Re: [RFC PATCH] drm/virtio: Export resource handles via DMA-buf API
On Wed, Oct 16, 2019 at 12:19:02PM +0900, Tomasz Figa wrote:
> On Wed, Oct 9, 2019 at 12:04 AM Daniel Vetter wrote:
> >
> > On Tue, Oct 08, 2019 at 07:49:39PM +0900, Tomasz Figa wrote:
> > > On Tue, Oct 8, 2019 at 7:03 PM Daniel Vetter wrote:
> > > >
> > > > On Sat, Oct 05, 2019 at 02:41:54PM +0900, Tomasz Figa wrote:
> > > > > Hi Daniel, Gerd,
> > > > >
> > > > > On Tue, Sep 17, 2019 at 10:23 PM Daniel Vetter
> > > > > wrote:
> > > > > >
> > > > > > On Thu, Sep 12, 2019 at 06:41:21PM +0900, Tomasz Figa wrote:
> > > > > > > This patch is an early RFC to judge the direction we are
> > > > > > > following in
> > > > > > > our virtualization efforts in Chrome OS. The purpose is to start a
> > > > > > > discussion on how to handle buffer sharing between multiple virtio
> > > > > > > devices.
> > > > > > >
> > > > > > > On a side note, we are also working on a virtio video decoder
> > > > > > > interface
> > > > > > > and implementation, with a V4L2 driver for Linux. Those will be
> > > > > > > posted
> > > > > > > for review in the near future as well.
> > > > > > >
> > > > > > > Any feedback will be appreciated! Thanks in advance.
> > > > > > >
> > > > > > > ===
> > > > > > >
> > > > > > > With the range of use cases for virtualization expanding, there
> > > > > > > is going
> > > > > > > to be more virtio devices added to the ecosystem. Devices such as
> > > > > > > video
> > > > > > > decoders, encoders, cameras, etc. typically work together with the
> > > > > > > display and GPU in a pipeline manner, which can only be
> > > > > > > implemented
> > > > > > > efficiently by sharing the buffers between producers and
> > > > > > > consumers.
> > > > > > >
> > > > > > > Existing buffer management framework in Linux, such as the
> > > > > > > videobuf2
> > > > > > > framework in V4L2, implements all the DMA-buf handling inside
> > > > > > > generic
> > > > > > > code and do not expose any low level information about the
> > > > > > > buffers to
> > > > > > > the drivers.
> > > > > > >
> > > > > > > To seamlessly enable buffer sharing with drivers using such
> > > > > > > frameworks,
> > > > > > > make the virtio-gpu driver expose the resource handle as the DMA
> > > > > > > address
> > > > > > > of the buffer returned from the DMA-buf mapping operation.
> > > > > > > Arguably, the
> > > > > > > resource handle is a kind of DMA address already, as it is the
> > > > > > > buffer
> > > > > > > identifier that the device needs to access the backing memory,
> > > > > > > which is
> > > > > > > exactly the same role a DMA address provides for native devices.
> > > > > > >
> > > > > > > A virtio driver that does memory management fully on its own
> > > > > > > would have
> > > > > > > code similar to following. The code is identical to what a regular
> > > > > > > driver for real hardware would do to import a DMA-buf.
> > > > > > >
> > > > > > > static int virtio_foo_get_resource_handle(struct virtio_foo *foo,
> > > > > > > struct dma_buf *dma_buf,
> > > > > > > u32 *id)
> > > > > > > {
> > > > > > > struct dma_buf_attachment *attach;
> > > > > > > struct sg_table *sgt;
> > > > > > > int ret = 0;
> > > > > > >
> > > > > > > attach = dma_buf_attach(dma_buf, foo->dev);
> > > > > > > if (IS_ERR(attach))
> > > > > > > return PTR_ERR(attach);
> > > > > > >
> > > > > > > sgt = dma_buf_map_attachment(attach, DMA_BIDIRECTIONAL);
> > > > > > > if (IS_ERR(sgt)) {
> > > > > > > ret = PTR_ERR(sgt);
> > > > > > > goto err_detach;
> > > > > > > }
> > > > > > >
> > > > > > > if (sgt->nents != 1) {
> > > > > > > ret = -EINVAL;
> > > > > > > goto err_unmap;
> > > > > > > }
> > > > > > >
> > > > > > > *id = sg_dma_address(sgt->sgl);
> > > > > >
> > > > > > I agree with Gerd, this looks pretty horrible to me.
> > > > > >
> > > > > > The usual way we've done these kind of special dma-bufs is:
> > > > > >
> > > > > > - They all get allocated at the same place, through some library or
> > > > > > whatever.
> > > > > >
> > > > > > - You add a dma_buf_is_virtio(dma_buf) function, or maybe something
> > > > > > that
> > > > > > also upcasts or returns NULL, which checks for dma_buf->ops.
> > > > > >
> > > > >
> > > > > Thanks for a lot of valuable feedback and sorry for the late reply.
> > > > >
> > > > > While I agree that stuffing the resource ID in sg_dma_address() is
> > > > > quite ugly (for example, the regular address arithmetic doesn't work),
> > > > > I still believe we need to convey information about these buffers
> > > > > using regular kernel interfaces.
> > > > >
> > > > > Drivers in some subsystems like DRM tend to open code any buffer
> > > > > management and then it wouldn't be any problem to do what you
> > > > > suggested. However, other subsystems have generic frameworks for
> > > > > buffer management, like videobuf2 for
Re: [RFC PATCH] drm/virtio: Export resource handles via DMA-buf API
Hi, > up later when given a buffer index. But we would still need to make > the DMA-buf itself importable. For virtio-gpu I guess that would mean > returning an sg_table backed by the shadow buffer pages. The virtio-gpu driver in drm-misc-next supports dma-buf exports. cheers, Gerd ___ Virtualization mailing list Virtualization@lists.linux-foundation.org https://lists.linuxfoundation.org/mailman/listinfo/virtualization
Re: [RFC PATCH] drm/virtio: Export resource handles via DMA-buf API
On Sat, Oct 05, 2019 at 02:41:54PM +0900, Tomasz Figa wrote:
> Hi Daniel, Gerd,
>
> On Tue, Sep 17, 2019 at 10:23 PM Daniel Vetter wrote:
> >
> > On Thu, Sep 12, 2019 at 06:41:21PM +0900, Tomasz Figa wrote:
> > > This patch is an early RFC to judge the direction we are following in
> > > our virtualization efforts in Chrome OS. The purpose is to start a
> > > discussion on how to handle buffer sharing between multiple virtio
> > > devices.
> > >
> > > On a side note, we are also working on a virtio video decoder interface
> > > and implementation, with a V4L2 driver for Linux. Those will be posted
> > > for review in the near future as well.
> > >
> > > Any feedback will be appreciated! Thanks in advance.
> > >
> > > ===
> > >
> > > With the range of use cases for virtualization expanding, there is going
> > > to be more virtio devices added to the ecosystem. Devices such as video
> > > decoders, encoders, cameras, etc. typically work together with the
> > > display and GPU in a pipeline manner, which can only be implemented
> > > efficiently by sharing the buffers between producers and consumers.
> > >
> > > Existing buffer management framework in Linux, such as the videobuf2
> > > framework in V4L2, implements all the DMA-buf handling inside generic
> > > code and do not expose any low level information about the buffers to
> > > the drivers.
> > >
> > > To seamlessly enable buffer sharing with drivers using such frameworks,
> > > make the virtio-gpu driver expose the resource handle as the DMA address
> > > of the buffer returned from the DMA-buf mapping operation. Arguably, the
> > > resource handle is a kind of DMA address already, as it is the buffer
> > > identifier that the device needs to access the backing memory, which is
> > > exactly the same role a DMA address provides for native devices.
> > >
> > > A virtio driver that does memory management fully on its own would have
> > > code similar to following. The code is identical to what a regular
> > > driver for real hardware would do to import a DMA-buf.
> > >
> > > static int virtio_foo_get_resource_handle(struct virtio_foo *foo,
> > > struct dma_buf *dma_buf, u32 *id)
> > > {
> > > struct dma_buf_attachment *attach;
> > > struct sg_table *sgt;
> > > int ret = 0;
> > >
> > > attach = dma_buf_attach(dma_buf, foo->dev);
> > > if (IS_ERR(attach))
> > > return PTR_ERR(attach);
> > >
> > > sgt = dma_buf_map_attachment(attach, DMA_BIDIRECTIONAL);
> > > if (IS_ERR(sgt)) {
> > > ret = PTR_ERR(sgt);
> > > goto err_detach;
> > > }
> > >
> > > if (sgt->nents != 1) {
> > > ret = -EINVAL;
> > > goto err_unmap;
> > > }
> > >
> > > *id = sg_dma_address(sgt->sgl);
> >
> > I agree with Gerd, this looks pretty horrible to me.
> >
> > The usual way we've done these kind of special dma-bufs is:
> >
> > - They all get allocated at the same place, through some library or
> > whatever.
> >
> > - You add a dma_buf_is_virtio(dma_buf) function, or maybe something that
> > also upcasts or returns NULL, which checks for dma_buf->ops.
> >
>
> Thanks for a lot of valuable feedback and sorry for the late reply.
>
> While I agree that stuffing the resource ID in sg_dma_address() is
> quite ugly (for example, the regular address arithmetic doesn't work),
> I still believe we need to convey information about these buffers
> using regular kernel interfaces.
>
> Drivers in some subsystems like DRM tend to open code any buffer
> management and then it wouldn't be any problem to do what you
> suggested. However, other subsystems have generic frameworks for
> buffer management, like videobuf2 for V4L2. Those assume regular
> DMA-bufs that can be handled with regular dma_buf_() API and described
> using sgtables and/or pfn vectors and/or DMA addresses.
"other subsystem sucks" doesn't sound like a good design paradigm to me.
Forced midlayers are a bad design decision isn't really new at all ...
> > - Once you've upcasted at runtime by checking for ->ops, you can add
> > whatever fancy interfaces you want. Including a real interface to
> > get at whatever underlying id you need to for real buffer sharing
> > between virtio devices.
> >
> > In a way virtio buffer/memory ids are a kind of private bus, entirely
> > distinct from the dma_addr_t bus. So can't really stuff them under this
> > same thing like we e.g. do with pci peer2peer.
>
> As I mentioned earlier, there is no single "dma_addr_t bus". Each
> device (as in struct device) can be on its own different DMA bus, with
> a different DMA address space. There is not even a guarantee that a
> DMA address obtained for one PCI device will be valid for another if
> they are on different buses, which could have different address
> mappings.
>
> Putting that aside, we're thinking about a different approach, as Gerd
> suggested in another thread,
Re: [RFC PATCH] drm/virtio: Export resource handles via DMA-buf API
On Thu, Sep 12, 2019 at 06:41:21PM +0900, Tomasz Figa wrote:
> This patch is an early RFC to judge the direction we are following in
> our virtualization efforts in Chrome OS. The purpose is to start a
> discussion on how to handle buffer sharing between multiple virtio
> devices.
>
> On a side note, we are also working on a virtio video decoder interface
> and implementation, with a V4L2 driver for Linux. Those will be posted
> for review in the near future as well.
>
> Any feedback will be appreciated! Thanks in advance.
>
> ===
>
> With the range of use cases for virtualization expanding, there is going
> to be more virtio devices added to the ecosystem. Devices such as video
> decoders, encoders, cameras, etc. typically work together with the
> display and GPU in a pipeline manner, which can only be implemented
> efficiently by sharing the buffers between producers and consumers.
>
> Existing buffer management framework in Linux, such as the videobuf2
> framework in V4L2, implements all the DMA-buf handling inside generic
> code and do not expose any low level information about the buffers to
> the drivers.
>
> To seamlessly enable buffer sharing with drivers using such frameworks,
> make the virtio-gpu driver expose the resource handle as the DMA address
> of the buffer returned from the DMA-buf mapping operation. Arguably, the
> resource handle is a kind of DMA address already, as it is the buffer
> identifier that the device needs to access the backing memory, which is
> exactly the same role a DMA address provides for native devices.
>
> A virtio driver that does memory management fully on its own would have
> code similar to following. The code is identical to what a regular
> driver for real hardware would do to import a DMA-buf.
>
> static int virtio_foo_get_resource_handle(struct virtio_foo *foo,
> struct dma_buf *dma_buf, u32 *id)
> {
> struct dma_buf_attachment *attach;
> struct sg_table *sgt;
> int ret = 0;
>
> attach = dma_buf_attach(dma_buf, foo->dev);
> if (IS_ERR(attach))
> return PTR_ERR(attach);
>
> sgt = dma_buf_map_attachment(attach, DMA_BIDIRECTIONAL);
> if (IS_ERR(sgt)) {
> ret = PTR_ERR(sgt);
> goto err_detach;
> }
>
> if (sgt->nents != 1) {
> ret = -EINVAL;
> goto err_unmap;
> }
>
> *id = sg_dma_address(sgt->sgl);
I agree with Gerd, this looks pretty horrible to me.
The usual way we've done these kind of special dma-bufs is:
- They all get allocated at the same place, through some library or
whatever.
- You add a dma_buf_is_virtio(dma_buf) function, or maybe something that
also upcasts or returns NULL, which checks for dma_buf->ops.
- Once you've upcasted at runtime by checking for ->ops, you can add
whatever fancy interfaces you want. Including a real interface to
get at whatever underlying id you need to for real buffer sharing
between virtio devices.
In a way virtio buffer/memory ids are a kind of private bus, entirely
distinct from the dma_addr_t bus. So can't really stuff them under this
same thing like we e.g. do with pci peer2peer.
-Daniel
>
> err_unmap:
> dma_buf_unmap_attachment(attach, sgt, DMA_BIDIRECTIONAL);
> err_detach:
> dma_buf_detach(dma_buf, attach);
>
> return ret;
> }
>
> On the other hand, a virtio driver that uses an existing kernel
> framework to manage buffers would not need to explicitly handle anything
> at all, as the framework part responsible for importing DMA-bufs would
> already do the work. For example, a V4L2 driver using the videobuf2
> framework would just call thee vb2_dma_contig_plane_dma_addr() function
> to get what the above open-coded function would return.
>
> Signed-off-by: Tomasz Figa
> ---
> drivers/gpu/drm/virtio/virtgpu_drv.c | 2 +
> drivers/gpu/drm/virtio/virtgpu_drv.h | 4 ++
> drivers/gpu/drm/virtio/virtgpu_prime.c | 81 ++
> 3 files changed, 87 insertions(+)
>
> diff --git a/drivers/gpu/drm/virtio/virtgpu_drv.c
> b/drivers/gpu/drm/virtio/virtgpu_drv.c
> index 0fc32fa0b3c0..ac095f813134 100644
> --- a/drivers/gpu/drm/virtio/virtgpu_drv.c
> +++ b/drivers/gpu/drm/virtio/virtgpu_drv.c
> @@ -210,6 +210,8 @@ static struct drm_driver driver = {
> #endif
> .prime_handle_to_fd = drm_gem_prime_handle_to_fd,
> .prime_fd_to_handle = drm_gem_prime_fd_to_handle,
> + .gem_prime_export = virtgpu_gem_prime_export,
> + .gem_prime_import = virtgpu_gem_prime_import,
> .gem_prime_get_sg_table = virtgpu_gem_prime_get_sg_table,
> .gem_prime_import_sg_table = virtgpu_gem_prime_import_sg_table,
> .gem_prime_vmap = virtgpu_gem_prime_vmap,
> diff --git a/drivers/gpu/drm/virtio/virtgpu_drv.h
> b/drivers/gpu/drm/virtio/virtgpu_drv.h
> index e28829661724..687cfce91885 100644
> --- a/drivers/gpu/drm/virtio/virtgpu_drv.h
> +++
Re: [RFC PATCH] drm/virtio: Export resource handles via DMA-buf API
Hi, > > No. DMA master address space in virtual machines is pretty much the > > same it is on physical machines. So, on x86 without iommu, identical > > to (guest) physical address space. You can't re-define it like that. > > That's not true. Even on x86 without iommu the DMA address space can > be different from the physical address space. On a standard pc (like the ones emulated by qemu) that is the case. It's different on !x86, it also changes with a iommu being present. But that is not the main point here. The point is the dma master address already has a definition and you can't simply change that. > That could be still just > a simple addition/subtraction by constant, but still, the two are > explicitly defined without any guarantees about a simple mapping > between one or another existing. Sure. > "A CPU cannot reference a dma_addr_t directly because there may be > translation between its physical > address space and the DMA address space." Also note that dma address space is device-specific. In case a iommu is present in the system you can have *multiple* dma address spaces, separating (groups of) devices from each other. So passing a dma address from one device to another isn't going to work. > > > However, we could as well introduce a separate DMA address > > > space if resource handles are not the right way to refer to the memory > > > from other virtio devices. > > > > s/other virtio devices/other devices/ > > > > dma-bufs are for buffer sharing between devices, not limited to virtio. > > You can't re-define that in some virtio-specific way. > > > > We don't need to limit this to virtio devices only. In fact I actually > foresee this having a use case with the emulated USB host controller, > which isn't a virtio device. What exactly? > That said, I deliberately referred to virtio to keep the scope of the > problem in control. If there is a solution that could work without > such assumption, I'm more than open to discuss it, of course. But it might lead to taking virtio-specific (or virtualization-specific) shortcuts which will hurt in the long run ... > As per my understanding of the DMA address, anything that lets the DMA > master access the target memory would qualify and there would be no > need for an IOMMU in between. Yes. But that DMA address is already defined by the platform, you can't freely re-define it. Well, you sort-of can when you design your own virtual iommu for qemu. But even then you can't just pass around magic cookies masqueraded as dma address. You have to make sure they actually form an address space, without buffers overlapping, ... > Exactly. The very specific first scenario that we want to start with > is allocating host memory through virtio-gpu and using that memory > both as output of a video decoder and as input (texture) to Virgil3D. > The memory needs to be specifically allocated by the host as only the > host can know the requirements for memory allocation of the video > decode accelerator hardware. So you probably have some virtio-video-decoder. You allocate a gpu buffer, export it as dma-buf, import it into the decoder, then let the video decoder render to it. Right? Using dmabufs makes sense for sure. But we need an separate field in struct dma_buf for an (optional) host dmabuf identifier, we can't just hijack the dma address. cheers, Gerd ___ Virtualization mailing list Virtualization@lists.linux-foundation.org https://lists.linuxfoundation.org/mailman/listinfo/virtualization
Re: [RFC PATCH] drm/virtio: Export resource handles via DMA-buf API
Hi, > > > To seamlessly enable buffer sharing with drivers using such frameworks, > > > make the virtio-gpu driver expose the resource handle as the DMA address > > > of the buffer returned from the DMA-buf mapping operation. Arguably, the > > > resource handle is a kind of DMA address already, as it is the buffer > > > identifier that the device needs to access the backing memory, which is > > > exactly the same role a DMA address provides for native devices. > > First of all, thanks for taking a look at this. > > > No. A scatter list has guest dma addresses, period. Stuffing something > > else into a scatterlist is asking for trouble, things will go seriously > > wrong when someone tries to use such a fake scatterlist as real scatterlist. > > What is a "guest dma address"? The definition of a DMA address in the > Linux DMA API is an address internal to the DMA master address space. > For virtio, the resource handle namespace may be such an address > space. No. DMA master address space in virtual machines is pretty much the same it is on physical machines. So, on x86 without iommu, identical to (guest) physical address space. You can't re-define it like that. > However, we could as well introduce a separate DMA address > space if resource handles are not the right way to refer to the memory > from other virtio devices. s/other virtio devices/other devices/ dma-bufs are for buffer sharing between devices, not limited to virtio. You can't re-define that in some virtio-specific way. > > Also note that "the DMA address of the buffer" is bonkers in virtio-gpu > > context. virtio-gpu resources are not required to be physically > > contigous in memory, so typically you actually need a scatter list to > > describe them. > > There is no such requirement even on a bare metal system, see any > system that has an IOMMU, which is typical on ARM/ARM64. The DMA > address space must be contiguous only from the DMA master point of > view. Yes, the iommu (if present) could remap your scatterlist that way. You can't depend on that though. What is the plan here? Host-side buffer sharing I guess? So you are looking for some way to pass buffer handles from the guest to the host, even in case those buffers are not created by your driver but imported from somewhere else? cheers, Gerd ___ Virtualization mailing list Virtualization@lists.linux-foundation.org https://lists.linuxfoundation.org/mailman/listinfo/virtualization
Re: [RFC PATCH] drm/virtio: Export resource handles via DMA-buf API
Hi, > To seamlessly enable buffer sharing with drivers using such frameworks, > make the virtio-gpu driver expose the resource handle as the DMA address > of the buffer returned from the DMA-buf mapping operation. Arguably, the > resource handle is a kind of DMA address already, as it is the buffer > identifier that the device needs to access the backing memory, which is > exactly the same role a DMA address provides for native devices. No. A scatter list has guest dma addresses, period. Stuffing something else into a scatterlist is asking for trouble, things will go seriously wrong when someone tries to use such a fake scatterlist as real scatterlist. Also note that "the DMA address of the buffer" is bonkers in virtio-gpu context. virtio-gpu resources are not required to be physically contigous in memory, so typically you actually need a scatter list to describe them. cheers, Gerd ___ Virtualization mailing list Virtualization@lists.linux-foundation.org https://lists.linuxfoundation.org/mailman/listinfo/virtualization
