On Wed, Jul 17, 2024 at 6:44 PM Priya Bala Govindasamy <pgovind2@xxxxxxx> wrote:
>
> bpf_percpu_lru_pop_free uses raw_spin_lock_irqsave. This function is
> used by htab_percpu_lru_map_update_elem() which can be called from an
> NMI. A deadlock can happen if a bpf program holding the lock is
> interrupted by the same program in NMI. Use raw_spin_trylock_irqsave if
> in NMI.
And there is a htab_lock_bucket() protection and bpf prog
recursion protection logic that should prevent such deadlock.
Pls share the splat if this deadlock is real.
> - raw_spin_lock_irqsave(&l->lock, flags);
> + if (in_nmi()) {
> + if (!raw_spin_trylock_irqsave(&l->lock, flags))
> + return NULL;
> + } else {
> + raw_spin_lock_irqsave(&l->lock, flags);
> + }
We cannot do this, since it will make map behavior 'random'.
There are lots of other raw_spin_lock_irqsave() in that file.
Somehow they're not deadlocking?
pw-bot: cr
