Re: [PATCH v1 06/11] mm: support GUP-triggered unsharing via FAULT_FLAG_UNSHARE (!hugetlb)

[Jan Kara]

On Wed 22-12-21 15:48:34, David Hildenbrand wrote:
> On 22.12.21 15:42, Jan Kara wrote:
> > On Wed 22-12-21 14:09:41, David Hildenbrand wrote:
> >>>> IIUC, our COW logic makes sure that a shared anonymous page that might
> >>>> still be used by a R/O FOLL_GET cannot be modified, because any attempt
> >>>> to modify it would result in a copy.
> >>>
> >>> Well, we defined FOLL_PIN to mean the intent that the caller wants to access
> >>> not only page state (for which is enough FOLL_GET and there are some users
> >>> - mostly inside mm - who need this) but also page data. Eventually, we even
> >>> wanted to make FOLL_GET unavailable to broad areas of kernel (and keep it
> >>> internal to only MM for its dirty deeds ;)) to reduce the misuse of GUP.
> >>>
> >>> For file pages we need this data vs no-data access distinction so that
> >>> filesystems can detect when someone can be accessing page data although the
> >>> page is unmapped.  Practically, filesystems care most about when someone
> >>> can be *modifying* page data (we need to make sure data is stable e.g. when
> >>> writing back data to disk or doing data checksumming or other operations)
> >>> so using FOLL_GET when wanting to only read page data should be OK for
> >>> filesystems but honestly I would be reluctant to break the rule of "use
> >>> FOLL_PIN when wanting to access page data" to keep things simple and
> >>> reasonably easy to understand for parties such as filesystem developers or
> >>> driver developers who all need to interact with pinned pages...
> >>
> >> Right, from an API perspective we really want people to use FOLL_PIN.
> >>
> >> To optimize this case in particular it would help if we would have the
> >> FOLL flags on the unpin path. Then we could just decide internally
> >> "well, short-term R/O FOLL_PIN can be really lightweight, we can treat
> >> this like a FOLL_GET instead". And we would need that as well if we were
> >> to keep different counters for R/O vs. R/W pinned.
> > 
> > Well, I guess the question here is: Which GUP user needs only R/O access to
> > page data and is so performance critical that it would be worth it to
> > sacrifice API clarity for speed? I'm not aware of any but I was not looking
> > really hard...
> 
> I'd be interested in examples as well. Maybe databases that use O_DIRECT
> after fork()?

Well, but O_DIRECT reads must use FOLL_PIN in any case because they modify
page data (and so we need to detect them both for COW and filesystem needs).
O_DIRECT writes could use FOLL_GET but at this point I'm not convinced it
is worth it.

								Honza
-- 
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR