@@ -86,10 +86,12 @@ static DEFINE_MUTEX(nf_conntrack_mutex);
/* clamp timeouts to this value (TCP unacked) */
#define GC_SCAN_INTERVAL_CLAMP (300ul * HZ)
-/* large initial bias so that we don't scan often just because we have
- * three entries with a 1s timeout.
+/* Initial bias pretending we have 100 entries at the upper bound so we don't
+ * wakeup often just because we have three entries with a 1s timeout while still
+ * allowing non-idle machines to wakeup more often when needed.
*/
-#define GC_SCAN_INTERVAL_INIT INT_MAX
+#define GC_SCAN_INITIAL_COUNT 100
+#define GC_SCAN_INTERVAL_INIT GC_SCAN_INTERVAL_MAX
#define GC_SCAN_MAX_DURATION msecs_to_jiffies(10)
#define GC_SCAN_EXPIRED_MAX (64000u / HZ)
@@ -1477,7 +1479,7 @@ static void gc_worker(struct work_struct *work)
if (i == 0) {
gc_work->avg_timeout = GC_SCAN_INTERVAL_INIT;
- gc_work->count = 1;
+ gc_work->count = GC_SCAN_INITIAL_COUNT;
gc_work->start_time = start_time;
}
The previous commit changed the way the rescheduling delay is computed which has a side effect: the bias is now represented as much as the other entries in the rescheduling delay which makes the logic to kick in only with very large sets, as the initial interval is very large (INT_MAX). Revisit the GC initial bias to allow more frequent GC for smaller sets while still avoiding wakeups when a machine is mostly idle. We're moving from a large initial value to pretending we have 100 entries expiring at the upper bound. This way only a few entries having a small timeout won't impact much the rescheduling delay and non-idle machines will have enough entries to lower the delay when needed. This also improves readability as the initial bias is now linked to what is computed instead of being an arbitrary large value. Fixes: 2cfadb761d3d ("netfilter: conntrack: revisit gc autotuning") Suggested-by: Florian Westphal <fw@strlen.de> Signed-off-by: Antoine Tenart <atenart@kernel.org> --- net/netfilter/nf_conntrack_core.c | 10 ++++++---- 1 file changed, 6 insertions(+), 4 deletions(-)