On Fri, Feb 03, 2023 at 08:00:16AM -0500, Brian Foster wrote:
> On Thu, Feb 02, 2023 at 08:44:23PM +0000, Matthew Wilcox (Oracle) wrote:
> > POSIX requires that "If the file size is increased, the extended area
> > shall appear as if it were zero-filled". It is possible to use mmap to
> > write past EOF and that data will become visible instead of zeroes.
> > This fixes the problem for the filesystems which simply call
> > truncate_setsize(). More complex filesystems will need their own
> > patches.
> >
> > Signed-off-by: Matthew Wilcox (Oracle) <willy@xxxxxxxxxxxxx>
> > ---
> > mm/truncate.c | 7 +++++--
> > 1 file changed, 5 insertions(+), 2 deletions(-)
> >
> > diff --git a/mm/truncate.c b/mm/truncate.c
> > index 7b4ea4c4a46b..cebfc5415e9a 100644
> > --- a/mm/truncate.c
> > +++ b/mm/truncate.c
> > @@ -763,9 +763,12 @@ void truncate_setsize(struct inode *inode, loff_t newsize)
> > loff_t oldsize = inode->i_size;
> >
> > i_size_write(inode, newsize);
> > - if (newsize > oldsize)
> > + if (newsize > oldsize) {
> > pagecache_isize_extended(inode, oldsize, newsize);
> > - truncate_pagecache(inode, newsize);
> > + truncate_pagecache(inode, oldsize);
> > + } else {
> > + truncate_pagecache(inode, newsize);
> > + }
>
> I don't think this alone quite addresses the problem. Looking at ext4
> for example, if the eof page is dirty and writeback occurs between the
> i_size update (because writeback also zeroes the post-eof portion of the
> page) and the truncate_setsize() call, we end up with pagecache
> inconsistency because pagecache truncate doesn't dirty the page it
> zeroes.
>
> So for example, with this series plus a nefariously placed
> filemap_flush() in ext4_setattr():
>
> # xfs_io -fc "truncate 1" -c "mmap 0 1k" -c "mwrite 0 10" -c "truncate 5" -c "mread -v 0 5" /mnt/file
> 00000000: 58 00 00 00 00 X....
> # umount /mnt/; mount <dev> /mnt/
> # xfs_io -c "mmap 0 1k" -c "mread -v 0 5" /mnt/file
> 00000000: 58 58 58 58 58 XXXXX
Hm, so switch the order of i_size_write() and truncate_pagecache()?
There could still be a store between old-EOF and new-EOF from another
thread, which would then be visible, but I don't think you could prove
that store should have been zeroed. Not from the thread doing the
ftruncate() anyway -- I think the thread doing the store could prove
it, but that thread is relying on undefined behaviour anyway.
