[RFC] Explicit synchronization for Nouveau

[James Jones]

On 9/29/14 8:42 AM, Jerome Glisse wrote:
> On Mon, Sep 29, 2014 at 09:43:02AM +0200, Daniel Vetter wrote:
>> On Fri, Sep 26, 2014 at 01:00:05PM +0300, Lauri Peltonen wrote:
>>>
>>> Hi guys,
>>>
>>>
>>> I'd like to start a new thread about explicit fence synchronization.  This time
>>> with a Nouveau twist. :-)
>>>
>>> First, let me define what I understand by implicit/explicit sync:
>>>
>>> Implicit synchronization
>>> * Fences are attached to buffers
>>> * Kernel manages fences automatically based on buffer read/write access
>>>
>>> Explicit synchronization
>>> * Fences are passed around independently
>>> * Kernel takes and emits fences to/from user space when submitting work
>>>
>>> Implicit synchronization is already implemented in open source drivers, and
>>> works well for most use cases.  I don't seek to change any of that.  My
>>> proposal aims at allowing some drm drivers to operate in explicit sync mode to
>>> get maximal performance, while still remaining fully compatible with the
>>> implicit paradigm.
>>
>> Yeah, pretty much what we have in mind on the i915 side too. I didn't look
>> too closely at your patches, so just a few high level comments on your rfc
>> here.
>>
>>> I will try to explain why I think we should support the explicit model as well.
>>>
>>>
>>> 1. Bindless graphics
>>>
>>> Bindless graphics is a central concept when trying to reduce the OpenGL driver
>>> overhead.  The idea is that the application can bind a large set of buffers to
>>> the working set up front using extensions such as GL_ARB_bindless_texture, and
>>> they remain resident until the application releases them (note that compute
>>> APIs have typically similar semantics).  These working sets can be huge,
>>> hundreds or even thousands of buffers, so we would like to opt out from the
>>> per-submit overhead of acquiring locks, waiting for fences, and storing fences.
>>> Automatically synchronizing these working sets in kernel will also prevent
>>> parallelism between channels that are sharing the working set (in fact sharing
>>> just one buffer from the working set will cause the jobs of the two channels to
>>> be serialized).
>>>
>>> 2. Evolution of graphics APIs
>>>
>>> The graphics API evolution seems to be going to a direction where game engine
>>> and middleware vendors demand more control over work submission and
>>> synchronization.  We expect that this trend will continue, and more and more
>>> synchronization decisions will be pushed to the API level.  OpenGL and EGL
>>> already provide good explicit command stream level synchronization primitives:
>>> glFenceSync and EGL_KHR_wait_sync.  Their use is also encouraged - for example
>>> EGL_KHR_image_base spec clearly states that the application is responsible for
>>> synchronizing accesses to EGLImages.  If the API that is exposed to developers
>>> gives the control over synchronization to the developer, then implicit waits
>>> that are inserted by the kernel are unnecessary and unexpected, and can
>>> severely hurt performance.  It also makes it easy for the developer to write
>>> code that happens to work on Linux because of implicit sync, but will fail on
>>> other platforms.
>>>
>>> 3. Suballocation
>>>
>>> Using user space suballocation can help reduce the overhead when a large number
>>> of small textures are used.  Synchronizing suballocated surfaces implicitly in
>>> kernel doesn't make sense - many channels should be able to access the same
>>> kernel-level buffer object simultaneously.
>>>
>>> 4. Buffer sharing complications
>>>
>>> This is not really an argument for explicit sync as such, but I'd like to point
>>> out that sharing buffers across SoC engines is often much more complex than
>>> just exporting and importing a dma-buf and waiting for the dma-buf fences.
>>> Sometimes we need to do color format or tiling layout conversion.  Sometimes,
>>> at least on Tegra, we need to decompress buffers when we pass them from the GPU
>>> to an engine that doesn't support framebuffer compression.  These things are
>>> not uncommon, particularly when we have SoC's that combine licensed IP blocks
>>> from different vendors.  My point is that user space is already heavily
>>> involved when sharing buffers between drivers, and giving it some more control
>>> over synchronization is not adding that much complexity.
>>>
>>>
>>> Because of the above arguments, I think it makes sense to let some user space
>>> drm drivers opt out from implicit synchronization, while allowing them to still
>>> remain fully compatible with the rest of the drm world that uses implicit
>>> synchronization.  In practice, this would require three things:
>>>
>>> (1) Support passing fences (that are not tied to buffer objects) between kernel
>>>      and user space.
>>>
>>> (2) Stop automatically storing fences to the buffers that user space wants to
>>>      synchronize explicitly.
>>
>> The problem with this approach is that you then need hw faulting to make
>> sure the memory is there. Implicit fences aren't just used for syncing,
>> but also to make sure that the gpu still has access to the buffer as long
>> as it needs it. So you need at least a non-exclusive fence attached for
>> each command submission.
>>
>> Of course on Android you don't have swap (would kill the puny mmc within
>> seconds) and you don't care for letting userspace pin most of memory for
>> gfx. So you'll get away with no fences at all. But for upstream I don't
>> see a good solution unfortunately. Ideas very much welcome.
>
> Well i am gonna repeat myself. But yes you can do explicit without associating
> fence (at least no struct alloc) just associate a unique per command stream
> number and have it be global ie irrespective of different execution pipeline
> your hw have.
>
> For non scheduling GPU, today generation roughly, you keep buffer on lru and
> you know you can not evict buffer to swap for those that have an active id
> (hw did not yet write the sequence number back). You update the lru with each
> command stream ioctl.
>
> For binding to GPU GART you can do that as a preamble to the command stream
> which most hardware (AMD, Intel, NVidia) should be able to do.
>
> For VRAM you have several choice that depends on how you want to manage VRAM.
> For instance you might want to use it more like a cache and have each command
> stream preamble with a bunch of copy to VRAM and posibly bunch of post copy
> back to RAM. Or you can hold to current scheme but buffer move now becomes
> preamble to command stream (ie buffer move are scheduled as a preamble to
> command stream).

Additionally, I think the goal is to move to a model where some 
higher-level object such as a working set, rather than individual 
buffers, are assigned counters or sync primitives on a per-submission 
basis.  Versioning off tags for individual buffers then moves to working 
set modification time.  This is more feasible if the only thing that 
needs precise fencing of individual surfaces is lifetime management.

The trend seems to be towards establishing a relatively large working 
set up front and then submitting many command buffers against it, 
perhaps with incremental modifications to the working set along the way. 
  This may be what's referred to as the Android model above, but I view 
it as the "non-glitchy graphic" model going forward.

Thanks,
-James

> So i do not see what you would consider rocket science about this ?
>
> Cheers,
> Jérôme
>
>>
>>> (3) Allow user space to attach an explicit fence to dma-buf when exporting to
>>>      another driver that uses implicit sync.
>>>
>>> There are still some open issues beyond these.  For example, can we skip
>>> acquiring the ww mutex for explicitly synchronized buffers?  I think we could
>>> eventually, at least on unified memory systems where we don't need to migrate
>>> between heaps (our downstream Tegra GPU driver does not lock any buffers at
>>> submit, it just grabs refcounts for hw).  Another quirk is that now Nouveau
>>> waits on the buffer fences when closing the gem object to ensure that it
>>> doesn't unmap too early.  We need to rework that for explicit sync, but that
>>> shouldn't be difficult.
>>
>> See above, but you can't avoid to attach fences as long as we still use a
>> buffer-object based gfx memory management model. At least afaics. Which
>> means you need the ordering guarantees imposed by ww mutexes to ensure
>> that the oddball implicit ordered client can't deadlock the kernel's
>> memory management code.
>>
>>> I have written a prototype that demonstrates (1) by adding explicit sync fd
>>> support to Nouveau.  It's not a lot of code, because I only use a relatively
>>> small subset of the android sync driver functionality.  Thanks to Maarten's
>>> rewrite, all I need to do is to allow creating a sync_fence from a drm fence in
>>> order to pass it to user space.  I don't need to use sync_pt or sync_timeline,
>>> or fill in sync_timeline_ops.
>>>
>>> I can see why the upstream has been reluctant to de-stage the android sync
>>> driver in its current form, since (even though it now builds on struct fence)
>>> it still duplicates some of the drm fence concepts.  I'd like to think that my
>>> patches only use the parts of the android sync driver that genuinely are
>>> missing from the drm fence model: allowing user space to operate on fence
>>> objects that are independent of buffer objects.
>>
>> Imo de-staging the android syncpt stuff needs to happen first, before
>> drivers can use it. Since non-staging stuff really shouldn't depend upon
>> code from staging.
>>
>>> The last two patches are mocks that show how (2) and (3) might work out.  I
>>> haven't done any testing with them yet.  Before going any further, I'd like to
>>> get your feedback.  Can you see the benefits of explicit sync as an alternative
>>> synchronization model?  Do you think we could use the android sync_fence for
>>> passing fences between user space?  Or did you have something else in mind for
>>> explicit sync in the drm world?
>>
>> I'm all for adding explicit syncing. Our plans are roughly.
>> - Add both an in and and out fence to execbuf to sync with other rendering
>>    and give userspace a fence back. Needs to different flags probably.
>>
>> - Maybe add an ioctl to dma-bufs to get at the current implicit fences
>>    attached to them (both an exclusive and non-exclusive version). This
>>    should help with making explicit and implicit sync work together nicely.
>>
>> - Add fence support to kms. Probably only worth it together with the new
>>    atomic stuff. Again we need an in fence to wait for (one for each
>>    buffer) and an out fence. The later can easily be implemented by
>>    extending struct drm_event, which means not a single driver code line
>>    needs to be changed for this.
>>
>> - For de-staging android syncpts we need to de-clutter the internal
>>    interfaces and also review all the ioctls exposed. Like you say it
>>    should be just the userspace interface for struct drm_fence. Also, it
>>    needs testcases and preferrably manpages.
>>
>> Unfortunately it looks like Intel won't do this all for you due to a bunch
>> of hilarious internal reasons :( At least not anytime soon.
>>
>> Cheers, Daniel
>> --
>> Daniel Vetter
>> Software Engineer, Intel Corporation
>> +41 (0) 79 365 57 48 - http://blog.ffwll.ch
>> _______________________________________________
>> dri-devel mailing list
>> dri-devel at lists.freedesktop.org
>> http://lists.freedesktop.org/mailman/listinfo/dri-devel