Implement uncached prefetch

Previously the primarycache property was handled only in the dbuf
layer. Since the speculative prefetcher is implemented in the ARC,
it had to be disabled for uncacheable buffers.

This change gives the ARC knowledge about uncacheable buffers
via  arc_read() and arc_write(). So when remove_reference() drops
the last reference on the ARC header, it can either immediately destroy
it, or if it is marked as prefetch, put it into a new arc_uncached state. 
That state is scanned every second, evicting stale buffers that were
not demand read.

This change also tracks dbufs that were read from the beginning,
but not to the end.  It is assumed that such buffers may receive further
reads, and so they are stored in dbuf cache. If a following
reads reaches the end of the buffer, it is immediately evicted.
Otherwise it will follow regular dbuf cache eviction.  Since the dbuf
layer does not know actual file sizes, this logic is not applied to
the final buffer of a dnode.

Since uncacheable buffers should no longer stay in the ARC for long,
this patch also tries to optimize I/O by allocating ARC physical
buffers as linear to allow buffer sharing.

Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: George Wilson <[email protected]>
Reviewed-by: Ryan Moeller <[email protected]>
Signed-off-by: Alexander Motin <[email protected]>
Sponsored by: iXsystems, Inc.
Closes #14243

5 files changed, 196 insertions, 129 deletions

diff --git a/module/zfs/arc.c b/module/zfs/arc.c
index 1521caa6e..0d8aff947 100644
--- a/module/zfs/arc.c
+++ b/module/zfs/arc.c
@@ -333,6 +333,7 @@ static int arc_state_evict_marker_count;
 
 static kmutex_t arc_evict_lock;
 static boolean_t arc_evict_needed = B_FALSE;
+static clock_t arc_last_uncached_flush;
 
 /*
  * Count of bytes evicted since boot.
@@ -473,13 +474,14 @@ static uint_t zfs_arc_lotsfree_percent = 10;
  */
 static int zfs_arc_prune_task_threads = 1;
 
-/* The 6 states: */
+/* The 7 states: */
 arc_state_t ARC_anon;
 arc_state_t ARC_mru;
 arc_state_t ARC_mru_ghost;
 arc_state_t ARC_mfu;
 arc_state_t ARC_mfu_ghost;
 arc_state_t ARC_l2c_only;
+arc_state_t ARC_uncached;
 
 arc_stats_t arc_stats = {
 	{ "hits",			KSTAT_DATA_UINT64 },
@@ -501,6 +503,7 @@ arc_stats_t arc_stats = {
 	{ "mru_ghost_hits",		KSTAT_DATA_UINT64 },
 	{ "mfu_hits",			KSTAT_DATA_UINT64 },
 	{ "mfu_ghost_hits",		KSTAT_DATA_UINT64 },
+	{ "uncached_hits",		KSTAT_DATA_UINT64 },
 	{ "deleted",			KSTAT_DATA_UINT64 },
 	{ "mutex_miss",			KSTAT_DATA_UINT64 },
 	{ "access_skip",		KSTAT_DATA_UINT64 },
@@ -549,6 +552,9 @@ arc_stats_t arc_stats = {
 	{ "mfu_ghost_size",		KSTAT_DATA_UINT64 },
 	{ "mfu_ghost_evictable_data",	KSTAT_DATA_UINT64 },
 	{ "mfu_ghost_evictable_metadata", KSTAT_DATA_UINT64 },
+	{ "uncached_size",		KSTAT_DATA_UINT64 },
+	{ "uncached_evictable_data",	KSTAT_DATA_UINT64 },
+	{ "uncached_evictable_metadata", KSTAT_DATA_UINT64 },
 	{ "l2_hits",			KSTAT_DATA_UINT64 },
 	{ "l2_misses",			KSTAT_DATA_UINT64 },
 	{ "l2_prefetch_asize",		KSTAT_DATA_UINT64 },
@@ -702,6 +708,7 @@ taskq_t *arc_prune_taskq;
 	((hdr)->b_flags & ARC_FLAG_COMPRESSED_ARC)
 
 #define	HDR_L2CACHE(hdr)	((hdr)->b_flags & ARC_FLAG_L2CACHE)
+#define	HDR_UNCACHED(hdr)	((hdr)->b_flags & ARC_FLAG_UNCACHED)
 #define	HDR_L2_READING(hdr)	\
 	(((hdr)->b_flags & ARC_FLAG_IO_IN_PROGRESS) &&	\
 	((hdr)->b_flags & ARC_FLAG_HAS_L2HDR))
@@ -854,6 +861,7 @@ enum arc_hdr_alloc_flags {
 	ARC_HDR_ALLOC_RDATA = 0x1,
 	ARC_HDR_DO_ADAPT = 0x2,
 	ARC_HDR_USE_RESERVE = 0x4,
+	ARC_HDR_ALLOC_LINEAR = 0x8,
 };
 
 
@@ -866,8 +874,10 @@ static void arc_free_data_impl(arc_buf_hdr_t *hdr, uint64_t size,
     const void *tag);
 static void arc_hdr_free_abd(arc_buf_hdr_t *, boolean_t);
 static void arc_hdr_alloc_abd(arc_buf_hdr_t *, int);
+static void arc_hdr_destroy(arc_buf_hdr_t *);
 static void arc_access(arc_buf_hdr_t *, arc_flags_t, boolean_t);
 static void arc_buf_watch(arc_buf_t *);
+static void arc_change_state(arc_state_t *, arc_buf_hdr_t *);
 
 static arc_buf_contents_t arc_buf_type(arc_buf_hdr_t *);
 static uint32_t arc_bufc_to_flags(arc_buf_contents_t);
@@ -2321,18 +2331,23 @@ remove_reference(arc_buf_hdr_t *hdr, const void *tag)
 
 	ASSERT(HDR_HAS_L1HDR(hdr));
 	ASSERT(state == arc_anon || MUTEX_HELD(HDR_LOCK(hdr)));
-	ASSERT(!GHOST_STATE(state));
+	ASSERT(!GHOST_STATE(state));	/* arc_l2c_only counts as a ghost. */
 
-	/*
-	 * arc_l2c_only counts as a ghost state so we don't need to explicitly
-	 * check to prevent usage of the arc_l2c_only list.
-	 */
-	if (((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) == 0) &&
-	    (state != arc_anon)) {
-		multilist_insert(&state->arcs_list[arc_buf_type(hdr)], hdr);
-		arc_evictable_space_increment(hdr, state);
+	if ((cnt = zfs_refcount_remove(&hdr->b_l1hdr.b_refcnt, tag)) != 0)
+		return (cnt);
+
+	if (state == arc_anon) {
+		arc_hdr_destroy(hdr);
+		return (0);
 	}
-	return (cnt);
+	if (state == arc_uncached && !HDR_PREFETCH(hdr)) {
+		arc_change_state(arc_anon, hdr);
+		arc_hdr_destroy(hdr);
+		return (0);
+	}
+	multilist_insert(&state->arcs_list[arc_buf_type(hdr)], hdr);
+	arc_evictable_space_increment(hdr, state);
+	return (0);
 }
 
 /*
@@ -2439,8 +2454,12 @@ arc_change_state(arc_state_t *new_state, arc_buf_hdr_t *hdr)
 	if (refcnt == 0) {
 		if (old_state != arc_anon && old_state != arc_l2c_only) {
 			ASSERT(HDR_HAS_L1HDR(hdr));
-			multilist_remove(&old_state->arcs_list[buftype], hdr);
-			arc_evictable_space_decrement(hdr, old_state);
+			/* remove_reference() saves on insert. */
+			if (multilist_link_active(&hdr->b_l1hdr.b_arc_node)) {
+				multilist_remove(&old_state->arcs_list[buftype],
+				    hdr);
+				arc_evictable_space_decrement(hdr, old_state);
+			}
 		}
 		if (new_state != arc_anon && new_state != arc_l2c_only) {
 			/*
@@ -3845,7 +3864,6 @@ arc_buf_destroy(arc_buf_t *buf, const void *tag)
 		ASSERT3U(hdr->b_l1hdr.b_bufcnt, ==, 1);
 		ASSERT(!HDR_IO_IN_PROGRESS(hdr));
 		VERIFY0(remove_reference(hdr, tag));
-		arc_hdr_destroy(hdr);
 		return;
 	}
 
@@ -3858,8 +3876,8 @@ arc_buf_destroy(arc_buf_t *buf, const void *tag)
 	ASSERT3P(hdr->b_l1hdr.b_state, !=, arc_anon);
 	ASSERT3P(buf->b_data, !=, NULL);
 
-	(void) remove_reference(hdr, tag);
 	arc_buf_destroy_impl(buf);
+	(void) remove_reference(hdr, tag);
 	mutex_exit(hash_lock);
 }
 
@@ -3874,6 +3892,7 @@ arc_buf_destroy(arc_buf_t *buf, const void *tag)
  *    - arc_mru_ghost -> deleted
  *    - arc_mfu_ghost -> arc_l2c_only
  *    - arc_mfu_ghost -> deleted
+ *    - arc_uncached -> deleted
  *
  * Return total size of evicted data buffers for eviction progress tracking.
  * When evicting from ghost states return logical buffer size to make eviction
@@ -3941,8 +3960,9 @@ arc_evict_hdr(arc_buf_hdr_t *hdr, uint64_t *real_evicted)
 		return (bytes_evicted);
 	}
 
-	ASSERT(state == arc_mru || state == arc_mfu);
-	evicted_state = (state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost;
+	ASSERT(state == arc_mru || state == arc_mfu || state == arc_uncached);
+	evicted_state = (state == arc_uncached) ? arc_anon :
+	    ((state == arc_mru) ? arc_mru_ghost : arc_mfu_ghost);
 
 	/* prefetch buffers have a minimum lifespan */
 	if ((hdr->b_flags & (ARC_FLAG_PREFETCH | ARC_FLAG_INDIRECT)) &&
@@ -4013,9 +4033,13 @@ arc_evict_hdr(arc_buf_hdr_t *hdr, uint64_t *real_evicted)
 			arc_hdr_free_abd(hdr, B_TRUE);
 
 		arc_change_state(evicted_state, hdr);
-		ASSERT(HDR_IN_HASH_TABLE(hdr));
-		arc_hdr_set_flags(hdr, ARC_FLAG_IN_HASH_TABLE);
 		DTRACE_PROBE1(arc__evict, arc_buf_hdr_t *, hdr);
+		if (evicted_state == arc_anon) {
+			arc_hdr_destroy(hdr);
+			*real_evicted += HDR_FULL_SIZE;
+		} else {
+			ASSERT(HDR_IN_HASH_TABLE(hdr));
+		}
 	}
 
 	return (bytes_evicted);
@@ -4779,6 +4803,9 @@ arc_flush(spa_t *spa, boolean_t retry)
 
 	(void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_DATA, retry);
 	(void) arc_flush_state(arc_mfu_ghost, guid, ARC_BUFC_METADATA, retry);
+
+	(void) arc_flush_state(arc_uncached, guid, ARC_BUFC_DATA, retry);
+	(void) arc_flush_state(arc_uncached, guid, ARC_BUFC_METADATA, retry);
 }
 
 void
@@ -4910,7 +4937,16 @@ arc_evict_cb_check(void *arg, zthr_t *zthr)
 	 * which is held before this function is called, and is held by
 	 * arc_wait_for_eviction() when it calls zthr_wakeup().
 	 */
-	return (arc_evict_needed);
+	if (arc_evict_needed)
+		return (B_TRUE);
+
+	/*
+	 * If we have buffers in uncached state, evict them periodically.
+	 */
+	return ((zfs_refcount_count(&arc_uncached->arcs_esize[ARC_BUFC_DATA]) +
+	    zfs_refcount_count(&arc_uncached->arcs_esize[ARC_BUFC_METADATA]) &&
+	    ddi_get_lbolt() - arc_last_uncached_flush >
+	    MSEC_TO_TICK(arc_min_prefetch_ms / 2)));
 }
 
 /*
@@ -4925,8 +4961,14 @@ arc_evict_cb(void *arg, zthr_t *zthr)
 	uint64_t evicted = 0;
 	fstrans_cookie_t cookie = spl_fstrans_mark();
 
-	/* Evict from cache */
-	evicted = arc_evict();
+	/* Always try to evict from uncached state. */
+	arc_last_uncached_flush = ddi_get_lbolt();
+	evicted += arc_flush_state(arc_uncached, 0, ARC_BUFC_DATA, B_FALSE);
+	evicted += arc_flush_state(arc_uncached, 0, ARC_BUFC_METADATA, B_FALSE);
+
+	/* Evict from other states only if told to. */
+	if (arc_evict_needed)
+		evicted += arc_evict();
 
 	/*
 	 * If evicted is zero, we couldn't evict anything
@@ -5205,12 +5247,10 @@ arc_get_data_abd(arc_buf_hdr_t *hdr, uint64_t size, const void *tag,
 	arc_buf_contents_t type = arc_buf_type(hdr);
 
 	arc_get_data_impl(hdr, size, tag, alloc_flags);
-	if (type == ARC_BUFC_METADATA) {
-		return (abd_alloc(size, B_TRUE));
-	} else {
-		ASSERT(type == ARC_BUFC_DATA);
-		return (abd_alloc(size, B_FALSE));
-	}
+	if (alloc_flags & ARC_HDR_ALLOC_LINEAR)
+		return (abd_alloc_linear(size, type == ARC_BUFC_METADATA));
+	else
+		return (abd_alloc(size, type == ARC_BUFC_METADATA));
 }
 
 static void *
@@ -5475,16 +5515,22 @@ arc_access(arc_buf_hdr_t *hdr, arc_flags_t arc_flags, boolean_t hit)
 
 	clock_t now = ddi_get_lbolt();
 	if (hdr->b_l1hdr.b_state == arc_anon) {
+		arc_state_t	*new_state;
 		/*
-		 * This buffer is not in the cache, and does not
-		 * appear in our "ghost" list.  Add the new buffer
-		 * to the MRU state.
+		 * This buffer is not in the cache, and does not appear in
+		 * our "ghost" lists.  Add it to the MRU or uncached state.
 		 */
 		ASSERT0(hdr->b_l1hdr.b_arc_access);
 		hdr->b_l1hdr.b_arc_access = now;
-		DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
-		arc_change_state(arc_mru, hdr);
-
+		if (HDR_UNCACHED(hdr)) {
+			new_state = arc_uncached;
+			DTRACE_PROBE1(new_state__uncached, arc_buf_hdr_t *,
+			    hdr);
+		} else {
+			new_state = arc_mru;
+			DTRACE_PROBE1(new_state__mru, arc_buf_hdr_t *, hdr);
+		}
+		arc_change_state(new_state, hdr);
 	} else if (hdr->b_l1hdr.b_state == arc_mru) {
 		/*
 		 * This buffer has been accessed once recently and either
@@ -5556,6 +5602,14 @@ arc_access(arc_buf_hdr_t *hdr, arc_flags_t arc_flags, boolean_t hit)
 		hdr->b_l1hdr.b_arc_access = now;
 		DTRACE_PROBE1(new_state__mfu, arc_buf_hdr_t *, hdr);
 		arc_change_state(arc_mfu, hdr);
+	} else if (hdr->b_l1hdr.b_state == arc_uncached) {
+		/*
+		 * This buffer is uncacheable, but we got a hit.  Probably
+		 * a demand read after prefetch.  Nothing more to do here.
+		 */
+		if (!HDR_IO_IN_PROGRESS(hdr))
+			ARCSTAT_BUMP(arcstat_uncached_hits);
+		hdr->b_l1hdr.b_arc_access = now;
 	} else if (hdr->b_l1hdr.b_state == arc_l2c_only) {
 		/*
 		 * This buffer is on the 2nd Level ARC and was not accessed
@@ -5603,7 +5657,8 @@ arc_buf_access(arc_buf_t *buf)
 	mutex_exit(&buf->b_evict_lock);
 
 	ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
-	    hdr->b_l1hdr.b_state == arc_mfu);
+	    hdr->b_l1hdr.b_state == arc_mfu ||
+	    hdr->b_l1hdr.b_state == arc_uncached);
 
 	DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
 	arc_access(hdr, 0, B_TRUE);
@@ -5671,7 +5726,6 @@ arc_read_done(zio_t *zio)
 	kmutex_t	*hash_lock = NULL;
 	arc_callback_t	*callback_list;
 	arc_callback_t	*acb;
-	boolean_t	freeable = B_FALSE;
 
 	/*
 	 * The hdr was inserted into hash-table and removed from lists
@@ -5821,9 +5875,6 @@ arc_read_done(zio_t *zio)
 	 */
 	ASSERT(callback_cnt < 2 || hash_lock != NULL);
 
-	arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
-	(void) remove_reference(hdr, hdr); /* For IO_IN_PROGRESS. */
-
 	if (zio->io_error == 0) {
 		arc_hdr_verify(hdr, zio->io_bp);
 	} else {
@@ -5832,7 +5883,6 @@ arc_read_done(zio_t *zio)
 			arc_change_state(arc_anon, hdr);
 		if (HDR_IN_HASH_TABLE(hdr))
 			buf_hash_remove(hdr);
-		freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
 	}
 
 	/*
@@ -5842,18 +5892,11 @@ arc_read_done(zio_t *zio)
 	 */
 	cv_broadcast(&hdr->b_l1hdr.b_cv);
 
-	if (hash_lock != NULL) {
+	arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
+	(void) remove_reference(hdr, hdr);
+
+	if (hash_lock != NULL)
 		mutex_exit(hash_lock);
-	} else {
-		/*
-		 * This block was freed while we waited for the read to
-		 * complete.  It has been removed from the hash table and
-		 * moved to the anonymous state (so that it won't show up
-		 * in the cache).
-		 */
-		ASSERT3P(hdr->b_l1hdr.b_state, ==, arc_anon);
-		freeable = zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt);
-	}
 
 	/* execute each callback and free its structure */
 	while ((acb = callback_list) != NULL) {
@@ -5889,9 +5932,6 @@ arc_read_done(zio_t *zio)
 			kmem_free(acb, sizeof (arc_callback_t));
 		}
 	}
-
-	if (freeable)
-		arc_hdr_destroy(hdr);
 }
 
 /*
@@ -6072,7 +6112,8 @@ top:
 		}
 
 		ASSERT(hdr->b_l1hdr.b_state == arc_mru ||
-		    hdr->b_l1hdr.b_state == arc_mfu);
+		    hdr->b_l1hdr.b_state == arc_mfu ||
+		    hdr->b_l1hdr.b_state == arc_uncached);
 
 		DTRACE_PROBE1(arc__hit, arc_buf_hdr_t *, hdr);
 		arc_access(hdr, *arc_flags, B_TRUE);
@@ -6099,9 +6140,9 @@ top:
 				}
 			}
 			if (rc != 0) {
-				(void) remove_reference(hdr, private);
 				arc_buf_destroy_impl(buf);
 				buf = NULL;
+				(void) remove_reference(hdr, private);
 			}
 
 			/* assert any errors weren't due to unloaded keys */
@@ -6191,6 +6232,11 @@ top:
 				goto top;
 			}
 		}
+		if (*arc_flags & ARC_FLAG_UNCACHED) {
+			arc_hdr_set_flags(hdr, ARC_FLAG_UNCACHED);
+			if (!encrypted_read)
+				alloc_flags |= ARC_HDR_ALLOC_LINEAR;
+		}
 
 		/*
 		 * Call arc_adapt() explicitly before arc_access() to allow
@@ -6624,7 +6670,7 @@ arc_release(arc_buf_t *buf, const void *tag)
 		VERIFY3U(hdr->b_type, ==, type);
 
 		ASSERT(hdr->b_l1hdr.b_buf != buf || buf->b_next != NULL);
-		(void) remove_reference(hdr, tag);
+		VERIFY3S(remove_reference(hdr, tag), >, 0);
 
 		if (arc_buf_is_shared(buf) && !ARC_BUF_COMPRESSED(buf)) {
 			ASSERT3P(hdr->b_l1hdr.b_buf, !=, buf);
@@ -6897,7 +6943,8 @@ arc_write_ready(zio_t *zio)
 		arc_hdr_alloc_abd(hdr, ARC_HDR_DO_ADAPT | ARC_HDR_ALLOC_RDATA |
 		    ARC_HDR_USE_RESERVE);
 		abd_copy(hdr->b_crypt_hdr.b_rabd, zio->io_abd, psize);
-	} else if (!abd_size_alloc_linear(arc_buf_size(buf)) ||
+	} else if (!(HDR_UNCACHED(hdr) ||
+	    abd_size_alloc_linear(arc_buf_size(buf))) ||
 	    !arc_can_share(hdr, buf)) {
 		/*
 		 * Ideally, we would always copy the io_abd into b_pabd, but the
@@ -7025,17 +7072,16 @@ arc_write_done(zio_t *zio)
 			}
 		}
 		arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
-		(void) remove_reference(hdr, hdr); /* For IO_IN_PROGRESS. */
+		VERIFY3S(remove_reference(hdr, hdr), >, 0);
 		/* if it's not anon, we are doing a scrub */
 		if (exists == NULL && hdr->b_l1hdr.b_state == arc_anon)
 			arc_access(hdr, 0, B_FALSE);
 		mutex_exit(hash_lock);
 	} else {
 		arc_hdr_clear_flags(hdr, ARC_FLAG_IO_IN_PROGRESS);
-		(void) remove_reference(hdr, hdr); /* For IO_IN_PROGRESS. */
+		VERIFY3S(remove_reference(hdr, hdr), >, 0);
 	}
 
-	ASSERT(!zfs_refcount_is_zero(&hdr->b_l1hdr.b_refcnt));
 	callback->awcb_done(zio, buf, callback->awcb_private);
 
 	abd_free(zio->io_abd);
@@ -7044,7 +7090,7 @@ arc_write_done(zio_t *zio)
 
 zio_t *
 arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
-    blkptr_t *bp, arc_buf_t *buf, boolean_t l2arc,
+    blkptr_t *bp, arc_buf_t *buf, boolean_t uncached, boolean_t l2arc,
     const zio_prop_t *zp, arc_write_done_func_t *ready,
     arc_write_done_func_t *children_ready, arc_write_done_func_t *physdone,
     arc_write_done_func_t *done, void *private, zio_priority_t priority,
@@ -7061,7 +7107,9 @@ arc_write(zio_t *pio, spa_t *spa, uint64_t txg,
 	ASSERT(!HDR_IO_IN_PROGRESS(hdr));
 	ASSERT3P(hdr->b_l1hdr.b_acb, ==, NULL);
 	ASSERT3U(hdr->b_l1hdr.b_bufcnt, >, 0);
-	if (l2arc)
+	if (uncached)
+		arc_hdr_set_flags(hdr, ARC_FLAG_UNCACHED);
+	else if (l2arc)
 		arc_hdr_set_flags(hdr, ARC_FLAG_L2CACHE);
 
 	if (ARC_BUF_ENCRYPTED(buf)) {
@@ -7283,6 +7331,8 @@ arc_kstat_update(kstat_t *ksp, int rw)
 	    wmsum_value(&arc_sums.arcstat_mfu_hits);
 	as->arcstat_mfu_ghost_hits.value.ui64 =
 	    wmsum_value(&arc_sums.arcstat_mfu_ghost_hits);
+	as->arcstat_uncached_hits.value.ui64 =
+	    wmsum_value(&arc_sums.arcstat_uncached_hits);
 	as->arcstat_deleted.value.ui64 =
 	    wmsum_value(&arc_sums.arcstat_deleted);
 	as->arcstat_mutex_miss.value.ui64 =
@@ -7352,6 +7402,10 @@ arc_kstat_update(kstat_t *ksp, int rw)
 	    &as->arcstat_mfu_ghost_size,
 	    &as->arcstat_mfu_ghost_evictable_data,
 	    &as->arcstat_mfu_ghost_evictable_metadata);
+	arc_kstat_update_state(arc_uncached,
+	    &as->arcstat_uncached_size,
+	    &as->arcstat_uncached_evictable_data,
+	    &as->arcstat_uncached_evictable_metadata);
 
 	as->arcstat_dnode_size.value.ui64 =
 	    aggsum_value(&arc_sums.arcstat_dnode_size);
@@ -7660,6 +7714,10 @@ arc_state_init(void)
 	    arc_state_multilist_index_func, &num_sublists);
 	arc_state_multilist_init(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
 	    arc_state_multilist_index_func, &num_sublists);
+	arc_state_multilist_init(&arc_uncached->arcs_list[ARC_BUFC_METADATA],
+	    arc_state_multilist_index_func, &num_sublists);
+	arc_state_multilist_init(&arc_uncached->arcs_list[ARC_BUFC_DATA],
+	    arc_state_multilist_index_func, &num_sublists);
 
 	/*
 	 * L2 headers should never be on the L2 state list since they don't
@@ -7689,6 +7747,8 @@ arc_state_init(void)
 	zfs_refcount_create(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
 	zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
 	zfs_refcount_create(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
+	zfs_refcount_create(&arc_uncached->arcs_esize[ARC_BUFC_METADATA]);
+	zfs_refcount_create(&arc_uncached->arcs_esize[ARC_BUFC_DATA]);
 
 	zfs_refcount_create(&arc_anon->arcs_size);
 	zfs_refcount_create(&arc_mru->arcs_size);
@@ -7696,6 +7756,7 @@ arc_state_init(void)
 	zfs_refcount_create(&arc_mfu->arcs_size);
 	zfs_refcount_create(&arc_mfu_ghost->arcs_size);
 	zfs_refcount_create(&arc_l2c_only->arcs_size);
+	zfs_refcount_create(&arc_uncached->arcs_size);
 
 	wmsum_init(&arc_sums.arcstat_hits, 0);
 	wmsum_init(&arc_sums.arcstat_iohits, 0);
@@ -7716,6 +7777,7 @@ arc_state_init(void)
 	wmsum_init(&arc_sums.arcstat_mru_ghost_hits, 0);
 	wmsum_init(&arc_sums.arcstat_mfu_hits, 0);
 	wmsum_init(&arc_sums.arcstat_mfu_ghost_hits, 0);
+	wmsum_init(&arc_sums.arcstat_uncached_hits, 0);
 	wmsum_init(&arc_sums.arcstat_deleted, 0);
 	wmsum_init(&arc_sums.arcstat_mutex_miss, 0);
 	wmsum_init(&arc_sums.arcstat_access_skip, 0);
@@ -7800,6 +7862,7 @@ arc_state_init(void)
 	arc_mfu->arcs_state = ARC_STATE_MFU;
 	arc_mfu_ghost->arcs_state = ARC_STATE_MFU_GHOST;
 	arc_l2c_only->arcs_state = ARC_STATE_L2C_ONLY;
+	arc_uncached->arcs_state = ARC_STATE_UNCACHED;
 }
 
 static void
@@ -7817,6 +7880,8 @@ arc_state_fini(void)
 	zfs_refcount_destroy(&arc_mfu_ghost->arcs_esize[ARC_BUFC_DATA]);
 	zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_METADATA]);
 	zfs_refcount_destroy(&arc_l2c_only->arcs_esize[ARC_BUFC_DATA]);
+	zfs_refcount_destroy(&arc_uncached->arcs_esize[ARC_BUFC_METADATA]);
+	zfs_refcount_destroy(&arc_uncached->arcs_esize[ARC_BUFC_DATA]);
 
 	zfs_refcount_destroy(&arc_anon->arcs_size);
 	zfs_refcount_destroy(&arc_mru->arcs_size);
@@ -7824,6 +7889,7 @@ arc_state_fini(void)
 	zfs_refcount_destroy(&arc_mfu->arcs_size);
 	zfs_refcount_destroy(&arc_mfu_ghost->arcs_size);
 	zfs_refcount_destroy(&arc_l2c_only->arcs_size);
+	zfs_refcount_destroy(&arc_uncached->arcs_size);
 
 	multilist_destroy(&arc_mru->arcs_list[ARC_BUFC_METADATA]);
 	multilist_destroy(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA]);
@@ -7835,6 +7901,8 @@ arc_state_fini(void)
 	multilist_destroy(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA]);
 	multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA]);
 	multilist_destroy(&arc_l2c_only->arcs_list[ARC_BUFC_DATA]);
+	multilist_destroy(&arc_uncached->arcs_list[ARC_BUFC_METADATA]);
+	multilist_destroy(&arc_uncached->arcs_list[ARC_BUFC_DATA]);
 
 	wmsum_fini(&arc_sums.arcstat_hits);
 	wmsum_fini(&arc_sums.arcstat_iohits);
@@ -7855,6 +7923,7 @@ arc_state_fini(void)
 	wmsum_fini(&arc_sums.arcstat_mru_ghost_hits);
 	wmsum_fini(&arc_sums.arcstat_mfu_hits);
 	wmsum_fini(&arc_sums.arcstat_mfu_ghost_hits);
+	wmsum_fini(&arc_sums.arcstat_uncached_hits);
 	wmsum_fini(&arc_sums.arcstat_deleted);
 	wmsum_fini(&arc_sums.arcstat_mutex_miss);
 	wmsum_fini(&arc_sums.arcstat_access_skip);
@@ -8039,8 +8108,8 @@ arc_init(void)
 
 	arc_state_evict_markers =
 	    arc_state_alloc_markers(arc_state_evict_marker_count);
-	arc_evict_zthr = zthr_create("arc_evict",
-	    arc_evict_cb_check, arc_evict_cb, NULL, defclsyspri);
+	arc_evict_zthr = zthr_create_timer("arc_evict",
+	    arc_evict_cb_check, arc_evict_cb, NULL, SEC2NSEC(1), defclsyspri);
 	arc_reap_zthr = zthr_create_timer("arc_reap",
 	    arc_reap_cb_check, arc_reap_cb, NULL, SEC2NSEC(1), minclsyspri);
 
diff --git a/module/zfs/dbuf.c b/module/zfs/dbuf.c
index 5e1ed8386..efaa13317 100644
--- a/module/zfs/dbuf.c
+++ b/module/zfs/dbuf.c
@@ -1608,7 +1608,9 @@ dbuf_read_impl(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags,
 	DTRACE_SET_STATE(db, "read issued");
 	mutex_exit(&db->db_mtx);
 
-	if (dbuf_is_l2cacheable(db))
+	if (!DBUF_IS_CACHEABLE(db))
+		aflags |= ARC_FLAG_UNCACHED;
+	else if (dbuf_is_l2cacheable(db))
 		aflags |= ARC_FLAG_L2CACHE;
 
 	dbuf_add_ref(db, NULL);
@@ -1736,10 +1738,13 @@ dbuf_read(dmu_buf_impl_t *db, zio_t *zio, uint32_t flags)
 	dn = DB_DNODE(db);
 
 	prefetch = db->db_level == 0 && db->db_blkid != DMU_BONUS_BLKID &&
-	    (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL &&
-	    DBUF_IS_CACHEABLE(db);
+	    (flags & DB_RF_NOPREFETCH) == 0 && dn != NULL;
 
 	mutex_enter(&db->db_mtx);
+	if (flags & DB_RF_PARTIAL_FIRST)
+		db->db_partial_read = B_TRUE;
+	else if (!(flags & DB_RF_PARTIAL_MORE))
+		db->db_partial_read = B_FALSE;
 	if (db->db_state == DB_CACHED) {
 		/*
 		 * Ensure that this block's dnode has been decrypted if
@@ -3463,8 +3468,9 @@ dbuf_prefetch_impl(dnode_t *dn, int64_t level, uint64_t blkid,
 	dpa->dpa_cb = cb;
 	dpa->dpa_arg = arg;
 
-	/* flag if L2ARC eligible, l2arc_noprefetch then decides */
-	if (dnode_level_is_l2cacheable(&bp, dn, level))
+	if (!DNODE_LEVEL_IS_CACHEABLE(dn, level))
+		dpa->dpa_aflags |= ARC_FLAG_UNCACHED;
+	else if (dnode_level_is_l2cacheable(&bp, dn, level))
 		dpa->dpa_aflags |= ARC_FLAG_L2CACHE;
 
 	/*
@@ -3853,59 +3859,38 @@ dbuf_rele_and_unlock(dmu_buf_impl_t *db, const void *tag, boolean_t evicting)
 			 * This dbuf has anonymous data associated with it.
 			 */
 			dbuf_destroy(db);
-		} else {
-			boolean_t do_arc_evict = B_FALSE;
-			blkptr_t bp;
-			spa_t *spa = dmu_objset_spa(db->db_objset);
-
-			if (!DBUF_IS_CACHEABLE(db) &&
-			    db->db_blkptr != NULL &&
-			    !BP_IS_HOLE(db->db_blkptr) &&
-			    !BP_IS_EMBEDDED(db->db_blkptr)) {
-				do_arc_evict = B_TRUE;
-				bp = *db->db_blkptr;
-			}
-
-			if (!DBUF_IS_CACHEABLE(db) ||
-			    db->db_pending_evict) {
-				dbuf_destroy(db);
-			} else if (!multilist_link_active(&db->db_cache_link)) {
-				ASSERT3U(db->db_caching_status, ==,
-				    DB_NO_CACHE);
-
-				dbuf_cached_state_t dcs =
-				    dbuf_include_in_metadata_cache(db) ?
-				    DB_DBUF_METADATA_CACHE : DB_DBUF_CACHE;
-				db->db_caching_status = dcs;
-
-				multilist_insert(&dbuf_caches[dcs].cache, db);
-				uint64_t db_size = db->db.db_size;
-				size = zfs_refcount_add_many(
-				    &dbuf_caches[dcs].size, db_size, db);
-				uint8_t db_level = db->db_level;
-				mutex_exit(&db->db_mtx);
-
-				if (dcs == DB_DBUF_METADATA_CACHE) {
-					DBUF_STAT_BUMP(metadata_cache_count);
-					DBUF_STAT_MAX(
-					    metadata_cache_size_bytes_max,
-					    size);
-				} else {
-					DBUF_STAT_BUMP(cache_count);
-					DBUF_STAT_MAX(cache_size_bytes_max,
-					    size);
-					DBUF_STAT_BUMP(cache_levels[db_level]);
-					DBUF_STAT_INCR(
-					    cache_levels_bytes[db_level],
-					    db_size);
-				}
+		} else if (!(DBUF_IS_CACHEABLE(db) || db->db_partial_read) ||
+		    db->db_pending_evict) {
+			dbuf_destroy(db);
+		} else if (!multilist_link_active(&db->db_cache_link)) {
+			ASSERT3U(db->db_caching_status, ==, DB_NO_CACHE);
+
+			dbuf_cached_state_t dcs =
+			    dbuf_include_in_metadata_cache(db) ?
+			    DB_DBUF_METADATA_CACHE : DB_DBUF_CACHE;
+			db->db_caching_status = dcs;
+
+			multilist_insert(&dbuf_caches[dcs].cache, db);
+			uint64_t db_size = db->db.db_size;
+			size = zfs_refcount_add_many(
+			    &dbuf_caches[dcs].size, db_size, db);
+			uint8_t db_level = db->db_level;
+			mutex_exit(&db->db_mtx);
 
-				if (dcs == DB_DBUF_CACHE && !evicting)
-					dbuf_evict_notify(size);
+			if (dcs == DB_DBUF_METADATA_CACHE) {
+				DBUF_STAT_BUMP(metadata_cache_count);
+				DBUF_STAT_MAX(metadata_cache_size_bytes_max,
+				    size);
+			} else {
+				DBUF_STAT_BUMP(cache_count);
+				DBUF_STAT_MAX(cache_size_bytes_max, size);
+				DBUF_STAT_BUMP(cache_levels[db_level]);
+				DBUF_STAT_INCR(cache_levels_bytes[db_level],
+				    db_size);
 			}
 
-			if (do_arc_evict)
-				arc_freed(spa, &bp);
+			if (dcs == DB_DBUF_CACHE && !evicting)
+				dbuf_evict_notify(size);
 		}
 	} else {
 		mutex_exit(&db->db_mtx);
@@ -5083,8 +5068,8 @@ dbuf_write(dbuf_dirty_record_t *dr, arc_buf_t *data, dmu_tx_t *tx)
 			children_ready_cb = dbuf_write_children_ready;
 
 		dr->dr_zio = arc_write(pio, os->os_spa, txg,
-		    &dr->dr_bp_copy, data, dbuf_is_l2cacheable(db),
-		    &zp, dbuf_write_ready,
+		    &dr->dr_bp_copy, data, !DBUF_IS_CACHEABLE(db),
+		    dbuf_is_l2cacheable(db), &zp, dbuf_write_ready,
 		    children_ready_cb, dbuf_write_physdone,
 		    dbuf_write_done, db, ZIO_PRIORITY_ASYNC_WRITE,
 		    ZIO_FLAG_MUSTSUCCEED, &zb);
diff --git a/module/zfs/dmu.c b/module/zfs/dmu.c
index 45304e7dd..d6a9f813c 100644
--- a/module/zfs/dmu.c
+++ b/module/zfs/dmu.c
@@ -549,14 +549,14 @@ dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
 		    ZIO_FLAG_CANFAIL);
 	blkid = dbuf_whichblock(dn, 0, offset);
 	if ((flags & DMU_READ_NO_PREFETCH) == 0 &&
-	    DNODE_META_IS_CACHEABLE(dn) && length <= zfetch_array_rd_sz) {
+	    length <= zfetch_array_rd_sz) {
 		/*
 		 * Prepare the zfetch before initiating the demand reads, so
 		 * that if multiple threads block on same indirect block, we
 		 * base predictions on the original less racy request order.
 		 */
-		zs = dmu_zfetch_prepare(&dn->dn_zfetch, blkid, nblks,
-		    read && DNODE_IS_CACHEABLE(dn), B_TRUE);
+		zs = dmu_zfetch_prepare(&dn->dn_zfetch, blkid, nblks, read,
+		    B_TRUE);
 	}
 	for (i = 0; i < nblks; i++) {
 		dmu_buf_impl_t *db = dbuf_hold(dn, blkid + i, tag);
@@ -579,6 +579,14 @@ dmu_buf_hold_array_by_dnode(dnode_t *dn, uint64_t offset, uint64_t length,
 		 * state will not yet be CACHED.
 		 */
 		if (read) {
+			if (i == nblks - 1 && blkid + i < dn->dn_maxblkid &&
+			    offset + length < db->db.db_offset +
+			    db->db.db_size) {
+				if (offset <= db->db.db_offset)
+					dbuf_flags |= DB_RF_PARTIAL_FIRST;
+				else
+					dbuf_flags |= DB_RF_PARTIAL_MORE;
+			}
 			(void) dbuf_read(db, zio, dbuf_flags);
 			if (db->db_state != DB_CACHED)
 				missed = B_TRUE;
@@ -1850,8 +1858,8 @@ dmu_sync(zio_t *pio, uint64_t txg, dmu_sync_cb_t *done, zgd_t *zgd)
 	dsa->dsa_zgd = zgd;
 	dsa->dsa_tx = NULL;
 
-	zio_nowait(arc_write(pio, os->os_spa, txg,
-	    zgd->zgd_bp, dr->dt.dl.dr_data, dbuf_is_l2cacheable(db),
+	zio_nowait(arc_write(pio, os->os_spa, txg, zgd->zgd_bp,
+	    dr->dt.dl.dr_data, !DBUF_IS_CACHEABLE(db), dbuf_is_l2cacheable(db),
 	    &zp, dmu_sync_ready, NULL, NULL, dmu_sync_done, dsa,
 	    ZIO_PRIORITY_SYNC_WRITE, ZIO_FLAG_CANFAIL, &zb));
 
diff --git a/module/zfs/dmu_objset.c b/module/zfs/dmu_objset.c
index c17c829a0..5083b1763 100644
--- a/module/zfs/dmu_objset.c
+++ b/module/zfs/dmu_objset.c
@@ -1694,7 +1694,7 @@ dmu_objset_sync(objset_t *os, zio_t *pio, dmu_tx_t *tx)
 	}
 
 	zio = arc_write(pio, os->os_spa, tx->tx_txg,
-	    blkptr_copy, os->os_phys_buf, dmu_os_is_l2cacheable(os),
+	    blkptr_copy, os->os_phys_buf, B_FALSE, dmu_os_is_l2cacheable(os),
 	    &zp, dmu_objset_write_ready, NULL, NULL, dmu_objset_write_done,
 	    os, ZIO_PRIORITY_ASYNC_WRITE, ZIO_FLAG_MUSTSUCCEED, &zb);
 
diff --git a/module/zfs/dmu_tx.c b/module/zfs/dmu_tx.c
index 28f64369d..a7a554691 100644
--- a/module/zfs/dmu_tx.c
+++ b/module/zfs/dmu_tx.c
@@ -214,7 +214,12 @@ dmu_tx_check_ioerr(zio_t *zio, dnode_t *dn, int level, uint64_t blkid)
 	rw_exit(&dn->dn_struct_rwlock);
 	if (db == NULL)
 		return (SET_ERROR(EIO));
-	err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH);
+	/*
+	 * PARTIAL_FIRST allows caching for uncacheable blocks.  It will
+	 * be cleared after dmu_buf_will_dirty() call dbuf_read() again.
+	 */
+	err = dbuf_read(db, zio, DB_RF_CANFAIL | DB_RF_NOPREFETCH |
+	    (level == 0 ? DB_RF_PARTIAL_FIRST : 0));
 	dbuf_rele(db, FTAG);
 	return (err);
 }