Re: [linus:master] [x86] 4817f70c25: stress-ng.mmapaddr.ops_per_sec 63.0% regression

[Rik van Riel <riel@surriel.com>]

On Wed, 29 Jan 2025 16:14:01 +0800
Qi Zheng <zhengqi.arch@bytedance.com> wrote:

>
> It seems that the pcp lock is held when doing tlb_remove_table_rcu(), so
> trylock fails, then bypassing PCP and calling free_one_page() directly,
> which leads to the hot spot of zone lock.

Below is a tentative fix for the issue. It is kind of a big hammer,
and maybe the RCU people have a better idea on how to solve this
problem, but it may be worth giving this a try to see if it helps
with the regression you identified.

---8<---

From 2b0302f821d1fc94c968ac533dcc62b9ffe00c38 Mon Sep 17 00:00:00 2001
From: Rik van Riel <riel@surriel.com>
Date: Wed, 29 Jan 2025 10:51:51 -0500
Subject: [PATCH 2/2] mm,rcu: prevent RCU callbacks from running with pcp lock
 held

Enabling MMU_GATHER_RCU_TABLE_FREE can create contention on the
zone->lock.  This turns out to be because in some configurations
RCU callbacks are called when IRQs are re-enabled inside
rmqueue_bulk, while the CPU is still holding the per-cpu pages lock.

That results in the RCU callbacks being unable to grab the
PCP lock, and taking the slow path with the zone->lock for
each item freed.

Speed things up by blocking RCU callbacks while holding the
PCP lock.

Signed-off-by: Rik van Riel <riel@surriel.com>
Reported-by: Qi Zheng <zhengqi.arch@bytedance.com>
---
 mm/page_alloc.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/mm/page_alloc.c b/mm/page_alloc.c
index 6e469c7ef9a4..b3c4002ab0ab 100644
--- a/mm/page_alloc.c
+++ b/mm/page_alloc.c
@@ -3036,6 +3036,13 @@ static struct page *rmqueue_pcplist(struct zone *preferred_zone,
 		return NULL;
 	}
 
+	/*
+	 * Prevent RCU callbacks from being run from the spin_lock_irqrestore
+	 * inside rmqueue_bulk, while the pcp lock is held; that would result
+	 * in each RCU free taking the zone->lock, which can be very slow.
+	 */
+	rcu_read_lock();
+
 	/*
 	 * On allocation, reduce the number of pages that are batch freed.
 	 * See nr_pcp_free() where free_factor is increased for subsequent
@@ -3046,6 +3053,7 @@ static struct page *rmqueue_pcplist(struct zone *preferred_zone,
 	page = __rmqueue_pcplist(zone, order, migratetype, alloc_flags, pcp, list);
 	pcp_spin_unlock(pcp);
 	pcp_trylock_finish(UP_flags);
+	rcu_read_unlock();
 	if (page) {
 		__count_zid_vm_events(PGALLOC, page_zonenum(page), 1 << order);
 		zone_statistics(preferred_zone, zone, 1);
-- 
2.47.1