Re: [patch 0/4] uaccess: Provide and use helpers for user masked access

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

On Thu, Aug 21, 2025 at 09:45:22AM +0200, Christian Brauner wrote:
> On Thu, Aug 21, 2025 at 12:48:15AM +0100, Al Viro wrote:
> > On Tue, Aug 19, 2025 at 01:39:09AM +0100, Al Viro wrote:
> > > I'm still trying to come up with something edible for lock_mount() -
> > > the best approximation I've got so far is
> > > 
> > > 	CLASS(lock_mount, mp)(path);
> > > 	if (IS_ERR(mp.mp))
> > > 		bugger off
> > 
> > ... and that does not work, since DEFINE_CLASS() has constructor return
> > a value that gets copied into the local variable in question.
> > 
> > Which is unusable for situations when a part of what constructor is
> > doing is insertion of that local variable into a list.
> > 
> > __cleanup() per se is still usable, but... no DEFINE_CLASS for that kind
> > of data structures ;-/
> 
> Just add the custom infrastructure that we need for this to work out imho.

Obviously...  I'm going to put that into a branch on top of -rc3 and keep
the more infrastructural parts in the beginning, so they could be merged
into other branches in vfs/vfs.git without disrupting things on reordering.

> If it's useful outside of our own realm then we can add it to cleanup.h
> and if not we can just add our own header...

lock_mount() et.al. are purely fs/namespace.c, so no header is needed at
all.  FWIW, existing guards in there have problems - I ended up with

DEFINE_LOCK_GUARD_0(namespace_excl, namespace_lock(), namespace_unlock())
DEFINE_LOCK_GUARD_0(namespace_shared, down_read(&namespace_sem),
				      up_read(&namespace_sem))
in fs/namespace.c and
DEFINE_LOCK_GUARD_0(mount_writer, write_seqlock(&mount_lock),
		    write_sequnlock(&mount_lock))
DEFINE_LOCK_GUARD_0(mount_locked_reader, read_seqlock_excl(&mount_lock),
		    read_sequnlock_excl(&mount_lock))
in fs/mount.h; I'm doing conversions to those where they clearly are
good fit and documenting as I go.

mount_lock ones really should not be done in a blanket way - right
now they are wrong in quite a few cases, where writer is used instead
of the locked reader; we'll need to sort that out and I'd rather
keep the open-coded ones for the stuff yet to be considered and/or
tricky.

BTW, the comments I'm using for functions are along the lines of
 * locks: mount_locked_reader || namespace_shared && is_mounted(mnt)
this one - for is_path_reachable().  If you look through the comments
there you'll see things like "vfsmount lock must be held for write" and
the rwlock those are refering to had been gone for more than a decade...

DEFINE_LOCK_GUARD_0 vs. DEFINE_GUARD makes for saner code generation;
having it essenitally check IS_ERR_OR_NULL(&namespace_sem) is already
ridiculous, but when it decides to sacrifice a register for that, complete
with a bunch of spills...