On 7/22/24 4:10 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(), to ensure no pages are left in the pcp list after
> zone_pcp_disable()
>
> [1] https://lore.kernel.org/linux-mm/6a07125f-e720-404c-b2f9-e55f3f166e85@xxxxxxxxxxx/
>
> Cc: David Hildenbrand <david@xxxxxxxxxx>
> Cc: Vlastimil Babka (SUSE) <vbabka@xxxxxxxxxx>
> Reported-by: Yao Xingtao <yaoxt.fnst@xxxxxxxxxxx>
> Signed-off-by: Li Zhijian <lizhijian@xxxxxxxxxxx>
Can we find a breaking commit for Fixes: ?
> ---
> V2:
> - Narrow down the scope of the spin_lock() to limit the draining latency. # Vlastimil and David
> - In above scenario, it's sufficient to read pcp->count once with lock held, and it fully fixed
> my issue[1] in thounds runs(It happened in more than 5% before).
That should be ok indeed, but...
> RFC:
> https://lore.kernel.org/linux-mm/20240716073929.843277-1-lizhijian@xxxxxxxxxxx/
> ---
> mm/page_alloc.c | 5 ++++-
> 1 file changed, 4 insertions(+), 1 deletion(-)
>
> diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> index 9ecf99190ea2..5388a35c4e9c 100644
> --- a/mm/page_alloc.c
> +++ b/mm/page_alloc.c
> @@ -2323,8 +2323,11 @@ 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;
> + spin_unlock(&pcp->lock);
> while (count) {
> int to_drain = min(count, pcp->batch << CONFIG_PCP_BATCH_SCALE_MAX);
> count -= to_drain;
It's wasteful to do a lock/unlock cycle just to read the count. It could
rather look something like this:
while (true)
spin_lock(&pcp->lock);
count = pcp->count;
...
count -= to_drain;
if (to_drain)
drain_zone_pages(...)
...
spin_unlock(&pcp->lock);
if (count)
break;
