pvr_gem_object_create() clears every new object through a CPU mapping so that userspace never sees data left by a previous owner of the pages [1]. shmem already guarantees that: it clears each page it allocates, and the dma_sync_sgtable_for_device() call that follows makes those zeroes visible to the GPU. The second pass, write-combined on non-coherent devices, only adds latency.
Link: https://lore.kernel.org/[email protected] [1] Signed-off-by: Guillaume La Roque <[email protected]> --- pvr_gem_object_create() clears every new object through a CPU mapping, on top of the zeroing shmem already does. Dropping that pass speeds up object creation, most visibly where it goes through a write-combined mapping on a non-coherent platform. Testing: drm-misc-next (41ecf852c073) with a Debian rootfs, on a TI J722S EVM (BXS-4-64, arm64) and a SpacemiT K3 Pico-ITX (BXM-4-64, riscv64). The K3 also needs local patches that are not upstream yet: its GPU DT node, an RPMI power domain fix for its firmware and BVNC 36.56.104.183. Each board ran two small test programs, once without and once with the patch: - The first one only uses the powervr DRM ioctls. It fills 250 objects with a pattern through their CPU mapping, frees them, creates new objects of the same size and checks they read back as zero. It also times DRM_IOCTL_PVR_CREATE_BO. Result, without -> with the patch: J722S: 0 -> 0 non-zero bytes, 16 MiB object in 20.6 -> 17.4 ms K3: 0 -> 0 non-zero bytes, 16 MiB object in 41.6 -> 9.2 ms - The second one does the same check through Mesa's Vulkan driver, and also reads the new memory from the GPU: the GPU copies it into a buffer that the CPU then checks. 80 allocations, on the J722S only, as Mesa is not built on the K3 yet. Result: 0 non-zero bytes without and with the patch. The test programs were written and run with the help of an LLM. --- drivers/gpu/drm/imagination/pvr_gem.c | 34 ---------------------------------- 1 file changed, 34 deletions(-) diff --git a/drivers/gpu/drm/imagination/pvr_gem.c b/drivers/gpu/drm/imagination/pvr_gem.c index d8660d6a8e01..01b8e08d12f0 100644 --- a/drivers/gpu/drm/imagination/pvr_gem.c +++ b/drivers/gpu/drm/imagination/pvr_gem.c @@ -275,34 +275,6 @@ pvr_gem_object_vunmap(struct pvr_gem_object *pvr_obj) dma_resv_unlock(obj->resv); } -/** - * pvr_gem_object_zero() - Zeroes the physical memory behind an object. - * @pvr_obj: Target PowerVR GEM object. - * - * Return: - * * 0 on success, or - * * Any error encountered while attempting to map @pvr_obj to the CPU (see - * pvr_gem_object_vmap()). - */ -static int -pvr_gem_object_zero(struct pvr_gem_object *pvr_obj) -{ - void *cpu_ptr; - - cpu_ptr = pvr_gem_object_vmap(pvr_obj); - if (IS_ERR(cpu_ptr)) - return PTR_ERR(cpu_ptr); - - memset(cpu_ptr, 0, pvr_gem_object_size(pvr_obj)); - - /* Make sure the zero-ing is done before vumap-ing the object. */ - wmb(); - - pvr_gem_object_vunmap(pvr_obj); - - return 0; -} - /** * pvr_gem_create_object() - Allocate and pre-initializes a pvr_gem_object * @drm_dev: DRM device creating this object. @@ -378,12 +350,6 @@ pvr_gem_object_create(struct pvr_device *pvr_dev, size_t size, u64 flags) dma_sync_sgtable_for_device(drm_dev->dev, sgt, DMA_BIDIRECTIONAL); - /* - * Do this last because pvr_gem_object_zero() requires a fully - * configured instance of struct pvr_gem_object. - */ - pvr_gem_object_zero(pvr_obj); - return pvr_obj; err_shmem_object_free: --- base-commit: 41b5458595c60143d0ebaa1e032c9cda5d764dff change-id: 20261105-pvr-gem-drop-zeroing-67cc98380a55 Best regards, -- Guillaume La Roque <[email protected]>
