Introduce a new helper framework for buffer synchronization

Inki Dae inki.dae at samsung.com
Tue May 21 02:22:49 PDT 2013



> -----Original Message-----
> From: Daniel Vetter [mailto:daniel.vetter at ffwll.ch] On Behalf Of Daniel
> Vetter
> Sent: Tuesday, May 21, 2013 4:45 PM
> To: Inki Dae
> Cc: 'Daniel Vetter'; 'Rob Clark'; 'linux-fbdev'; 'DRI mailing list';
> 'Kyungmin Park'; 'myungjoo.ham'; 'YoungJun Cho'; linux-arm-
> kernel at lists.infradead.org; linux-media at vger.kernel.org
> Subject: Re: Introduce a new helper framework for buffer synchronization
> 
> On Tue, May 21, 2013 at 04:03:06PM +0900, Inki Dae wrote:
> > > - Integration of fence syncing into dma_buf. Imo we should have a
> > >   per-attachment mode which decides whether map/unmap (and the new
> sync)
> > >   should wait for fences or whether the driver takes care of syncing
> > >   through the new fence framework itself (for direct hw sync).
> >
> > I think it's a good idea to have per-attachment mode for buffer sync.
> But
> > I'd like to say we make sure what is the purpose of map
> > (dma_buf_map_attachment)first. This interface is used to get a sgt;
> > containing pages to physical memory region, and map that region with
> > device's iommu table. The important thing is that this interface is
> called
> > just one time when user wants to share an allocated buffer with dma. But
> cpu
> > will try to access the buffer every time as long as it wants. Therefore,
> we
> > need cache control every time cpu and dma access the shared buffer:
> cache
> > clean when cpu goes to dma and cache invalidate when dma goes to cpu.
> That
> > is why I added new interfaces, DMA_BUF_GET_FENCE and DMA_BUF_PUT_FENCE,
> to
> > dma buf framework. Of course, Those are ugly so we need a better way: I
> just
> > wanted to show you that in such way, we can synchronize the shared
> buffer
> > between cpu and dma. By any chance, is there any way that kernel can be
> > aware of when cpu accesses the shared buffer or is there any point I
> didn't
> > catch up?
> 
> Well dma_buf_map/unmap_attachment should also flush/invalidate any caches,
> and with the current dma_buf spec those two functions are the _only_ means

I know that dma buf exporter should make sure that cache clean/invalidate
are done when dma_buf_map/unmap_attachement is called. For this, already we
do so. However, this function is called when dma buf import is requested by
user to map a dmabuf fd with dma. This means that dma_buf_map_attachement()
is called ONCE when user wants to share the dmabuf fd with dma. Actually, in
case of exynos drm, dma_map_sg_attrs(), performing cache operation
internally, is called when dmabuf import is requested by user.

> you have to do so. Which strictly means that if you interleave device dma
> and cpu acccess you need to unmap/map every time.
> 
> Which is obviously horribly inefficient, but a known gap in the dma_buf

Right, and also this has big overhead.

> interface. Hence why I proposed to add dma_buf_sync_attachment similar to
> dma_sync_single_for_device:
> 
> /**
>  * dma_buf_sync_sg_attachment - sync caches for dma access
>  * @attach: dma-buf attachment to sync
>  * @sgt: the sg table to sync (returned by dma_buf_map_attachement)
>  * @direction: dma direction to sync for
>  *
>  * This function synchronizes caches for device dma through the given
>  * dma-buf attachment when interleaving dma from different devices and the
>  * cpu. Other device dma needs to be synced also by calls to this
>  * function (or a pair of dma_buf_map/unmap_attachment calls), cpu access
>  * needs to be synced with dma_buf_begin/end_cpu_access.
>  */
> void dma_buf_sync_sg_attachment(struct dma_buf_attachment *attach,
> 				struct sg_table *sgt,
> 				enum dma_data_direction direction)
> 
> Note that "sync" here only means to synchronize caches, not wait for any
> outstanding fences. This is simply to be consistent with the established
> lingo of the dma api. How the dma-buf fences fit into this is imo a
> different topic, but my idea is that all the cache coherency barriers
> (i.e. dma_buf_map/unmap_attachment, dma_buf_sync_sg_attachment and
> dma_buf_begin/end_cpu_access) would automatically block for any attached
> fences (i.e. block for write fences when doing read-only access, block for
> all fences otherwise).

As I mentioned already, kernel can't aware of when cpu accesses a shared
buffer: user can access a shared buffer after mmap anytime and the shared
buffer should be synchronized between cpu and dma. Therefore, the above
cache coherency barriers should be called every time cpu and dma tries to
access a shared buffer, checking before and after of cpu and dma access. And
exactly, the proposed way do such thing. For this, you can refer to the
below link,
	
http://www.mail-archive.com/linux-media@vger.kernel.org/msg62124.html

My point is that how kernel can aware of when those cache coherency barriers
should be called to synchronize caches and buffer access between cpu and
dma.

> 
> Then we could do a new dma_buf_attach_flags interface for special cases
> (might also be useful for other things, similar to the recently added
> flags in the dma api for wc/no-kernel-mapping/...). A new flag like
> DMA_BUF_ATTACH_NO_AUTOFENCE would then mean that the driver takes care of
> all
> fencing for this attachment and the dma-buf functions should not do the
> automagic fence blocking.
> 
> > In Linux kernel, especially embedded system, we had tried to implement
> > generic interfaces for buffer management; how to allocate a buffer and
> how
> > to share a buffer. As a result, now we have CMA (Contiguous Memory
> > Allocator) for buffer allocation, and we have DMA-BUF for buffer sharing
> > between cpu and dma. And then how to synchronize a buffer between cpu
> and
> > dma? I think now it's best time to discuss buffer synchronization
> mechanism,
> > and that is very natural.
> 
> I think it's important to differentiate between the two meanings of sync:
> - synchronize caches (i.e. cpu cache flushing in most cases)
> - and synchronize in-flight dma with fences.
> 

Agree, and I meant that the buffer synchronization mechanism includes the
above two things. And I think the two things should be addressed together.

Thanks,
Inki Dae

> Cheers, Daniel
> --
> Daniel Vetter
> Software Engineer, Intel Corporation
> +41 (0) 79 365 57 48 - http://blog.ffwll.ch



More information about the dri-devel mailing list