[Nouveau] [PATCH] drm/nouveau: do not move buffers when not needed

[Ben Skeggs]

On Wed, Sep 4, 2013 at 11:25 PM, Maarten Lankhorst
<maarten.lankhorst at canonical.com> wrote:
> Op 04-09-13 03:24, Ben Skeggs schreef:
>> On Mon, Jul 15, 2013 at 6:39 PM, Maarten Lankhorst
>> <maarten.lankhorst at canonical.com> wrote:
>>> Op 15-07-13 08:05, Ben Skeggs schreef:
>>>> On Fri, Jul 12, 2013 at 10:45 PM, Maarten Lankhorst
>>>> <maarten.lankhorst at canonical.com> wrote:
>>>>> I have no idea what this bogus restriction on placement is, but it breaks decoding 1080p
>>>>> VDPAU at boot speed. With this patch applied I only need to bump the vdec clock to
>>>>> get real-time 1080p decoding. It prevents a lot of VRAM <-> VRAM buffer moves.
>>>> It's not bogus, and is required for pre-GF8 boards with VRAM > BAR size.
>>>>
>>>> What configuration does the buffer that's getting moved here have
>>>> exactly?  The placement restriction isn't necessary on GF8, the rest
>>>> of the restrictions may currently be required still however.
>>>>
>>>> = vdpau on NVC0 with tiling. I upload the raw bitstream to a tiling bo. This is ok because
>>> the vm hides all the tiling translations, and the engines will read the raw bitstream correctly.
>> Why would you be doing such a thing in the first place?  It seems
>> pointless, and quite possibly counter-productive to use a tiled layout
>> for a linear data structure...
> Initially I just allocated everything I didn't need to access directly tiled, and it seems I did the same for
> the bitstream bo. I only found out later about the bug with excessive moves causing a major slowdown.
>
>>> 8<---
>>> This prevents buffer moves from being done on NV50+, where remapping is not needed because
>>> the bar has its own VM, instead of only having the first BAR1-size chunk of VRAM accessible.
>>> nouveau_bo_tile_layout is always 0 on < NV_50.
>>>
>>> Signed-off-by: Maarten Lankhorst <maarten.lankhorst at canonical.com>
>>> ---
>>> diff --git a/drivers/gpu/drm/nouveau/nouveau_bo.c b/drivers/gpu/drm/nouveau/nouveau_bo.c
>>> index d506da5..762bfcd 100644
>>> --- a/drivers/gpu/drm/nouveau/nouveau_bo.c
>>> +++ b/drivers/gpu/drm/nouveau/nouveau_bo.c
>>> @@ -1346,14 +1361,13 @@ nouveau_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
>>>         struct nouveau_device *device = nv_device(drm->device);
>>>         u32 mappable = pci_resource_len(device->pdev, 1) >> PAGE_SHIFT;
>>>
>>> -       /* as long as the bo isn't in vram, and isn't tiled, we've got
>>> -        * nothing to do here.
>>> +       /*
>>> +        * if the bo is not in vram, or remapping can be done (nv50+)
>>> +        * do not worry about placement, any location is valid
>>>          */
>>> -       if (bo->mem.mem_type != TTM_PL_VRAM) {
>>> -               if (nv_device(drm->device)->card_type < NV_50 ||
>>> -                   !nouveau_bo_tile_layout(nvbo))
>>> -                       return 0;
>>> -       }
>>> +       if (nv_device(drm->device)->card_type >= NV_50 ||
>>> +           bo->mem.mem_type != TTM_PL_VRAM)
>>> +               return 0;
>> I get what you're trying to do here, and we should definitely avoid
>> the "mappable vram" check on GF8, but I suspect this condition is too
>> broad.  I'll think about it more after I finish reviewing the rest of
>> the patches on the list..
>>
> I think this relaxed check is fine. If it's !VRAM, the host can always access it because it has direct access to the
> pages without needing anything from the gpu. On >= NV50 the move can always be skipped too because the
> memory is mapped to the vm, and always accessible.
Yeah, I think the check is sane, after thinking it through.  We shall
find out :)

But first, if you're going to not move tiled system memory to vram
first, you're going to need to deal with mapping it into bar1 so the
vm deals with some of the reordering.  See what's done in
io_mem_reserve() for TTM_PL_VRAM.

Ben.

>
> ~Maarten
>