neigh_forced_gc() already counts only same-VE reclaims into "shrunk" and
is reached from neigh_alloc() when a container hits its per-VE gc_thresh3
limit, but its reclaim target was still derived from the global counter:
max_clean = atomic_read(&tbl->gc_entries) - gc_thresh2;
With several containers tbl->gc_entries is the sum across all of them, so
max_clean balloons far beyond the number of entries the allocating
container actually owns. The "shrunk >= max_clean" stop condition is then
never reached and forced GC walks the entire global gc_list on every call
(until the 16-iteration / 1 ms budget runs out), churning other VEs'
entries instead of stopping once the offending VE is back under its limit.
Derive max_clean from the per-VE counter so the trigger, the reclaim
target and the counted reclaims are all per-VE; fall back to the global
counter when there is no per-VE counter for the table family.
Fixes: 5f0a2a6f78f7 ("ve/net/neighbour: per-ct limit for neighbour entries")
https://virtuozzo.atlassian.net/browse/VSTOR-132310
Feature: net: make the neighbor entries limit per-CT
Reported-by: Pavel Tikhomirov <[email protected]>
Signed-off-by: Konstantin Khorenko <[email protected]>
---
net/core/neighbour.c | 4 +++-
1 file changed, 3 insertions(+), 1 deletion(-)
diff --git a/net/core/neighbour.c b/net/core/neighbour.c
index cf1f11a596d8..f90deb17fb25 100644
--- a/net/core/neighbour.c
+++ b/net/core/neighbour.c
@@ -275,7 +275,9 @@ bool neigh_remove_one(struct neighbour *ndel, struct
neigh_table *tbl)
static int neigh_forced_gc(struct neigh_table *tbl, struct ve_struct *ve)
{
- int max_clean = atomic_read(&tbl->gc_entries) -
+ atomic_t *cnt = get_perve_tbl_entries_counter(tbl, ve);
+ int max_clean = (cnt ? atomic_read(cnt) :
+ atomic_read(&tbl->gc_entries)) -
READ_ONCE(tbl->gc_thresh2);
u64 tmax = ktime_get_ns() + NSEC_PER_MSEC;
unsigned long tref = jiffies - 5 * HZ;
--
2.47.1
_______________________________________________
Devel mailing list
[email protected]
https://lists.openvz.org/mailman/listinfo/devel
