On 18/07/2024 22:02, David Hildenbrand wrote:
> On 18.07.24 13:16, Vlastimil Babka (SUSE) wrote:
>> On 7/16/24 9:39 AM, Li Zhijian wrote:
>>> It's expected that no page should be left in pcp_list after calling
>>> zone_pcp_disable() in offline_pages(). Previously, it's observed that
>>> offline_pages() gets stuck [1] due to some pages remaining in pcp_list.
>>>
>>> Cause:
>>> There is a race condition between drain_pages_zone() and __rmqueue_pcplist()
>>> involving the pcp->count variable. See below scenario:
>>>
>>> CPU0 CPU1
>>> ---------------- ---------------
>>> spin_lock(&pcp->lock);
>>> __rmqueue_pcplist() {
>>> zone_pcp_disable() {
>>> /* list is empty */
>>> if (list_empty(list)) {
>>> /* add pages to pcp_list */
>>> alloced = rmqueue_bulk()
>>> mutex_lock(&pcp_batch_high_lock)
>>> ...
>>> __drain_all_pages() {
>>> drain_pages_zone() {
>>> /* read pcp->count, it's 0 here */
>>> count = READ_ONCE(pcp->count)
>>> /* 0 means nothing to drain */
>>> /* update pcp->count */
>>> pcp->count += alloced << order;
>>> ...
>>> ...
>>> spin_unlock(&pcp->lock);
>>>
>>> In this case, after calling zone_pcp_disable() though, there are still some
>>> pages in pcp_list. And these pages in pcp_list are neither movable nor
>>> isolated, offline_pages() gets stuck as a result.
>>>
>>> Solution:
>>> Expand the scope of the pcp->lock to also protect pcp->count in
>>> drain_pages_zone(), ensuring no pages are left in the pcp list.
>>>
>>> [1] https://lore.kernel.org/linux-mm/6a07125f-e720-404c-b2f9-e55f3f166e85@xxxxxxxxxxx/
>>>
>>> Cc: David Hildenbrand <david@xxxxxxxxxx>
>>> Reported-by: Yao Xingtao <yaoxt.fnst@xxxxxxxxxxx>
>>> Signed-off-by: Li Zhijian <lizhijian@xxxxxxxxxxx>
>>> ---
>>> mm/page_alloc.c | 7 ++++---
>>> 1 file changed, 4 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
>>> index 9ecf99190ea2..1780df31d5f5 100644
>>> --- a/mm/page_alloc.c
>>> +++ b/mm/page_alloc.c
>>> @@ -2323,16 +2323,17 @@ void drain_zone_pages(struct zone *zone, struct per_cpu_pages *pcp)
>>> static void drain_pages_zone(unsigned int cpu, struct zone *zone)
>>> {
>>> struct per_cpu_pages *pcp = per_cpu_ptr(zone->per_cpu_pageset, cpu);
>>> - int count = READ_ONCE(pcp->count);
>>> + int count;
>>> + spin_lock(&pcp->lock);
>>> + count = pcp->count;
>>> while (count) {
>>> int to_drain = min(count, pcp->batch << CONFIG_PCP_BATCH_SCALE_MAX);
>>> count -= to_drain;
>>> - spin_lock(&pcp->lock);
>>> free_pcppages_bulk(zone, to_drain, pcp, 0);
>>> - spin_unlock(&pcp->lock);
>>> }
>>> + spin_unlock(&pcp->lock);
>>
>> This way seems to be partially going against the purpose of 55f77df7d715
>> ("mm: page_alloc: control latency caused by zone PCP draining") - the zone
>> lock hold time will still be limited by the batch, but not the pcp lock
>> time. It should still be possible to relock between the iterations? To
>> prevent the race I think the main part is determining pcp->count under the
>> lock, but release/retake should still be ok if the pcp->count is reread
>> after relocking.
Okay, I will try it.
>
> Agreed, had the smame thing in mind when skimming over this patch.
>
> @Li, with this patch the problems you have been seeing are fully resolved, correct?
>
Yeah, It worked in my thousand test runs.
P.S, Previously, It occurred more than 5%.
