Re: [PATCH v3] mm/swap: strengthen locking assertions and invariants in cluster allocation

[Barry Song]

On Tue, Mar 10, 2026 at 9:57 AM Hui Zhu <hui.zhu@xxxxxxxxx> wrote:
>
> From: Hui Zhu <zhuhui@xxxxxxxxxx>
>
> The swap_cluster_alloc_table() function requires several locks to be held
> by its callers: ci->lock, the per-CPU swap_cluster lock, and, for
> non-solid-state devices (non-SWP_SOLIDSTATE), the si->global_cluster_lock.
>
> While most call paths (e.g., via cluster_alloc_swap_entry() or
> alloc_swap_scan_list()) correctly acquire these locks before invocation,
> the path through swap_reclaim_work() -> swap_reclaim_full_clusters() ->
> isolate_lock_cluster() is distinct. This path operates exclusively on
> si->full_clusters, where the swap allocation tables are guaranteed to be
> already allocated. Consequently, isolate_lock_cluster() should never
> trigger a call to swap_cluster_alloc_table() for these clusters.
>
> Strengthen the locking and state assertions to formalize these invariants:
>
> 1. Add a lockdep_assert_held() for si->global_cluster_lock in
>    swap_cluster_alloc_table() for non-SWP_SOLIDSTATE devices.
> 2. Reorder existing lockdep assertions in swap_cluster_alloc_table() to
>    match the actual lock acquisition order (per-CPU lock, then global lock,
>    then cluster lock).
> 3. Add a VM_WARN_ON_ONCE() in isolate_lock_cluster() to ensure that table
>    allocations are only attempted for clusters being isolated from the
>    free list. Attempting to allocate a table for a cluster from other
>    lists (like the full list during reclaim) indicates a violation of
>    subsystem invariants.
>
> These changes ensure locking consistency and help catch potential
> synchronization or logic issues during development.
>
> Changelog:
> v3:
> According to the comments of Kairui Song, squash patches and fix logic
> bug in isolate_lock_cluster() where flags were cleared before check.
> v2:
> According to the comments of YoungJun Park, Kairui Song and Chris Li,
> change acquire locks in swap_reclaim_work() to adds a VM_WARN_ON in
> isolate_lock_cluster().
> According to the comments of YoungJun Park, add code in patch 2 to Change
> the order of lockdep_assert_held() to match the actual lock acquisition
> order.
>
> Reviewed-by: Youngjun Park <youngjun.park@xxxxxxx>
> Signed-off-by: Hui Zhu <zhuhui@xxxxxxxxxx>

Just a few nits; otherwise:

Reviewed-by: Barry Song <baohua@xxxxxxxxxx>

> ---
>  mm/swapfile.c | 9 ++++++++-
>  1 file changed, 8 insertions(+), 1 deletion(-)
>
> diff --git a/mm/swapfile.c b/mm/swapfile.c
> index 94af29d1de88..4e0fb1ce5245 100644
> --- a/mm/swapfile.c
> +++ b/mm/swapfile.c
> @@ -476,8 +476,10 @@ swap_cluster_alloc_table(struct swap_info_struct *si,
>          * Only cluster isolation from the allocator does table allocation.
>          * Swap allocator uses percpu clusters and holds the local lock.
>          */
> -       lockdep_assert_held(&ci->lock);
>         lockdep_assert_held(&this_cpu_ptr(&percpu_swap_cluster)->lock);
> +       if (!(si->flags & SWP_SOLIDSTATE))
> +               lockdep_assert_held(&si->global_cluster_lock);
> +       lockdep_assert_held(&ci->lock);
>
>         /* The cluster must be free and was just isolated from the free list. */
>         VM_WARN_ON_ONCE(ci->flags || !cluster_is_empty(ci));
> @@ -577,6 +579,7 @@ static struct swap_cluster_info *isolate_lock_cluster(
>                 struct swap_info_struct *si, struct list_head *list)
>  {
>         struct swap_cluster_info *ci, *found = NULL;
> +       u8 flags;
>
>         spin_lock(&si->lock);
>         list_for_each_entry(ci, list, list) {
> @@ -589,6 +592,7 @@ static struct swap_cluster_info *isolate_lock_cluster(
>                           ci->flags != CLUSTER_FLAG_FULL);
>
>                 list_del(&ci->list);
> +               flags = ci->flags;
>                 ci->flags = CLUSTER_FLAG_NONE;
>                 found = ci;
>                 break;
> @@ -596,6 +600,9 @@ static struct swap_cluster_info *isolate_lock_cluster(
>         spin_unlock(&si->lock);
>
>         if (found && !cluster_table_is_alloced(found)) {
> +               /* Table of non-free cluster must be allocated. */

I feel this comment is somewhat redundant with the one below,
and its wording might be inaccurate. It could be rewritten as:

/* Non-free clusters must have their swap table already allocated. */

> +               VM_WARN_ON_ONCE(flags != CLUSTER_FLAG_FREE);
> +
>                 /* Only an empty free cluster's swap table can be freed. */
>                 VM_WARN_ON_ONCE(list != &si->free_clusters);
>                 VM_WARN_ON_ONCE(!cluster_is_empty(found));

I'd prefer moving VM_WARN_ON_ONCE(flags != CLUSTER_FLAG_FREE)
here directly, without the redundant comment you're adding. Doesn't
it already convey "empty free cluster"?

Thanks
Barry