when allocating a resource in place it is common to free the buffer's
resource, then allocate a new resource in a different placement.
e.g. amdgpu_bo_create_kernel_at calls ttm_resource_free, then calls
ttm_bo_mem_space.
In this situation, bo->resource will be null as it is cleared during
the initial freeing of the previous resource.
This leads to a null deref.
Fixes: d3116756a710 (drm/ttm: rename bo->mem and make it a pointer)
Signed-off-by: Robert Beckett <[email protected]>
---
drivers/gpu/drm/ttm/ttm_bo.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
index db3dc7ef5382..62b29ee7d040 100644
--- a/drivers/gpu/drm/ttm/ttm_bo.c
+++ b/drivers/gpu/drm/ttm/ttm_bo.c
@@ -875,7 +875,7 @@ int ttm_bo_mem_space(struct ttm_buffer_object *bo,
}
error:
- if (bo->resource->mem_type == TTM_PL_SYSTEM && !bo->pin_count)
+ if (bo->resource && bo->resource->mem_type == TTM_PL_SYSTEM && !bo->pin_count)
ttm_bo_move_to_lru_tail_unlocked(bo);
return ret;
--
2.25.1
Am 18.03.22 um 20:50 schrieb Robert Beckett:
> when allocating a resource in place it is common to free the buffer's
> resource, then allocate a new resource in a different placement.
>
> e.g. amdgpu_bo_create_kernel_at calls ttm_resource_free, then calls
> ttm_bo_mem_space.
Well yes I'm working the drivers towards this, but NAK at the moment.
Currently bo->resource is never expected to be NULL.
And yes I'm searching for this bug in amdgpu for quite a while. Where
exactly does that happen?
Amdgpu is supposed to allocate a new resource first, then do a swap and
the free the old one.
Thanks,
Christian.
>
> In this situation, bo->resource will be null as it is cleared during
> the initial freeing of the previous resource.
> This leads to a null deref.
>
> Fixes: d3116756a710 (drm/ttm: rename bo->mem and make it a pointer)
>
> Signed-off-by: Robert Beckett <[email protected]>
> ---
> drivers/gpu/drm/ttm/ttm_bo.c | 2 +-
> 1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
> index db3dc7ef5382..62b29ee7d040 100644
> --- a/drivers/gpu/drm/ttm/ttm_bo.c
> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
> @@ -875,7 +875,7 @@ int ttm_bo_mem_space(struct ttm_buffer_object *bo,
> }
>
> error:
> - if (bo->resource->mem_type == TTM_PL_SYSTEM && !bo->pin_count)
> + if (bo->resource && bo->resource->mem_type == TTM_PL_SYSTEM && !bo->pin_count)
> ttm_bo_move_to_lru_tail_unlocked(bo);
>
> return ret;
On 21/03/2022 09:51, Christian König wrote:
> Am 18.03.22 um 20:50 schrieb Robert Beckett:
>> when allocating a resource in place it is common to free the buffer's
>> resource, then allocate a new resource in a different placement.
>>
>> e.g. amdgpu_bo_create_kernel_at calls ttm_resource_free, then calls
>> ttm_bo_mem_space.
>
> Well yes I'm working the drivers towards this, but NAK at the moment.
> Currently bo->resource is never expected to be NULL.
>
> And yes I'm searching for this bug in amdgpu for quite a while. Where
> exactly does that happen?
in my case, I am writing new code for i915 that does this. I will switch
it to allocate the new resource first, then free the old one if successful.
For the existing amd case, see
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/gpu/drm/amd/amdgpu/amdgpu_object.c?h=v5.17#n384
amdgpu_bo_create_kernel_at calls ttm_resource_free, then calls
ttm_bo_mem_space. If the ttm_bo_mem_space call fails (e.g. due to memory
pressure), then the error path will try to deref bo->resource, which
will be null at that point.
to fix this, I honestly don't see a reason to not also have the safety
check for null there. It could check early and return an error if it is
null. I think that defensive programming here makes sense, better than a
null deref if someone programs it wrong.
>
> Amdgpu is supposed to allocate a new resource first, then do a swap and
> the free the old one.
>
> Thanks,
> Christian.
>
>>
>> In this situation, bo->resource will be null as it is cleared during
>> the initial freeing of the previous resource.
>> This leads to a null deref.
>>
>> Fixes: d3116756a710 (drm/ttm: rename bo->mem and make it a pointer)
>>
>> Signed-off-by: Robert Beckett <[email protected]>
>> ---
>> drivers/gpu/drm/ttm/ttm_bo.c | 2 +-
>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>
>> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c b/drivers/gpu/drm/ttm/ttm_bo.c
>> index db3dc7ef5382..62b29ee7d040 100644
>> --- a/drivers/gpu/drm/ttm/ttm_bo.c
>> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
>> @@ -875,7 +875,7 @@ int ttm_bo_mem_space(struct ttm_buffer_object *bo,
>> }
>> error:
>> - if (bo->resource->mem_type == TTM_PL_SYSTEM && !bo->pin_count)
>> + if (bo->resource && bo->resource->mem_type == TTM_PL_SYSTEM &&
>> !bo->pin_count)
>> ttm_bo_move_to_lru_tail_unlocked(bo);
>> return ret;
>
Am 21.03.22 um 16:44 schrieb Robert Beckett:
>
>
> On 21/03/2022 09:51, Christian König wrote:
>> Am 18.03.22 um 20:50 schrieb Robert Beckett:
>>> when allocating a resource in place it is common to free the buffer's
>>> resource, then allocate a new resource in a different placement.
>>>
>>> e.g. amdgpu_bo_create_kernel_at calls ttm_resource_free, then calls
>>> ttm_bo_mem_space.
>>
>> Well yes I'm working the drivers towards this, but NAK at the moment.
>> Currently bo->resource is never expected to be NULL.
>>
>> And yes I'm searching for this bug in amdgpu for quite a while. Where
>> exactly does that happen?
>
> in my case, I am writing new code for i915 that does this. I will
> switch it to allocate the new resource first, then free the old one if
> successful.
>
> For the existing amd case, see
> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Fgit.kernel.org%2Fpub%2Fscm%2Flinux%2Fkernel%2Fgit%2Ftorvalds%2Flinux.git%2Ftree%2Fdrivers%2Fgpu%2Fdrm%2Famd%2Famdgpu%2Famdgpu_object.c%3Fh%3Dv5.17%23n384&data=04%7C01%7Cchristian.koenig%40amd.com%7C81f91d39683e4991181008da0b51a8d0%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637834742606744919%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000&sdata=mjHPvBStFM7tsYf%2BL9fYCWqddZuIaeza6BsRF3fAmao%3D&reserved=0
>
>
> amdgpu_bo_create_kernel_at calls ttm_resource_free, then calls
> ttm_bo_mem_space. If the ttm_bo_mem_space call fails (e.g. due to
> memory pressure), then the error path will try to deref bo->resource,
> which will be null at that point.
Yeah, but that's a special handling only used during driver startup. We
somehow have this on systems with DMA-buf sharing as well.
>
> to fix this, I honestly don't see a reason to not also have the safety
> check for null there. It could check early and return an error if it
> is null. I think that defensive programming here makes sense, better
> than a null deref if someone programs it wrong.
Having it here is fine, the problem is you need to have that at tons of
other places as well.
Maybe I should send you my WIP patch set for this? If you handle all the
other cases as well I'm perfectly fine with this.
Regards,
Christian.
>
>
>
>>
>> Amdgpu is supposed to allocate a new resource first, then do a swap
>> and the free the old one.
>>
>> Thanks,
>> Christian.
>>
>>>
>>> In this situation, bo->resource will be null as it is cleared during
>>> the initial freeing of the previous resource.
>>> This leads to a null deref.
>>>
>>> Fixes: d3116756a710 (drm/ttm: rename bo->mem and make it a pointer)
>>>
>>> Signed-off-by: Robert Beckett <[email protected]>
>>> ---
>>> drivers/gpu/drm/ttm/ttm_bo.c | 2 +-
>>> 1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/drivers/gpu/drm/ttm/ttm_bo.c
>>> b/drivers/gpu/drm/ttm/ttm_bo.c
>>> index db3dc7ef5382..62b29ee7d040 100644
>>> --- a/drivers/gpu/drm/ttm/ttm_bo.c
>>> +++ b/drivers/gpu/drm/ttm/ttm_bo.c
>>> @@ -875,7 +875,7 @@ int ttm_bo_mem_space(struct ttm_buffer_object *bo,
>>> }
>>> error:
>>> - if (bo->resource->mem_type == TTM_PL_SYSTEM && !bo->pin_count)
>>> + if (bo->resource && bo->resource->mem_type == TTM_PL_SYSTEM &&
>>> !bo->pin_count)
>>> ttm_bo_move_to_lru_tail_unlocked(bo);
>>> return ret;
>>