drm/amdkfd: Allow CPU access for all VRAM BOs
Subsystems (5)
Alex Deucher
maintainerPATCH 1/2 drm/amdkfd: Allow CPU access for all VRAM BOs
From: Felix Kuehling <[email protected]>
The thunk needs to mmap all BOs for CPU access to allow the debugger to access them. Invisible ones are mapped with PROT_NONE.
--- drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c | 3 +-- 1 file changed, 1 insertion(+), 2 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_amdkfd_gpuvm.c +1 -2
@@ -1404,8 +1404,7 @@ int amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu( domain = alloc_domain =AMDGPU_GEM_DOMAIN_VRAM ; alloc_flags =AMDGPU_GEM_CREATE_VRAM_WIPE_ON_RELEASE ; alloc_flags |= (flags &KFD_IOC_ALLOC_MEM_FLAGS_PUBLIC ) ? -AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED : -AMDGPU_GEM_CREATE_NO_CPU_ACCESS ; +AMDGPU_GEM_CREATE_CPU_ACCESS_REQUIRED :0 ; }else if (flags &KFD_IOC_ALLOC_MEM_FLAGS_GTT ) { domain = alloc_domain =AMDGPU_GEM_DOMAIN_GTT ; alloc_flags =0 ;
Signature · 6 lines
2.31.1 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Alex Deucher
maintainerPATCH 2/2 drm/ttm: Fix COW check
From: Felix Kuehling <[email protected]>
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.
---drivers/gpu/drm/ttm/ttm_bo_vm.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/ttm/ttm_bo_vm.c +1 -1
@@ -552,7 +552,7 @@ static const struct vm_operations_struct ttm_bo_vm_ops = {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);
Signature · 6 lines
2.31.1 _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH 2/2] drm/ttm: Fix COW check
Am 08.07.21 um 21:36 schrieb Alex Deucher:
Quoted 6 lines
From: Felix Kuehling <[email protected]>
KFD Thunk maps invisible VRAM BOs with
PROT_NONE,MAP_PRIVATE.is_cow_mappingreturns true for these mappings. Add a check forvm_flags&VM_WRITEto avoid mmap failures on private read-only orPROT_NONEmappings.
I'm pretty sure that this is not working as expected.
Quoted 13 lines
Fixes f91142c62161 ("drm/ttm: nuke VM_MIXEDMAP on BO mappings v3")Signed-off-by Felix Kuehling <[email protected]>Signed-off-by Alex Deucher <[email protected]>---
drivers/gpu/drm/ttm/ttm_bo_vm.c| 2 +-1 file changed, 1 insertion(+), 1 deletion(-)drivers/gpu/drm/ttm/ttm_bo_vm.c +0
-0
drivers/gpu/drm/ttm/ttm_bo_vm.c static const struct vm_operations_struct ttm_bo_vm_ops = { · line 555
Quoted 5 lines
drivers/gpu/drm/ttm/ttm_bo_vm.c +1
-1@@ -552,7 +552,7 @@ static const struct vm_operations_struct ttm_bo_vm_ops = {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 ))
is_cow_mapping() already checks for VM_MAYWRITE, so this here shouldn't be necessary.
Christian.
drivers/gpu/drm/ttm/ttm_bo_vm.c static const struct vm_operations_struct ttm_bo_vm_ops = { · line 558
Quoted 3 lines
drivers/gpu/drm/ttm/ttm_bo_vm.c +0
-0return -EINVAL ;ttm_bo_get (bo);
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH 2/2] drm/ttm: Fix COW check
Quoted 21 lines · Christian König
Am 08.07.21 um 21:36 schrieb Alex Deucher:
Quoted 6 lines
From: Felix Kuehling <[email protected]>
KFD Thunk maps invisible VRAM BOs with
PROT_NONE,MAP_PRIVATE.is_cow_mappingreturns true for these mappings. Add a check forvm_flags&VM_WRITEto avoid mmap failures on private read-only orPROT_NONEmappings.I'm pretty sure that this is not working as expected.
Quoted 8 lines
Fixes f91142c62161 ("drm/ttm: nuke VM_MIXEDMAP on BO mappings v3")Signed-off-by Felix Kuehling <[email protected]>Signed-off-by Alex Deucher <[email protected]>---
drivers/gpu/drm/ttm/ttm_bo_vm.c| 2 +-1 file changed, 1 insertion(+), 1 deletion(-)
Quoted 13 lines
drivers/gpu/drm/ttm/ttm_bo_vm.c static const struct vm_operations_struct ttm_bo_vm_ops = { · line 555
Quoted 10 lines
drivers/gpu/drm/ttm/ttm_bo_vm.c +1
-1@@ -552,7 +552,7 @@ static const struct vm_operations_struct ttm_bo_vm_ops = {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 ))
is_cow_mapping()already checks forVM_MAYWRITE, so this here shouldn't be necessary.
MAYWRITE != WRITE
But then you need to make sure you do catch mprotect() calls to catch the cow, and I'm not sure that's even possible. Otherwise it'll go boom again on the page refcount.
-Daniel
Quoted 7 lines
Christian.
Quoted 3 lines
drivers/gpu/drm/ttm/ttm_bo_vm.c +0
-0return -EINVAL ;ttm_bo_get (bo);
Signature · 7 lines
Daniel Vetter Software Engineer, Intel Corporation http://blog.ffwll.ch _______________________________________________ amd-gfx mailing list [email protected] https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Felix Kuehling
maintainerRe: [PATCH 2/2] drm/ttm: Fix COW check
In reply to Christian KönigQuoted 11 lines · Christian König
Am 08.07.21 um 21:36 schrieb Alex Deucher:
Quoted 6 lines
From: Felix Kuehling <[email protected]>
KFD Thunk maps invisible VRAM BOs with
PROT_NONE,MAP_PRIVATE.is_cow_mappingreturns true for these mappings. Add a check forvm_flags&VM_WRITEto avoid mmap failures on private read-only orPROT_NONEmappings.I'm pretty sure that this is not working as expected.
Not sure what you mean. Debugger access to the memory through the PROT_NONE VMAs is definitely working, with both ptrace and /proc/<pid>/mem.
Quoted 9 lines
Quoted 9 lines
Fixes f91142c62161 ("drm/ttm: nuke VM_MIXEDMAP on BO mappings v3")Signed-off-by Felix Kuehling <[email protected]>Signed-off-by Alex Deucher <[email protected]>---
drivers/gpu/drm/ttm/ttm_bo_vm.c| 2 +-1 file changed, 1 insertion(+), 1 deletion(-)
Quoted 13 lines
drivers/gpu/drm/ttm/ttm_bo_vm.c static const struct vm_operations_struct · line 556
Quoted 11 lines
drivers/gpu/drm/ttm/ttm_bo_vm.c +1
-1@@ -552,7 +552,7 @@ static const struct vm_operations_struct ttm_bo_vm_ops = { 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))
is_cow_mapping()already checks forVM_MAYWRITE, so this here shouldn't be necessary.
AFAICT, VM_MAYWRITE is not based on the PROT_... bits used to create the VMA, but based on the permissions of the file. So as long as the render node is writable, VM_MAYWRITE is set for all VMAs that map it.
I would agree that it's probably a bad idea for the Thunk to map these VMAs with MAP_PRIVATE. We can try changing the Thunk to use MAP_SHARED for these PROT_NONE mappings. But that doesn't change the fact that this kernel patch broke existing usermode.
Regards,
Felix
Quoted 6 lines
Christian.
Quoted 2 lines
drivers/gpu/drm/ttm/ttm_bo_vm.c +0
-0return -EINVAL; ttm_bo_get(bo);
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH 2/2] drm/ttm: Fix COW check
Am 09.07.21 um 21:31 schrieb Felix Kuehling:
Quoted 39 lines · Christian König
Quoted 11 lines
Am 08.07.21 um 21:36 schrieb Alex Deucher:
Quoted 6 lines
From: Felix Kuehling <[email protected]>
KFD Thunk maps invisible VRAM BOs with
PROT_NONE,MAP_PRIVATE.is_cow_mappingreturns true for these mappings. Add a check forvm_flags&VM_WRITEto avoid mmap failures on private read-only orPROT_NONEmappings.I'm pretty sure that this is not working as expected.
Not sure what you mean. Debugger access to the memory through the
PROT_NONEVMAs is definitely working, with both ptrace and /proc/<pid>/mem.Quoted 9 lines
Quoted 9 lines
Fixes f91142c62161 ("drm/ttm: nuke VM_MIXEDMAP on BO mappings v3")Signed-off-by Felix Kuehling <[email protected]>Signed-off-by Alex Deucher <[email protected]>---
drivers/gpu/drm/ttm/ttm_bo_vm.c| 2 +-1 file changed, 1 insertion(+), 1 deletion(-)Quoted 13 lines
drivers/gpu/drm/ttm/ttm_bo_vm.c static const struct vm_operations_struct · line 556
Quoted 11 lines
drivers/gpu/drm/ttm/ttm_bo_vm.c +1
-1@@ -552,7 +552,7 @@ static const struct vm_operations_struct ttm_bo_vm_ops = { 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))
is_cow_mapping()already checks forVM_MAYWRITE, so this here shouldn't be necessary.
Quoted 3 lines
AFAICT,
VM_MAYWRITEis not based on thePROT_... bits used to create the VMA, but based on the permissions of the file. So as long as the render node is writable,VM_MAYWRITEis set for all VMAs that map it.I would agree that it's probably a bad idea for the Thunk to map these VMAs with
MAP_PRIVATE. We can try changing the Thunk to useMAP_SHAREDfor thesePROT_NONEmappings. But that doesn't change the fact that this kernel patch broke existing usermode.
Yeah, but see the discussion around MAP_PRIVATE and COW regarding this. What Thunk did here was a really bad idea.
I think we have only two options here:
1. Accept the breakage of thunk.
2. Add a workaround in amdgpu/amdkfd to force MAP_PRIVATE into MAP_SHARED in the kernel.
Regards,
Christian.
Quoted 11 lines
Regards,
FelixQuoted 6 lines
Christian.
Quoted 2 lines
drivers/gpu/drm/ttm/ttm_bo_vm.c +0
-0return -EINVAL; ttm_bo_get(bo);
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Felix Kuehling
maintainerRe: [PATCH 2/2] drm/ttm: Fix COW check
Quoted 43 lines · Christian König
Am 09.07.21 um 21:31 schrieb Felix Kuehling:
Quoted 42 lines · Christian König
Quoted 11 lines
Am 08.07.21 um 21:36 schrieb Alex Deucher:
Quoted 6 lines
From: Felix Kuehling <[email protected]>
KFD Thunk maps invisible VRAM BOs with
PROT_NONE,MAP_PRIVATE.is_cow_mappingreturns true for these mappings. Add a check forvm_flags&VM_WRITEto avoid mmap failures on private read-only orPROT_NONEmappings.I'm pretty sure that this is not working as expected.
Not sure what you mean. Debugger access to the memory through the
PROT_NONEVMAs is definitely working, with both ptrace and /proc/<pid>/mem.Quoted 9 lines
Quoted 9 lines
Fixes f91142c62161 ("drm/ttm: nuke VM_MIXEDMAP on BO mappings v3")Signed-off-by Felix Kuehling <[email protected]>Signed-off-by Alex Deucher <[email protected]>---
drivers/gpu/drm/ttm/ttm_bo_vm.c| 2 +-1 file changed, 1 insertion(+), 1 deletion(-)Quoted 13 lines
drivers/gpu/drm/ttm/ttm_bo_vm.c static const struct vm_operations_struct · line 556
Quoted 11 lines
drivers/gpu/drm/ttm/ttm_bo_vm.c +1
-1@@ -552,7 +552,7 @@ static const struct vm_operations_struct ttm_bo_vm_ops = { 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))
is_cow_mapping()already checks forVM_MAYWRITE, so this here shouldn't be necessary.AFAICT,
VM_MAYWRITEis not based on thePROT_... bits used to create the VMA, but based on the permissions of the file. So as long as the render node is writable,VM_MAYWRITEis set for all VMAs that map it.I would agree that it's probably a bad idea for the Thunk to map these VMAs with
MAP_PRIVATE. We can try changing the Thunk to useMAP_SHAREDfor thesePROT_NONEmappings. But that doesn't change the fact that this kernel patch broke existing usermode.
For the record, changing the Thunk to use MAP_SHARED works.
Quoted 2 lines
Yeah, but see the discussion around
MAP_PRIVATEand COW regarding this. What Thunk did here was a really bad idea.
Hindsight ... Which part was a bad idea?
* Creating a PROT_NONE VMA? That's necessary to let ptrace or /proc/<pid>/mem access the memory. It's a brilliant idea, IMHO
* Making that VMA MAP_PRIVATE? Probably a bad idea in hindsight. The process itself can't access this memory, let alone write to it. So I didn't think too much about whether to make it shared or private. Either way, we are not causing any actual COW on these mappings because they are not writable, and we never make them writable with mprotect either
* Introducing a COW check after letting it slide for 15 years? It's a reasonable check. In this case it catches a false positive. Had this check been in place 4 or 5 years ago, it would have easily prevented this mess
Quoted 3 lines
I think we have only two options here:
1. Accept the breakage of thunk.
Really?
Quoted 1 line
2. Add a workaround in amdgpu/amdkfd to force
MAP_PRIVATEintoMAP_SHAREDin the kernel.
I tried to do this in amdgpu_gem_object_mmap and spent 4 hours debugging why it doesn't work. It breaks because the mapping->i_mmap_writable gets unbalanced and causes subsequent mappings to fail when that atomic counter becomes negative (indicating DENYWRITE). I could fix it by calling mapping_map_writable. But I don't think this is intended as driver API. I see no precedent of any driver calling this. For the record this works, but it feels fragile and ugly:
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +14 -0
@@ -255,6 +255,20 @@ static int amdgpu_gem_object_mmap(struct drm_gem_object *obj, struct vm_area_strif (bo->flags &AMDGPU_GEM_CREATE_NO_CPU_ACCESS )return -EPERM ; +/* Workaround for Thunk bug creating PROT_NONE,MAP_PRIVATE mappings +* for debugger access to invisible VRAM. Should have used MAP_SHARED +* instead. +*/ +if (is_cow_mapping (vma->vm_flags) && + !(vma->vm_flags & (VM_READ |VM_WRITE |VM_EXEC ))) { +int ret; + + ret =mapping_map_writable (vma->vm_file->f_mapping); +if (ret) +return ret; + vma->vm_flags |=VM_SHARED |VM_MAYSHARE ; + } +return drm_gem_ttm_mmap (obj, vma); }
3. Improve my previous workaround by adding a similar check for COW in ttm_bo_vm_ops.mprotect. I'll send an updated patch.
Regards,
Felix
Quoted 16 lines
Regards,
Christian.Quoted 11 lines
Regards,
FelixQuoted 6 lines
Christian.
Quoted 2 lines
drivers/gpu/drm/ttm/ttm_bo_vm.c +0
-0return -EINVAL; ttm_bo_get(bo);
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH 2/2] drm/ttm: Fix COW check
Am 10.07.21 um 02:29 schrieb Felix Kuehling:
Quoted 54 lines · Christian König
Quoted 43 lines
Am 09.07.21 um 21:31 schrieb Felix Kuehling:
Quoted 42 lines · Christian König
Quoted 11 lines
Am 08.07.21 um 21:36 schrieb Alex Deucher:
Quoted 6 lines
From: Felix Kuehling <[email protected]>
KFD Thunk maps invisible VRAM BOs with
PROT_NONE,MAP_PRIVATE.is_cow_mappingreturns true for these mappings. Add a check forvm_flags&VM_WRITEto avoid mmap failures on private read-only orPROT_NONEmappings.I'm pretty sure that this is not working as expected.
Not sure what you mean. Debugger access to the memory through the
PROT_NONEVMAs is definitely working, with both ptrace and /proc/<pid>/mem.Quoted 9 lines
Quoted 9 lines
Fixes f91142c62161 ("drm/ttm: nuke VM_MIXEDMAP on BO mappings v3")Signed-off-by Felix Kuehling <[email protected]>Signed-off-by Alex Deucher <[email protected]>---
drivers/gpu/drm/ttm/ttm_bo_vm.c| 2 +-1 file changed, 1 insertion(+), 1 deletion(-)Quoted 13 lines
drivers/gpu/drm/ttm/ttm_bo_vm.c static const struct vm_operations_struct · line 556
Quoted 11 lines
drivers/gpu/drm/ttm/ttm_bo_vm.c +1
-1@@ -552,7 +552,7 @@ static const struct vm_operations_struct ttm_bo_vm_ops = { 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))
is_cow_mapping()already checks forVM_MAYWRITE, so this here shouldn't be necessary.AFAICT,
VM_MAYWRITEis not based on thePROT_... bits used to create the VMA, but based on the permissions of the file. So as long as the render node is writable,VM_MAYWRITEis set for all VMAs that map it.I would agree that it's probably a bad idea for the Thunk to map these VMAs with
MAP_PRIVATE. We can try changing the Thunk to useMAP_SHAREDfor thesePROT_NONEmappings. But that doesn't change the fact that this kernel patch broke existing usermode.For the record, changing the Thunk to use
MAP_SHAREDworks.Quoted 2 lines
Yeah, but see the discussion around
MAP_PRIVATEand COW regarding this. What Thunk did here was a really bad idea.Hindsight ... Which part was a bad idea?
Quoted 12 lines
* Creating a
PROT_NONEVMA? That's necessary to let ptrace or /proc/<pid>/mem access the memory. It's a brilliant idea, IMHO
* Making that VMAMAP_PRIVATE? Probably a bad idea in hindsight. The process itself can't access this memory, let alone write to it. So I didn't think too much about whether to make it shared or private. Either way, we are not causing any actual COW on these mappings because they are not writable, and we never make them writable with mprotect either
* Introducing a COW check after letting it slide for 15 years? It's a reasonable check. In this case it catches a false positive. Had this check been in place 4 or 5 years ago, it would have easily prevented this mess
A combination of everything.
MAP_PRIVATE would have previously caused very odd behavior and/or a BUG_ON()/WARN_ON() in the VMA code (depending on platform).
MAP_PRIVATE + PROT_NONE kind of worked because it never faulted a page so we never actually triggered a COW.
Quoted 6 lines
Quoted 4 lines
I think we have only two options here:
1. Accept the breakage of thunk.Really?
Only as last resort, I'm running out of ideas how to fix this cleanly.
Quoted 21 lines
Quoted 3 lines
2. Add a workaround in amdgpu/amdkfd to force
MAP_PRIVATEintoMAP_SHAREDin the kernel.I tried to do this in
amdgpu_gem_object_mmapand spent 4 hours debugging why it doesn't work. It breaks because themapping->i_mmap_writablegets unbalanced and causes subsequent mappings to fail when that atomic counter becomes negative (indicating DENYWRITE). I could fix it by callingmapping_map_writable. But I don't think this is intended as driver API. I see no precedent of any driver calling this. For the record this works, but it feels fragile and ugly:drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +0
-0@@ -255,6 +255,20 @@ static int amdgpu_gem_object_mmap(struct
drm_gem_object*obj, structvm_area_str
if (bo->flags&AMDGPU_GEM_CREATE_NO_CPU_ACCESS) return-EPERM;+ /* Workaround for Thunk bug creating PROT_NONE,MAP_PRIVATE mappings + * for debugger access to invisible VRAM. Should have used
MAP_SHARED
+ * instead.
+ */
+ if (is_cow_mapping(vma->vm_flags)&&
+ !(vma->vm_flags & (VM_READ|VM_WRITE|VM_EXEC))) { + int ret;
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c static int amdgpu_gem_object_mmap(struct drm_gem_object *obj, struct vm_area_str · line 271
Quoted 7 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_gem.c +7
-0+ + ret = mapping_map_writable(vma->vm_file->f_mapping); + if (ret) + return ret; + vma->vm_flags |= VM_SHARED | VM_MAYSHARE; + } +
Yeah, I'm on another thread where we are discussing mapping_map_writable() for vma_set_file() as well.
Quoted 4 lines
return
drm_gem_ttm_mmap(obj, vma);
}3. Improve my previous workaround by adding a similar check for COW in ttm_bo_vm_ops.mprotect. I'll send an updated patch.
That has it's own issues, but going to reply there in detail.
Regards,
Christian.
Quoted 21 lines
Regards,
FelixQuoted 16 lines
Regards,
Christian.Quoted 11 lines
Regards,
FelixQuoted 6 lines
Christian.
Quoted 2 lines
drivers/gpu/drm/ttm/ttm_bo_vm.c +0
-0return -EINVAL; ttm_bo_get(bo);
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx