On Wed 22-12-21 10:40:18, Linus Torvalds wrote:

On Wed, Dec 22, 2021 at 8:08 AM Jan Kara [off-list ref] wrote:

quoted

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).

Well, O_DIRECT reads do, but not necessarily writes.

I agree.

And hey, even reads have been dodgy in the past when we didn't really
have the pinning logic - there's been a lot of users that just wanted
it to work for their particular use-case rather than in general and in
all situations..

Yes, but currently a malicious user can take the system down (BUG_ON) or
cause DIF/DIX failures if he is nasty and tries hard enough with O_DIRECT
reads (practically, the window is small so I haven't really seen a report
that I could trace to O_DIRECT reads but in principle the problem is the
same as with pinning & dirtying done e.g. by video capture drivers and
there we've seen these problem happen). So forcing pinning for O_DIRECT
reads is IMO mandatory.

								Honza

-- 
Jan Kara [off-list ref]
SUSE Labs, CR