The PGDAT_RECLAIM_LOCKED bit is used to provide mutual exclusion of
node reclaim for struct pglist_data using a single bit.

It is "locked" with a test_and_set_bit (similarly to a try lock) which
provides full ordering with respect to loads and stores done within
__node_reclaim().

It is "unlocked" with clear_bit(), which does not provide any ordering
with respect to loads and stores done before clearing the bit.

The lack of clear_bit() memory ordering with respect to stores within
__node_reclaim() can cause a subsequent CPU to fail to observe stores
from a prior node reclaim. This is not an issue in practice on TSO (e.g.
x86), but it is an issue on weakly-ordered architectures (e.g. arm64).

Fix this with following changes:

A) Use clear_bit_unlock rather than clear_bit to clear PGDAT_RECLAIM_LOCKED
   with a release memory ordering semantic.

This provides stronger memory ordering (release rather than relaxed).

B) Use test_and_set_bit_lock rather than test_and_set_bit to test-and-set
   PGDAT_RECLAIM_LOCKED with an acquire memory ordering semantic.

This changes the "lock" acquisition from a full barrier to an acquire
memory ordering, which is weaker. The acquire semi-permeable barrier
paired with the release on unlock is sufficient for this mutual
exclusion use-case.

Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
Cc: Andrew Morton <akpm@linux-foundation.org>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Alan Stern <stern@rowland.harvard.edu>
Cc: Andrea Parri <parri.andrea@gmail.com>
Cc: Will Deacon <will@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Boqun Feng <boqun.feng@gmail.com>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: David Howells <dhowells@redhat.com>
Cc: Jade Alglave <j.alglave@ucl.ac.uk>
Cc: Luc Maranget <luc.maranget@inria.fr>
Cc: "Paul E. McKenney" <paulmck@kernel.org>
Cc: linux-mm@kvack.org
---
 mm/vmscan.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/mm/vmscan.c b/mm/vmscan.c
index c22175120f5d..021b25bdba91 100644
--- a/mm/vmscan.c
+++ b/mm/vmscan.c
@@ -7567,11 +7567,11 @@ int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
 	if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
 		return NODE_RECLAIM_NOSCAN;
 
-	if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
+	if (test_and_set_bit_lock(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
 		return NODE_RECLAIM_NOSCAN;
 
 	ret = __node_reclaim(pgdat, gfp_mask, order);
-	clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
+	clear_bit_unlock(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
 
 	if (ret)
 		count_vm_event(PGSCAN_ZONE_RECLAIM_SUCCESS);
-- 
2.25.1

On Fri, Mar 07, 2025 at 02:30:47PM -0500, Mathieu Desnoyers wrote:
> The PGDAT_RECLAIM_LOCKED bit is used to provide mutual exclusion of
> node reclaim for struct pglist_data using a single bit.
> 
> It is "locked" with a test_and_set_bit (similarly to a try lock) which
> provides full ordering with respect to loads and stores done within
> __node_reclaim().
> 
> It is "unlocked" with clear_bit(), which does not provide any ordering
> with respect to loads and stores done before clearing the bit.
> 
> The lack of clear_bit() memory ordering with respect to stores within
> __node_reclaim() can cause a subsequent CPU to fail to observe stores
> from a prior node reclaim. This is not an issue in practice on TSO (e.g.
> x86), but it is an issue on weakly-ordered architectures (e.g. arm64).
> 
> Fix this with following changes:
> 
> A) Use clear_bit_unlock rather than clear_bit to clear PGDAT_RECLAIM_LOCKED
>    with a release memory ordering semantic.
> 
> This provides stronger memory ordering (release rather than relaxed).
> 
> B) Use test_and_set_bit_lock rather than test_and_set_bit to test-and-set
>    PGDAT_RECLAIM_LOCKED with an acquire memory ordering semantic.
> 
> This changes the "lock" acquisition from a full barrier to an acquire
> memory ordering, which is weaker. The acquire semi-permeable barrier
> paired with the release on unlock is sufficient for this mutual
> exclusion use-case.

FWIW, this aligns with my understanding.

Is (A) intended to be (submitted separately and) backported?

  Andrea


> Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
> Cc: Andrew Morton <akpm@linux-foundation.org>
> Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
> Cc: Matthew Wilcox <willy@infradead.org>
> Cc: Alan Stern <stern@rowland.harvard.edu>
> Cc: Andrea Parri <parri.andrea@gmail.com>
> Cc: Will Deacon <will@kernel.org>
> Cc: Peter Zijlstra <peterz@infradead.org>
> Cc: Boqun Feng <boqun.feng@gmail.com>
> Cc: Nicholas Piggin <npiggin@gmail.com>
> Cc: David Howells <dhowells@redhat.com>
> Cc: Jade Alglave <j.alglave@ucl.ac.uk>
> Cc: Luc Maranget <luc.maranget@inria.fr>
> Cc: "Paul E. McKenney" <paulmck@kernel.org>
> Cc: linux-mm@kvack.org
> ---
>  mm/vmscan.c | 4 ++--
>  1 file changed, 2 insertions(+), 2 deletions(-)
> 
> diff --git a/mm/vmscan.c b/mm/vmscan.c
> index c22175120f5d..021b25bdba91 100644
> --- a/mm/vmscan.c
> +++ b/mm/vmscan.c
> @@ -7567,11 +7567,11 @@ int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
>  	if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
>  		return NODE_RECLAIM_NOSCAN;
>  
> -	if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
> +	if (test_and_set_bit_lock(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
>  		return NODE_RECLAIM_NOSCAN;
>  
>  	ret = __node_reclaim(pgdat, gfp_mask, order);
> -	clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
> +	clear_bit_unlock(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
>  
>  	if (ret)
>  		count_vm_event(PGSCAN_ZONE_RECLAIM_SUCCESS);
> -- 
> 2.25.1
>