Re: [PATCH] drm/amdgpu: fix userq page count in input_validate

Ping Christian.


On 7/20/2026 3:49 PM, Zhu, Lingshan wrote:

> On 7/17/2026 4:06 PM, Christian König wrote:
>
>> On 7/17/26 04:25, Zhu Lingshan wrote:
>>> amdgpu_userq_input_va_validate() converts userq
>>> expected_size to page count by right shift.
>>>
>>> This undercounts va buffers that less than a page.
>>> For example, in the AMDGPU_HW_IP_COMPUTE userq
>>> creation case, the expected_size 2048,
>>> so size(page count) becomes zero page and the
>>> "va_map->last - user_addr + 1 >= size" is always true
>>> once the start addr locates in the va_map mapping.
>>>
>>> A similar case is AMDGPU_HW_IP_DMA, where expected_size is 32.
>>>
>>> This commit calculates a proper page count by subtracting
>>> the start_addr and end_addr in pages.
>>>
>>> Signed-off-by: Zhu Lingshan <[email protected]>
>>> ---
>>>  drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c | 14 +++++++++++---
>>>  1 file changed, 11 insertions(+), 3 deletions(-)
>>>
>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
>>> index 969f49592450..f9f11bdf2cef 100644
>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq.c
>>> @@ -25,6 +25,7 @@
>>>  #include <drm/drm_auth.h>
>>>  #include <drm/drm_exec.h>
>>>  #include <linux/pm_runtime.h>
>>> +#include <linux/overflow.h>
>>>  #include <drm/drm_drv.h>
>>>  
>>>  #include "amdgpu.h"
>>> @@ -238,14 +239,21 @@ int amdgpu_userq_input_va_validate(struct amdgpu_device *adev,
>>>  {
>>>  	struct amdgpu_bo_va_mapping *va_map;
>>>  	struct amdgpu_vm *vm = queue->vm;
>>> -	u64 user_addr;
>>> +	u64 user_addr, end;
>>>  	u64 size;
>>>  
>>>  	/* Caller must hold vm->root.bo reservation */
>>>  	dma_resv_assert_held(queue->vm->root.bo->tbo.base.resv);
>>>  
>>> -	user_addr = (addr & AMDGPU_GMC_HOLE_MASK) >> AMDGPU_GPU_PAGE_SHIFT;
>>> -	size = expected_size >> AMDGPU_GPU_PAGE_SHIFT;
>> Good catch, but I would just use ALIGN(expected_size, AMDGPU_GPU_PAGE_SIZE) here.
> Hello Christian
>
> ALIGN is more concise, however it is only correct when start_addr is PAGE aligned.
>
> It undercounts one page when start_addr is not page aligned and "start_addr + expected_size" crosses page boundary.
>
> An example is, if start_addr is 0xC00 (1024 X 3), and expected_size is 2048, so the size should be two pages,
> but the result of ALIGN() is one page, which is wrong. 
>
> I think this "start - end" is a more general solution.
>
> Thanks
> Lingshan 
>
>> Regards,
>> Christian.
>>
>>> +	if (!expected_size)
>>> +		return -EINVAL;
>>> +
>>> +	user_addr = addr & AMDGPU_GMC_HOLE_MASK;
>>> +	if (check_add_overflow(user_addr, expected_size - 1, &end))
>>> +		return -EINVAL;
>>> +
>>> +	user_addr >>= AMDGPU_GPU_PAGE_SHIFT;
>>> +	size = (end >> AMDGPU_GPU_PAGE_SHIFT) - user_addr + 1;
>>>  
>>>  	va_map = amdgpu_vm_bo_lookup_mapping(vm, user_addr);
>>>  	if (!va_map)