On Wed, Apr 15, 2020 at 01:25:37PM -0700, Ira Weiny wrote:
> This fundamentally changes the behavior from forcing the dax mode to be the
> same across the remount to only failing if we are going from non-dax to dax,
> adding -o dax on the remount?
>
> But going from -o dax to 'not -o dax' would be ok?
>
> FWIW after thinking about it some I _think_ it would be ok to allow the dax
> mode to change on a remount and let the inodes in memory stay in the mode they
> are at. And newly loaded inodes would get the new mode... Unfortunately
> without the STATX patch I have proposed the user does not have any way of
> knowing which files are in which mode.
We don't currently support mount -o nodax. So the intention of the
current code is that the dax mode can't change in either direction
(enabling or disabling) as a remount option.
The syzkaller report was because changing dax mode racing with other
operations given the current code base, could result in a kernel OOPS.
So we *do* need to rule it out at least for now.
I certainly don't object to allowing changing dax mode as a remount
--- so long as we have tests to make sure that if we stress opening,
reading, writing, mmap'ing files, etc., while another thread is
flipping back and forth between dax=never and dax=always is mount -o
remount --- and make sure that we don't end up crashing.
And this test needs to be in xfstests, because trying to figure out
what triggers a syzkaller failures in file system land is a pain in
the *ss so we really want a dedicated xfstests for this case. Have
you tested your patch series to make sure we don't have some potential
races here?
- Ted
