Re: [PATCH][RFC] get rid of busy-wait in shrink_dcache_tree()

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

On Sat, Jan 24, 2026 at 09:45:54AM -0800, Linus Torvalds wrote:
> On Fri, 23 Jan 2026 at 21:34, Al Viro <viro@zeniv.linux.org.uk> wrote:
> >
> > The only trouble is that as soon as some joker slaps __randomize_layout
> > on struct hlist_node they'll start flipping from sharing with ->next to
> > sharing with ->pprev, at random.
> 
> If somebody starts using randomize_layout on core data structures,
> they get what they deserve.
> 
> We have tons of data structures that are *NOT* randomizable.
> 
> In fact, RANDSTRUCT is so broken in general that we actually taint the
> kernel if you enable that crazy option in the first place. So no,
> "what if somebody enables it on random things" is not even remotely
> worth worrying about.

Very much agreed, but we *do* have that on e.g. struct path (two pointers),
as well as struct inode, struct file, struct mount, etc.  As far as VFS goes,
those are core data structures...

While we are at it, does anybody have objections to making dentry->d_u anonymous?
We are already using anonymous union members in struct dentry, so compiler support
is no longer a consideration.

Another thing in the same area:

#define for_each_alias(dentry, inode) \
	hlist_for_each_entry(dentry, &(inode)->i_dentry, d_u.d_alias)

to avoid the boilerplate.  Objections?