enforce arc_dnode_limit

Linux kernel shrinker in the context of null/root memcg does not scan
dentry and inode caches added by a task running in non-root memcg. For
ZFS this means that dnode cache routinely overflows, evicting valuable
meta/data and putting additional memory pressure on the system.

This patch restores zfs_prune_aliases as fallback when the kernel
shrinker does nothing, enabling zfs to actually free dnodes. Moreover,
it (indirectly) calls arc_evict when dnode_size > dnode_limit.

Reviewed-by: Rob Norris <robn@despairlabs.com>
Reviewed-by: Alexander Motin <mav@FreeBSD.org>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: Gionatan Danti <g.danti@assyoma.it>
Closes openzfs#17487
Closes openzfs#17542

Author: shodanshok

include/sys/arc_impl.h

@@ -954,7 +954,7 @@ typedef struct arc_sums {
 	wmsum_t arcstat_data_size;
 	wmsum_t arcstat_metadata_size;
 	wmsum_t arcstat_dbuf_size;
-	wmsum_t arcstat_dnode_size;
+	aggsum_t arcstat_dnode_size;
 	wmsum_t arcstat_bonus_size;
 	wmsum_t arcstat_l2_hits;
 	wmsum_t arcstat_l2_misses;

module/os/linux/zfs/zfs_vfsops.c

@@ -1216,6 +1216,63 @@ zfs_root(zfsvfs_t *zfsvfs, struct inode **ipp)
 	return (error);
 }
 
+/*
+ * Dentry and inode caches referenced by a task in non-root memcg are
+ * not going to be scanned by the kernel-provided shrinker. So, if
+ * kernel prunes nothing, fall back to this manual walk to free dnodes.
+ * To avoid scanning the same znodes multiple times they are always rotated
+ * to the end of the z_all_znodes list. New znodes are inserted at the
+ * end of the list so we're always scanning the oldest znodes first.
+ */
+static int
+zfs_prune_aliases(zfsvfs_t *zfsvfs, unsigned long nr_to_scan)
+{
+	znode_t **zp_array, *zp;
+	int max_array = MIN(nr_to_scan, PAGE_SIZE * 8 / sizeof (znode_t *));
+	int objects = 0;
+	int i = 0, j = 0;
+
+	zp_array = vmem_zalloc(max_array * sizeof (znode_t *), KM_SLEEP);
+
+	mutex_enter(&zfsvfs->z_znodes_lock);
+	while ((zp = list_head(&zfsvfs->z_all_znodes)) != NULL) {
+
+		if ((i++ > nr_to_scan) || (j >= max_array))
+			break;
+
+		ASSERT(list_link_active(&zp->z_link_node));
+		list_remove(&zfsvfs->z_all_znodes, zp);
+		list_insert_tail(&zfsvfs->z_all_znodes, zp);
+
+		/* Skip active znodes and .zfs entries */
+		if (MUTEX_HELD(&zp->z_lock) || zp->z_is_ctldir)
+			continue;
+
+		if (igrab(ZTOI(zp)) == NULL)
+			continue;
+
+		zp_array[j] = zp;
+		j++;
+	}
+	mutex_exit(&zfsvfs->z_znodes_lock);
+
+	for (i = 0; i < j; i++) {
+		zp = zp_array[i];
+
+		ASSERT3P(zp, !=, NULL);
+		d_prune_aliases(ZTOI(zp));
+
+		if (atomic_read(&ZTOI(zp)->i_count) == 1)
+			objects++;
+
+		zrele(zp);
+	}
+
+	vmem_free(zp_array, max_array * sizeof (znode_t *));
+
+	return (objects);
+}
+
 /*
  * The ARC has requested that the filesystem drop entries from the dentry
  * and inode caches.  This can occur when the ARC needs to free meta data
@@ -1267,6 +1324,14 @@ zfs_prune(struct super_block *sb, unsigned long nr_to_scan, int *objects)
 	*objects = (*shrinker->scan_objects)(shrinker, &sc);
 #endif
 
+	/*
+	 * Fall back to zfs_prune_aliases if kernel's shrinker did nothing
+	 * due to dentry and inode caches being referenced by a task running
+	 * in non-root memcg.
+	 */
+	if (*objects == 0)
+		*objects = zfs_prune_aliases(zfsvfs, nr_to_scan);
+
 	zfs_exit(zfsvfs, FTAG);
 
 	dprintf_ds(zfsvfs->z_os->os_dsl_dataset,

module/zfs/arc.c

@@ -2631,7 +2631,7 @@ arc_space_consume(uint64_t space, arc_space_type_t type)
 		ARCSTAT_INCR(arcstat_bonus_size, space);
 		break;
 	case ARC_SPACE_DNODE:
-		ARCSTAT_INCR(arcstat_dnode_size, space);
+		aggsum_add(&arc_sums.arcstat_dnode_size, space);
 		break;
 	case ARC_SPACE_DBUF:
 		ARCSTAT_INCR(arcstat_dbuf_size, space);
@@ -2677,7 +2677,7 @@ arc_space_return(uint64_t space, arc_space_type_t type)
 		ARCSTAT_INCR(arcstat_bonus_size, -space);
 		break;
 	case ARC_SPACE_DNODE:
-		ARCSTAT_INCR(arcstat_dnode_size, -space);
+		aggsum_add(&arc_sums.arcstat_dnode_size, -space);
 		break;
 	case ARC_SPACE_DBUF:
 		ARCSTAT_INCR(arcstat_dbuf_size, -space);
@@ -4490,7 +4490,7 @@ arc_evict(void)
 	 * target is not evictable or if they go over arc_dnode_limit.
 	 */
 	int64_t prune = 0;
-	int64_t dn = wmsum_value(&arc_sums.arcstat_dnode_size);
+	int64_t dn = aggsum_value(&arc_sums.arcstat_dnode_size);
 	int64_t nem = zfs_refcount_count(&arc_mru->arcs_size[ARC_BUFC_METADATA])
 	    + zfs_refcount_count(&arc_mfu->arcs_size[ARC_BUFC_METADATA])
 	    - zfs_refcount_count(&arc_mru->arcs_esize[ARC_BUFC_METADATA])
@@ -5082,11 +5082,13 @@ arc_is_overflowing(boolean_t lax, boolean_t use_reserve)
 	 * in the ARC. In practice, that's in the tens of MB, which is low
 	 * enough to be safe.
 	 */
-	int64_t over = aggsum_lower_bound(&arc_sums.arcstat_size) - arc_c -
+	int64_t arc_over = aggsum_lower_bound(&arc_sums.arcstat_size) - arc_c -
 	    zfs_max_recordsize;
+	int64_t dn_over = aggsum_lower_bound(&arc_sums.arcstat_dnode_size) -
+	    arc_dnode_limit;
 
 	/* Always allow at least one block of overflow. */
-	if (over < 0)
+	if (arc_over < 0 && dn_over <= 0)
 		return (ARC_OVF_NONE);
 
 	/* If we are under memory pressure, report severe overflow. */
@@ -5097,7 +5099,7 @@ arc_is_overflowing(boolean_t lax, boolean_t use_reserve)
 	int64_t overflow = (arc_c >> zfs_arc_overflow_shift) / 2;
 	if (use_reserve)
 		overflow *= 3;
-	return (over < overflow ? ARC_OVF_SOME : ARC_OVF_SEVERE);
+	return (arc_over < overflow ? ARC_OVF_SOME : ARC_OVF_SEVERE);
 }
 
 static abd_t *
@@ -7326,7 +7328,7 @@ arc_kstat_update(kstat_t *ksp, int rw)
 #if defined(COMPAT_FREEBSD11)
 	as->arcstat_other_size.value.ui64 =
 	    wmsum_value(&arc_sums.arcstat_bonus_size) +
-	    wmsum_value(&arc_sums.arcstat_dnode_size) +
+	    aggsum_value(&arc_sums.arcstat_dnode_size) +
 	    wmsum_value(&arc_sums.arcstat_dbuf_size);
 #endif
 
@@ -7368,7 +7370,7 @@ arc_kstat_update(kstat_t *ksp, int rw)
 	    &as->arcstat_uncached_evictable_metadata);
 
 	as->arcstat_dnode_size.value.ui64 =
-	    wmsum_value(&arc_sums.arcstat_dnode_size);
+	    aggsum_value(&arc_sums.arcstat_dnode_size);
 	as->arcstat_bonus_size.value.ui64 =
 	    wmsum_value(&arc_sums.arcstat_bonus_size);
 	as->arcstat_l2_hits.value.ui64 =
@@ -7738,7 +7740,7 @@ arc_state_init(void)
 	wmsum_init(&arc_sums.arcstat_data_size, 0);
 	wmsum_init(&arc_sums.arcstat_metadata_size, 0);
 	wmsum_init(&arc_sums.arcstat_dbuf_size, 0);
-	wmsum_init(&arc_sums.arcstat_dnode_size, 0);
+	aggsum_init(&arc_sums.arcstat_dnode_size, 0);
 	wmsum_init(&arc_sums.arcstat_bonus_size, 0);
 	wmsum_init(&arc_sums.arcstat_l2_hits, 0);
 	wmsum_init(&arc_sums.arcstat_l2_misses, 0);
@@ -7897,7 +7899,7 @@ arc_state_fini(void)
 	wmsum_fini(&arc_sums.arcstat_data_size);
 	wmsum_fini(&arc_sums.arcstat_metadata_size);
 	wmsum_fini(&arc_sums.arcstat_dbuf_size);
-	wmsum_fini(&arc_sums.arcstat_dnode_size);
+	aggsum_fini(&arc_sums.arcstat_dnode_size);
 	wmsum_fini(&arc_sums.arcstat_bonus_size);
 	wmsum_fini(&arc_sums.arcstat_l2_hits);
 	wmsum_fini(&arc_sums.arcstat_l2_misses);