Moving a GpuVaAlloc or GpuVmBo between threads currently forces drivers
to write their own unsafe Send and Sync impls. Provide the markers in
the abstraction instead.
GpuVaAlloc wraps only uninitialised memory and exposes none of it.
GpuVmBo hands out the driver data and GEM object by shared reference and
drops them in its deferred put; the DriverGpuVm trait already guarantees
both are Send + Sync, so both impls are unconditional.
Signed-off-by: Sami Tolvanen <samitolvanen@google.com>
Link: https://patch.msgid.link/20260611-gpuvm-sync-send-v4-2-6c7f4ab2778a@google.com
Signed-off-by: Alice Ryhl <aliceryhl@google.com>
DriverGpuVm permitted !Send/!Sync associated data on an abstraction whose
handles are shared and dropped across threads: obtain() runs from many
threads and the VA API performs deferred cross-thread drops. That is
unsound.
Require Send + Sync on the trait and its associated data so the GpuVm and
UniqueRefGpuVm handle impls need no per-impl bounds.
Fixes: 82b78182ea ("rust: drm: add base GPUVM immediate mode abstraction")
Signed-off-by: Sami Tolvanen <samitolvanen@google.com>
Link: https://patch.msgid.link/20260611-gpuvm-sync-send-v4-1-6c7f4ab2778a@google.com
Signed-off-by: Alice Ryhl <aliceryhl@google.com>
If a DriverGpuVm implementation is lifetime-parameterized, its
SmContext<'ctx> type may depend on lifetimes carried by that driver
implementation. In this case, SmContext<'ctx> is only valid if the
driver implementation outlives 'ctx.
Add a Self: 'ctx bound to DriverGpuVm::SmContext<'ctx> to express that
requirement. Then propagate the corresponding T: 'ctx bound to the GPUVM
state machine helper types that store T::SmContext<'ctx>.
This allows drivers to provide lifetime-parameterized implementations of
DriverGpuVm.
Signed-off-by: Boris Brezillon <boris.brezillon@collabora.com>
Signed-off-by: Deborah Brouwer <deborah.brouwer@collabora.com>
Reviewed-by: Alice Ryhl <aliceryhl@google.com>
Link: https://patch.msgid.link/20260610-gpuvm_smcontext_lifetime_bound_v1-v1-1-531e7d2ee7b4@collabora.com
Signed-off-by: Danilo Krummrich <dakr@kernel.org>
Since the introduction of DeviceContext, there is no longer a single
driver object type to equate with the GPUVM object type.
Instead of threading DeviceContext through GPUVM, remove the strict
identity between DriverGpuVm::Object and drm::Driver::Object and
instead tighten the requirement that the DriverGpuVm::Object be an
allocatable GEM object associated with the same DRM driver.
Also, make GpuVm::new() generic over DeviceContext so it can accept a
drm::Device<T::Driver, Ctx>.
Fixes: 0023a1e8d0 ("rust/drm/gem: Use DeviceContext with GEM objects")
Signed-off-by: Deborah Brouwer <deborah.brouwer@collabora.com>
Reviewed-by: Alice Ryhl <aliceryhl@google.com>
Reviewed-by: Sami Tolvanen <samitolvanen@google.com>
Link: https://patch.msgid.link/20260610-gpuvm_device_context_v1-v1-1-01a890b17448@collabora.com
Signed-off-by: Danilo Krummrich <dakr@kernel.org>
This provides a mechanism to create (or look up) VMBO instances, which
represent the mapping between GPUVM and GEM objects.
The GpuVmBoRegistered<T> type can be considered like ARef<GpuVm<T>>,
except that no way to increment the refcount is provided.
The GpuVmBoAlloc<T> type is more akin to a pre-allocated GpuVmBo<T>, so
it's not really a GpuVmBo<T> yet. Its destructor could call
drm_gpuvm_bo_destroy_not_in_lists(), but as the type is currently
private and never called anywhere, this perf optimization does not need
to happen now.
Pre-allocating and obtaining the gpuvm_bo object is exposed as a single
step. This could theoretically be a problem if one wanted to call
drm_gpuvm_bo_obtain_prealloc() during the fence signalling critical
path, but that's not a possibility because:
1. Adding the BO to the extobj list requires the resv lock, so it cannot
happen during the fence signalling critical path.
2. obtain() requires that the BO is not in the extobj list, so obtain()
must be called before adding the BO to the extobj list.
Thus, drm_gpuvm_bo_obtain_prealloc() cannot be called during the fence
signalling critical path. (For extobjs at least.)
Reviewed-by: Daniel Almeida <daniel.almeida@collabora.com>
Signed-off-by: Alice Ryhl <aliceryhl@google.com>
Link: https://patch.msgid.link/20260409-gpuvm-rust-v6-2-b16e6ada7261@google.com
Signed-off-by: Danilo Krummrich <dakr@kernel.org>
Add a GPUVM abstraction to be used by Rust GPU drivers.
GPUVM keeps track of a GPU's virtual address (VA) space and manages the
corresponding virtual mappings represented by "GPU VA" objects. It also
keeps track of the gem::Object<T> used to back the mappings through
GpuVmBo<T>.
This abstraction is only usable by drivers that wish to use GPUVM in
immediate mode. This allows us to build the locking scheme into the API
design. It means that the GEM mutex is used for the GEM gpuva list, and
that the resv lock is used for the extobj list. The evicted list is not
yet used in this version.
This abstraction provides a special handle called the UniqueRefGpuVm,
which is a wrapper around ARef<GpuVm> that provides access to the
interval tree. Generally, all changes to the address space requires
mutable access to this unique handle.
Signed-off-by: Asahi Lina <lina+kernel@asahilina.net>
Co-developed-by: Daniel Almeida <daniel.almeida@collabora.com>
Signed-off-by: Daniel Almeida <daniel.almeida@collabora.com>
Reviewed-by: Daniel Almeida <daniel.almeida@collabora.com>
Co-developed-by: Alice Ryhl <aliceryhl@google.com>
Signed-off-by: Alice Ryhl <aliceryhl@google.com>
Link: https://patch.msgid.link/20260409-gpuvm-rust-v6-1-b16e6ada7261@google.com
Signed-off-by: Danilo Krummrich <dakr@kernel.org>
