Using drm_exec for TTMs BO eviction

[Christian König]

Am 03.07.24 um 15:59 schrieb Thomas Hellström:
> On Wed, 2024-07-03 at 15:53 +0200, Thomas Hellström wrote:
>> On Wed, 2024-07-03 at 15:40 +0200, Thomas Hellström wrote:
>>> Hi, Christian,
>>>
>>> On Wed, 2024-07-03 at 15:25 +0200, Christian König wrote:
>>>> Hi guys,
>>>>
>>>> We recently ran into a problem with deadlocks during eviction and
>>>> while back Thomas worked on a patch set which was going into the
>>>> direction of solving this.
>>>>
>>>> So I simplified it to just the functionality needed to resolve
>>>> this
>>>> issue at. The resulting patch set is just the initial first step
>>>> of
>>>> using drm_exec in TTM for locking BOs during eviction.
>>>>
>>>> Should a deadlock happen the drm_exec object is now used to
>>>> resolve
>>>> it and prelock the contended object. This approach solves this
>>>> the
>>>> ENOMEM issue on contending evictions quite nicely.
>>>>
>>>> Please review and comment,
>>>> Christian.
>>> Overall it looks sane, but I think it makes sense to review and
>>> land
>>> the part of the shrinker series first that touches this eviction
>>> path
>>> and gets rid of a lot of code that's hard to understand and
>>> simplifies
>>> the locking a lot. (That part doesn't touch drm_exec), and it has
>>> been
>>> pending reviews for some time.

That's actually exactly what I wanted to avoid.

>>>
>>> I don't think it's correct to bypass that. Then we could work out
>>> the
>>> drm_exec implications.
>> It's
>>
>> https://patchwork.freedesktop.org/series/131815/
>>
>> And in particular patch 7 there brings in the restartable LRU
>> functionality and sipmlifies eviction immensely and would make the
>> usage of this patchset's rudimentary drm_exec support easier to
>> understand and review.

Yeah, seen Mathews comments on that stuff.

Looked like 99% sane to me the last time I checked, the only thing I'm 
still not very keen at is still the bulk and cursor interaction.

>>
>> /Thomas
> Hm. I actually think all review comments have been sorted out up to
> that patch, so what's missing is a resend of the new version, RB from
> Matt and Review / Ack from you, then that part could be partially
> merged.

Going to take another look at that.

Regards,
Christian.

>
> /Thomas
>
>
>
>>
>>> /Thomas
>>>
>>>
>>>>