Re: [PATCH 1/2] fs/dcache: introduce d_alloc_parallel_check_existing

[Al Viro]

On Wed, Aug 21, 2024 at 07:22:39PM -0400, Gabriel Krisman Bertazi wrote:

> Would it be acceptable to just change the dentry->d_name here in a new
> flavor of d_add_ci used only by these filesystems? We are inside the
> creation path, so the dentry has never been hashed.  Concurrent lookups
> will be stuck in d_wait_lookup() until we are done and will never become
> invalid after the change because the lookup was already done
> case-insensitively, so they all match the same dentry, per-definition,
> and we know there is no other matching dentries in the directory.  We'd
> only need to be careful not to expose partial names to concurrent
> parallel lookups.

*Ow*

->d_name stability rules are already convoluted as hell; that would make
them even more painful.

What locking are you going to use there?