.compress_chunk);
 
-				fill_random_buf_percentage(rs, buf, perc,
-					this_write, this_write,
-					o->buffer_pattern,
-					o->buffer_pattern_bytes);
-			} else {
-				fill_random_buf(rs, buf, min_write);
-				this_write = min_write;
-			}
+			fill_random_buf_percentage(rs, buf, perc,
+				this_write, this_write,
+				o->buffer_pattern,
+				o->buffer_pattern_bytes);
 
 			buf += this_write;
 			left -= this_write;
@@ -2293,7 +2351,11 @@ int do_io_u_sync(const struct thread_data *td, struct io_u *io_u)
 	int ret;
 
 	if (io_u->ddir == DDIR_SYNC) {
+#ifdef CONFIG_FCNTL_SYNC
+		ret = fcntl(io_u->file->fd, F_FULLFSYNC);
+#else
 		ret = fsync(io_u->file->fd);
+#endif
 	} else if (io_u->ddir == DDIR_DATASYNC) {
 #ifdef CONFIG_FDATASYNC
 		ret = fdatasync(io_u->file->fd);