Re: [PATCH] drm/amdgpu: fix an underflow on non-4KB-page systems

["Christian König" <ckoenig.leichtzumerken@gmail.com>]

Am 29.03.21 um 21:27 schrieb Xi Ruoyao:
> Hi Christian,
>
> I don't think there is any constraint implemented to ensure `num_entries %
> AMDGPU_GPU_PAGES_IN_CPU_PAGE == 0`.  For example, in `amdgpu_vm_bo_map()`:
>
>          /* validate the parameters */
>          if (saddr & AMDGPU_GPU_PAGE_MASK || offset & AMDGPU_GPU_PAGE_MASK ||
>              size == 0 || size & AMDGPU_GPU_PAGE_MASK)
>                  return -EINVAL;
>
> /* snip */
>
>          saddr /= AMDGPU_GPU_PAGE_SIZE;
>          eaddr /= AMDGPU_GPU_PAGE_SIZE;
>
> /* snip */
>
>          mapping->start = saddr;
>          mapping->last = eaddr;
>
>
> If we really want to ensure (mapping->last - mapping->start + 1) %
> AMDGPU_GPU_PAGES_IN_CPU_PAGE == 0, then we should replace "AMDGPU_GPU_PAGE_MASK"
> in "validate the parameters" with "PAGE_MASK".

Yeah, good point.

> I tried it and it broke userspace: Xorg startup fails with EINVAL with this
> change.

Well in theory it is possible that we always fill the GPUVM on a 4k 
basis while the native page size of the CPU is larger. Let me double 
check the code.

BTW: What code base are you based on? The code your post here is quite 
outdated.

Christian.

>
> On 2021-03-30 02:30 +0800, Xi Ruoyao wrote:
>> On 2021-03-30 02:21 +0800, Xi Ruoyao wrote:
>>> On 2021-03-29 20:10 +0200, Christian König wrote:
>>>> You need to identify the root cause of this, most likely start or last
>>>> are not a multiple of AMDGPU_GPU_PAGES_IN_CPU_PAGE.
>>> I printk'ed the value of start & last, they are all a multiple of 4
>>> (AMDGPU_GPU_PAGES_IN_CPU_PAGE).
>>>
>>> However... `num_entries = last - start + 1` so it became some irrational
>>> thing...  Either this line is wrong, or someone called
>>> amdgpu_vm_bo_update_mapping with [start, last) instead of [start, last], which
>>> is unexpected.
>> I added BUG_ON(num_entries % AMDGPU_GPU_PAGES_IN_CPU_PAGE != 0), get:
>>
>>> Mar 30 02:28:27 xry111-A1901 kernel: [<ffffffff83c90750>]
>>> amdgpu_vm_bo_update_mapping.constprop.0+0x218/0xae8
>>> Mar 30 02:28:27 xry111-A1901 kernel: [<ffffffff83c922b8>]
>>> amdgpu_vm_bo_update+0x270/0x4c0
>>> Mar 30 02:28:27 xry111-A1901 kernel: [<ffffffff83c823ec>]
>>> amdgpu_gem_va_ioctl+0x40c/0x430
>>> Mar 30 02:28:27 xry111-A1901 kernel: [<ffffffff83c1cae4>]
>>> drm_ioctl_kernel+0xcc/0x120
>>> Mar 30 02:28:27 xry111-A1901 kernel: [<ffffffff83c1cfd8>]
>>> drm_ioctl+0x220/0x408
>>> Mar 30 02:28:27 xry111-A1901 kernel: [<ffffffff83c5ea48>]
>>> amdgpu_drm_ioctl+0x58/0x98
>>> Mar 30 02:28:27 xry111-A1901 kernel: [<ffffffff838feac4>] sys_ioctl+0xcc/0xe8
>>> Mar 30 02:28:27 xry111-A1901 kernel: [<ffffffff836e74f0>]
>>> syscall_common+0x34/0x58
>>>
>>>>>>> BugLink: https://gitlab.freedesktop.org/drm/amd/-/issues/1549
>>>>>>> Fixes: a39f2a8d7066 ("drm/amdgpu: nuke amdgpu_vm_bo_split_mapping v2")
>>>>>>> Reported-by: Xi Ruoyao <xry111@mengyan1223.wang>
>>>>>>> Reported-by: Dan Horák <dan@danny.cz>
>>>>>>> Cc: stable@vger.kernel.org
>>>>>>> Signed-off-by: Xi Ruoyao <xry111@mengyan1223.wang>
>>>>>>> ---
>>>>>>>     drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 2 +-
>>>>>>>     1 file changed, 1 insertion(+), 1 deletion(-)
>>>>>>>
>>>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>>>>> index ad91c0c3c423..cee0cc9c8085 100644
>>>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
>>>>>>> @@ -1707,7 +1707,7 @@ static int amdgpu_vm_bo_update_mapping(struct
>>>>>>> amdgpu_device *adev,
>>>>>>>                   }
>>>>>>>                   start = tmp;
>>>>>>>     
>>>>>>> -       } while (unlikely(start != last + 1));
>>>>>>> +       } while (unlikely(start < last + 1));
>>>>>>>     
>>>>>>>           r = vm->update_funcs->commit(&params, fence);
>>>>>>>     
>>>>>>>
>>>>>>> base-commit: a5e13c6df0e41702d2b2c77c8ad41677ebb065b3