On Wed, Nov 22, 2023 at 04:18:56PM -0800, Linus Torvalds wrote: > On Wed, 22 Nov 2023 at 13:19, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote: > > > > The serious gap, AFAICS, is the interplay with open-by-fhandle. > > So I'm obviously not a fan of igncase filesystems, but I don't think > this series actually changes any of that. > > > It's not unfixable, but we need to figure out what to do when > > lookup runs into a disconnected directory alias. d_splice_alias() > > will move it in place, all right, but any state ->lookup() has > > hung off the dentry that had been passed to it will be lost. > > I guess this migth be about the new DCACHE_CASEFOLDED_NAME bit. > > At least for now, that is only used by generic_ci_d_revalidate() for > negative dentries, so it shouldn't matter for that d_splice_alias() > that only does positive dentries. No? > > Or is there something else you worry about? Dentries created by d_obtain_alias() will never go anywhere near generic_set_encrypted_ci_d_ops(). They do *not* get ->d_op set that way. When ext4_lookup() does a lookup in c-i directory it does have ->d_op set on dentry it got from the caller. Which is promptly discarded when d_splice_alias() finds a preexisting alias for it. Positive dentries eventually become negative; not invalidating them when that happens is a large part of the point of this series. ->d_revalidate() is taught to check if they are marked with that bit, but... they won't have that ->d_revalidate() associated with them, will they? ->d_hash() and ->d_compare() come from the parent, but ->d_revalidate() comes from dentry itself. In other words, generic_ci_d_revalidate() won't see the lack of that bit on dentry, etc. - it won't be called for that dentry in the first place.
