Hi Maíra, On 18/11/26 09:54, Maíra Canal wrote: > The issue you reported is real and makes sense, but your fix only treats > the symptom, not the problem. > > V3D allocates the GPU address inside the gem_prime_import_sg_table() > callback, before the core marks the object as imported. No teardown > would be correct at that point.
Thanks for the review. v2 moves the allocation out of the callback: https://lore.kernel.org/all/[email protected]/ > I'd rather move the allocation after the import completes. Implement > .gem_prime_import as a wrapper around drm_gem_prime_import() and call > v3d_bo_create_finish() once it returns. At that point import_attach is > set, so drm_gem_object_put() takes the regular import teardown path, > which unmaps the sg_table once and detaches. v2 follows this and differs from it in two places. 1. Same-device imports For a dma-buf that this device exported, drm_gem_prime_import() returns the BO it was exported from, with one more reference. That BO already has its node in v3d->mm, so the wrapper returns it without calling v3d_bo_create_finish() again. A compositor importing its clients' buffers takes this path: in a glmark2-es2-wayland run under labwc all 5 calls of the wrapper were same-device imports. 2. drm_gem_shmem_free() on failure, not drm_gem_object_put() drm_gem_object_put() ends in v3d_free_object(), which expects a BO that v3d_bo_create_finish() completed. When the node was never inserted, node.start is 0, so v3d_mmu_remove_ptes() would clear the page table entries from page 0 up to the size of the BO, and those can belong to other BOs. bo_stats would be decremented without having been incremented, and drm_mm_remove_node() would run on a node that is not allocated. With drm_gem_object_put() in the wrapper, the failing import oopses on the same board: Unable to handle kernel NULL pointer dereference at virtual address 00000000000000c8 pc : drm_mm_remove_node+0x34/0x388 lr : v3d_free_object+0x88/0xc0 [v3d] Call trace: drm_mm_remove_node+0x34/0x388 (P) v3d_free_object+0x88/0xc0 [v3d] drm_gem_object_free+0x24/0x40 v3d_prime_import+0xc8/0xe0 [v3d] drm_gem_prime_fd_to_handle+0xd4/0x280 and bo_stats then counts 14 BOs while 15 are alive. The cleared page table entries I did not observe: no other BO was alive in that test. So the wrapper frees the object the way v3d_bo_create() does when v3d_bo_create_finish() fails. With import_attach set, drm_gem_shmem_free() goes through drm_prime_gem_destroy(): the sg_table is unmapped once, the attachment is detached and the reference on the dma-buf is dropped. Would you rather have drm_gem_object_put() there, with v3d_free_object() checking drm_mm_node_allocated()? One side effect of moving the allocation: v3d_bo_create_finish() now runs after obj->resv has been pointed at the reservation object of the dma-buf, so drm_gem_shmem_get_pages_sgt() takes that lock instead of the private lock of a new object. It is taken interruptibly, so an import can return -EINTR if a signal arrives while it waits. libdrm's drmPrimeFDToHandle() retries on EINTR. Tested on a Raspberry Pi 5, v7.4-rc2 with slub_debug=FZPU: - the import that exhausts the V3D address space returns -ENOSPC in three consecutive fills, on two boots; - after 50 such failures /sys/kernel/debug/dma_buf/bufinfo lists no dma-buf (15 udmabufs attached to v3d while the test holds them); - a v3d BO re-imported through a second file is the same BO at the same offset over 1000 iterations, and bo_stats returns to its starting value; - a vc4 dumb buffer imported into v3d can be mapped, written, read back and closed; - under labwc, glmark2-es2-wayland runs, and mpv playing 4K HEVC from the hardware decoder (--vo=gpu and --vo=dmabuf-wayland, --hwdec=drm) imports the decoder's dma-heap buffers into v3d about 2500 times; none of them is left attached afterwards. The median time of an import and the number of dropped frames are the same as with the unpatched module, with and without slub_debug. The kernel log stays clean in all of them. Not tested on other V3D generations. Best regards, Aleksei
