Re: [git pull] more vfs bits

From: Linus Torvalds
Date: Sat Feb 21 2015 - 21:20:03 EST

On Sat, Feb 21, 2015 at 6:02 PM, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> I'm somewhat tempted to do this:
> fs_inode -> d_inode
> fs_inode_once ->d_inode_rcu (it's not quite ->d_revalidate()-only, there's
> a bit in autofs ->d_manage() as well)

Ok, those at least match our existing naming logic (ie "d_inode()"
would match what we did to "d_count()").

I'm not sure about d_inode_rcu(), for the simple reason that even when
we're doing RCU walking, most of the time we have *not* used the
ACCESS_ONCE() model, we instead end up just using the regular d_inode
and then check the sequence count.

I think.

So the ACCESS_ONCE() thing is more special than just "done under RCU".
It's more like "really special case done without any of the normal
locking _or_ any of the normal RCU checks".

That said, the overhead of using ACCESS_ONCE() is basically nil, so
it's not like we couldn't just start doing more of them, and make it
be more of a "any time we're under RCU" kind of thing.

> dentry_inode -> something. d_opened_inode() might do, but I'm not sure -
> still sounds a bit wrong to me. What it's about is "the actual fs object
> behind this name, maybe from upper fs, maybe showing through from underlying
> layer"

Yeah, I think "d_backing_store_inode()" would probably be more along
the lines, but that's a mouthful. Maybe shortened to

To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at
Please read the FAQ at