[PATCH v4 1/2] ext4: fix shrinker scan budget accounting in ext4_es_scan()

[Qiliang Yuan]

do_shrink_slab() reads count_objects() once per invocation to derive
a one-shot scan budget, then calls scan_objects() repeatedly until
that budget is exhausted or scan_objects() returns SHRINK_STOP.
include/linux/shrinker.h documents that scan_objects() "should track
its actual progress" in sc->nr_scanned, so do_shrink_slab() can tell
when there is nothing left to examine and stop early.

ext4_es_scan() never updates sc->nr_scanned, so it defaults to the
full sc->nr_to_scan on every call. do_shrink_slab() therefore always
believes a full batch was examined, regardless of what __es_shrink()
actually did, and keeps calling scan_objects() until the budget
derived from the (possibly stale) percpu extent_status count is
drained, even after sbi->s_es_list has nothing left to examine.

Make __es_shrink() report the number of extent_status objects it
actually examined through a new nr_scanned output parameter, derived
from the existing per-extent nr_to_scan counter that
es_reclaim_extents() already decrements as it walks the tree. Have
ext4_es_scan() copy this into sc->nr_scanned, and return SHRINK_STOP
once it comes back zero. Key this off nr_scanned rather than
nr_shrunk: a batch that finds every extent still referenced
correctly ages them without freeing any, and that is real progress
nr_shrunk would miss.

nr_scanned can come back zero for more than a genuinely empty list:
es_stats_shk_cnt is a stale percpu count, an inode can be
momentarily skipped (precached, lock contended), or durably have
nothing shrinkable (i_es_shk_nr == 0) without es_reclaim_extents()
ever touching nr_to_scan. Treat all of these alike: not stopping
doesn't help the transient cases either, since do_shrink_slab()'s
budget only decrements by nr_scanned, so retrying forever at
nr_scanned == 0 never terminates.

Tested by fallocate(2)-ing 10000 4K files (to populate the shrinker
with reclaimable unwritten extents without also exercising the
extent_status "referenced" second-chance path, which needs a
separate two-pass accounting of its own) and triggering
"echo 2 > /proc/sys/vm/drop_caches", while tracing the
ext4_es_shrink* tracepoints:

                      total scan_objects()   calls with
                      calls                  nr_shrunk == 0
  before this patch   429                    189 (44%)
  after this patch    239                      1 (0.4%)

Fixes: 1ab6c4997e04 ("fs: convert fs shrinkers to new scan/count API")
Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Qiliang Yuan <odys.yuan@xxxxxxxxx>
Reviewed-by: Jan Kara <jack@xxxxxxx>
---
 fs/ext4/extents_status.c | 32 ++++++++++++++++++++++++++++----
 1 file changed, 28 insertions(+), 4 deletions(-)

diff --git a/fs/ext4/extents_status.c b/fs/ext4/extents_status.c
index 6e4a191e82191..b8821b93693a9 100644
--- a/fs/ext4/extents_status.c
+++ b/fs/ext4/extents_status.c
@@ -184,7 +184,7 @@ static int __es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
 			      struct extent_status *prealloc);
 static int es_reclaim_extents(struct ext4_inode_info *ei, int *nr_to_scan);
 static int __es_shrink(struct ext4_sb_info *sbi, int nr_to_scan,
-		       struct ext4_inode_info *locked_ei);
+		       struct ext4_inode_info *locked_ei, int *nr_scanned);
 static int __revise_pending(struct inode *inode, ext4_lblk_t lblk,
 			    ext4_lblk_t len,
 			    struct pending_reservation **prealloc);
@@ -1670,7 +1670,7 @@ void ext4_es_remove_extent(struct inode *inode, ext4_lblk_t lblk,
 }
 
 static int __es_shrink(struct ext4_sb_info *sbi, int nr_to_scan,
-		       struct ext4_inode_info *locked_ei)
+		       struct ext4_inode_info *locked_ei, int *nr_scanned)
 {
 	struct ext4_inode_info *ei;
 	struct ext4_es_stats *es_stats;
@@ -1679,6 +1679,7 @@ static int __es_shrink(struct ext4_sb_info *sbi, int nr_to_scan,
 	int nr_to_walk;
 	int nr_shrunk = 0;
 	int retried = 0, nr_skipped = 0;
+	int orig_nr_to_scan = nr_to_scan;
 
 	es_stats = &sbi->s_es_stats;
 	start_time = ktime_get();
@@ -1738,6 +1739,8 @@ static int __es_shrink(struct ext4_sb_info *sbi, int nr_to_scan,
 		nr_shrunk = es_reclaim_extents(locked_ei, &nr_to_scan);
 
 out:
+	*nr_scanned = orig_nr_to_scan - nr_to_scan;
+
 	scan_time = ktime_to_ns(ktime_sub(ktime_get(), start_time));
 	if (likely(es_stats->es_stats_scan_time))
 		es_stats->es_stats_scan_time = (scan_time +
@@ -1774,15 +1777,36 @@ static unsigned long ext4_es_scan(struct shrinker *shrink,
 {
 	struct ext4_sb_info *sbi = shrink->private_data;
 	int nr_to_scan = sc->nr_to_scan;
-	int ret, nr_shrunk;
+	int ret, nr_shrunk, nr_scanned;
 
 	ret = percpu_counter_read_positive(&sbi->s_es_stats.es_stats_shk_cnt);
 	trace_ext4_es_shrink_scan_enter(sbi->s_sb, nr_to_scan, ret);
 
-	nr_shrunk = __es_shrink(sbi, nr_to_scan, NULL);
+	nr_shrunk = __es_shrink(sbi, nr_to_scan, NULL, &nr_scanned);
+	sc->nr_scanned = nr_scanned;
 
 	ret = percpu_counter_read_positive(&sbi->s_es_stats.es_stats_shk_cnt);
 	trace_ext4_es_shrink_scan_exit(sbi->s_sb, nr_shrunk, ret);
+
+	/*
+	 * Key SHRINK_STOP off nr_scanned (extents actually examined), not
+	 * nr_shrunk (extents actually freed): a batch that finds every
+	 * extent still referenced correctly ages them without freeing
+	 * any, and that is real progress nr_shrunk would miss.
+	 *
+	 * nr_scanned can come back 0 for more than a genuinely empty
+	 * sbi->s_es_list: es_stats_shk_cnt is a stale percpu count, an
+	 * inode can be momentarily skipped (precached, lock contended),
+	 * or durably have nothing shrinkable (i_es_shk_nr == 0) without
+	 * es_reclaim_extents() ever touching nr_to_scan. Treat all of
+	 * these alike: not stopping doesn't help the transient cases
+	 * either, since do_shrink_slab()'s budget only decrements by
+	 * nr_scanned, so retrying forever at nr_scanned == 0 never
+	 * terminates.
+	 */
+	if (nr_scanned == 0)
+		return SHRINK_STOP;
+
 	return nr_shrunk;
 }
 

-- 
2.43.0