Re: [PATCH v2 4/6] mm/migrate_device: convert __migrate_device_pages() to folios

[Alistair Popple <apopple@nvidia.com>]

Matthew Wilcox <willy@infradead.org> writes:

> On Fri, Feb 16, 2024 at 02:00:31PM -0800, Sidhartha Kumar wrote:
>> On 2/16/24 1:55 PM, Matthew Wilcox wrote:
>> > On Fri, Feb 16, 2024 at 01:13:18PM -0800, Sidhartha Kumar wrote:
>> > > Use migrate_pfn_to_folio() so we can work with folios directly in
>> > > __migrate_device_pages().
>> > 
>> > i don't understand why this would be correct if we have multipage
>> > folios.
>> > 
>> 
>> Alistair mentioned that he is working on order > 0 device page support so I
>> was under the impression that currently device pages are only order 0.

Right, at the moment we only create order 0 device private pages.

> That might well be true, but I'm *very* uncomfortable with doing a folio
> conversion in core MM that won't work with large folios.  We need to
> consider what will happen with large device folios.

Yes. Perhaps it would be better to post these fixes as part of a series
introducing multi-page folio support for deivce private pages after
all. I don't think we can address your other comments here without first
describing how large folios for device private pages would work, and the
best way to describe that is with the patches.

I doubt I will get those posted this week, but will aim to do so in the
next 2-3 weeks.

> (for filesystems, I am less bothered.  Individual filesystems control
> whether they see large folios or not, and for lesser filesystems it may
> never be worth converting them to support large folios)