L2ARC: Lazy sublist reset flags for persistent markers

Replace direct marker-to-tail manipulation with per-sublist boolean
flags consumed lazily by feed threads.  Each scanning thread resets its
own marker when it sees the flag, rather than having another thread
manipulate the marker directly.

Reviewed-by: Brian Behlendorf <[email protected]>
Reviewed-by: Alexander Motin <[email protected]>
Signed-off-by: Ameer Hamza <[email protected]>
Closes #18289

Author: ixhamza

include/sys/arc_impl.h

@@ -56,10 +56,12 @@ typedef struct l2arc_info {
 	uint64_t	l2arc_total_capacity;	/* total L2ARC capacity */
 	uint64_t	l2arc_smallest_capacity; /* smallest device capacity */
 	/*
-	 * Per-device thread coordination for sublist processing
+	 * Per-device thread coordination for sublist processing.
+	 * reset: flags sublist marker for lazy reset to tail.
 	 */
 	boolean_t	*l2arc_sublist_busy[L2ARC_FEED_TYPES];
-	kmutex_t	l2arc_sublist_lock;	/* protects busy flags */
+	boolean_t	*l2arc_sublist_reset[L2ARC_FEED_TYPES];
+	kmutex_t	l2arc_sublist_lock;	/* protects busy/reset flags */
 	int		l2arc_next_sublist[L2ARC_FEED_TYPES]; /* round-robin */
 } l2arc_info_t;
 

module/zfs/arc.c

@@ -9073,6 +9073,8 @@ l2arc_pool_markers_init(spa_t *spa)
 		    arc_state_alloc_markers(num_sublists);
 		spa->spa_l2arc_info.l2arc_sublist_busy[pass] =
 		    kmem_zalloc(num_sublists * sizeof (boolean_t), KM_SLEEP);
+		spa->spa_l2arc_info.l2arc_sublist_reset[pass] =
+		    kmem_zalloc(num_sublists * sizeof (boolean_t), KM_SLEEP);
 
 		for (int i = 0; i < num_sublists; i++) {
 			multilist_sublist_t *mls =
@@ -9117,12 +9119,18 @@ l2arc_pool_markers_fini(spa_t *spa)
 		    num_sublists);
 		spa->spa_l2arc_info.l2arc_markers[pass] = NULL;
 
-		/* Free sublist busy flags for this pass */
+		/* Free sublist busy and reset flags for this pass */
 		ASSERT3P(spa->spa_l2arc_info.l2arc_sublist_busy[pass], !=,
 		    NULL);
 		kmem_free(spa->spa_l2arc_info.l2arc_sublist_busy[pass],
 		    num_sublists * sizeof (boolean_t));
 		spa->spa_l2arc_info.l2arc_sublist_busy[pass] = NULL;
+
+		ASSERT3P(spa->spa_l2arc_info.l2arc_sublist_reset[pass], !=,
+		    NULL);
+		kmem_free(spa->spa_l2arc_info.l2arc_sublist_reset[pass],
+		    num_sublists * sizeof (boolean_t));
+		spa->spa_l2arc_info.l2arc_sublist_reset[pass] = NULL;
 	}
 
 	mutex_destroy(&spa->spa_l2arc_info.l2arc_sublist_lock);
@@ -9608,6 +9616,19 @@ l2arc_write_sublist(spa_t *spa, l2arc_dev_t *dev, int pass, int sublist_idx,
 	persistent_marker = spa->spa_l2arc_info.
 	    l2arc_markers[pass][sublist_idx];
 
+	/*
+	 * Check if this sublist's marker was flagged for reset to tail.
+	 * This handles depth cap resets and global resets without needing
+	 * to coordinate with actively-scanning threads.
+	 */
+	if (save_position &&
+	    spa->spa_l2arc_info.l2arc_sublist_reset[pass][sublist_idx]) {
+		multilist_sublist_remove(mls, persistent_marker);
+		multilist_sublist_insert_tail(mls, persistent_marker);
+		spa->spa_l2arc_info.l2arc_sublist_reset[pass][sublist_idx] =
+		    B_FALSE;
+	}
+
 	if (save_position && persistent_marker == multilist_sublist_head(mls)) {
 		multilist_sublist_unlock(mls);
 		return (B_FALSE);
@@ -9798,14 +9819,24 @@ l2arc_write_sublist(spa_t *spa, l2arc_dev_t *dev, int pass, int sublist_idx,
 	}
 
 	/*
-	 * Position persistent marker for next iteration. In case of
-	 * save_position, validate that prev_hdr still belongs to the current
-	 * sublist. The sublist lock is dropped during L2ARC write I/O, allowing
-	 * ARC eviction to potentially free prev_hdr. If freed, we can't do much
-	 * except to reset the marker.
+	 * Position persistent marker for next iteration.
+	 *
+	 * If a reset was flagged during our scan (sublist lock was dropped
+	 * for I/O, allowing another thread to set the flag), honor it by
+	 * moving the marker to tail instead of advancing.
+	 *
+	 * Otherwise, validate that prev_hdr still belongs to the current
+	 * sublist.  The sublist lock is dropped during L2ARC write I/O,
+	 * allowing ARC eviction to potentially free prev_hdr.  If freed,
+	 * we can't do much except to reset the marker.
 	 */
 	multilist_sublist_remove(mls, persistent_marker);
 	if (save_position &&
+	    spa->spa_l2arc_info.l2arc_sublist_reset[pass][sublist_idx]) {
+		multilist_sublist_insert_tail(mls, persistent_marker);
+		spa->spa_l2arc_info.l2arc_sublist_reset[pass][sublist_idx] =
+		    B_FALSE;
+	} else if (save_position &&
 	    multilist_link_active(&prev_hdr->b_l1hdr.b_arc_node)) {
 		if (hdr != NULL) {
 			/*
@@ -9845,40 +9876,33 @@ l2arc_blk_fetch_done(zio_t *zio)
 }
 
 /*
- * Reset all L2ARC markers to tail position for the given spa.
+ * Flag all sublists for a single pass for lazy marker reset to tail.
+ * Each sublist's marker will be reset when next visited by a feed thread.
  */
 static void
-l2arc_reset_all_markers(spa_t *spa)
+l2arc_flag_pass_reset(spa_t *spa, int pass)
 {
-	ASSERT(spa->spa_l2arc_info.l2arc_markers != NULL);
 	ASSERT(MUTEX_HELD(&spa->spa_l2arc_info.l2arc_sublist_lock));
 
-	for (int pass = 0; pass < L2ARC_FEED_TYPES; pass++) {
-		if (spa->spa_l2arc_info.l2arc_markers[pass] == NULL)
-			continue;
-
-		multilist_t *ml = l2arc_get_list(pass);
-		int num_sublists = multilist_get_num_sublists(ml);
-
-		for (int i = 0; i < num_sublists; i++) {
-			ASSERT3P(spa->spa_l2arc_info.l2arc_markers[pass][i],
-			    !=, NULL);
-			multilist_sublist_t *mls =
-			    multilist_sublist_lock_idx(ml, i);
-
-			/* Remove from current position */
-			ASSERT(multilist_link_active(&spa->spa_l2arc_info.
-			    l2arc_markers[pass][i]->b_l1hdr.b_arc_node));
-			multilist_sublist_remove(mls, spa->spa_l2arc_info.
-			    l2arc_markers[pass][i]);
+	multilist_t *ml = l2arc_get_list(pass);
+	int num_sublists = multilist_get_num_sublists(ml);
 
-			/* Insert at tail (like initialization) */
-			multilist_sublist_insert_tail(mls,
-			    spa->spa_l2arc_info.l2arc_markers[pass][i]);
-
-			multilist_sublist_unlock(mls);
-		}
+	for (int i = 0; i < num_sublists; i++) {
+		multilist_sublist_t *mls = multilist_sublist_lock_idx(ml, i);
+		spa->spa_l2arc_info.l2arc_sublist_reset[pass][i] = B_TRUE;
+		multilist_sublist_unlock(mls);
 	}
+}
+
+/*
+ * Flag all L2ARC markers for lazy reset to tail for the given spa.
+ * Each sublist's marker will be reset when next visited by a feed thread.
+ */
+static void
+l2arc_reset_all_markers(spa_t *spa)
+{
+	for (int pass = 0; pass < L2ARC_FEED_TYPES; pass++)
+		l2arc_flag_pass_reset(spa, pass);
 
 	/* Reset write counter */
 	spa->spa_l2arc_info.l2arc_total_writes = 0;