Re: [PATCH] VFS: Add back check for !inode in walk_component()

[Al Viro <viro@ZenIV.linux.org.uk>]

On Thu, May 07, 2015 at 12:52:41PM -0400, Steven Rostedt wrote:
> 
> Commit 698934df8b45 "VFS: Combine inode checks with d_is_negative() and
> d_is_positive() in pathwalk" removed a check for inode being NULL in
> walk_component() where the type is tested. Stressing my tracefs create
> and remove instances while reading the files now triggers this:

So you get NULL ->d_inode with stale flags?  The thing is, ->d_inode
becoming NULL should happen via d_delete(), which goes throug this:
        unsigned flags = READ_ONCE(dentry->d_flags);

        flags &= ~(DCACHE_ENTRY_TYPE | DCACHE_FALLTHRU);
        WRITE_ONCE(dentry->d_flags, flags);
        smp_wmb();
        dentry->d_inode = NULL;

and after that assignment to ->d_flags you'll see d_is_negative() being
true.  OTOH, we have
                *inode = dentry->d_inode;
                if (read_seqcount_retry(&dentry->d_seq, seq))
in lookup_fast(), and read_seqcount_retry() is
{
        smp_rmb();
        return __read_seqcount_retry(s, start);
}

IOW, we have smp_rmb() between fetching ->d_inode and checking ->d_flags.

If you can reproduce that at will, could you make it dump nd->flags along with
dentry involved?