Re: [git pull] vfs.git spurious ENOTDIR fix

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

On Sat, Aug 01, 2015 at 09:06:55PM -0700, Hugh Dickins wrote:

> (I don't actually understand why the clearing of DCACHE_ENTRY_TYPE in
> dentry_iput() is not of continuing concern; but don't worry, there's
> plenty I don't understand - so long as you're both satisfied that
> it's not a concern, no need to persuade me.)

Because before we even get to dentry_iput(), we evict the fucker from hash.
And that will do dentry_rcuwalk_invalidate(dentry), which will bump ->d_seq
*AFTER* having it unhashed.  Now look at __d_lookup_rcu() - there we fetch
->d_seq, then verify that it's still hashed.

So having hit dentry_iput() means that everyone who'd found it via RCU
lookup will be guaranteed a ->d_seq mismatch.  The same goes for things
like d_drop() and d_instantiate().  Look for dentry_rcuwalk_invalidate()
callers in there...

Clearing DCACHE_ENTRY_TYPE there is fine - dentry *is* made negative there,
after all.  What we want is to have ->d_inode stable at least as long as
->d_seq remains so.