[PATCH v12 21/31] mm: Introduce find_vma_rcu()

[Laurent Dufour]

Le 23/04/2019 à 11:27, Peter Zijlstra a écrit :
> On Tue, Apr 16, 2019 at 03:45:12PM +0200, Laurent Dufour wrote:
>> This allows to search for a VMA structure without holding the mmap_sem.
>>
>> The search is repeated while the mm seqlock is changing and until we found
>> a valid VMA.
>>
>> While under the RCU protection, a reference is taken on the VMA, so the
>> caller must call put_vma() once it not more need the VMA structure.
>>
>> At the time a VMA is inserted in the MM RB tree, in vma_rb_insert(), a
>> reference is taken to the VMA by calling get_vma().
>>
>> When removing a VMA from the MM RB tree, the VMA is not release immediately
>> but at the end of the RCU grace period through vm_rcu_put(). This ensures
>> that the VMA remains allocated until the end the RCU grace period.
>>
>> Since the vm_file pointer, if valid, is released in put_vma(), there is no
>> guarantee that the file pointer will be valid on the returned VMA.
> 
> What I'm missing here, and in the previous patch introducing the
> refcount (also see refcount_t), is _why_ we need the refcount thing at
> all.

The need for the VMA's refcount is to ensure that the VMA will remain 
until the end of the SPF handler. This is a consequence of the use of 
RCU instead of SRCU to protect the RB tree.

I was not aware of the refcount_t type, it would be better here to avoid 
wrapping.

> My original plan was to use SRCU, which at the time was not complete
> enough so I abused/hacked preemptible RCU, but that is no longer the
> case, SRCU has all the required bits and pieces.

When I did test using SRCU it was involving a lot a scheduling to run 
the SRCU callback mechanism. In some workload the impact on the 
perfomance was significant [1].

I can't see this overhead using RCU.

> 
> Also; the initial motivation was prefaulting large VMAs and the
> contention on mmap was killing things; but similarly, the contention on
> the refcount (I did try that) killed things just the same.

Doing prefaulting should be doable, I'll try to think further about that.

Regarding the refcount, I should I missed something, this is an atomic 
counter, so there should not be contention on it but cache exclusivity, 
not ideal I agree but I can't see what else to use here.

> So I'm really sad to see the refcount return; and without any apparent
> justification.

I'm not opposed to use another mechanism here, but SRCU didn't show good 
performance with some workload, and I can't see how to use RCU without a 
reference counter here. So please, advise.

Thanks,
Laurent.

[1] 
https://lore.kernel.org/linux-mm/7ca80231-fe02-a3a7-84bc-ce81690ea051@intel.com/