Re: [PATCH] mm/percpu: avoid IO dependency in alloc_percpu

[Khazhy Kumykov]

On Mon, Jun 15, 2026 at 6:23 PM Khazhismel Kumykov <khazhy@xxxxxxxxxxxx> wrote:
>
> percpu_alloc users within noio context may block on other percpu_alloc
> users doing IO due to the shared pcpu_alloc_mutex. pcpu_balance_workfn
> already grabs under noio context - update remaining holders to also run
> under noio.
>
> Lockdep reported this as a circular locking dependency, with one caller
> having entered fs_reclaim holding pcpu_alloc_mutex, and the second
> caller (elevator init) grabbing pcpu_alloc_mutex in noio context.
>
> Fixes: 9a5b183941b5 ("mm, percpu: do not consider sleepable allocations atomic")
>
> Signed-off-by: Khazhismel Kumykov <khazhy@xxxxxxxxxx>
> ---
>  mm/percpu.c | 9 +++++++--
>  1 file changed, 7 insertions(+), 2 deletions(-)
>
> diff --git a/mm/percpu.c b/mm/percpu.c
> index b0676b8054ed..50057403ae79 100644
> --- a/mm/percpu.c
> +++ b/mm/percpu.c
> @@ -1749,8 +1749,13 @@ void __percpu *pcpu_alloc_noprof(size_t size, size_t align, bool reserved,
>         size_t bits, bit_align;
>
>         gfp = current_gfp_context(gfp);
> -       /* whitelisted flags that can be passed to the backing allocators */
> -       pcpu_gfp = gfp & (GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN);
> +       /*
> +        * whitelisted flags that can be passed to the backing allocators.
> +        *
> +        * Allocations must be GFP_NOIO, else users in noio/nofs contexts may
> +        * form dependency through pcpu_alloc_mutex
> +        */
> +       pcpu_gfp = gfp & (GFP_NOIO | __GFP_NORETRY | __GFP_NOWARN);
ah, this is incomplete since we have subsequent allocations that use
"GFP_KERNEL", I'll send a version with memalloc_noio instead.
Apologies...

>         is_atomic = !gfpflags_allow_blocking(gfp);
>         do_warn = !(gfp & __GFP_NOWARN);
>
> --
> 2.54.0.1189.g8c84645362-goog
>