[PATCH v5] jbd2: fix shrinker scan budget accounting in jbd2_journal_shrink_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.

jbd2_journal_shrink_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 jbd2_journal_shrink_checkpoint_list() actually did, and keeps
calling scan_objects() until the budget derived from the (possibly
stale) percpu checkpoint count is drained, even after the checkpoint
list has already been fully drained.

jbd2_journal_shrink_checkpoint_list() already tracks examined
buffers accurately: it decrements its nr_to_scan in/out parameter
for every journal_head walked, whether or not it gets freed. Derive
sc->nr_scanned from the difference, and return SHRINK_STOP once it
comes back zero, since that only happens when the checkpoint list
had nothing left to walk. Key this off sc->nr_scanned rather than
nr_shrunk: a batch can find every buffer in its transactions busy
and free none, while later transactions may still hold buffers
whose writeback has completed; sc->nr_scanned reflects that real
work either way.

Tested by creating 20000 small files, running sync, and letting
checkpointing settle for a few seconds (with systemd-journald and
cron stopped, since their own writes to the same filesystem keep
re-populating the checkpoint list and mask the effect being
measured), then triggering "echo 2 > /proc/sys/vm/drop_caches" while
tracing jbd2_shrink_scan_enter/exit:

                      total scan_objects()   calls with
                      calls                  nr_scanned == 0
  before this patch   22                     10 (45%)
  after this patch     8                      1 (12%)

Fixes: 4ba3fcdde7e3 ("jbd2,ext4: add a shrinker to release checkpointed buffers")
Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Qiliang Yuan <odys.yuan@xxxxxxxxx>
Reviewed-by: Jan Kara <jack@xxxxxxx>
---
V4 -> V5:
- Revert to keying SHRINK_STOP off sc->nr_scanned (buffers actually
  examined) rather than nr_shrunk (buffers actually freed), as v1
  originally did. Jan Kara rejected the nr_shrunk-based v4, pointing
  out that failing to reclaim the buffers in one batch doesn't mean
  later transactions won't have freeable ones; Sashiko AI review
  independently flagged the same issue. Restore Jan Kara's v1
  Reviewed-by: the diff in jbd2_journal_shrink_scan() is unchanged
  from what he reviewed there (only the explanatory comment reads
  differently).
- Redesign the test: the old methodology (no sync, buffers kept busy
  by design) exercises exactly the busy-but-productive case Jan Kara
  describes, so it isn't a useful before/after comparison for this
  change. Replace it with one that lets the checkpoint list actually
  drain via sync, then measures scan_objects() calls wasted after
  draining due to percpu-counter staleness.

V3 -> V4:
- Shorten the SHRINK_STOP comment; no functional change (the trigger
  has been nr_shrunk == 0 since v2, not sc->nr_scanned).
- Carry forward Jan Kara's v3 Reviewed-by.

V2 -> V3:
- Rebase onto v7.3-rc5. It already includes Max Kellermann's
  "jbd2: bound shrinker scans by examined checkpoint buffers"
  (15cb16496446), which independently fixes the busy-buffer
  accounting in journal_shrink_one_cp_list()/
  jbd2_journal_shrink_checkpoint_list() that v2 also touched. Drop
  the now-redundant checkpoint.c changes; this revision only touches
  journal.c, reusing the accurate nr_to_scan tracking Kellermann's
  fix already provides.
- Add a Fixes: tag for the commit that introduced
  jbd2_journal_shrink_scan() without ever setting sc->nr_scanned.
- Re-measure test data against the new baseline (176 -> 8 calls,
  vs the old baseline's 168 -> 7).

V1 -> V2:
- Count examined buffers, not just freed ones, in
  journal_shrink_one_cp_list() (Sashiko AI review finding).
- Trigger SHRINK_STOP on nr_shrunk == 0 instead of
  sc->nr_scanned == 0.

v4: https://lore.kernel.org/r/20260930-fix-jbd2-shrink-scan-nr-scanned-v4-1-eb9a3ccc2524@xxxxxxxxx
v3: https://lore.kernel.org/r/20260928-fix-jbd2-shrink-scan-nr-scanned-v3-1-916caaa53420@xxxxxxxxx
v2: https://lore.kernel.org/r/20260928-fix-jbd2-shrink-scan-nr-scanned-v2-1-7e1efec3afdf@xxxxxxxxx
v1: https://lore.kernel.org/r/20260928-fix-jbd2-shrink-scan-nr-scanned-v1-1-e6f4016ec699@xxxxxxxxx
---
 fs/jbd2/journal.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/fs/jbd2/journal.c b/fs/jbd2/journal.c
index 00f5a98f3d4fe..eac5f84a1afc5 100644
--- a/fs/jbd2/journal.c
+++ b/fs/jbd2/journal.c
@@ -1263,10 +1263,22 @@ static unsigned long jbd2_journal_shrink_scan(struct shrinker *shrink,
 	trace_jbd2_shrink_scan_enter(journal, sc->nr_to_scan, count);
 
 	nr_shrunk = jbd2_journal_shrink_checkpoint_list(journal, &nr_to_scan);
+	sc->nr_scanned = sc->nr_to_scan - nr_to_scan;
 
 	count = percpu_counter_read_positive(&journal->j_checkpoint_jh_count);
 	trace_jbd2_shrink_scan_exit(journal, nr_to_scan, nr_shrunk, count);
 
+	/*
+	 * Key SHRINK_STOP off sc->nr_scanned (buffers actually examined),
+	 * not nr_shrunk (buffers actually freed): a batch can find every
+	 * buffer in its transactions busy and free none, while later
+	 * transactions may still hold buffers whose writeback has
+	 * completed. sc->nr_scanned only comes back 0 when the checkpoint
+	 * list had nothing left to walk.
+	 */
+	if (sc->nr_scanned == 0)
+		return SHRINK_STOP;
+
 	return nr_shrunk;
 }
 

---
base-commit: 72d3fcf802c45d00b300f25b848a93c3a2bd7c7e
change-id: 20260928-fix-jbd2-shrink-scan-nr-scanned-c3098cd901e2

Best regards,
-- 
Qiliang Yuan <odys.yuan@xxxxxxxxx>