Re: [PATCH 04/10] drm/ttm: change fence_lock to inner lock, v3

[Maarten Lankhorst <maarten.lankhorst@canonical.com>]

Op 21-11-12 13:42, Thomas Hellstrom schreef:
> On 11/21/2012 12:38 PM, Maarten Lankhorst wrote:
>> Hey,
>>
>> Op 20-11-12 16:08, Thomas Hellstrom schreef:
>>> On 11/20/2012 02:13 PM, Maarten Lankhorst wrote:
>>>> Op 20-11-12 13:03, Thomas Hellstrom schreef:
>>>>> On 11/20/2012 12:33 PM, Maarten Lankhorst wrote:
>>>>>> Op 20-11-12 08:48, Thomas Hellstrom schreef:
>>>>>>> On 11/19/2012 04:33 PM, Maarten Lankhorst wrote:
>>>>>>>> Op 19-11-12 16:04, Thomas Hellstrom schreef:
>>>>>>>>> On 11/19/2012 03:17 PM, Thomas Hellstrom wrote:
>>>>>>>>>> Hi,
>>>>>>>>>>
>>>>>>>>>> This patch looks mostly good, although I think ttm_bo_cleanup_refs becomes overly complicated:
>>>>>>>>>> Could this do, or am I missing something?
>>>>>>>>>>
>>>>>>>>> Actually, my version is bad, because ttm_bo_wait() is called with the lru lock held.
>>>>>>>>>
>>>>>>>>> /Thomas
>>>>>>>> Oh digging through it made me remember why I had to release the reservation early and
>>>>>>>> had to allow move_notify to be called without reservation.
>>>>>>>>
>>>>>>>> Fortunately move_notify has a NULL parameter, which is the only time that happens,
>>>>>>>> so you can still check do BUG_ON(mem != NULL && !ttm_bo_reserved(bo)); in your
>>>>>>>> move_notify handler.
>>>>>>>>
>>>>>>>> 05/10 removed the loop and assumed no new fence could be attached after the driver has
>>>>>>>> declared the bo dead.
>>>>>>>>
>>>>>>>> However, at that point it may no longer hold a reservation to confirm this, that's why
>>>>>>>> I moved the cleanup to be done in the release_list handler. It could still be done in
>>>>>>>> ttm_bo_release, but we no longer have a reservation after we waited. Getting
>>>>>>>> a reservation can fail if the bo is imported for example.
>>>>>>>>
>>>>>>>> While it would be true that in that case a new fence may be attached as well, that
>>>>>>>> would be less harmful since that operation wouldn't involve this device, so the
>>>>>>>> ttm bo can still be removed in that case. When that time comes I should probably
>>>>>>>> fix up that WARN_ON(ret) in ttm_bo_cleanup_refs. :-)
>>>>>>>>
>>>>>>>> I did add a WARN_ON(!atomic_read(&bo->kref.refcount)); to
>>>>>>>> ttm_bo_reserve and ttm_eu_reserve_buffers to be sure nothing is done on the device
>>>>>>>> itself. If that is too paranoid, those WARN_ON's could be dropped. I prefer to leave them
>>>>>>>> in for a kernel release or 2. But according to the rules that would be the only time you
>>>>>>>> could attach a new fence and trigger the WARN_ON for now..
>>>>>>> Hmm, I'd appreciate if you could group patches with functional changes that depend on eachother togeteher,
>>>>>>> and "this is done because ...", which makes it much easier to review, (and to follow the commit history in case
>>>>>>> something goes terribly wrong and we need to revert).
>>>>>>>
>>>>>>> Meanwhile I'll take a look at the final ttm_bo.c and see if I can spot any culprits.
>>>>>>>
>>>>>>> In general, as long as a bo is on a LRU list, we must be able to attach fences because of accelerated eviction.
>>>>>> I thought it was deliberately designed in such a way that it was kept on the lru list,
>>>>>> but since it's also on the ddestroy list it won't start accelerated eviction,
>>>>>> since it branches into cleanup_refs early, and lru_lock still protects all the list entries.
>>>>> I used bad wording. I meant that unbinding might be accelerated, but  currently (quite inefficiently)
>>>>> do synchronized unbinding, assuming that only the CPU can do that. When we start to support
>>>>> unsynchronized moves, we need to be able to attach fences at least at the last move_notify(bo, NULL);
>>>> Would you need to wait in that case on fence_wait being completed before calling move_notify?
>>>>
>>>> If not, you would still only need to perform one wait, but you'd have to make sure move_notify only gets
>>>> called by 1 thread before checking the fence pointer and performing a wait. At that point you still hold the
>>>> lru_lock though, so it shouldn't be too hard to make something safe.
>>> I think typically a driver that wants to implement asynchronous moves don't want to wait before calling
>>> move_notify, but may wait in move_notify or move. Typically (upcoming vmwgfx) it would invalidate the buffer in move_notify(bo, NULL), attach a fence and then use the normal delayed destroy to wait on that fence before destroying the buffer.
>>>
>>> Otherwise, since binds / unbinds are handled in the GPU command stream there's never any need to wait for moves except when there's a CPU
>>> access.
>> Well, nouveau actually needs fence_wait to finish first, since vm changes are out of band.
>> But I guess it should be possible to attach it as work to the fence when it's signaled, and I
>> may want to do something like that already for performance reasons in a different place,
>> so I guess it doesn't matter.
>
> Actions to be performed on fence signaling tend to be very cpu consuming, I think due to the context switches involved.
> We had to replace that in the old psb driver and batch things like TTM does instead.
>
> Also remember that TTM fences are not required to signal in finite time unless fence_flush is called.
>
> I think nouveau doesn't use fence irqs to signal its fences.
>
>>
>> Is calling move_notify(bo, NULL) legal and a noop the second time?
>
> I see no fundamental reason why it shouldn't be OK, although we might need to patch drivers to cope with it.
>
>>   That would save a flag in the bo to check if it's called already,
>> although I suppose we could always define a TTM_BO_PRIV_FLAG_* for it otherwise.
>>
>> move_notify might end up being called with the lru_lock held, but that shouldn't be a problem.
>
> I don't think that's a good idea. Drivers sleeping in move_notify will need to release the spinlock, and that means it's
> better to release it before move_notify is called.
Is the only sleeping being done on fences? In that case we might wish to split it up in 2 pieces for destruction,
the piece that runs immediately, and a piece to run after the new fence has signaled (current behavior).

Nouveau needs the final move_notify unmap to be called after object is idle, like it is now. It doesn't need
to attach a new fence.

~Maarten