Re: [PATCH RFC 06/24] vfs: break parent dir delegations in open(..., O_CREAT) codepath
From: Jeff Layton <jlayton@kernel.org>
Date: 2024-03-17 12:23:33
Also in:
linux-cifs, linux-fsdevel, linux-nfs, linux-unionfs, lkml, netfs
On Sun, 2024-03-17 at 00:19 +0000, Al Viro wrote:
On Fri, Mar 15, 2024 at 12:52:57PM -0400, Jeff Layton wrote:quoted
In order to add directory delegation support, we need to break delegations on the parent whenever there is going to be a change in the directory. Add a delegated_inode parameter to lookup_open and have it break the delegation. Then, open_last_lookups can wait for the delegation break and retry the call to lookup_open once it's done.quoted
@@ -3490,6 +3490,11 @@ static struct dentry *lookup_open(struct nameidata *nd, struct file *file,Wait a sec - are you going to do anything to the atomic_open side of things?
Hmm good point. I was thinking that all of the filesystems that had atomic_open didn't support leases. I'm wrong though -- there are some that currently do: 9p: It's a network filesystem, and I don't think it has any sort of asynchronous notification or delegation-like object, does it? It might be best though to just make it call simple_nosetlease. fuse: fuse allows leases today. I doubt we can get away with turning that off now. There probably ought to be a way for the userland driver to opt-in or out of allowing built-in lease support maybe a flag or something? ntfs3: IDGI. Why does ntfs3 (which is a local filesystem, unless I'm mistaken) have an atomic_open? Shouldn't lookup+open be fine, like with most local filesystems? vboxsf: Probably the same situation as 9p. Can we just disable leases? I'll spin up a patchset soon to add proper setlease handlers to all of the above. Then we can then guard against allowing generic_setlease on filesystems by default on filesystems with an atomic_open handler. Another (maybe better) idea might be to require filesystems to specify a setlease handler if they want them enabled. We could just set the existing local filesystems to generic_setlease. That would make lease support a strictly opt-in thing, which is probably the best idea for avoiding surprises with them.
quoted
/* Negative dentry, just create the file */ if (!dentry->d_inode && (open_flag & O_CREAT)) { + /* but break the directory lease first! */ + error = try_break_deleg(dir_inode, delegated_inode); + if (error) + goto out_dput;
-- Jeff Layton [off-list ref]