Hello,
This series enables Vulkan Venus context support on virtio-gpu.
All virglrender and almost all Linux kernel prerequisite changes
needed by Venus are already in upstream. For kernel there is a pending
KVM patchset that fixes mapping of compound pages needed for DRM drivers
using TTM [1], othewrwise hostmem blob mapping will fail with a KVM error
from Qemu.
[1] https://lore.kernel.org/kvm/20240229025759.1187910-1-stevensd@google.com/
Example Qemu cmdline that enables Venus for virtio-gpu:
qemu-system-x86_64 -device virtio-vga-gl,hostmem=4G,blob=true,vulkan=true
Changes from V8 to V9
- Added resuming of cmdq processing when hostmem MR is freed,
as was suggested by Akihiko Odaki.
- Added more error messages, suggested by Akihiko Odaki
- Dropped superfluous 'res->async_unmap_completed', suggested
by Akihiko Odaki.
- Kept using cmd->suspended flag. Akihiko Odaki suggested to make
virtio_gpu_virgl_process_cmd() return false if cmd processing is
suspended, but it's not easy to implement due to ubiquitous
VIRTIO_GPU_FILL_CMD() macros that returns void, requiring to change
all the virtio-gpu processing code.
- Added back virtio_gpu_virgl_resource as was requested by Akihiko Odaki,
though I'm not convinced it's really needed.
- Switched to use GArray, renamed capset2_max_ver/size vars and moved
"vulkan" property definition to the virtio-gpu-gl device in the Venus
patch, like was suggested by Akihiko Odaki.
- Moved UUID to virtio_gpu_virgl_resource and dropped UUID save/restore
since it will require bumping VM version and virgl device isn't miratable
anyways.
- Fixed exposing UUID feature with Rutabaga
- Dropped linux-headers update patch because headers were already updated
in Qemu/staging.
- Added patch that updates virtio migration doc with a note about virtio-gpu
migration specifics, suggested by Akihiko Odaki.
- Addressed coding style issue noticed by Akihiko Odaki
Changes from V7 to V8
- Supported suspension of virtio-gpu commands processing and made
unmapping of hostmem region asynchronous by blocking/suspending
cmd processing until region is unmapped. Suggested by Akihiko Odaki.
- Fixed arm64 building of x86 targets using updated linux-headers.
Corrected the update script. Thanks to Rob Clark for reporting
the issue.
- Added new patch that makes registration of virgl capsets dynamic.
Requested by Antonio Caggiano and Pierre-Eric Pelloux-Prayer.
- Venus capset now isn't advertised if Vulkan is disabled with vulkan=false
Changes from V6 to V7
- Used scripts/update-linux-headers.sh to update Qemu headers based
on Linux v6.8-rc3 that adds Venus capset definition to virtio-gpu
protocol, was requested by Peter Maydel
- Added r-bs that were given to v6 patches. Corrected missing s-o-bs
- Dropped context_init Qemu's virtio-gpu device configuration flag,
was suggested by Marc-André Lureau
- Added missing error condition checks spotted by Marc-André Lureau
and Akihiko Odaki, and few more
- Returned back res->mr referencing to memory_region_init_ram_ptr() like
was suggested by Akihiko Odaki. Incorporated fix suggested by Pierre-Eric
to specify the MR name
- Dropped the virgl_gpu_resource wrapper, cleaned up and simplified
patch that adds blob-cmd support
- Fixed improper blob resource removal from resource list on resource_unref
that was spotted by Akihiko Odaki
- Change order of the blob patches, was suggested by Akihiko Odaki.
The cmd_set_scanout_blob support is enabled first
- Factored out patch that adds resource management support to virtio-gpu-gl,
was requested by Marc-André Lureau
- Simplified and improved the UUID support patch, dropped the hash table
as we don't need it for now. Moved QemuUUID to virtio_gpu_simple_resource.
This all was suggested by Akihiko Odaki and Marc-André Lureau
- Dropped console_has_gl() check, suggested by Akihiko Odaki
- Reworked Meson cheking of libvirglrender features, made new features
available based on virglrender pkgconfig version instead of checking
symbols in header. This should fix build error using older virglrender
version, reported by Alex Bennée
- Made enabling of Venus context configrable via new virtio-gpu device
"vulkan=true" flag, suggested by Marc-André Lureau. The flag is disabled
by default because it requires blob and hostmem options to be enabled
and configured
Changes from V5 to V6
- Move macros configurations under virgl.found() and rename
HAVE_VIRGL_CONTEXT_CREATE_WITH_FLAGS.
- Handle the case while context_init is disabled.
- Enable context_init by default.
- Move virtio_gpu_virgl_resource_unmap() into
virgl_cmd_resource_unmap_blob().
- Introduce new struct virgl_gpu_resource to store virgl specific members.
- Remove erro handling of g_new0, because glib will abort() on OOM.
- Set resource uuid as option.
- Implement optional subsection of vmstate_virtio_gpu_resource_uuid_state
for virtio live migration.
- Use g_int_hash/g_int_equal instead of the default
- Add scanout_blob function for virtio-gpu-virgl
- Resolve the memory leak on virtio-gpu-virgl
- Remove the unstable API flags check because virglrenderer is already 1.0
- Squash the render server flag support into "Initialize Venus"
Changes from V4 (virtio gpu V4) to V5
- Inverted patch 5 and 6 because we should configure
HAVE_VIRGL_CONTEXT_INIT firstly.
- Validate owner of memory region to avoid slowing down DMA.
- Use memory_region_init_ram_ptr() instead of
memory_region_init_ram_device_ptr().
- Adjust sequence to allocate gpu resource before virglrender resource
creation
- Add virtio migration handling for uuid.
- Send kernel patch to define VIRTIO_GPU_CAPSET_VENUS.
https://lore.kernel.org/lkml/20230915105918.3763061-1-ray.huang@amd.com/
- Add meson check to make sure unstable APIs defined from 0.9.0.
Changes from V1 to V2 (virtio gpu V4)
- Remove unused #include "hw/virtio/virtio-iommu.h"
- Add a local function, called virgl_resource_destroy(), that is used
to release a vgpu resource on error paths and in resource_unref.
- Remove virtio_gpu_virgl_resource_unmap from
virtio_gpu_cleanup_mapping(),
since this function won't be called on blob resources and also because
blob resources are unmapped via virgl_cmd_resource_unmap_blob().
- In virgl_cmd_resource_create_blob(), do proper cleanup in error paths
and move QTAILQ_INSERT_HEAD(&g->reslist, res, next) after the resource
has been fully initialized.
- Memory region has a different life-cycle from virtio gpu resources
i.e. cannot be released synchronously along with the vgpu resource.
So, here the field "region" was changed to a pointer and is allocated
dynamically when the blob is mapped.
Also, since the pointer can be used to indicate whether the blob
is mapped, the explicite field "mapped" was removed.
- In virgl_cmd_resource_map_blob(), add check on the value of
res->region, to prevent beeing called twice on the same resource.
- Add a patch to enable automatic deallocation of memory regions to resolve
use-after-free memory corruption with a reference.
Antonio Caggiano (3):
virtio-gpu: Handle resource blob commands
virtio-gpu: Resource UUID
virtio-gpu: Support Venus context
Dmitry Osipenko (4):
virtio-gpu: Use pkgconfig version to decide which virgl features are
available
virtio-gpu: Don't require udmabuf when blobs and virgl are enabled
virtio-gpu: Support suspension of commands processing
migration/virtio: Add virtio-gpu section
Huang Rui (2):
virtio-gpu: Support context-init feature with virglrenderer
virtio-gpu: Add virgl resource management
Pierre-Eric Pelloux-Prayer (1):
virtio-gpu: Register capsets dynamically
Robert Beckett (1):
virtio-gpu: Support blob scanout using dmabuf fd
docs/devel/migration/virtio.rst | 7 +
hw/display/trace-events | 1 +
hw/display/virtio-gpu-base.c | 3 +
hw/display/virtio-gpu-gl.c | 7 +
hw/display/virtio-gpu-rutabaga.c | 1 +
hw/display/virtio-gpu-virgl.c | 561 ++++++++++++++++++++++++++++++-
hw/display/virtio-gpu.c | 35 +-
include/hw/virtio/virtio-gpu.h | 13 +
meson.build | 10 +-
9 files changed, 610 insertions(+), 28 deletions(-)
--
2.44.0