Am 03.09.2018 um 09:52 schrieb Daniel Vetter: > On Sat, Sep 01, 2018 at 01:32:54PM +0100, Chris Wilson wrote: >> Quoting Jia-Ju Bai (2018-09-01 13:20:41) >>> The driver may sleep with holding a spinlock. >>> >>> The function call paths (from bottom to top) in Linux-4.16 are: >>> >>> [FUNC] kmalloc(GFP_KERNEL) >>> drivers/gpu/drm/drm_mm.c, 130: >>> kmalloc in show_leaks >>> drivers/gpu/drm/drm_mm.c, 913: >>> show_leaks in drm_mm_takedown >>> drivers/gpu/drm/drm_vma_manager.c, 107: >>> drm_mm_takedown in drm_vma_offset_manager_destroy >>> drivers/gpu/drm/drm_vma_manager.c, 106: >>> _raw_write_lock in drm_vma_offset_manager_destroy >>> >>> [FUNC] kmalloc(GFP_KERNEL) >>> drivers/gpu/drm/drm_mm.c, 130: >>> kmalloc in show_leaks >>> drivers/gpu/drm/drm_mm.c, 913: >>> show_leaks in drm_mm_takedown >>> drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c, 71: >>> drm_mm_takedown in amdgpu_vram_mgr_fini >>> drivers/gpu/drm/amd/amdgpu/amdgpu_vram_mgr.c, 70: >>> spin_lock in amdgpu_vram_mgr_fini >>> >>> [FUNC] kmalloc(GFP_KERNEL) >>> drivers/gpu/drm/drm_mm.c, 130: >>> kmalloc in show_leaks >>> drivers/gpu/drm/drm_mm.c, 913: >>> show_leaks in drm_mm_takedown >>> drivers/gpu/drm/ttm/ttm_bo_manager.c, 128: >>> drm_mm_takedown in ttm_bo_man_takedown >>> drivers/gpu/drm/ttm/ttm_bo_manager.c, 126: >>> spin_lock in ttm_bo_man_takedown >>> >>> To fix this bug, GFP_KERNEL is replaced with GFP_ATOMIC. >> The bug are above, since those spinlocks do not protect the data and >> imply use-after-free. > Adding amdgpu, since that's where the bug seems to be. When we have use after free we might have concurrent uses as well. I think taking the lock here is probably a good idea if you don't want to accidentally access freed memory in show_leaks. So Chris change sounds valid to me. Christian. > -Daniel
