Re: [PATCH 6/8] drm: Add a drm_get_unmapped_area() helper

["Thomas Hellström (VMware)" <thomas_os@shipmail.org>]

On 12/4/19 3:40 PM, Christian König wrote:
> Am 04.12.19 um 13:32 schrieb Thomas Hellström (VMware):
>> On 12/4/19 1:08 PM, Christian König wrote:
>>> Am 04.12.19 um 12:36 schrieb Thomas Hellström (VMware):
>>>> On 12/4/19 12:11 PM, Christian König wrote:
>>>>> Am 03.12.19 um 14:22 schrieb Thomas Hellström (VMware):
>>>>>> From: Thomas Hellstrom <thellstrom@vmware.com>
>>>>>>
>>>>>> This helper is used to align user-space buffer object addresses to
>>>>>> huge page boundaries, minimizing the chance of alignment mismatch
>>>>>> between user-space addresses and physical addresses.
>>>>>
>>>>> Mhm, I'm wondering if that is really such a good idea.
>>>>
>>>> Could you elaborate? What drawbacks do you see?
>>>
>>> Main problem for me seems to be that I don't fully understand what 
>>> the get_unmapped_area callback is doing.
>>
>> It makes sure that, if there is a chance that we could use huge 
>> page-table entries, virtual address huge page boundaries are 
>> perfectly aligned to physical address huge page boundaries, which is 
>> if not a CPU hardware requirement, at least a kernel requirement 
>> currently.
>>
>>
>>>
>>> For example why do we need to use drm_vma_offset_lookup_locked() to 
>>> adjust the pgoff?
>>>
>>> The mapped offset should be completely irrelevant for finding some 
>>> piece of userspace address space or am I totally off here?
>>
>>
>> Because the unmodified pgoff assumes that physical address boundaries 
>> are perfectly aligned with file offset boundaries, which is typical 
>> for all other subsystems.
>>
>> That's not true for TTM, however, where a buffer object start 
>> physical address may be huge page aligned, but the file offset is 
>> always page aligned. We could of course change that to align also 
>> file offsets to huge page size boundaries, but with the above 
>> adjustment, that's not needed. I opted for the adjustment.
>
> I would opt for aligning the file offsets instead.

Yes but that adds additional complexity and considerations which made me 
think that lookup was the by far simplest choice:

- We need to modify the vma manager to care about alignments.
- Fragmentation issues. Do we want to align > 1G BOs
- For which drivers do we want to do this, how do we handle drivers that 
want to opt out in TTM mmap()?
- Non TTM drivers. Could they still reuse the same get_unmapped_area.

>
> Now that you explained it that the rest of the kernel enforces this 
> actually makes sense.

So is that an ack?


Thanks,

Thomas



>
> Regards,
> Christian.
>
>>
>> Thanks,
>>
>> Thomas
>>
>>