Re: [PATCH bpf 1/2] bpf: use rqspinlock for lru map

On Wed, Oct 29, 2025 at 8:38 PM bot+bpf-ci@kernel.org wrote:

commit 8ae9fd98fe53c45a1a89adf6d254c2db90d26544
Author: Menglong Dong <menglong8.dong@gmail.com>

bpf: use rqspinlock for lru map

This patch converts raw_spinlock_t to rqspinlock_t in BPF LRU maps to
prevent NMI context deadlocks. The new lock API can fail gracefully
instead of deadlocking, returning an error code.

Link: https://lore.kernel.org/bpf/CAEf4BzbTJCUx0D=zjx6+5m5iiGhwLzaP94hnw36ZMDHAf4-U_w@mail.gmail.com/

> diff --git a/kernel/bpf/bpf_lru_list.c b/kernel/bpf/bpf_lru_list.c
> index e7a2fc605..38fddcb1e 100644
> --- a/kernel/bpf/bpf_lru_list.c
> +++ b/kernel/bpf/bpf_lru_list.c

[ ... ]

> @@ -466,23 +470,26 @@ static struct bpf_lru_node *bpf_common_lru_pop_free(struct bpf_lru *lru,
>       do {
>               steal_loc_l = per_cpu_ptr(clru->local_list, steal);
>
> -             raw_spin_lock_irqsave(&steal_loc_l->lock, flags);
> +             if (raw_res_spin_lock_irqsave(&steal_loc_l->lock, flags))
> +                     goto out_next;
>
>               node = __local_list_pop_free(steal_loc_l);
>               if (!node)
>                       node = __local_list_pop_pending(lru, steal_loc_l);
>
> -             raw_spin_unlock_irqrestore(&steal_loc_l->lock, flags);
> +             raw_res_spin_unlock_irqrestore(&steal_loc_l->lock, flags);
>
> +out_next:
>               steal = cpumask_next_wrap(steal, cpu_possible_mask);
>       } while (!node && steal != first_steal);
>
>       loc_l->next_steal = steal;
>
>       if (node) {
> -             raw_spin_lock_irqsave(&loc_l->lock, flags);
> +             if (raw_res_spin_lock_irqsave(&loc_l->lock, flags))
> +                     return NULL;
                        ^^^^^^^^^^^

Can this leak the node? At this point in bpf_common_lru_pop_free(),
a node was successfully stolen from another CPU's list. The
__local_list_pop_free() or __local_list_pop_pending() call removed
the node from steal_loc_l, and steal_loc_l was unlocked. The node
now exists but is not in any list.

When raw_res_spin_lock_irqsave() fails here, the function returns
NULL, but the stolen node is never added to the local pending list
and never returned to any list. The node becomes orphaned.

AI is right. Here and in other places you can just leak the objects. res_spin_lock() is not a drop-in replacement. The whole thing needs to be thought through.