Re: [RFC] metadata updates vs. fetches (was Re: [PATCH v4] fs: Fix data race in inode_set_ctime_to_ts)

[Matthew Wilcox <willy@infradead.org>]

On Sun, Nov 24, 2024 at 02:43:58PM -0800, Linus Torvalds wrote:
> On Sun, 24 Nov 2024 at 14:34, Matthew Wilcox <willy@infradead.org> wrote:
> >
> > Could we just do:
> >
> > again:
> >         nsec = READ_ONCE(inode->nsec)
> >         sec = READ_ONCE(inode->sec)
> >         if (READ_ONCE(inode->nsec) != nsec)
> >                 goto again;
> 
> No. You would need to use the right memory ordering barriers.
> 
> And make sure the writes are in the right order.
> 
> And even then it wouldn't protect against the race in theory, since
> two (separate) time writes could make that nsec check work, even when
> the 'sec' read wouldn't necessarily match *either* of the matching
> nsec cases.

But if we assume that time only goes forwards (ie nobody's calling
utime()), I don't think there's a sequence of updates which let you see
a file time which is newer than the actual time of the file.  I tried
to construct an example, and I couldn't.  eg:

A:	WRITE_ONCE(inode->sec, 5)
A:	WRITE_ONCE(inode->nsec, 950)
A:	WRITE_ONCE(inode->sec, 6)
B:	READ_ONCE(inode->nsec)
B:	READ_ONCE(inode->sec)
A:	WRITE_ONCE(inode->sec, 170)
A:	WRITE_ONCE(inode->sec, 7)
A:	WRITE_ONCE(inode->sec, 950)
B:	READ_ONCE(inode->nsec)

Now we have a time of 6:950 which is never a time that this file had,
but it's intermediate in time between two times that the file _did_
have, so it won't break make.

Or did I not try hard enough to construct a counterexample that
would break make?

(assume the appropriate read/write barriers are in there)