lkml.org 
[lkml]   [2022]   [Nov]   [29]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
SubjectRe: [net-next] bpf: avoid hashtab deadlock with try_lock
From
On 11/29/22 22:32, Tonghao Zhang wrote:
> On Wed, Nov 30, 2022 at 11:07 AM Waiman Long <longman@redhat.com> wrote:
>> On 11/29/22 21:47, Tonghao Zhang wrote:
>>> On Wed, Nov 30, 2022 at 9:50 AM Hou Tao <houtao@huaweicloud.com> wrote:
>>>> Hi Hao,
>>>>
>>>> On 11/30/2022 3:36 AM, Hao Luo wrote:
>>>>> On Tue, Nov 29, 2022 at 9:32 AM Boqun Feng <boqun.feng@gmail.com> wrote:
>>>>>> Just to be clear, I meant to refactor htab_lock_bucket() into a try
>>>>>> lock pattern. Also after a second thought, the below suggestion doesn't
>>>>>> work. I think the proper way is to make htab_lock_bucket() as a
>>>>>> raw_spin_trylock_irqsave().
>>>>>>
>>>>>> Regards,
>>>>>> Boqun
>>>>>>
>>>>> The potential deadlock happens when the lock is contended from the
>>>>> same cpu. When the lock is contended from a remote cpu, we would like
>>>>> the remote cpu to spin and wait, instead of giving up immediately. As
>>>>> this gives better throughput. So replacing the current
>>>>> raw_spin_lock_irqsave() with trylock sacrifices this performance gain.
>>>>>
>>>>> I suspect the source of the problem is the 'hash' that we used in
>>>>> htab_lock_bucket(). The 'hash' is derived from the 'key', I wonder
>>>>> whether we should use a hash derived from 'bucket' rather than from
>>>>> 'key'. For example, from the memory address of the 'bucket'. Because,
>>>>> different keys may fall into the same bucket, but yield different
>>>>> hashes. If the same bucket can never have two different 'hashes' here,
>>>>> the map_locked check should behave as intended. Also because
>>>>> ->map_locked is per-cpu, execution flows from two different cpus can
>>>>> both pass.
>>>> The warning from lockdep is due to the reason the bucket lock A is used in a
>>>> no-NMI context firstly, then the same bucke lock is used a NMI context, so
>>> Yes, I tested lockdep too, we can't use the lock in NMI(but only
>>> try_lock work fine) context if we use them no-NMI context. otherwise
>>> the lockdep prints the warning.
>>> * for the dead-lock case: we can use the
>>> 1. hash & min(HASHTAB_MAP_LOCK_MASK, htab->n_buckets -1)
>>> 2. or hash bucket address.
>>>
>>> * for lockdep warning, we should use in_nmi check with map_locked.
>>>
>>> BTW, the patch doesn't work, so we can remove the lock_key
>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=c50eb518e262fa06bd334e6eec172eaf5d7a5bd9
>>>
>>> static inline int htab_lock_bucket(const struct bpf_htab *htab,
>>> struct bucket *b, u32 hash,
>>> unsigned long *pflags)
>>> {
>>> unsigned long flags;
>>>
>>> hash = hash & min(HASHTAB_MAP_LOCK_MASK, htab->n_buckets -1);
>>>
>>> preempt_disable();
>>> if (unlikely(__this_cpu_inc_return(*(htab->map_locked[hash])) != 1)) {
>>> __this_cpu_dec(*(htab->map_locked[hash]));
>>> preempt_enable();
>>> return -EBUSY;
>>> }
>>>
>>> if (in_nmi()) {
>>> if (!raw_spin_trylock_irqsave(&b->raw_lock, flags))
>>> return -EBUSY;
>> That is not right. You have to do the same step as above by decrementing
>> the percpu count and enable preemption. So you may want to put all these
>> busy_out steps after the return 0 and use "goto busy_out;" to jump there.
> Yes, thanks Waiman, I should add the busy_out label.
>>> } else {
>>> raw_spin_lock_irqsave(&b->raw_lock, flags);
>>> }
>>>
>>> *pflags = flags;
>>> return 0;
>>> }
>> BTW, with that change, I believe you can actually remove all the percpu
>> map_locked count code.
> there are some case, for example, we run the bpf_prog A B in task
> context on the same cpu.
> bpf_prog A
> update map X
> htab_lock_bucket
> raw_spin_lock_irqsave()
> lookup_elem_raw()
> // bpf prog B is attached on lookup_elem_raw()
> bpf prog B
> update map X again and update the element
> htab_lock_bucket()
> // dead-lock
> raw_spinlock_irqsave()

I see, so nested locking is possible in this case. Beside using the
percpu map_lock, another way is to have cpumask associated with each
bucket lock and use each bit in the cpumask for to control access using
test_and_set_bit() for each cpu. That will allow more concurrency and
you can actually find out how contended is the lock. Anyway, it is just
a thought.

Cheers,
Longman


\
 
 \ /
  Last update: 2022-11-30 05:09    [W:0.051 / U:2.092 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site