Re: [PATCH v5 6/7] locking/lockdep: Reuse freed chain_hlocks entries

[Waiman Long]

On 2/4/20 11:26 AM, Waiman Long wrote:
> On 2/4/20 11:12 AM, Waiman Long wrote:
>> On 2/4/20 10:42 AM, Peter Zijlstra wrote:
>>> On Mon, Feb 03, 2020 at 11:41:46AM -0500, Waiman Long wrote:
>>>> +	/*
>>>> +	 * We require a minimum of 2 (u16) entries to encode a freelist
>>>> +	 * 'pointer'.
>>>> +	 */
>>>> +	req = max(req, 2);
>>> Would something simple like the below not avoid that whole 1 entry
>>> 'chain' nonsense?
>>>
>>> It boots and passes the selftests, so it must be perfect :-)
>>>
>>> --- a/kernel/locking/lockdep.c
>>> +++ b/kernel/locking/lockdep.c
>>> @@ -3163,7 +3163,7 @@ static int validate_chain(struct task_st
>>>  	 * (If lookup_chain_cache_add() return with 1 it acquires
>>>  	 * graph_lock for us)
>>>  	 */
>>> -	if (!hlock->trylock && hlock->check &&
>>> +	if (!chain_head && !hlock->trylock && hlock->check &&
>>>  	    lookup_chain_cache_add(curr, hlock, chain_key)) {
>>>  		/*
>>>  		 * Check whether last held lock:
>>>
>> Well, I think that will eliminate the 1-entry chains for the process
>> context. However, we can still have 1-entry chain in the irq context, I
>> think, as long as there are process context locks in front of it.
>>
>> I think this fix is still worthwhile as it will eliminate some of the
>> 1-entry chains.
> Sorry, I think I mis-read the code. This patch will eliminate some
> cross-context check. How about something like
>
> diff --git a/kernel/locking/lockdep.c b/kernel/locking/lockdep.c
> index 32406ef0d6a2..d746897b638f 100644
> --- a/kernel/locking/lockdep.c
> +++ b/kernel/locking/lockdep.c
> @@ -2931,7 +2931,7 @@ static int validate_chain(struct task_struct *curr,
> ÂÂÂÂÂÂÂÂ * (If lookup_chain_cache_add() return with 1 it acquires
> ÂÂÂÂÂÂÂÂ * graph_lock for us)
> ÂÂÂÂÂÂÂÂ */
> -ÂÂÂÂÂÂ if (!hlock->trylock && hlock->check &&
> +ÂÂÂÂÂÂ if ((chain_head != 1) && !hlock->trylock && hlock->check &&
> ÂÂÂÂÂÂÂÂÂÂÂ lookup_chain_cache_add(curr, hlock, chain_key)) {
> ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ /*
> ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ * Check whether last held lock:
> @@ -3937,7 +3937,7 @@ static int __lock_acquire(struct lockdep_map
> *lock, unsign
> ÂÂÂÂÂÂÂ hlock->prev_chain_key = chain_key;
> ÂÂÂÂÂÂÂ if (separate_irq_context(curr, hlock)) {
> ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ chain_key = INITIAL_CHAIN_KEY;
> -ÂÂÂÂÂÂÂÂÂÂÂÂÂÂ chain_head = 1;
> +ÂÂÂÂÂÂÂÂÂÂÂÂÂÂ chain_head = 2; /* Head of irq context chain */
> ÂÂÂÂÂÂÂ }
> ÂÂÂÂÂÂÂ chain_key = iterate_chain_key(chain_key, class_idx);

Wait, it is possible that we can have deadlock like this:

 cpu 0 cpu 1
 ----- -----
 lock A lock B
 <irq> <irq>
 lock B lock A
Â
If we eliminate 1-entry chain, will that impact our ability to detect this
kind of deadlock?

Thanks,
Longman