[PATCH] jbd2: avoid false sharing with j_state_lock

[JonasZhou-oc]

is_journal_aborted() reads journal->j_flags from ext4 and JBD2 hot
paths. At the same time, transaction start and stop paths repeatedly
acquire and release j_state_lock. read_lock() and read_unlock() update
the lock word, so this traffic invalidates the cache line containing
j_flags.

On the tested x86-64 build, j_flags, j_errno, and j_state_lock occupy
the same 64-byte cache line. j_flags starts at offset 0, j_errno at
offset 8, and j_state_lock at offset 56.

Align j_state_lock to a cacheline boundary on SMP builds. This moves it
to offset 64 and prevents lock traffic from invalidating j_flags and
j_errno. Tests were run on Linux 7.3-rc1.

On a two-socket Intel Xeon Silver 4208 system with SMT disabled, CPUs
fixed at 2 GHz, and 16 stress-ng workers, ten-run average bogo-ops/s for
stress-ng.utime, stress-ng.access, stress-ng.file-ioctl, and
stress-ng.open increased by 9.60%, 5.42%, 2.05%, and 4.65%,
respectively. In separate five-run perf c2c measurements, average
Remote HITM samples for the affected j_flags/j_errno line decreased
from 1996, 2041, 1555, and 383 to zero, respectively.

Signed-off-by: JonasZhou-oc <jonaszhou-oc@xxxxxxxxxxx>
---
 include/linux/jbd2.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/include/linux/jbd2.h b/include/linux/jbd2.h
index 1b42fe47c26b..e3886963ba00 100644
--- a/include/linux/jbd2.h
+++ b/include/linux/jbd2.h
@@ -795,7 +795,7 @@ struct journal_s
 	/**
 	 * @j_state_lock: Protect the various scalars in the journal.
 	 */
-	rwlock_t		j_state_lock;
+	rwlock_t		j_state_lock ____cacheline_aligned_in_smp;
 
 	/**
 	 * @j_barrier_count:

base-commit: cee9395acd8043be0644b25c34bfa86623f2b935
-- 
2.43.0