Re: [PATCH v2 0/5] Test overlayfs readdir cache

[Amir Goldstein <amir73il@gmail.com>]

On Mon, Apr 26, 2021 at 1:07 PM Miklos Szeredi <miklos@szeredi.hu> wrote:
>
> On Sun, Apr 25, 2021 at 9:14 AM Amir Goldstein <amir73il@gmail.com> wrote:
> >
> > Eryu,
> >
> > This extends the generic t_dir_offset2 helper program to verify
> > some cases of missing/stale entries and adds a new generic test which
> > passes on overlayfs (and other fs) on upstream kernel.
> >
> > The overlayfs specific test fails on upstream kernel and the fix commit
> > is currently in linux-next.  As usual, you may want to wait with merging
> > until the fix commit hits upstream.
> >
> > Based on feedback from Miklos, I changed the test to check for the
> > missing/stale entries on a new fd, while old fd is kept open, because
> > POSIX allows for stale/missing entries in the old fd.
> >
> > I was looking into another speculated bug in overlayfs which involves
> > multiple calls to getdents.  Although it turned out that overlayfs does
> > not have the speculated bug, I left both generic and overlay test with
> > multiple calls to getdents in order to excersize the relevant code.
> >
> > The attached patch was used to verify that the overlayfs test excercises
> > the call to ovl_cache_update_ino() with stale entries.
> > Overlayfs populates the merge dir readdir cache with a list of files in
> > the first getdents call, but updates d_ino of files on the list in
> > subsequent getdents calls.  By that time, the last entry is stale and the
> > following warning is printed (on linux-next with patch below applied):
> > [   ] overlayfs: failed to look up (m100) for ino (0)
> > [   ] overlayfs: failed to look up (f100) for ino (0)
> >
> > Miklos,
> >
> > Do you think it is worth the trouble to set p->is_whiteout and skip
> > dir_emit() in this case? and do we need to worry about lookup_one_len()
> > returning -ENOENT in this case?
>
> So lookup_one_len() first does a cached lookup, and if found returns
> that.  If not then it calls the filesystem's ->lookup() callback,
> which in this case is ovl_lookup().  AFAICS ovl_lookup() will never
> return ENOENT, even if the underlying filesystem does.
>
> Which means it's not necessary to worry about that case.
>
> The other case you found it that in case of a stale direntry the i_ino
> update will be skipped and so it will return an inconsistent result,
> right?

Right. It returns a stale entry with the old real ino.
Not sure if that is an "inconsistent" result.
inconsistent w.r.t what?

> Fixing that looks worthwhile, yes.
>

Will look into it.

Thanks,
Amir.