Re: [RFC PATCH bpf-next v2 06/12] bpf: Change local_storage->lock and b->lock to rqspinlock

[Martin KaFai Lau <martin.lau@linux.dev>]

On 10/6/25 10:58 AM, Alexei Starovoitov wrote:
> On Fri, Oct 3, 2025 at 3:03 PM Amery Hung <ameryhung@gmail.com> wrote:
>>
>> On Thu, Oct 2, 2025 at 4:37 PM Alexei Starovoitov
>> <alexei.starovoitov@gmail.com> wrote:
>>>
>>> On Thu, Oct 2, 2025 at 3:54 PM Amery Hung <ameryhung@gmail.com> wrote:
>>>>
>>>>          bpf_selem_free_list(&old_selem_free_list, false);
>>>>          if (alloc_selem) {
>>>>                  mem_uncharge(smap, owner, smap->elem_size);
>>>> @@ -791,7 +812,7 @@ void bpf_local_storage_destroy(struct bpf_local_storage *local_storage)
>>>>           * when unlinking elem from the local_storage->list and
>>>>           * the map's bucket->list.
>>>>           */
>>>> -       raw_spin_lock_irqsave(&local_storage->lock, flags);
>>>> +       while (raw_res_spin_lock_irqsave(&local_storage->lock, flags));
>>>
>>> This pattern and other while(foo) doesn't make sense to me.
>>> res_spin_lock will fail only on deadlock or timeout.
>>> We should not spin, since retry will likely produce the same
>>> result. So the above pattern just enters into infinite spin.
>>
>> I only spin in destroy() and map_free(), which cannot deadlock with
>> itself or each other. However, IIUC, a head waiter that detects
>> deadlock will cause other queued waiters to also return -DEADLOCK. I
>> think they should be able to make progress with a retry.
> 
> If it's in map_free() path then why are we taking the lock at all?
> There are supposed to be no active users of it.

There is no user from the syscall or from the bpf prog.

There are still kernel users, bpf_cgrp_storage_free() and bpf_sk_storage_free(), 
that can race with map_free().

> If there are users and we actually need that lock then the deadlock
> is possible and retrying will deadlock the same way.
> I feel AI explained it better:
> "
> raw_res_spin_lock_irqsave() can return -ETIMEDOUT (after 250ms) or
> -EDEADLK. Both are non-zero, so the while() loop continues. The commit
> message says "it cannot deadlock with itself or
> bpf_local_storage_map_free", but:
> 
> 1. If -ETIMEDOUT is returned because the lock holder is taking too long,
>     retrying immediately won't help. The timeout means progress isn't
>     being made, and spinning in a retry loop without any backoff or
>     limit will prevent other work from proceeding.
> 
> 2. If -EDEADLK is returned, it means the deadlock detector found a
>     cycle. Retrying immediately without any state change won't break the
>     deadlock cycle.
> "
> 
>> Or better if
>> rqspinlock does not force queued waiters to exit the queue if it is
>> deadlock not timeout.
> 
> If a deadlock is detected, it's the same issue for all waiters.
> I don't see the difference between timeout and deadlock.
> Both are in the "do-not-retry" category.
> Both mean that there is a bug somewhere.

Both bpf_cgrp_storage_free() and map_free() are the only remaining kernel users 
of the locks, so no deadlock is expected unless there is a bug. The busy percpu 
counter is currently not used in both of them also. Theoretically, the 
res_spin_lock (and the current regular spin_lock) should never fail here in 
bpf_cgrp_storage_free() and map_free(). If res_spin_lock returns error, there is 
a bug somewhere.

> 
>>>
>>> If it should never fail in practice then pr_warn_once and goto out
>>> leaking memory. Better yet defer to irq_work and cleanup there.>>

Not sure how to handle the bug. yeah, maybe just leak it and then splat.
I think deferring it still need to take the lock.

>> Hmm, both functions are already called in some deferred callbacks.
>> Even if we defer the cleanup again, they still need to grab locks and
>> still might fail, no?
> 
> If it's a map destroy path and we waited for RCU GP, there shouldn't be
> a need to take a lock.
> The css_free_rwork_fn() -> bpf_cgrp_storage_free() path
> is currently implemented like it's similar to:
> bpf_cgrp_storage_delete() which needs a lock.
> But bpf_cgrp_storage_free() doesn't have to.
> In css_free_rwork_fn() no prog or user space
> should see 'cgrp' pointer, since we're about to kfree(cgrp); it.
> I could be certainly missing something.