Split dmu_zfetch() speculation and execution parts

To make better predictions on parallel workloads dmu_zfetch() should
be called as early as possible to reduce possible request reordering.
In particular, it should be called before dmu_buf_hold_array_by_dnode()
calls dbuf_hold(), which may sleep waiting for indirect blocks, waking
up multiple threads same time on completion, that can significantly
reorder the requests, making the stream look like random.  But we
should not issue prefetch requests before the on-demand ones, since
they may get to the disks first despite the I/O scheduler, increasing
on-demand request latency.

This patch splits dmu_zfetch() into two functions: dmu_zfetch_prepare()
and dmu_zfetch_run().  The first can be executed as early as needed.
It only updates statistics and makes predictions without issuing any
I/Os.  The I/O issuance is handled by dmu_zfetch_run(), which can be
called later when all on-demand I/Os are already issued.  It even
tracks the activity of other concurrent threads, issuing the prefetch
only when _all_ on-demand requests are issued.

For many years it was a big problem for storage servers, handling
deeper request queues from their clients, having to either serialize
consequential reads to make ZFS prefetcher usable, or execute the
incoming requests as-is and get almost no prefetch from ZFS, relying
only on deep enough prefetch by the clients.  Benefits of those ways
varied, but neither was perfect.  With this patch deeper queue
sequential read benchmarks with CrystalDiskMark from Windows via
iSCSI to FreeBSD target show me much better throughput with almost
100% prefetcher hit rate, comparing to almost zero before.

While there, I also removed per-stream zs_lock as useless, completely
covered by parent zf_lock.  Also I reused zs_blocks refcount to track
zf_stream linkage of the stream, since I believe previous zs_fetch ==
NULL check in dmu_zfetch_stream_done() was racy.

Delete prefetch streams when they reach ends of files.  It saves up
to 1KB of RAM per file, plus reduces searches through the stream list.

Block data prefetch (speculation and indirect block prefetch is still
done since they are cheaper) if all dbufs of the stream are already
in DMU cache.  First cache miss immediately fires all the prefetch
that would be done for the stream by that time.  It saves some CPU
time if same files within DMU cache capacity are read over and over.

Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Adam Moss <[email protected]>
Reviewed-by: Matthew Ahrens <[email protected]>
Signed-off-by: Alexander Motin <[email protected]>
Sponsored-By: iXsystems, Inc.
Closes #11652 

1 files changed, 147 insertions, 103 deletions

diff --git a/module/zfs/dmu_zfetch.c b/module/zfs/dmu_zfetch.c
index 5d061fe38..3d7407016 100644
--- a/module/zfs/dmu_zfetch.c
+++ b/module/zfs/dmu_zfetch.c
@@ -59,8 +59,6 @@ typedef struct zfetch_stats {
 	kstat_named_t zfetchstat_hits;
 	kstat_named_t zfetchstat_misses;
 	kstat_named_t zfetchstat_max_streams;
-	kstat_named_t zfetchstat_max_completion_us;
-	kstat_named_t zfetchstat_last_completion_us;
 	kstat_named_t zfetchstat_io_issued;
 } zfetch_stats_t;
 
@@ -68,8 +66,6 @@ static zfetch_stats_t zfetch_stats = {
 	{ "hits",			KSTAT_DATA_UINT64 },
 	{ "misses",			KSTAT_DATA_UINT64 },
 	{ "max_streams",		KSTAT_DATA_UINT64 },
-	{ "max_completion_us",		KSTAT_DATA_UINT64 },
-	{ "last_completion_us",		KSTAT_DATA_UINT64 },
 	{ "io_issued",		KSTAT_DATA_UINT64 },
 };
 
@@ -129,7 +125,7 @@ dmu_zfetch_init(zfetch_t *zf, dnode_t *dno)
 static void
 dmu_zfetch_stream_fini(zstream_t *zs)
 {
-	mutex_destroy(&zs->zs_lock);
+	ASSERT(!list_link_active(&zs->zs_node));
 	kmem_free(zs, sizeof (*zs));
 }
 
@@ -138,17 +134,10 @@ dmu_zfetch_stream_remove(zfetch_t *zf, zstream_t *zs)
 {
 	ASSERT(MUTEX_HELD(&zf->zf_lock));
 	list_remove(&zf->zf_stream, zs);
-	dmu_zfetch_stream_fini(zs);
-	zf->zf_numstreams--;
-}
-
-static void
-dmu_zfetch_stream_orphan(zfetch_t *zf, zstream_t *zs)
-{
-	ASSERT(MUTEX_HELD(&zf->zf_lock));
-	list_remove(&zf->zf_stream, zs);
-	zs->zs_fetch = NULL;
 	zf->zf_numstreams--;
+	membar_producer();
+	if (zfs_refcount_remove(&zs->zs_refs, NULL) == 0)
+		dmu_zfetch_stream_fini(zs);
 }
 
 /*
@@ -161,12 +150,8 @@ dmu_zfetch_fini(zfetch_t *zf)
 	zstream_t *zs;
 
 	mutex_enter(&zf->zf_lock);
-	while ((zs = list_head(&zf->zf_stream)) != NULL) {
-		if (zfs_refcount_count(&zs->zs_blocks) != 0)
-			dmu_zfetch_stream_orphan(zf, zs);
-		else
-			dmu_zfetch_stream_remove(zf, zs);
-	}
+	while ((zs = list_head(&zf->zf_stream)) != NULL)
+		dmu_zfetch_stream_remove(zf, zs);
 	mutex_exit(&zf->zf_lock);
 	list_destroy(&zf->zf_stream);
 	mutex_destroy(&zf->zf_lock);
@@ -195,9 +180,9 @@ dmu_zfetch_stream_create(zfetch_t *zf, uint64_t blkid)
 	    zs != NULL; zs = zs_next) {
 		zs_next = list_next(&zf->zf_stream, zs);
 		/*
-		 * Skip gethrtime() call if there are still references
+		 * Skip if still active.  1 -- zf_stream reference.
 		 */
-		if (zfs_refcount_count(&zs->zs_blocks) != 0)
+		if (zfs_refcount_count(&zs->zs_refs) != 1)
 			continue;
 		if (((now - zs->zs_atime) / NANOSEC) >
 		    zfetch_min_sec_reap)
@@ -222,12 +207,17 @@ dmu_zfetch_stream_create(zfetch_t *zf, uint64_t blkid)
 
 	zstream_t *zs = kmem_zalloc(sizeof (*zs), KM_SLEEP);
 	zs->zs_blkid = blkid;
+	zs->zs_pf_blkid1 = blkid;
 	zs->zs_pf_blkid = blkid;
+	zs->zs_ipf_blkid1 = blkid;
 	zs->zs_ipf_blkid = blkid;
 	zs->zs_atime = now;
 	zs->zs_fetch = zf;
-	zfs_refcount_create(&zs->zs_blocks);
-	mutex_init(&zs->zs_lock, NULL, MUTEX_DEFAULT, NULL);
+	zs->zs_missed = B_FALSE;
+	zfs_refcount_create(&zs->zs_callers);
+	zfs_refcount_create(&zs->zs_refs);
+	/* One reference for zf_stream. */
+	zfs_refcount_add(&zs->zs_refs, NULL);
 	zf->zf_numstreams++;
 	list_insert_head(&zf->zf_stream, zs);
 }
@@ -237,48 +227,36 @@ dmu_zfetch_stream_done(void *arg, boolean_t io_issued)
 {
 	zstream_t *zs = arg;
 
-	if (zs->zs_start_time && io_issued) {
-		hrtime_t now = gethrtime();
-		hrtime_t delta = NSEC2USEC(now - zs->zs_start_time);
-
-		zs->zs_start_time = 0;
-		ZFETCHSTAT_SET(zfetchstat_last_completion_us, delta);
-		if (delta > ZFETCHSTAT_GET(zfetchstat_max_completion_us))
-			ZFETCHSTAT_SET(zfetchstat_max_completion_us, delta);
-	}
-
-	if (zfs_refcount_remove(&zs->zs_blocks, NULL) != 0)
-		return;
-
-	/*
-	 * The parent fetch structure has gone away
-	 */
-	if (zs->zs_fetch == NULL)
+	if (zfs_refcount_remove(&zs->zs_refs, NULL) == 0)
 		dmu_zfetch_stream_fini(zs);
 }
 
 /*
- * This is the predictive prefetch entry point.  It associates dnode access
- * specified with blkid and nblks arguments with prefetch stream, predicts
- * further accesses based on that stats and initiates speculative prefetch.
+ * This is the predictive prefetch entry point.  dmu_zfetch_prepare()
+ * associates dnode access specified with blkid and nblks arguments with
+ * prefetch stream, predicts further accesses based on that stats and returns
+ * the stream pointer on success.  That pointer must later be passed to
+ * dmu_zfetch_run() to initiate the speculative prefetch for the stream and
+ * release it.  dmu_zfetch() is a wrapper for simple cases when window between
+ * prediction and prefetch initiation is not needed.
  * fetch_data argument specifies whether actual data blocks should be fetched:
  *   FALSE -- prefetch only indirect blocks for predicted data blocks;
  *   TRUE -- prefetch predicted data blocks plus following indirect blocks.
  */
-void
-dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks, boolean_t fetch_data,
-    boolean_t have_lock)
+zstream_t *
+dmu_zfetch_prepare(zfetch_t *zf, uint64_t blkid, uint64_t nblks,
+    boolean_t fetch_data, boolean_t have_lock)
 {
 	zstream_t *zs;
-	int64_t pf_start, ipf_start, ipf_istart, ipf_iend;
+	int64_t pf_start, ipf_start;
 	int64_t pf_ahead_blks, max_blks;
-	int epbs, max_dist_blks, pf_nblks, ipf_nblks, issued;
-	uint64_t end_of_access_blkid;
+	int max_dist_blks, pf_nblks, ipf_nblks;
+	uint64_t end_of_access_blkid, maxblkid;
 	end_of_access_blkid = blkid + nblks;
 	spa_t *spa = zf->zf_dnode->dn_objset->os_spa;
 
 	if (zfs_prefetch_disable)
-		return;
+		return (NULL);
 	/*
 	 * If we haven't yet loaded the indirect vdevs' mappings, we
 	 * can only read from blocks that we carefully ensure are on
@@ -287,14 +265,14 @@ dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks, boolean_t fetch_data,
 	 * blocks (e.g. of the MOS's dnode object).
 	 */
 	if (!spa_indirect_vdevs_loaded(spa))
-		return;
+		return (NULL);
 
 	/*
 	 * As a fast path for small (single-block) files, ignore access
 	 * to the first block.
 	 */
 	if (!have_lock && blkid == 0)
-		return;
+		return (NULL);
 
 	if (!have_lock)
 		rw_enter(&zf->zf_dnode->dn_struct_rwlock, RW_READER);
@@ -303,10 +281,11 @@ dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks, boolean_t fetch_data,
 	 * A fast path for small files for which no prefetch will
 	 * happen.
 	 */
-	if (zf->zf_dnode->dn_maxblkid < 2) {
+	maxblkid = zf->zf_dnode->dn_maxblkid;
+	if (maxblkid < 2) {
 		if (!have_lock)
 			rw_exit(&zf->zf_dnode->dn_struct_rwlock);
-		return;
+		return (NULL);
 	}
 	mutex_enter(&zf->zf_lock);
 
@@ -317,45 +296,47 @@ dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks, boolean_t fetch_data,
 	 */
 	for (zs = list_head(&zf->zf_stream); zs != NULL;
 	    zs = list_next(&zf->zf_stream, zs)) {
-		if (blkid == zs->zs_blkid || blkid + 1 == zs->zs_blkid) {
-			mutex_enter(&zs->zs_lock);
-			/*
-			 * zs_blkid could have changed before we
-			 * acquired zs_lock; re-check them here.
-			 */
-			if (blkid == zs->zs_blkid) {
-				break;
-			} else if (blkid + 1 == zs->zs_blkid) {
-				blkid++;
-				nblks--;
-				if (nblks == 0) {
-					/* Already prefetched this before. */
-					mutex_exit(&zs->zs_lock);
-					mutex_exit(&zf->zf_lock);
-					if (!have_lock) {
-						rw_exit(&zf->zf_dnode->
-						    dn_struct_rwlock);
-					}
-					return;
-				}
-				break;
-			}
-			mutex_exit(&zs->zs_lock);
+		if (blkid == zs->zs_blkid) {
+			break;
+		} else if (blkid + 1 == zs->zs_blkid) {
+			blkid++;
+			nblks--;
+			break;
 		}
 	}
 
+	/*
+	 * If the file is ending, remove the matching stream if found.
+	 * If not found then it is too late to create a new one now.
+	 */
+	if (end_of_access_blkid >= maxblkid) {
+		if (zs != NULL)
+			dmu_zfetch_stream_remove(zf, zs);
+		mutex_exit(&zf->zf_lock);
+		if (!have_lock)
+			rw_exit(&zf->zf_dnode->dn_struct_rwlock);
+		return (NULL);
+	}
+
+	/* Exit if we already prefetched this block before. */
+	if (nblks == 0) {
+		mutex_exit(&zf->zf_lock);
+		if (!have_lock)
+			rw_exit(&zf->zf_dnode->dn_struct_rwlock);
+		return (NULL);
+	}
+
 	if (zs == NULL) {
 		/*
 		 * This access is not part of any existing stream.  Create
 		 * a new stream for it.
 		 */
-		ZFETCHSTAT_BUMP(zfetchstat_misses);
-
 		dmu_zfetch_stream_create(zf, end_of_access_blkid);
 		mutex_exit(&zf->zf_lock);
 		if (!have_lock)
 			rw_exit(&zf->zf_dnode->dn_struct_rwlock);
-		return;
+		ZFETCHSTAT_BUMP(zfetchstat_misses);
+		return (NULL);
 	}
 
 	/*
@@ -369,6 +350,10 @@ dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks, boolean_t fetch_data,
 	 * start just after the block we just accessed.
 	 */
 	pf_start = MAX(zs->zs_pf_blkid, end_of_access_blkid);
+	if (zs->zs_pf_blkid1 < end_of_access_blkid)
+		zs->zs_pf_blkid1 = end_of_access_blkid;
+	if (zs->zs_ipf_blkid1 < end_of_access_blkid)
+		zs->zs_ipf_blkid1 = end_of_access_blkid;
 
 	/*
 	 * Double our amount of prefetched data, but don't let the
@@ -407,49 +392,108 @@ dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks, boolean_t fetch_data,
 	 * (i.e. the amount read now + the amount of data prefetched now).
 	 */
 	pf_ahead_blks = zs->zs_ipf_blkid - blkid + nblks + pf_nblks;
-	max_blks = max_dist_blks - (ipf_start - end_of_access_blkid);
+	max_blks = max_dist_blks - (ipf_start - zs->zs_pf_blkid);
 	ipf_nblks = MIN(pf_ahead_blks, max_blks);
 	zs->zs_ipf_blkid = ipf_start + ipf_nblks;
 
-	epbs = zf->zf_dnode->dn_indblkshift - SPA_BLKPTRSHIFT;
-	ipf_istart = P2ROUNDUP(ipf_start, 1 << epbs) >> epbs;
-	ipf_iend = P2ROUNDUP(zs->zs_ipf_blkid, 1 << epbs) >> epbs;
-
-	zs->zs_atime = gethrtime();
-	/* no prior reads in progress */
-	if (zfs_refcount_count(&zs->zs_blocks) == 0)
-		zs->zs_start_time = zs->zs_atime;
 	zs->zs_blkid = end_of_access_blkid;
-	zfs_refcount_add_many(&zs->zs_blocks, pf_nblks + ipf_iend - ipf_istart,
-	    NULL);
-	mutex_exit(&zs->zs_lock);
+	/* Protect the stream from reclamation. */
+	zs->zs_atime = gethrtime();
+	zfs_refcount_add(&zs->zs_refs, NULL);
+	/* Count concurrent callers. */
+	zfs_refcount_add(&zs->zs_callers, NULL);
 	mutex_exit(&zf->zf_lock);
-	issued = 0;
+
+	if (!have_lock)
+		rw_exit(&zf->zf_dnode->dn_struct_rwlock);
+
+	ZFETCHSTAT_BUMP(zfetchstat_hits);
+	return (zs);
+}
+
+void
+dmu_zfetch_run(zstream_t *zs, boolean_t missed, boolean_t have_lock)
+{
+	zfetch_t *zf = zs->zs_fetch;
+	int64_t pf_start, pf_end, ipf_start, ipf_end;
+	int epbs, issued;
+
+	if (missed)
+		zs->zs_missed = missed;
 
 	/*
-	 * dbuf_prefetch() is asynchronous (even when it needs to read
-	 * indirect blocks), but we still prefer to drop our locks before
-	 * calling it to reduce the time we hold them.
+	 * Postpone the prefetch if there are more concurrent callers.
+	 * It happens when multiple requests are waiting for the same
+	 * indirect block.  The last one will run the prefetch for all.
 	 */
+	if (zfs_refcount_remove(&zs->zs_callers, NULL) != 0) {
+		/* Drop reference taken in dmu_zfetch_prepare(). */
+		if (zfs_refcount_remove(&zs->zs_refs, NULL) == 0)
+			dmu_zfetch_stream_fini(zs);
+		return;
+	}
 
-	for (int i = 0; i < pf_nblks; i++) {
-		issued += dbuf_prefetch_impl(zf->zf_dnode, 0, pf_start + i,
+	mutex_enter(&zf->zf_lock);
+	if (zs->zs_missed) {
+		pf_start = zs->zs_pf_blkid1;
+		pf_end = zs->zs_pf_blkid1 = zs->zs_pf_blkid;
+	} else {
+		pf_start = pf_end = 0;
+	}
+	ipf_start = MAX(zs->zs_pf_blkid1, zs->zs_ipf_blkid1);
+	ipf_end = zs->zs_ipf_blkid1 = zs->zs_ipf_blkid;
+	mutex_exit(&zf->zf_lock);
+	ASSERT3S(pf_start, <=, pf_end);
+	ASSERT3S(ipf_start, <=, ipf_end);
+
+	epbs = zf->zf_dnode->dn_indblkshift - SPA_BLKPTRSHIFT;
+	ipf_start = P2ROUNDUP(ipf_start, 1 << epbs) >> epbs;
+	ipf_end = P2ROUNDUP(ipf_end, 1 << epbs) >> epbs;
+	ASSERT3S(ipf_start, <=, ipf_end);
+	issued = pf_end - pf_start + ipf_end - ipf_start;
+	if (issued > 1) {
+		/* More references on top of taken in dmu_zfetch_prepare(). */
+		zfs_refcount_add_many(&zs->zs_refs, issued - 1, NULL);
+	} else if (issued == 0) {
+		/* Some other thread has done our work, so drop the ref. */
+		if (zfs_refcount_remove(&zs->zs_refs, NULL) == 0)
+			dmu_zfetch_stream_fini(zs);
+		return;
+	}
+
+	if (!have_lock)
+		rw_enter(&zf->zf_dnode->dn_struct_rwlock, RW_READER);
+
+	issued = 0;
+	for (int64_t blk = pf_start; blk < pf_end; blk++) {
+		issued += dbuf_prefetch_impl(zf->zf_dnode, 0, blk,
 		    ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH,
 		    dmu_zfetch_stream_done, zs);
 	}
-	for (int64_t iblk = ipf_istart; iblk < ipf_iend; iblk++) {
+	for (int64_t iblk = ipf_start; iblk < ipf_end; iblk++) {
 		issued += dbuf_prefetch_impl(zf->zf_dnode, 1, iblk,
 		    ZIO_PRIORITY_ASYNC_READ, ARC_FLAG_PREDICTIVE_PREFETCH,
 		    dmu_zfetch_stream_done, zs);
 	}
+
 	if (!have_lock)
 		rw_exit(&zf->zf_dnode->dn_struct_rwlock);
-	ZFETCHSTAT_BUMP(zfetchstat_hits);
 
 	if (issued)
 		ZFETCHSTAT_ADD(zfetchstat_io_issued, issued);
 }
 
+void
+dmu_zfetch(zfetch_t *zf, uint64_t blkid, uint64_t nblks, boolean_t fetch_data,
+    boolean_t missed, boolean_t have_lock)
+{
+	zstream_t *zs;
+
+	zs = dmu_zfetch_prepare(zf, blkid, nblks, fetch_data, have_lock);
+	if (zs)
+		dmu_zfetch_run(zs, missed, have_lock);
+}
+
 /* BEGIN CSTYLED */
 ZFS_MODULE_PARAM(zfs_prefetch, zfs_prefetch_, disable, INT, ZMOD_RW,
 	"Disable all ZFS prefetching");