On Wed, Nov 13, 2019 at 09:01:36AM +0200, Amir Goldstein wrote:
> > - if (d_really_is_negative(lower_dentry)) {
> > + /*
> > + * negative dentry can go positive under us here - its parent is not
> > + * locked. That's OK and that could happen just as we return from
> > + * ecryptfs_lookup() anyway. Just need to be careful and fetch
> > + * ->d_inode only once - it's not stable here.
> > + */
> > + lower_inode = READ_ONCE(lower_dentry->d_inode);
> > +
> > + if (!lower_inode) {
> > /* We want to add because we couldn't find in lower */
> > d_add(dentry, NULL);
> > return NULL;
>
> Sigh!
>
> Open coding a human readable macro to solve a subtle lookup race.
> That doesn't sound like a scalable solution.
> I have a feeling this is not the last patch we will be seeing along
> those lines.
>
> Seeing that developers already confused about when they should use
> d_really_is_negative() over d_is_negative() [1] and we probably
> don't want to add d_really_really_is_negative(), how about
> applying that READ_ONCE into d_really_is_negative() and
> re-purpose it as a macro to be used when races with lookup are
> a concern?
Would you care to explain what that "fix" would've achieved here,
considering the fact that barriers are no-ops on UP and this is
*NOT* an SMP race?
And it's very much present on UP - we have
fetch ->d_inode into local variable
do blocking allocation
check if ->d_inode is NULL now
if it is not, use the value in local variable and expect it to be non-NULL
That's not a case of missing barriers. At all. And no redefinition of
d_really_is_negative() is going to help - it can't retroactively affect
the value explicitly fetched into a local variable some time prior to
that.
There are other patches dealing with ->d_inode accesses, but they are
generally not along the same lines. The problem is rarely the same...
