Thanks Lucas for the explanation!

Lucas Stach wrote:

Hi Tobias,

Am Sonntag, den 16.08.2015, 14:48 +0200 schrieb Tobias Jakobi:

...

Hello,

some time ago I checked whether I could use the userptr functionality to do zero-copy from userspace allocated buffers via the G2D. This didn't work out so well, so kinda put this to the bottom of my TODO list.

Now that IOMMU support has landed and Jan Kara has rewrote page pinning using frame vectors (see [1]) I gave userptr another try.

The results are much better. I'm not experiencing any kernel lockups or sysmmu pagefaults anymore. However the image now suffers from visual artifacts. These images show the nature of the artifacts: http://i.imgur.com/nzT6g3Y.jpg http://i.imgur.com/wkuYI6X.jpg

The corruption always manifests itself in these pixel lines of fixed size and wrong color.

I have written a testcase as part of libdrm for this issue: https://github.com/tobiasjakobi/libdrm/commit/db8bf6844436598251f67a71fc334b...

It allocates N (N an even number) buffers which are aligned to the system pagesize. Then it does this each iteration:

1. Fill the first N/2 buffers with random data
2. Copy the first half to the second half of the buffers
3. memcmp() first and second half (verification pass)

Usually this verification already fails on the first iteration. An interesting observation is that increasing (!) the buffer size (so the amount of pixels that have to copied per buffer grows) makes this issue less likely to happen.

With the default 512x512 buffers however it happens, like I said above, almost immediately.

This is obviously a cache flush missing. The memory you get from userspace is normal cached memory, so to make it visible to the GPU you need to flush parts of the cache out to main memory.

The corruption you are seeing is just unflushed cachelines. This also explains why increasing the buffer size helps: the more memory the CPU touches the more cachelines will be flushed out to be replaced with new data.

I should point out that the snapshots I uploaded were done with a different setup. There only the source memory of the G2D operation is a userspace allocated buffer. The destination is a GEM buffer allocated through libdrm, which is then used as framebuffer. So the issue already appears when just the source is userspace allocated.

What works however is an operation between GEM to GEM. However this might be related to the default allocation flags libdrm uses.

So you need to go and have a look at dma_map() and dma_sync_*_for_*() and friends.

Regards, Lucas

With best wishes, Tobias

------------------------------------------------------------------------------ -- _______________________________________________ Dri-devel mailing list Dri-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/dri-devel

On Wed, Aug 19, 2015 at 03:53:44PM +0200, Tobias Jakobi wrote:

Adding Jérôme to Cc. I think he looked the userptr code before, so maybe he has some idea what is going wrong here.

I also had a look at the code, but my knowledge about the DMA API is almost nonexistant. However I can see that before doing any DMA via the G2D on the buffer the code calls dma_map() on it, and also unmaps it when the commandlist is finished.

With best wishes, Tobias

Tobias Jakobi wrote:

...

Thanks Lucas for the explanation!

Lucas Stach wrote:

...

Hi Tobias,

Am Sonntag, den 16.08.2015, 14:48 +0200 schrieb Tobias Jakobi:

...

Hello,

some time ago I checked whether I could use the userptr functionality to do zero-copy from userspace allocated buffers via the G2D. This didn't work out so well, so kinda put this to the bottom of my TODO list.

Now that IOMMU support has landed and Jan Kara has rewrote page pinning using frame vectors (see [1]) I gave userptr another try.

The results are much better. I'm not experiencing any kernel lockups or sysmmu pagefaults anymore. However the image now suffers from visual artifacts. These images show the nature of the artifacts: http://i.imgur.com/nzT6g3Y.jpg http://i.imgur.com/wkuYI6X.jpg

The corruption always manifests itself in these pixel lines of fixed size and wrong color.

I have written a testcase as part of libdrm for this issue: https://github.com/tobiasjakobi/libdrm/commit/db8bf6844436598251f67a71fc334b...

It allocates N (N an even number) buffers which are aligned to the system pagesize. Then it does this each iteration:

1. Fill the first N/2 buffers with random data
2. Copy the first half to the second half of the buffers
3. memcmp() first and second half (verification pass)

Usually this verification already fails on the first iteration. An interesting observation is that increasing (!) the buffer size (so the amount of pixels that have to copied per buffer grows) makes this issue less likely to happen.

With the default 512x512 buffers however it happens, like I said above, almost immediately.

This is obviously a cache flush missing. The memory you get from userspace is normal cached memory, so to make it visible to the GPU you need to flush parts of the cache out to main memory.

The corruption you are seeing is just unflushed cachelines. This also explains why increasing the buffer size helps: the more memory the CPU touches the more cachelines will be flushed out to be replaced with new data.

I should point out that the snapshots I uploaded were done with a different setup. There only the source memory of the G2D operation is a userspace allocated buffer. The destination is a GEM buffer allocated through libdrm, which is then used as framebuffer. So the issue already appears when just the source is userspace allocated.

This is still consistent with cachelines issue. Is your GPU & IOMMU cache coherent with the CPU ? If not then it means you need to cache flush the buffer before you use it with the GPU. The dma API provide few helpers for that.

Cheers, Jérôme

Hello,

I did some further investigation into this issue.

I extended my test application to also cover GEM/GEM and GEM/userptr transfers, and also putting GEM allocation flags into the mix.

You can find the current version here: https://github.com/tobiasjakobi/libdrm/blob/exynos/tests/exynos/exynos_fimg2...

What is interesting is that all GEM/GEM transfer variants work, regardless of the allocation flags.

In particular I can allocate a cacheable GEM buffer and the verification still works. This confuses me since everything point into the direction that this is a cache flush / cache coherency issue. So why doesn't this happen for GEM buffers as well if we allow caching to happen there?

Next I hacked up a custom ioctl to trigger flush_cache_all(), outer_flush_all() and outer_sync(), issuing this ioctl after filling the buffers and also after the g2d operations. This somehow helps, in the sense that the test application runs longer. But it also makes the system prone to lockups. Looks like calling outer_flush_all() a lot triggers these. I don't know why this happens but IIRC one shouldn't call these low-level functions from driver code anyway.

Next I checked the DMA-API and looked into dma_sync_sg_for_cpu and dma_sync_sg_for_device. But the driver code already does this: 1) dma_map before starting the engine 2) dma_unmap after the engine finishes

The DMA-API HOWTO explicitly says that the sync functions are only there if I modify the buffer with the CPU between map and unmap. Which is not the case here.

In particular I traced what dma_{map,unmap} do depending on the DMA direction and deep down the callstack some flush/invalidate stuff is called. And the GEM cases show that this is somehow working.

Next I looked into Jerome's question about whethere the G2D is cache coherent with the CPU. I looked into old Android code and found FIMG2D_AXI_MODE_REG, a register that currently isn't touched in the DRM code. It seems to manipulate signals to the AXI Master interface.

The register looks like this: [0:3] ARCACHE [4:7] AWCACHE [8:15] ARUSERS [16:23] AWUSERS [24:25] MaxBurstLength

If I understand ARM's AXI specifications correctly then you can set cache coherent operation by setting AxUSER[0]=1 and AxCACHE[1]=1.

The initialization values for AxCACHE and AxUSERS are all zero, so the default seems to be non-coherent operation. But still after setting the bits in FIMG2D_AXI_MODE_REG I don't see any change in behaviour -- the issue stays.

Any idea what I'm missing here?

With best wishes, Tobias

On 2015-08-19 16:41, Rob Clark wrote:

On Wed, Aug 19, 2015 at 10:08 AM, Jerome Glisse jglisse@redhat.com wrote:

...

On Wed, Aug 19, 2015 at 03:53:44PM +0200, Tobias Jakobi wrote:

...

Adding Jérôme to Cc. I think he looked the userptr code before, so maybe he has some idea what is going wrong here.

I also had a look at the code, but my knowledge about the DMA API is almost nonexistant. However I can see that before doing any DMA via the G2D on the buffer the code calls dma_map() on it, and also unmaps it when the commandlist is finished.

With best wishes, Tobias

Tobias Jakobi wrote:

...

Thanks Lucas for the explanation!

Lucas Stach wrote:

...

Hi Tobias,

Am Sonntag, den 16.08.2015, 14:48 +0200 schrieb Tobias Jakobi:

...

Hello,

some time ago I checked whether I could use the userptr functionality to do zero-copy from userspace allocated buffers via the G2D. This didn't work out so well, so kinda put this to the bottom of my TODO list.

Now that IOMMU support has landed and Jan Kara has rewrote page pinning using frame vectors (see [1]) I gave userptr another try.

The results are much better. I'm not experiencing any kernel lockups or sysmmu pagefaults anymore. However the image now suffers from visual artifacts. These images show the nature of the artifacts: http://i.imgur.com/nzT6g3Y.jpg http://i.imgur.com/wkuYI6X.jpg

The corruption always manifests itself in these pixel lines of fixed size and wrong color.

I have written a testcase as part of libdrm for this issue: https://github.com/tobiasjakobi/libdrm/commit/db8bf6844436598251f67a71fc334b...

It allocates N (N an even number) buffers which are aligned to the system pagesize. Then it does this each iteration:

1. Fill the first N/2 buffers with random data
2. Copy the first half to the second half of the buffers
3. memcmp() first and second half (verification pass)

Usually this verification already fails on the first iteration. An interesting observation is that increasing (!) the buffer size (so the amount of pixels that have to copied per buffer grows) makes this issue less likely to happen.

With the default 512x512 buffers however it happens, like I said above, almost immediately.

This is obviously a cache flush missing. The memory you get from userspace is normal cached memory, so to make it visible to the GPU you need to flush parts of the cache out to main memory.

The corruption you are seeing is just unflushed cachelines. This also explains why increasing the buffer size helps: the more memory the CPU touches the more cachelines will be flushed out to be replaced with new data.

I should point out that the snapshots I uploaded were done with a different setup. There only the source memory of the G2D operation is a userspace allocated buffer. The destination is a GEM buffer allocated through libdrm, which is then used as framebuffer. So the issue already appears when just the source is userspace allocated.

This is still consistent with cachelines issue. Is your GPU & IOMMU cache coherent with the CPU ? If not then it means you need to cache flush the buffer before you use it with the GPU. The dma API provide few helpers for that.

although I suspect dma-api probably not aware of any device caches (and I suspect a bit weak when it comes to devices that support mix of coherent and non-coherent mappings)..

BR, -R

...

Cheers, Jérôme _______________________________________________ dri-devel mailing list dri-devel@lists.freedesktop.org http://lists.freedesktop.org/mailman/listinfo/dri-devel