[PATCH] mm, oom: distinguish blockable mode for mmu notifiers

[Christian König]

Am 24.08.2018 um 15:40 schrieb Michal Hocko:
> On Fri 24-08-18 15:28:33, Christian König wrote:
>> Am 24.08.2018 um 15:24 schrieb Michal Hocko:
>>> On Fri 24-08-18 15:10:08, Christian König wrote:
>>>> Am 24.08.2018 um 15:01 schrieb Michal Hocko:
>>>>> On Fri 24-08-18 14:52:26, Christian König wrote:
>>>>>> Am 24.08.2018 um 14:33 schrieb Michal Hocko:
>>>>> [...]
>>>>>>> Thiking about it some more, I can imagine that a notifier callback which
>>>>>>> performs an allocation might trigger a memory reclaim and that in turn
>>>>>>> might trigger a notifier to be invoked and recurse. But notifier
>>>>>>> shouldn't really allocate memory. They are called from deep MM code
>>>>>>> paths and this would be extremely deadlock prone. Maybe Jerome can come
>>>>>>> up some more realistic scenario. If not then I would propose to simplify
>>>>>>> the locking here. We have lockdep to catch self deadlocks and it is
>>>>>>> always better to handle a specific issue rather than having a code
>>>>>>> without a clear indication how it can recurse.
>>>>>> Well I agree that we should probably fix that, but I have some concerns to
>>>>>> remove the existing workaround.
>>>>>>
>>>>>> See we added that to get rid of a real problem in a customer environment and
>>>>>> I don't want to that to show up again.
>>>>> It would really help to know more about that case and fix it properly
>>>>> rather than workaround it like this. Anyway, let me think how to handle
>>>>> the non-blocking notifier invocation then. I was not able to come up
>>>>> with anything remotely sane yet.
>>>> With avoiding allocating memory in the write lock path I don't see an issue
>>>> any more with that.
>>>>
>>>> All what the write lock path does now is adding items to a linked lists,
>>>> arrays etc....
>>> Can we change it to non-sleepable lock then?
>> No, the write side doesn't sleep any more, but the read side does.
>>
>> See amdgpu_mn_invalidate_node() and that is where you actually need to
>> handle the non-blocking flag correctly.
> Ohh, right you are. We already handle that by bailing out before calling
> amdgpu_mn_invalidate_node in !blockable mode.

Yeah, that is sufficient.

It could be improved because we have something like 90% chance that 
amdgpu_mn_invalidate_node() actually doesn't need to do anything.

But I can take care of that when the patch set has landed.

> So does this looks good to
> you?

Yeah, that looks perfect to me. Reviewed-by: Christian König 
<christian.koenig at amd.com>

Thanks,
Christian.

>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c
> index e55508b39496..48fa152231be 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_mn.c
> @@ -180,11 +180,15 @@ void amdgpu_mn_unlock(struct amdgpu_mn *mn)
>    */
>   static int amdgpu_mn_read_lock(struct amdgpu_mn *amn, bool blockable)
>   {
> -	if (blockable)
> -		mutex_lock(&amn->read_lock);
> -	else if (!mutex_trylock(&amn->read_lock))
> -		return -EAGAIN;
> -
> +	/*
> +	 * We can take sleepable lock even on !blockable mode because
> +	 * read_lock is only ever take from this path and the notifier
> +	 * lock never really sleeps. In fact the only reason why the
> +	 * later is sleepable is because the notifier itself might sleep
> +	 * in amdgpu_mn_invalidate_node but blockable mode is handled
> +	 * before calling into that path.
> +	 */
> +	mutex_lock(&amn->read_lock);
>   	if (atomic_inc_return(&amn->recursion) == 1)
>   		down_read_non_owner(&amn->lock);
>   	mutex_unlock(&amn->read_lock);