On Wed, Jul 14, 2021 at 12:51:15PM +0200, Christian König wrote:
Am 14.07.21 um 12:44 schrieb Daniel Vetter:
On Mon, Jul 12, 2021 at 06:06:36PM -0400, Felix Kuehling wrote:
KFD Thunk maps invisible VRAM BOs with PROT_NONE, MAP_PRIVATE. is_cow_mapping returns true for these mappings. Add a check for vm_flags & VM_WRITE to avoid mmap failures on private read-only or PROT_NONE mappings.
v2: protect against mprotect making a mapping writable after the fact v3: update driver-specific vm_operations_structs
Fixes: f91142c62161 ("drm/ttm: nuke VM_MIXEDMAP on BO mappings v3") Signed-off-by: Felix Kuehling Felix.Kuehling@amd.com Signed-off-by: Alex Deucher alexander.deucher@amd.com
So looking at vmf_insert_pfn_prot() and the comment there we can't have VM_PFNMAP and is_cow_mapping ever be true, or things break. On platforms without pte_special at least.
Key idea is that we never end up in vmf_insert_pfn_prot() because the vma is mapped with PROT_NONE.
Ah right if it's PROT_NONE then it's ok. But the code here only checks for VM_WRITE, not VM_READ, so PROT_READ can get through and go boom? Or something else I'm missing?
Maybe time for a few amdgpu mmap tests that go through the combos and make sure it works/fails all correctly. -Daniel
So I'm not sure this is a great idea, and definitely not for all drivers
Yeah, I'm absolutely not happy with this either but it seemed to be the least painful thing to do.
...
Can we clear VM_MAYWRITE instead to force this to be a non-cow mapping instead?
Well we have considered forcefully setting VM_SHARED, which won't work easily for a couple of reasons.
But clearing VM_MAYWRITE in amdgpu/amdkfd may actually work as well.
Felix can you test this?
Thanks, Christian.
-Daniel
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c | 3 ++- drivers/gpu/drm/nouveau/nouveau_gem.c | 3 ++- drivers/gpu/drm/radeon/radeon_gem.c | 3 ++- drivers/gpu/drm/ttm/ttm_bo_vm.c | 14 +++++++++++++- drivers/gpu/drm/vmwgfx/vmwgfx_ttm_glue.c | 1 + include/drm/ttm/ttm_bo_api.h | 4 ++++ 6 files changed, 24 insertions(+), 4 deletions(-)
diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c index b3404c43a911..1aa750a6a5d2 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c @@ -79,7 +79,8 @@ static const struct vm_operations_struct amdgpu_gem_vm_ops = { .fault = amdgpu_gem_fault, .open = ttm_bo_vm_open, .close = ttm_bo_vm_close,
- .access = ttm_bo_vm_access
- .access = ttm_bo_vm_access,
- .mprotect = ttm_bo_vm_mprotect }; static void amdgpu_gem_object_free(struct drm_gem_object *gobj)
diff --git a/drivers/gpu/drm/nouveau/nouveau_gem.c b/drivers/gpu/drm/nouveau/nouveau_gem.c index 5b27845075a1..164ea564bb7a 100644 --- a/drivers/gpu/drm/nouveau/nouveau_gem.c +++ b/drivers/gpu/drm/nouveau/nouveau_gem.c @@ -70,7 +70,8 @@ static const struct vm_operations_struct nouveau_ttm_vm_ops = { .fault = nouveau_ttm_fault, .open = ttm_bo_vm_open, .close = ttm_bo_vm_close,
- .access = ttm_bo_vm_access
- .access = ttm_bo_vm_access,
- .mprotect = ttm_bo_vm_mprotect }; void
diff --git a/drivers/gpu/drm/radeon/radeon_gem.c b/drivers/gpu/drm/radeon/radeon_gem.c index 458f92a70887..c19ad07eb7b5 100644 --- a/drivers/gpu/drm/radeon/radeon_gem.c +++ b/drivers/gpu/drm/radeon/radeon_gem.c @@ -77,7 +77,8 @@ static const struct vm_operations_struct radeon_gem_vm_ops = { .fault = radeon_gem_fault, .open = ttm_bo_vm_open, .close = ttm_bo_vm_close,
- .access = ttm_bo_vm_access
- .access = ttm_bo_vm_access,
- .mprotect = ttm_bo_vm_mprotect }; static void radeon_gem_object_free(struct drm_gem_object *gobj)
diff --git a/drivers/gpu/drm/ttm/ttm_bo_vm.c b/drivers/gpu/drm/ttm/ttm_bo_vm.c index f56be5bc0861..fb325bad5db6 100644 --- a/drivers/gpu/drm/ttm/ttm_bo_vm.c +++ b/drivers/gpu/drm/ttm/ttm_bo_vm.c @@ -542,17 +542,29 @@ int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr, } EXPORT_SYMBOL(ttm_bo_vm_access); +int ttm_bo_vm_mprotect(struct vm_area_struct *vma, unsigned long start,
unsigned long end, unsigned long newflags)+{
- /* Enforce no COW since would have really strange behavior with it. */
- if (is_cow_mapping(newflags) && (newflags & VM_WRITE))
return -EINVAL;- return 0;
+} +EXPORT_SYMBOL(ttm_bo_vm_mprotect);
- static const struct vm_operations_struct ttm_bo_vm_ops = { .fault = ttm_bo_vm_fault, .open = ttm_bo_vm_open, .close = ttm_bo_vm_close, .access = ttm_bo_vm_access,
- .mprotect = ttm_bo_vm_mprotect, }; int ttm_bo_mmap_obj(struct vm_area_struct *vma, struct ttm_buffer_object *bo) { /* Enforce no COW since would have really strange behavior with it. */
- if (is_cow_mapping(vma->vm_flags))
- if (is_cow_mapping(vma->vm_flags) && (vma->vm_flags & VM_WRITE)) return -EINVAL; ttm_bo_get(bo);
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_ttm_glue.c b/drivers/gpu/drm/vmwgfx/vmwgfx_ttm_glue.c index e6b1f98ec99f..e4bf7dc99320 100644 --- a/drivers/gpu/drm/vmwgfx/vmwgfx_ttm_glue.c +++ b/drivers/gpu/drm/vmwgfx/vmwgfx_ttm_glue.c @@ -61,6 +61,7 @@ int vmw_mmap(struct file *filp, struct vm_area_struct *vma) .fault = vmw_bo_vm_fault, .open = ttm_bo_vm_open, .close = ttm_bo_vm_close,
.mprotect = ttm_bo_vm_mprotect,diff --git a/include/drm/ttm/ttm_bo_api.h b/include/drm/ttm/ttm_bo_api.h index f681bbdbc698..40eb95875355 100644 --- a/include/drm/ttm/ttm_bo_api.h +++ b/include/drm/ttm/ttm_bo_api.h @@ -605,6 +605,10 @@ void ttm_bo_vm_close(struct vm_area_struct *vma); int ttm_bo_vm_access(struct vm_area_struct *vma, unsigned long addr, void *buf, int len, int write);
+int ttm_bo_vm_mprotect(struct vm_area_struct *vma, unsigned long start,
unsigned long end, unsigned long newflags);- bool ttm_bo_delayed_delete(struct ttm_device *bdev, bool remove_all); vm_fault_t ttm_bo_vm_dummy_page(struct vm_fault *vmf, pgprot_t prot);
-- 2.32.0