Avoid memory allocations in the ARC eviction thread

When the eviction thread goes to shrink an ARC state, it allocates a set
of marker buffers used to hold its place in the state's sublists.

This can be problematic in low memory conditions, since
1) the allocation can be substantial, as we allocate NCPU markers;
2) on at least FreeBSD, page reclamation can block in
   arc_wait_for_eviction()

In particular, in stress tests it's possible to hit a deadlock on
FreeBSD when the number of free pages is very low, wherein the system is
waiting for the page daemon to reclaim memory, the page daemon is
waiting for the ARC eviction thread to finish, and the ARC eviction
thread is blocked waiting for more memory.

Try to reduce the likelihood of such deadlocks by pre-allocating markers
for the eviction thread at ARC initialization time.  When evicting
buffers from an ARC state, check to see if the current thread is the ARC
eviction thread, and use the pre-allocated markers for that purpose
rather than dynamically allocating them.

Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: George Amanakis <gamanakis@gmail.com>
Signed-off-by: Mark Johnston <markj@FreeBSD.org>
Closes openzfs#12985

include/sys/zthr.h

@@ -38,6 +38,7 @@ extern void zthr_resume(zthr_t *t);
 extern void zthr_wait_cycle_done(zthr_t *t);
 
 extern boolean_t zthr_iscancelled(zthr_t *t);
+extern boolean_t zthr_iscurthread(zthr_t *t);
 extern boolean_t zthr_has_waiters(zthr_t *t);
 
 #endif /* _SYS_ZTHR_H */

module/zfs/arc.c

@@ -328,6 +328,8 @@ static zthr_t *arc_reap_zthr;
  * arc_evict(), which improves arc_is_overflowing().
  */
 static zthr_t *arc_evict_zthr;
+static arc_buf_hdr_t **arc_state_evict_markers;
+static int arc_state_evict_marker_count;
 
 static kmutex_t arc_evict_lock;
 static boolean_t arc_evict_needed = B_FALSE;
@@ -4166,6 +4168,38 @@ arc_evict_state_impl(multilist_t *ml, int idx, arc_buf_hdr_t *marker,
 	return (bytes_evicted);
 }
 
+/*
+ * Allocate an array of buffer headers used as placeholders during arc state
+ * eviction.
+ */
+static arc_buf_hdr_t **
+arc_state_alloc_markers(int count)
+{
+	arc_buf_hdr_t **markers;
+
+	markers = kmem_zalloc(sizeof (*markers) * count, KM_SLEEP);
+	for (int i = 0; i < count; i++) {
+		markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
+
+		/*
+		 * A b_spa of 0 is used to indicate that this header is
+		 * a marker. This fact is used in arc_evict_type() and
+		 * arc_evict_state_impl().
+		 */
+		markers[i]->b_spa = 0;
+
+	}
+	return (markers);
+}
+
+static void
+arc_state_free_markers(arc_buf_hdr_t **markers, int count)
+{
+	for (int i = 0; i < count; i++)
+		kmem_cache_free(hdr_full_cache, markers[i]);
+	kmem_free(markers, sizeof (*markers) * count);
+}
+
 /*
  * Evict buffers from the given arc state, until we've removed the
  * specified number of bytes. Move the removed buffers to the
@@ -4197,19 +4231,15 @@ arc_evict_state(arc_state_t *state, uint64_t spa, uint64_t bytes,
 	 * pick up where we left off for each individual sublist, rather
 	 * than starting from the tail each time.
 	 */
-	markers = kmem_zalloc(sizeof (*markers) * num_sublists, KM_SLEEP);
+	if (zthr_iscurthread(arc_evict_zthr)) {
+		markers = arc_state_evict_markers;
+		ASSERT3S(num_sublists, <=, arc_state_evict_marker_count);
+	} else {
+		markers = arc_state_alloc_markers(num_sublists);
+	}
 	for (int i = 0; i < num_sublists; i++) {
 		multilist_sublist_t *mls;
 
-		markers[i] = kmem_cache_alloc(hdr_full_cache, KM_SLEEP);
-
-		/*
-		 * A b_spa of 0 is used to indicate that this header is
-		 * a marker. This fact is used in arc_evict_type() and
-		 * arc_evict_state_impl().
-		 */
-		markers[i]->b_spa = 0;
-
 		mls = multilist_sublist_lock(ml, i);
 		multilist_sublist_insert_tail(mls, markers[i]);
 		multilist_sublist_unlock(mls);
@@ -4291,10 +4321,9 @@ arc_evict_state(arc_state_t *state, uint64_t spa, uint64_t bytes,
 		multilist_sublist_t *mls = multilist_sublist_lock(ml, i);
 		multilist_sublist_remove(mls, markers[i]);
 		multilist_sublist_unlock(mls);
-
-		kmem_cache_free(hdr_full_cache, markers[i]);
 	}
-	kmem_free(markers, sizeof (*markers) * num_sublists);
+	if (markers != arc_state_evict_markers)
+		arc_state_free_markers(markers, num_sublists);
 
 	return (total_evicted);
 }
@@ -7618,53 +7647,52 @@ arc_tuning_update(boolean_t verbose)
 	WARN_IF_TUNING_IGNORED(zfs_arc_sys_free, arc_sys_free, verbose);
 }
 
+static void
+arc_state_multilist_init(multilist_t *ml,
+    multilist_sublist_index_func_t *index_func, int *maxcountp)
+{
+	multilist_create(ml, sizeof (arc_buf_hdr_t),
+	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node), index_func);
+	*maxcountp = MAX(*maxcountp, multilist_get_num_sublists(ml));
+}
+
 static void
 arc_state_init(void)
 {
-	multilist_create(&arc_mru->arcs_list[ARC_BUFC_METADATA],
-	    sizeof (arc_buf_hdr_t),
-	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
-	    arc_state_multilist_index_func);
-	multilist_create(&arc_mru->arcs_list[ARC_BUFC_DATA],
-	    sizeof (arc_buf_hdr_t),
-	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
-	    arc_state_multilist_index_func);
-	multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
-	    sizeof (arc_buf_hdr_t),
-	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
-	    arc_state_multilist_index_func);
-	multilist_create(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
-	    sizeof (arc_buf_hdr_t),
-	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
-	    arc_state_multilist_index_func);
-	multilist_create(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
-	    sizeof (arc_buf_hdr_t),
-	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
-	    arc_state_multilist_index_func);
-	multilist_create(&arc_mfu->arcs_list[ARC_BUFC_DATA],
-	    sizeof (arc_buf_hdr_t),
-	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
-	    arc_state_multilist_index_func);
-	multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
-	    sizeof (arc_buf_hdr_t),
-	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
-	    arc_state_multilist_index_func);
-	multilist_create(&arc_mfu_ghost->arcs_list[ARC_BUFC_DATA],
-	    sizeof (arc_buf_hdr_t),
-	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
-	    arc_state_multilist_index_func);
+	int num_sublists = 0;
+
+	arc_state_multilist_init(&arc_mru->arcs_list[ARC_BUFC_METADATA],
+	    arc_state_multilist_index_func, &num_sublists);
+	arc_state_multilist_init(&arc_mru->arcs_list[ARC_BUFC_DATA],
+	    arc_state_multilist_index_func, &num_sublists);
+	arc_state_multilist_init(&arc_mru_ghost->arcs_list[ARC_BUFC_METADATA],
+	    arc_state_multilist_index_func, &num_sublists);
+	arc_state_multilist_init(&arc_mru_ghost->arcs_list[ARC_BUFC_DATA],
+	    arc_state_multilist_index_func, &num_sublists);
+	arc_state_multilist_init(&arc_mfu->arcs_list[ARC_BUFC_METADATA],
+	    arc_state_multilist_index_func, &num_sublists);
+	arc_state_multilist_init(&arc_mfu->arcs_list[ARC_BUFC_DATA],
+	    arc_state_multilist_index_func, &num_sublists);
+	arc_state_multilist_init(&arc_mfu_ghost->arcs_list[ARC_BUFC_METADATA],
+	    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);
+
 	/*
 	 * L2 headers should never be on the L2 state list since they don't
 	 * have L1 headers allocated.  Special index function asserts that.
 	 */
-	multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
-	    sizeof (arc_buf_hdr_t),
-	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
-	    arc_state_l2c_multilist_index_func);
-	multilist_create(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
-	    sizeof (arc_buf_hdr_t),
-	    offsetof(arc_buf_hdr_t, b_l1hdr.b_arc_node),
-	    arc_state_l2c_multilist_index_func);
+	arc_state_multilist_init(&arc_l2c_only->arcs_list[ARC_BUFC_METADATA],
+	    arc_state_l2c_multilist_index_func, &num_sublists);
+	arc_state_multilist_init(&arc_l2c_only->arcs_list[ARC_BUFC_DATA],
+	    arc_state_l2c_multilist_index_func, &num_sublists);
+
+	/*
+	 * Keep track of the number of markers needed to reclaim buffers from
+	 * any ARC state.  The markers will be pre-allocated so as to minimize
+	 * the number of memory allocations performed by the eviction thread.
+	 */
+	arc_state_evict_marker_count = num_sublists;
 
 	zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_METADATA]);
 	zfs_refcount_create(&arc_anon->arcs_esize[ARC_BUFC_DATA]);
@@ -8008,6 +8036,8 @@ arc_init(void)
 		kstat_install(arc_ksp);
 	}
 
+	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_reap_zthr = zthr_create_timer("arc_reap",
@@ -8087,6 +8117,8 @@ arc_fini(void)
 
 	(void) zthr_cancel(arc_evict_zthr);
 	(void) zthr_cancel(arc_reap_zthr);
+	arc_state_free_markers(arc_state_evict_markers,
+	    arc_state_evict_marker_count);
 
 	mutex_destroy(&arc_evict_lock);
 	list_destroy(&arc_evict_waiters);

module/zfs/zthr.c

@@ -469,6 +469,12 @@ zthr_iscancelled(zthr_t *t)
 	return (cancelled);
 }
 
+boolean_t
+zthr_iscurthread(zthr_t *t)
+{
+	return (t->zthr_thread == curthread);
+}
+
 /*
  * Wait for the zthr to finish its current function. Similar to
  * zthr_iscancelled, you can use zthr_has_waiters to have the zthr_func end