Received-SPF: pass (google.com: domain of linux-nfs-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Content-Type: text/plain; charset="utf-8"
Content-Transfer-Encoding: 7bit
MIME-Version: 1.0
From:   "NeilBrown" <neilb@suse.de>
To:     "Al Viro" <viro@zeniv.linux.org.uk>
Cc:     "Christian Brauner" <brauner@kernel.org>,
        "Jeff Layton" <jlayton@kernel.org>,
        "Dave Wysochanski" <dwysocha@redhat.com>,
        "linux-fsdevel" <linux-fsdevel@vger.kernel.org>,
        "linux-nfs" <linux-nfs@vger.kernel.org>,
        "David Howells" <dhowells@redhat.com>,
        "Christoph Hellwig" <hch@lst.de>
Subject: Re: [PATCH/RFC] VFS: LOOKUP_MOUNTPOINT should used cached info
 whenever possible.
In-reply-to: <20230420213529.GS3390869@ZenIV>
References: <95ee689c76bf034fa2fe9fade0bccdb311f3a04f.camel@kernel.org>,
 <168168683217.24821.6260957092725278201@noble.neil.brown.name>,
 <20230417-beisein-investieren-360fa20fb68a@brauner>,
 <168176679417.24821.211742267573907874@noble.neil.brown.name>,
 <20230420213529.GS3390869@ZenIV>
Date:   Fri, 21 Apr 2023 08:01:09 +1000
Message-id: <168202806952.24821.15445938161479912532@noble.neil.brown.name>
Precedence: bulk

On Fri, 21 Apr 2023, Al Viro wrote:
> On Tue, Apr 18, 2023 at 07:26:34AM +1000, NeilBrown wrote:
> 
> > MNT_FORCE is, I think, a good idea and a needed functionality that has
> > never been implemented well.
> > MNT_FORCE causes nfs_umount_begin to be called as you noted, which
> > aborts all pending RPCs on that filesystem.
> 
> Suppose it happens to be mounted in another namespace as well.  Or bound
> at different mountpoint, for that matter.  What should MNT_FORCE do?
> 

1/ set a "forced-unmount" flag on the vfs_mount which causes any syscall
   that uses the vfsmount (whether from an fd, or found in a path walk,
   or elsewhere), except for close(), to abort with an error;
2/ call ->umount_begin passing in the vfs_mount.  The fs can abort any
   outstanding transaction on any fd from that vfs_mount.   Possibly
   it might instead abort a wait rather than the whole transaction,
   particularly if requests using some other vfs_mount might also be
   interested in the transaction
3/ ->close() on a force-unmount vfs_mount would clean up without
   blocking indefinitely, discarding dirty data if necessary.

This still depends on the application to close all fds that return
errors (and to chdir out of a problematic directory).  But at least it
*allows* applications to do that without requiring that they be killed.

Thanks,
NeilBrown