[Devel] [PATCH vz10] fs/fuse kio: zero cs_stats accumulators on each seqlock retry

[Konstantin Khorenko]

do_show_cs_stats() sums the per-cpu latency/rate counters into the
on-stack accumulators (iolat, netlat, read_ops_rate, write_ops_rate,
sync_ops_rate) inside a read_seqretry() loop. The accumulators were
initialised only once, before the loop. If a writer
(pcs_cs_update_stat() / pcs_cs_stat_up(), both under
write_seqlock(&cs->stat.seqlock)) runs concurrently, read_seqretry()
re-executes the loop body, summing every per-cpu counter a second time
on top of the already-populated structs - producing over-counted values
in the cs_stats output.

Zero the accumulators at the top of each pass so a retry starts clean.

Fixes: 0fb245bf6f1a ("fs/fuse kio: implement cs_stats statistics info")
https://virtuozzo.atlassian.net/browse/VSTOR-132310
Signed-off-by: Konstantin Khorenko <khorenko at virtuozzo.com>
---
 fs/fuse/kio/pcs/fuse_stat.c | 11 +++++++++++
 1 file changed, 11 insertions(+)

diff --git a/fs/fuse/kio/pcs/fuse_stat.c b/fs/fuse/kio/pcs/fuse_stat.c
index 447e4a617655..3f57ef03c455 100644
--- a/fs/fuse/kio/pcs/fuse_stat.c
+++ b/fs/fuse/kio/pcs/fuse_stat.c
@@ -340,6 +340,17 @@ static int do_show_cs_stats(struct pcs_cs *cs, void *ctx)
 		int cpu;
 
 		seq = read_seqbegin(&cs->stat.seqlock);
+		/*
+		 * read_seqretry() may re-run this loop if a writer updated the
+		 * stats meanwhile; zero the accumulators on every pass so a
+		 * retry does not sum the per-cpu counters on top of the values
+		 * already gathered by the previous (aborted) pass.
+		 */
+		memset(&iolat, 0, sizeof(iolat));
+		memset(&netlat, 0, sizeof(netlat));
+		memset(&read_ops_rate, 0, sizeof(read_ops_rate));
+		memset(&write_ops_rate, 0, sizeof(write_ops_rate));
+		memset(&sync_ops_rate, 0, sizeof(sync_ops_rate));
 		for_each_possible_cpu(cpu) {
 			struct fuse_lat_stat *pcpu_iolat, *pcpu_netlat;
 			struct pcs_perf_rate_cnt *pcpu_read_rate,
-- 
2.47.1