Re: [PATCH 2/2] mm/hmm: make full use of walk_page_range()

[Ralph Campbell <rcampbell@nvidia.com>]

On 7/24/19 4:53 AM, Jason Gunthorpe wrote:
> On Wed, Jul 24, 2019 at 08:51:46AM +0200, Christoph Hellwig wrote:
>> On Tue, Jul 23, 2019 at 04:30:16PM -0700, Ralph Campbell wrote:
>>> hmm_range_snapshot() and hmm_range_fault() both call find_vma() and
>>> walk_page_range() in a loop. This is unnecessary duplication since
>>> walk_page_range() calls find_vma() in a loop already.
>>> Simplify hmm_range_snapshot() and hmm_range_fault() by defining a
>>> walk_test() callback function to filter unhandled vmas.
>>
>> I like the approach a lot!
>>
>> But we really need to sort out the duplication between hmm_range_fault
>> and hmm_range_snapshot first, as they are basically the same code.  I
>> have patches here:
>>
>> http://git.infradead.org/users/hch/misc.git/commitdiff/a34ccd30ee8a8a3111d9e91711c12901ed7dea74
>>
>> http://git.infradead.org/users/hch/misc.git/commitdiff/81f442ebac7170815af7770a1efa9c4ab662137e
> 
> Yeah, that is a straightforward improvement, maybe Ralph should grab
> these two as part of his series?

Sure, no problem.
I'll add them in v2 when I fix the other issues in the series.

>> That being said we don't really have any users for the snapshot mode
>> or non-blocking faults, and I don't see any in the immediate pipeline
>> either.
> 
> If this code was production ready I'd use it in ODP right away.
> 
> When we first create a ODP MR we'd want to snapshot to pre-load the
> NIC tables with something other than page fault, but not fault
> anything.
> 
> This would be a big performance improvement for ODP.
> 
> Jason
>