mirrors/haproxy
1
0
Fork 0
mirror of https://git.haproxy.org/git/haproxy.git/ synced 2025-09-20 21:31:28 +02:00
Code Issues Projects Releases Wiki Activity

BUG/MEDIUM: stick-tables: don't leave the expire loop with elements deleted

Browse Source 
In 3.2, the table expiration latency was improved by commit 994cc58576
("MEDIUM: stick-tables: Limit the number of entries we expire"), however
it introduced an issue by which it's possible to leave the loop after a
certain number of elements were expired, without requeuing the deleted
elements. The issue it causes is that other places with a non-null ref_cnt
will not necessarily delete it themselves, resulting in orphan elements in
the table. These ones will then pollute it and force recycling old ones
more often which in turn results in an increase of the contention.

Let's check for the expiration counter before deleting the element so
that it can be found upon next visit.

This fix must be backported to 3.2. It is directly related to GH
issue #3084. Thanks to Felipe and Ricardo for sharing precious info
and testing a candidate fix.
This commit is contained in:
Willy Tarreau 2025-09-03 11:46:30 +00:00
parent 554a15562f
commit 8da6ed6b6a
1 changed files with 9 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
src/stick_table.c
View File

@ -943,6 +943,15 @@ struct task *process_table_expire(struct task *task, void *context, unsigned int
if (HA_ATOMIC_LOAD(&ts->ref_cnt) != 0) if (HA_ATOMIC_LOAD(&ts->ref_cnt) != 0)
continue; continue;
if (updt_locked == 1) {
expired++;
if (expired == STKTABLE_MAX_UPDATES_AT_ONCE) {
need_resched = 1;
exp_next = TICK_ETERNITY;
goto out_unlock;
}
}
eb32_delete(&ts->exp); eb32_delete(&ts->exp);
if (!tick_is_expired(ts->expire, now_ms)) { if (!tick_is_expired(ts->expire, now_ms)) {
@ -968,14 +977,6 @@ struct task *process_table_expire(struct task *task, void *context, unsigned int
continue; continue;
} }
if (updt_locked == 1) {
expired++;
if (expired == STKTABLE_MAX_UPDATES_AT_ONCE) {
need_resched = 1;
exp_next = TICK_ETERNITY;
goto out_unlock;
}
}
/* if the entry is in the update list, we must be extremely careful /* if the entry is in the update list, we must be extremely careful
* because peers can see it at any moment and start to use it. Peers * because peers can see it at any moment and start to use it. Peers
* will take the table's updt_lock for reading when doing that, and * will take the table's updt_lock for reading when doing that, and