On Tuesday 08 July 2003 01:00, Andrew Morton wrote:
> Andrey Borzenkov <arvidjaar@xxxxxxx> wrote:
> > I finally hit a painfully trivial way to reproduce another long standing
> > devfs problem - deadlock between devfs_lookup and
> > devfs_d_revalidate_wait.
>
> uh.
>
> > The current fix is to move re-acquire of i_sem after all
> > devfs_d_revalidate_wait waiters have been waked up.
>
> Directory modifications appear to be under write_lock(&dir->u.dir.lock); so
> that obvious problem is covered. Your patch might introduce a race around
> _devfs_get_vfs_inode() - two CPUs running that against the same inode at
> the same time?
>
Actually it just makes it marginally more probable.
Normal open without O_CREATE runs ->d_revalidate outside of i_sem i.e. neither
devfs_lookup vs. devfs_d_revalidate_wait nor devfs_d_revalidate_wait vs.
itself are protected by i_sem and this is (should be) the most common case
for /dev access.
This race happens under non-trivial conditions. devfsd descendant (i.e. some
action) should be involved; and action triggered by devfs_lookup does not
race with it by definition because devfs_lookup waits for action to finish.
I.e. it needs another devfsd action that would access /dev entry after it
just has been created or two concurrent lookups in LOOKUP action itself.
Quite unlikely in real life and race window is very small.
I do not want to sound like it has to be ignored - but devfs code is so messy
that no trivial fix exists that would not make code even more messy. So I
would still apply original fixes and let this problem be solved later - it is
not so important as to delay two other.
-andrey
|