> On Jan 13, 2023, at 9:53 AM, Martin KaFai Lau <martin.lau@xxxxxxxxx> wrote:
>
> On 1/11/23 1:29 AM, tong@xxxxxxxxxxxxx wrote:
>> + /*
>> + * The lock may be taken in both NMI and non-NMI contexts.
>> + * There is a false lockdep warning (inconsistent lock state),
>> + * if lockdep enabled. The potential deadlock happens when the
>> + * lock is contended from the same cpu. map_locked rejects
>> + * concurrent access to the same bucket 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. atomic map_locked is necessary.
>> + * lockdep_off is invoked temporarily to fix the false warning.
>> + */
>> + lockdep_off();
>> raw_spin_lock_irqsave(&b->raw_lock, flags);
>> - *pflags = flags;
>> + lockdep_on();
>
> I am not very sure about the lockdep_off/on. Other than the false warning when using the very same htab map by both NMI and non-NMI context, I think the lockdep will still be useful to catch other potential issues. The commit c50eb518e262 ("bpf: Use separate lockdep class for each hashtab") has already solved this false alarm when NMI happens on one map and non-NMI happens on another map.
>
> Alexei, what do you think? May be only land the patch 1 fix for now.
Hi Martin
Patch 2 is used for patch 1 to test whether there is a deadlock. We should apply this two patches.
>
>> + *pflags = flags;
>> return 0;
>> }
>> @@ -172,7 +187,11 @@ static inline void htab_unlock_bucket(const struct bpf_htab *htab,
>> unsigned long flags)
>> {
>> hash = hash & min_t(u32, HASHTAB_MAP_LOCK_MASK, htab->n_buckets - 1);
>> +
>> + lockdep_off();
>> raw_spin_unlock_irqrestore(&b->raw_lock, flags);
>> + lockdep_on();
>> +
>> __this_cpu_dec(*(htab->map_locked[hash]));
>> preempt_enable();
>> }
>
>
