Message-ID: <20b8ad93-7a90-dc8c-581b-491d543423a5@huaweicloud.com>
Date:   Thu, 1 Dec 2022 10:53:37 +0800
From:   Hou Tao <houtao@...weicloud.com>
To:     Tonghao Zhang <xiangxia.m.yue@...il.com>
Cc:     Hao Luo <haoluo@...gle.com>, Waiman Long <longman@...hat.com>,
        Peter Zijlstra <peterz@...radead.org>,
        Ingo Molnar <mingo@...hat.com>, Will Deacon <will@...nel.org>,
        netdev@...r.kernel.org, Alexei Starovoitov <ast@...nel.org>,
        Daniel Borkmann <daniel@...earbox.net>,
        Andrii Nakryiko <andrii@...nel.org>,
        Martin KaFai Lau <martin.lau@...ux.dev>,
        Song Liu <song@...nel.org>, Yonghong Song <yhs@...com>,
        John Fastabend <john.fastabend@...il.com>,
        KP Singh <kpsingh@...nel.org>,
        Stanislav Fomichev <sdf@...gle.com>,
        Jiri Olsa <jolsa@...nel.org>, bpf <bpf@...r.kernel.org>,
        "houtao1@...wei.com" <houtao1@...wei.com>,
        LKML <linux-kernel@...r.kernel.org>,
        Boqun Feng <boqun.feng@...il.com>
Subject: Re: [net-next] bpf: avoid hashtab deadlock with try_lock

Hi,

On 11/30/2022 1:55 PM, Tonghao Zhang wrote:
> On Wed, Nov 30, 2022 at 12:13 PM Hou Tao <houtao@...weicloud.com> wrote:
>> Hi,
>>
>> On 11/30/2022 10:47 AM, Tonghao Zhang wrote:
>>> On Wed, Nov 30, 2022 at 9:50 AM Hou Tao <houtao@...weicloud.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@...il.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.
>> Use the computed hash will be better than hash bucket address, because the hash
>> buckets are allocated sequentially.
>>> * 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;
>> The only purpose of trylock here is to make lockdep happy and it may lead to
>> unnecessary -EBUSY error for htab operations in NMI context. I still prefer add
>> a virtual lock-class for map_locked to fix the lockdep warning. So could you use
> Hi, what is virtual lock-class ? Can you give me an example of what you mean?
If LOCKDEP is enabled, raw_spinlock will add dep_map in the definition and it
also calls lock_acquire() and lock_release() to assist the deadlock check. Now
map_locked is not a lock but it acts like a raw_spin_trylock, so we need to add
dep_map to it manually, and then also call lock_acquire(trylock=1) and
lock_release() before increasing and decreasing map_locked. You can reference
the implementation of raw_spin_trylock and raw_spin_unlock for more details.
>> separated patches to fix the potential dead-lock and the lockdep warning ? It
>> will be better you can also add a bpf selftests for deadlock problem as said before.
>>
>> Thanks,
>> Tao
>>>         } else {
>>>                 raw_spin_lock_irqsave(&b->raw_lock, flags);
>>>         }
>>>
>>>         *pflags = flags;
>>>         return 0;
>>> }
>>>
>>>
>>>> lockdep deduces that may be a dead-lock. I have already tried to use the same
>>>> map_locked for keys with the same bucket, the dead-lock is gone, but still got
>>>> lockdep warning.
>>>>> Hao
>>>>> .
>

Powered by blists - more mailing lists