diff options
Diffstat (limited to 'Documentation/gpu')
| -rw-r--r-- | Documentation/gpu/drm-mm.rst | 29 | ||||
| -rw-r--r-- | Documentation/gpu/todo.rst | 45 |
2 files changed, 68 insertions, 6 deletions
diff --git a/Documentation/gpu/drm-mm.rst b/Documentation/gpu/drm-mm.rst index d55751cad67c..f22433470c76 100644 --- a/Documentation/gpu/drm-mm.rst +++ b/Documentation/gpu/drm-mm.rst @@ -155,7 +155,12 @@ drm_gem_object_init() will create an shmfs file of the requested size and store it into the struct :c:type:`struct drm_gem_object <drm_gem_object>` filp field. The memory is used as either main storage for the object when the graphics hardware -uses system memory directly or as a backing store otherwise. +uses system memory directly or as a backing store otherwise. Drivers +can call drm_gem_huge_mnt_create() to create, mount and use a huge +shmem mountpoint instead of the default one ('shm_mnt'). For builds +with CONFIG_TRANSPARENT_HUGEPAGE enabled, further calls to +drm_gem_object_init() will let shmem allocate huge pages when +possible. Drivers are responsible for the actual physical pages allocation by calling shmem_read_mapping_page_gfp() for each page. @@ -290,15 +295,27 @@ The open and close operations must update the GEM object reference count. Drivers can use the drm_gem_vm_open() and drm_gem_vm_close() helper functions directly as open and close handlers. -The fault operation handler is responsible for mapping individual pages -to userspace when a page fault occurs. Depending on the memory -allocation scheme, drivers can allocate pages at fault time, or can -decide to allocate memory for the GEM object at the time the object is -created. +The fault operation handler is responsible for mapping pages to +userspace when a page fault occurs. Depending on the memory allocation +scheme, drivers can allocate pages at fault time, or can decide to +allocate memory for the GEM object at the time the object is created. Drivers that want to map the GEM object upfront instead of handling page faults can implement their own mmap file operation handler. +In order to reduce page table overhead, if the internal shmem mountpoint +"shm_mnt" is configured to use transparent huge pages (for builds with +CONFIG_TRANSPARENT_HUGEPAGE enabled) and if the shmem backing store +managed to allocate a huge page for a faulty address, the fault handler +will first attempt to insert that huge page into the VMA before falling +back to individual page insertion. mmap() user address alignment for GEM +objects is handled by providing a custom get_unmapped_area file +operation which forwards to the shmem backing store. For most drivers, +which don't create a huge mountpoint by default or through a module +parameter, transparent huge pages can be enabled by either setting the +"transparent_hugepage_shmem" kernel parameter or the +"/sys/kernel/mm/transparent_hugepage/shmem_enabled" sysfs knob. + For platforms without MMU the GEM core provides a helper method drm_gem_dma_get_unmapped_area(). The mmap() routines will call this to get a proposed address for the mapping. diff --git a/Documentation/gpu/todo.rst b/Documentation/gpu/todo.rst index 9013ced318cb..22487ac1b011 100644 --- a/Documentation/gpu/todo.rst +++ b/Documentation/gpu/todo.rst @@ -878,6 +878,51 @@ Contact: Christian König Level: Starter +DRM GPU Scheduler +================= + +Provide a universal successor for drm_sched_resubmit_jobs() +----------------------------------------------------------- + +drm_sched_resubmit_jobs() is deprecated. Main reason being that it leads to +reinitializing dma_fences. See that function's docu for details. The better +approach for valid resubmissions by amdgpu and Xe is (apparently) to figure out +which job (and, through association: which entity) caused the hang. Then, the +job's buffer data, together with all other jobs' buffer data currently in the +same hardware ring, must be invalidated. This can for example be done by +overwriting it. amdgpu currently determines which jobs are in the ring and need +to be overwritten by keeping copies of the job. Xe obtains that information by +directly accessing drm_sched's pending_list. + +Tasks: + +1. implement scheduler functionality through which the driver can obtain the + information which *broken* jobs are currently in the hardware ring. +2. Such infrastructure would then typically be used in + drm_sched_backend_ops.timedout_job(). Document that. +3. Port a driver as first user. +4. Document the new alternative in the docu of deprecated + drm_sched_resubmit_jobs(). + +Contact: Christian König <christian.koenig@amd.com> + Philipp Stanner <phasta@kernel.org> + +Level: Advanced + +Add locking for runqueues +------------------------- + +There is an old FIXME by Sima in include/drm/gpu_scheduler.h. It details that +struct drm_sched_rq is read at many places without any locks, not even with a +READ_ONCE. At XDC 2025 no one could really tell why that is the case, whether +locks are needed and whether they could be added. (But for real, that should +probably be locked!). Check whether it's possible to add locks everywhere, and +do so if yes. + +Contact: Philipp Stanner <phasta@kernel.org> + +Level: Intermediate + Outside DRM =========== |