Re: [PATCH] drm/amdgpu: Fix minmax error

[Luben Tuikov <luben.tuikov@amd.com>]

On 2022-11-25 04:57, Christian König wrote:
> 
> 
> Am 25.11.22 um 09:33 schrieb Luben Tuikov:
>> On 2022-11-25 02:59, Christian König wrote:
>>> Am 25.11.22 um 08:56 schrieb Luben Tuikov:
>>>> On 2022-11-25 02:45, Christian König wrote:
>>>>> Am 24.11.22 um 22:19 schrieb Luben Tuikov:
>>>>>> Fix minmax compilation error by using min_t()/max_t(), of the assignment type.
>>>>>>
>>>>>> Cc: James Zhu <James.Zhu@amd.com>
>>>>>> Cc: Felix Kuehling <Felix.Kuehling@amd.com>
>>>>>> Fixes: 58170a7a002ad6 ("drm/amdgpu: fix stall on CPU when allocate large system memory")
>>>>>> Signed-off-by: Luben Tuikov <luben.tuikov@amd.com>
>>>>>> ---
>>>>>>     drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c | 10 +++++++---
>>>>>>     1 file changed, 7 insertions(+), 3 deletions(-)
>>>>>>
>>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c
>>>>>> index 8a2e5716d8dba2..d22d14b0ef0c84 100644
>>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c
>>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_hmm.c
>>>>>> @@ -191,14 +191,18 @@ int amdgpu_hmm_range_get_pages(struct mmu_interval_notifier *notifier,
>>>>>>     	hmm_range->dev_private_owner = owner;
>>>>>>     
>>>>>>     	do {
>>>>>> -		hmm_range->end = min(hmm_range->start + MAX_WALK_BYTE, end);
>>>>>> +		hmm_range->end = min_t(typeof(hmm_range->end),
>>>>>> +				       hmm_range->start + MAX_WALK_BYTE,
>>>>>> +				       end);
>>>>> Since end is a local variable I would strongly prefer to just have it
>>>>> use the correct type for it.
>>>>>
>>>>> Otherwise we might end up using something which doesn't work on all
>>>>> architectures.
>>>> They all appear to be "unsigned long". I thought, since we assign to
>>>> hmm_range->end, we use that type.
>>> Mhm, then why does the compiler complain here?
>> Right... so MAX_WALK_BYTE is 2^36 ULL (diabolically defined as 64ULL<<30 :-) ),
>> and this is why the minmax check complains.
>>
>> So, since the left-hand expression is unsigned long,
>> i.e.,
>> 	hmm_range->end = min(hmm_range->start + MAX_WALK_BYTE, end);
>> is,
>> 	unsigned long = min(unsigned long long, unsigned long);
>> The compiler complains.
>>
>> I'd really prefer MAX_WALK_BYTE be less than or equal to ULONG_MAX,
> 
> That's not only a preference, but a must have. Otherwise the code maybe 
> won't work as expected on 32bit architectures.

Well, I don't know what to change MAX_WALK_BYTE to, and given the suggestion
below to change to min_t(u64, ...), I wonder if messing with MAX_WALK_BYTE
even matters--it wouldn't matter so long as the type in min_t() is u64.
It's a macro at the moment.

However, the LHS--struct hmm_range's members are all
unsigned long and then we're essentially doing (unsigned long) = (u64),
which on 32-bit is (u32) = (u64).

Regards,
Luben

> 
>> and be defined as <literal>UL. I mean, why is everything in struct hmm_range
>> "unsigned long", but we set a high limit of 10_0000_0000h for an end, and
>> compare it to "end" to find the smaller? If our "end" could potentially
>> be 10_0000_0000h then shouldn't the members in struct hmm_range be
>> unsigned long long as well?
> 
> No, that the hmm range depends on the address space bits of the CPU is 
> perfectly correct. Essentially this is just an userspace address range.
> 
> Our problem here is that this code needs to work on both 32bit and 64bit 
> systems. And on a 32bit system limiting the types wouldn't work 
> correctly as far as I can see.
> 
> So the compiler is complaining for rather good reasons and by using 
> "min_t(UL" we just hide that instead of fixing the problem.
> 
> I suggest to use "min_t(u64" instead. An intelligent compiler should 
> even be capable of optimizing this away by looking at the input types on 
> 32bit archs.
> 
>>
>> And for the timeout, we have the (now) obvious,
>>
>> 	timeout = max((hmm_range->end - hmm_range->start) >> 29, 1ULL);
>>
>> and I don't know why we necessarily need a "1ULL", when 1UL would do just fine,
>> and then compilation passes for that statement. I can set this to 1UL, instead
>> of using max_t().
> 
> I think just changing this to 1UL should be sufficient.
> 
> Regards,
> Christian.
> 
>>
>> Regards,
>> Luben
>>
>>
>>> As far as I can see "unsigned long" is correct here, but if we somehow
>>> have a typecast then something is not working as expected.
>>>
>>> Is MAX_WALK_BYTE maybe of signed type?
>>>
>>>> Would you prefer at the top of the function to define "timeout" and "end" as,
>>>> 	typeof(hmm_range->end) end, timeout;
>>> Well for end that might make sense, but timeout is independent of the
>>> hmm range.
>>>
>>> Regards,
>>> Christian.
>>>
>>>> Regards,
>>>> Luben
>>>>
>