Hi Michal,
On Fri 2021-03-26 09:16 +0100, Michal Hocko wrote:
> The oom killer is never triggered for costly allocation request.
Yes - I agree. Looking at __alloc_pages_may_oom() I can see for a costly
order allocation request the OOM killer is explicitly not used.
If I understand correctly, the patch I proposed was for the following
scenarios:
1. The costly order allocation request to fail when
"some" progress is made (i.e. order-0) and the last
compaction request was "skipped"
2. A non-costly order allocation request that is
unable to make any progress and failed over the
maximum reclaim retry count in a row and the last
known compaction result was skipped to consider
using the OOM killer for assistance
> Both reclaim and compaction maintain their own retries counters as they
> are targeting a different operation. Although the compaction really
> depends on the reclaim to do some progress.
Yes. Looking at should_compact_retry() if the last known compaction result
was skipped i.e. suggesting there was not enough order-0 pages to support
compaction, so assistance is needed from reclaim
(see __compaction_suitable()).
I noticed that the value of compaction_retries, compact_result and
compact_priority was 0, COMPACT_SKIPPED and 1 i.e. COMPACT_PRIO_SYNC_LIGHT,
respectively.
> OK, this sound unexpected as it indicates that the reclaim is able to
> make a forward progress but compaction doesn't want to give up and keeps
> retrying. Are you able to reproduce this or could you find out which
> specific condition keeps compaction retrying? I would expect that it is
> one of the 3 conditions before the max_retries is checked.
Unfortunately, I have been told it is not entirely reproducible.
I suspect it is the following in should_compact_retry() - as I indicated
above the last known value stored in compaction_retries was 0:
if (order > PAGE_ALLOC_COSTLY_ORDER)
max_retries /= 4;
if (*compaction_retries <= max_retries) {
ret = true;
goto out;
}
Kind regards,
--
Aaron Tomlin
