The new type struct dma_buf_map represents a mapping of dma-buf memory
into kernel space. It contains a flag, is_iomem, that signals users to
access the mapped memory with I/O operations instead of regular loads
and stores.
It was assumed that DMA buffer memory can be accessed with regular load
and store operations. Some architectures, such as sparc64, require the
use of I/O operations to access dma-map buffers that are located in I/O
memory. Providing struct dma_buf_map allows drivers to implement this.
This was specifically a problem when refreshing the graphics framebuffer
on such systems. [1]
As the first step, struct dma_buf stores an instance of struct dma_buf_map
internally. Afterwards, dma-buf's vmap and vunmap interfaces are be
converted. Finally, affected drivers can be fixed.
v3:
* moved documentation into separate patch
* test for NULL pointers with !<ptr>
[1] https://lore.kernel.org/dri-devel/20200725191012.GA434957@ravnborg.org/
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Acked-by: Sumit Semwal <sumit.semwal@linaro.org>
Acked-by: Sam Ravnborg <sam@ravnborg.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20200925115601.23955-2-tzimmermann@suse.de
The stride field has never been used, so repurpose it to be
"blob_mem". This way, userspace can know the memory properties
of the blob if it's passed between userspace processes and
no suitable userspace API exists to transmit that knowledge.
Signed-off-by: Gurchetan Singh <gurchetansingh@chromium.org>
Acked-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20200924003214.662-17-gurchetansingh@chromium.org
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
The old transfer ioctls may work on blob resources, and there is no
TRANSFER_BLOB hypercall now for simplicity.
The guest may have a image view on the blob resources such that the
stride is not equal to width * bytes_per_pixel.
For host-only blobs, we can repurpose the transfer ioctls to synchronize
caches as well. For guest-only blobs, these operations are undefined
for now so leave them out.
Also, with seamless Wayland integration between guest/host looking
increasingly attractive, it also makes sense to keep track of
one value for stride.
Signed-off-by: Gurchetan Singh <gurchetansingh@chromium.org>
Acked-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20200924003214.662-16-gurchetansingh@chromium.org
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
For upcoming blob resources, userspace can specify that the
resource will be used for cross-device sharing. This is mainly
for exportable blobs that will only shared with the virtgpu
display but not across devices.
Signed-off-by: Gurchetan Singh <gurchetansingh@chromium.org>
Acked-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20200924003214.662-15-gurchetansingh@chromium.org
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
SCANOUT_BLOB forwards the DRM framebuffer metadata to the host. The
modifier is intentionally left out -- it may be possible to query
the host for that.
We also assume one blob resource per DRM framebuffer. That too is
an intentional simplification.
Signed-off-by: Gurchetan Singh <gurchetansingh@chromium.org>
Acked-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20200924003214.662-14-gurchetansingh@chromium.org
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
A virtio-gpu vram object is based on range-based allocation.
No guest shmemfs backing, so we call drm_gem_private_object_init.
This is for host memory without any guest backing (atleast initially).
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
Acked-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20200924003214.662-12-gurchetansingh@chromium.org
Co-developed-by: Gurchetan Singh <gurchetansingh@chromium.org>
Signed-off-by: Gurchetan Singh <gurchetansingh@chromium.org>
The availability of the host visible region means host 3D
allocations can be directly mapped in the guest.
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
Acked-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20200924003214.662-9-gurchetansingh@chromium.org
Co-developed-by: Gurchetan Singh <gurchetansingh@chromium.org>
Signed-off-by: Gurchetan Singh <gurchetansingh@chromium.org>
This feature was recently added to virtio-gpu, lets make
it userspace queryable. It's an error to use
BLOB_FLAG_USE_CROSS_DEVICE when this feature is not present.
Signed-off-by: Gurchetan Singh <gurchetansingh@chromium.org>
Acked-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20200924003214.662-7-gurchetansingh@chromium.org
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
This exposes the host visible feature to userspace. Without it,
it is an error to specify BLOB_MEM_HOST3D with
BLOG_FLAG_USE_MAPPABLE.
Signed-off-by: Gurchetan Singh <gurchetansingh@chromium.org>
Acked-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Acked-by: Lingfeng Yang <lfy@google.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20200924003214.662-6-gurchetansingh@chromium.org
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
This makes blob resources available to guest userspace. They are needed
for GL4.5, Vulkan and zero-copy virtio-gpu.
For Mesa, blob resources have been tested with Piglit's ARB_buffer_storage
tests and apitraces. Apitraces of GL4.5 games show we're between 70%
to 80% of host performance on Iris, based on a apitrace of a 2013 GL4.5
game:
11.204 FPS (guest)
15.947 FPS (host)
This is still better than the status quo, when said game was unplayable
with Virgl due to an inefficient GL4.3 fallback. But there's still room
for improvement if we want to match HW-assisted virtualization.
For Vulkan, blob resources have been tested with dEQP.vk.memory* and
running Vulkan applications in production with the "Cuttlefish" virtual
Android device. This has been done with Lingfeng Yang's "gfxstream"
Vulkan implementation, which virtualizes Vulkan across many Google
products.
Signed-off-by: Gurchetan Singh <gurchetansingh@chromium.org>
Acked-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Acked-by: Chia-I Wu <olvaffe@gmail.com>
Acked-by: Lingfeng Yang <lfy@google.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20200924003214.662-5-gurchetansingh@chromium.org
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
This patch adds a new virtgpu feature that allows directly
mapping host allocated resources.
This is based on virtio shared memory regions, which allows
querying for memory regions using PCI transport. Each shared
memory region has an associated "shmid", the meaning of which
is device specific.
For virtio-gpu, we can define the shared memory region with id
VIRTIO_GPU_SHM_ID_HOST_VISIBLE to be the "host visible memory
region".
The presence of the host visible memory region means the following
hypercalls are supported:
1) VIRTIO_GPU_CMD_RESOURCE_MAP_BLOB
This hypercall tells the host to inject the host resource's
mapping in an offset into virtio-gpu's PCI address space.
This is typically done via KVM_SET_USER_MEMORY_REGION on Linux
hosts.
On success, VIRTIO_GPU_RESP_OK_MAP_INFO is returned, which
specifies the host buffer's caching type and possibly in the
future performance hints about the buffer..
2) VIRTIO_GPU_CMD_RESOURCE_UNMAP_BLOB
This hypercall tells the host to remove the host resource's
mapping from the guest VM.
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
Acked-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Acked-by: Lingfeng Yang <lfy@google.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20200924003214.662-4-gurchetansingh@chromium.org
Co-developed-by: Gurchetan Singh <gurchetansingh@chromium.org>
Signed-off-by: Gurchetan Singh <gurchetansingh@chromium.org>
A blob resource is a container for:
- VIRTIO_GPU_BLOB_MEM_GUEST: a guest memory allocation
(referred to as a "guest-only blob resource")
- VIRTIO_GPU_BLOB_MEM_HOST3D: a host3d memory allocation
(referred to as a "host-only blob resource")
- VIRTIO_GPU_BLOB_MEM_HOST3D_GUEST: a guest + host3d memory allocation
(referred to as a "default blob resource").
The memory properties of the blob resource must be described by
`blob_mem`.
For default and guest only blob resources set, `nents` guest system
pages are assigned to the resource. For default blob resources,
these guest pages are used for transfer operations. Attach/detach is
also possible to allow swap-in/swap-out, but isn't required since it
may not be applicable to future blob mem types
(shared guest/guest vram).
Host allocations depend on whether the 3D is supported. If 3D is not
supported, the only valid field for `blob_mem` is
VIRTIO_GPU_BLOB_MEM_GUEST.
If 3D is supported, the virtio-gpu resource is created from the
context local object identified by the `blob_id`. The actual host
allocation done by the CMD_SUBMIT_3D.
Userspace must specify if the blob resource is intended to be used
for userspace mapping, sharing between virtio-gpu contexts and/or
sharing between virtio devices. This is done via `blob_flags`.
For 3D hosts, both VIRTIO_GPU_CMD_TRANSFER_TO_HOST_3D and
VIRTIO_GPU_CMD_TRANSFER_FROM_HOST_3D may be used to update
the host resource. There is no restriction on the image/buffer
view the guest/host userspace has on the blob resource.
VIRTIO_GPU_CMD_SET_SCANOUT_BLOB / VIRTIO_GPU_CMD_RESOURCE_FLUSH may
be used with blob resources as well. The modifier is intentionally
left out of SCANOUT_BLOB, and auxilary blobs are also left out
as a simplification.
The use case for blob resources is zero-copy, needed for coherent
memory in virglrenderer. Host only blob resources are not mappable
without the feature described in the next patch, but are shareable.
Future work:
- Emulated coherent `blob_mem` type for QEMU/vhost-user
- A `blob_mem` type for guest-only resources imported in
cache-coherent FOSS GPU/display drivers.
- Display integration involving the blob model using seamless
Wayland windows.
Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
Acked-by: Tomeu Vizoso <tomeu.vizoso@collabora.com>
Acked-by: Chia-I Wu <olvaffe@gmail.com>
Acked-by: Lingfeng Yang <lfy@google.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20200924003214.662-3-gurchetansingh@chromium.org
Co-developed-by: Gurchetan Singh <gurchetansingh@chromium.org>
Signed-off-by: Gurchetan Singh <gurchetansingh@chromium.org>
We already implemented the fault handler ourself,
just open code what is necessary here.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Link: https://patchwork.freedesktop.org/patch/392323/
Implement the fault handler ourself using the provided TTM functions.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Link: https://patchwork.freedesktop.org/patch/392324/
We already implemented the fault handler ourself,
just open code what is necessary here.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Link: https://patchwork.freedesktop.org/patch/392322/
Just check earlier if a BO can be page faulted in the first place.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Link: https://patchwork.freedesktop.org/patch/392321/
Starting from the JZ4725B SoC, the primary and overlay planes support
24-bit pixel modes (8 bits per color component, without dummy byte).
Add support for these in the ingenic-drm driver.
Signed-off-by: Paul Cercueil <paul@crapouillou.net>
Reviewed-by: Sam Ravnborg <sam@ravnborg.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20200926170501.1109197-8-paul@crapouillou.net
Until now the ingenic-drm driver supported the same pixel formats on the
F0 and F1 planes, and across all SoCs. However, the F0 plane does support
paletted 8bpp, while the F1 plane doesn't.
Furthermore, the three SoCs currently supported all have different pixel
formats available; 24bpp was added in JZ4725B, 30bpp was added in
JZ4770.
Prepare the inclusion of paletted 8bpp, 24bpp and 30bpp support by
having separate pixel format lists for F0 and F1 planes.
Signed-off-by: Paul Cercueil <paul@crapouillou.net>
Reviewed-by: Sam Ravnborg <sam@ravnborg.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20200926170501.1109197-5-paul@crapouillou.net
Instead of calling dmam_alloc_coherent() once for each 4-bit DMA
hardware descriptor, we can have them both in a physical memory page, as
long as they are aligned to 16 bytes. This reduces memory consumption,
and will make it easier to add more DMA descriptors in the future.
Note that the old code would not create the F0 descriptor on SoCs that
don't support multiple planes. We don't care, because:
- we don't use more memory by allocating two descriptors instead of a
single one;
- the only SoC that does not support multiple planes (JZ4740) still has
two independent DMA channels, for an unknown reason.
Signed-off-by: Paul Cercueil <paul@crapouillou.net>
Reviewed-by: Sam Ravnborg <sam@ravnborg.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20200926170501.1109197-4-paul@crapouillou.net
Add support for static memory reserved from Device Tree. Since we're
using GEM buffers backed by CMA, it is interesting to have an option to
specify the CMA area where the GEM buffers will be allocated.
v2: Don't abort probe if reserved memory cannot be obtained. The driver
will still work fine provided the kernel configuration is sane.
Signed-off-by: Paul Cercueil <paul@crapouillou.net>
Reviewed-by: Sam Ravnborg <sam@ravnborg.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20200926170501.1109197-3-paul@crapouillou.net
Old Ingenic SoCs can overclock very well, up to +50% of their nominal
clock rate, whithout requiring overvolting or anything like that, just
by changing the rate of the main PLL. Unfortunately, all clocks on the
system are derived from that PLL, and when the PLL rate is updated, so
is our pixel clock.
To counter that issue, we make sure that the panel is in VBLANK before
the rate change happens, and we will then re-set the pixel clock rate
afterwards, once the PLL has been changed, to be as close as possible to
the pixel rate requested by the encoder.
v2: Add comment about mutex usage
Signed-off-by: Paul Cercueil <paul@crapouillou.net>
Reviewed-by: Sam Ravnborg <sam@ravnborg.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20200926170501.1109197-2-paul@crapouillou.net
This is one of the last users of compat_alloc_user_space()
and copy_in_user(). The actual handler is implemented in the
same file and could be shared, but as I couldn't test this
properly I leave the native case alone and just make a straight
copy of it for the compat case, with a minimum set of
modifications.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Sam Ravnborg <sam@ravnborg.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20200918100926.1447563-3-arnd@arndb.de
No driver implements FBIOSCURSOR, so this function has no purpose
and can be removed. Apparently it was added in linux-2.1.44 to handle
compatibility for drivers/sbus/char/sunfb.c but lost its purpose when
that driver got rewritten in linux-2.5.63.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Sam Ravnborg <sam@ravnborg.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20200918100926.1447563-2-arnd@arndb.de
This function is one of the remaining users of compat_alloc_user_space()
and copy_in_user().
Clean it up by copying to a local data structure copy instead,
which also leads to more readable code.
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Sam Ravnborg <sam@ravnborg.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20200918100926.1447563-1-arnd@arndb.de
Most drivers that use VRAM helpers have only a few MiB of framebuffer
memory available. To reduce fragmentation, new BOs are now put into
system memory by default. Only pin operations are allowed to move BOs
into VRAM.
v2:
* rebased onto latest drm-tip
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20200922145700.25413-4-tzimmermann@suse.de
Don't override the GEM object functions unconditionally. If the driver
sets the GEM functions, VRAM helpers will not set them. The idea has been
taken from SHMEM helpers.
v2:
* updated the commit message (Sam)
* document the new feature for drm_gem_vram_create()
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20200922145700.25413-3-tzimmermann@suse.de
The drm_gem_vram_create() function is the only caller of the internal
helper drm_gem_vram_init(). Streamline the code by putting all code into
the create function.
v2:
* rebased onto latest drm-tip
* fixed typo in commit message (Sam)
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Acked-by: Sam Ravnborg <sam@ravnborg.org>
Link: https://patchwork.freedesktop.org/patch/msgid/20200922145700.25413-2-tzimmermann@suse.de
Several GEM and PRIME callbacks have been deprecated in favor of
per-instance GEM object functions. Remove the callbacks as they are
now unused. The only exception is .gem_prime_mmap, which is still
in use by several drivers.
What is also gone is gem_vm_ops in struct drm_driver. All drivers now
use struct drm_gem_object_funcs.vm_ops instead.
While at it, the patch also improves error handling around calls
to .free and .get_sg_table callbacks.
v3:
* restore default call to drm_gem_prime_export() in
drm_gem_prime_handle_to_fd()
* return -ENOSYS if get_sg_table is not set
* drop all checks for obj->funcs
* clean up TODO list and documentation
v2:
* update related TODO item (Sam)
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Acked-by: Christian König <christian.koenig@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200923102159.24084-23-tzimmermann@suse.de
The xlnx driver uses CMA helpers with default callback functions.
Initialize the driver structure with the rsp CMA helper macro. The
driver is being converted to use GEM object functions as part of
this change.
Two callbacks, .dumb_destroy and .gem_prime_import, were initialized
to their default implementations, so they are just kept empty now.
v2:
* initialize with DRM_GEM_CMA_DRIVER_OPS_WITH_DUMB_CREATE (Laurent)
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Hyun Kwon <hyun.kwon@xilinx.com>
Acked-by: Christian König <christian.koenig@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200923102159.24084-22-tzimmermann@suse.de
GEM object functions deprecate several similar callback interfaces in
struct drm_driver. This patch replaces the per-driver callbacks with
per-instance callbacks in xen. The only exception is gem_prime_mmap,
which is non-trivial to convert.
v2:
* convert xen_drm_drv_free_object_unlocked() to static
callback (Oleksandr)
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Acked-by: Oleksandr Andrushchenko <oleksandr_andrushchenko@epam.com>
Acked-by: Christian König <christian.koenig@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200923102159.24084-21-tzimmermann@suse.de
GEM object functions deprecate several similar callback interfaces in
struct drm_driver. This patch replaces the per-driver callbacks with
per-instance callbacks in vkms.
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Melissa Wen <melissa.srw@gmail.com>
Acked-by: Christian König <christian.koenig@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200923102159.24084-20-tzimmermann@suse.de
GEM object functions deprecate several similar callback interfaces in
struct drm_driver. This patch replaces virtgpu's per-driver PRIME export
function with a per-object function.
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Acked-by: Christian König <christian.koenig@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200923102159.24084-19-tzimmermann@suse.de
GEM object functions deprecate several similar callback interfaces in
struct drm_driver. This patch replaces the per-driver callbacks with
per-instance callbacks in vgem. The only exception is gem_prime_mmap,
which is non-trivial to convert.
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Melissa Wen <melissa.srw@gmail.com>
Acked-by: Christian König <christian.koenig@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200923102159.24084-18-tzimmermann@suse.de
GEM object functions deprecate several similar callback interfaces in
struct drm_driver. This patch replaces the per-driver callbacks with
per-instance callbacks in vc4. The only exception is gem_prime_mmap,
which is non-trivial to convert.
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Eric Anholt <eric@anholt.net>
Acked-by: Christian König <christian.koenig@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200923102159.24084-17-tzimmermann@suse.de
GEM object functions deprecate several similar callback interfaces in
struct drm_driver. This patch replaces the per-driver callbacks with
per-instance callbacks in tegra.
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Acked-by: Thierry Reding <treding@nvidia.com>
Acked-by: Christian König <christian.koenig@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200923102159.24084-16-tzimmermann@suse.de
GEM object functions deprecate several similar callback interfaces in
struct drm_driver. This patch replaces the per-driver callbacks with
per-instance callbacks in rockchip. The only exception is gem_prime_mmap,
which is non-trivial to convert.
v3:
* update documentation
Signed-off-by: Thomas Zimmermann <tzimmermann@suse.de>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Acked-by: Christian König <christian.koenig@amd.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20200923102159.24084-15-tzimmermann@suse.de
