On Wed, 2016-09-21 at 07:00 +0800, Ian Kent wrote: > On Wed, 2016-09-21 at 06:44 +0800, Ian Kent wrote: > > On Tue, 2016-09-20 at 11:50 -0500, Eric W. Biederman wrote: > > > Ian Kent <ra...@themaw.net> writes: > > > > > > > On Sat, 2016-09-17 at 22:10 +0200, Mateusz Guzik wrote: > > > > > On Wed, Sep 14, 2016 at 02:14:45PM +0800, Ian Kent wrote: > > > > > > If an automount mount is clone(2)ed into a file system that is > > > > > > propagation private, when it later expires in the originating > > > > > > namespace subsequent calls to autofs ->d_automount() for that > > > > > > dentry in the original namespace will return ELOOP until the > > > > > > mount is manually umounted in the cloned namespace. > > > > > > > > > > > > In the same way, if an autofs mount is triggered by automount(8) > > > > > > running within a container the dentry will be seen as mounted in > > > > > > the root init namespace and calls to ->d_automount() in that > > > > > > namespace > > > > > > will return ELOOP until the mount is umounted within the container. > > > > > > > > > > > > Also, have_submounts() can return an incorect result when a mount > > > > > > exists in a namespace other than the one being checked. > > > > > > > > > > > > @@ -460,7 +460,7 @@ static int autofs4_d_manage(struct dentry > > > > > > *dentry, > > > > > > bool > > > > > > rcu_walk) > > > > > > > > > > > > if (ino->flags & AUTOFS_INF_WANT_EXPIRE) > > > > > > return 0; > > > > > > - if (d_mountpoint(dentry)) > > > > > > + if (is_local_mountpoint(dentry)) > > > > > > return 0; > > > > > > inode = d_inode_rcu(dentry); > > > > > > if (inode && S_ISLNK(inode->i_mode)) > > > > > > > > > > This change is within RCU lookup. > > > > > > > > > > is_local_mountpoint may end up calling __is_local_mountpoint, which > > > > > will > > > > > optionally take the namespace_sem lock, resulting in a splat: > > > > > > > > Yes, that's a serious problem I missed. > > > > > > > > snip ... > > > > > > > > > I don't know this code. Perhaps it will be perfectly fine performance > > > > > wise > > > > > to > > > > > just drop out of RCU lookup in this case. > > > > > > > > It's a bit worse than that. > > > > > > > > I think being able to continue the rcu-walk for an already mounted > > > > dentry > > > > that > > > > is not being expired is an important part of the performance improvement > > > > given > > > > by the series that added this. > > > > > > > > Can you confirm that Neil? > > > > > > > > But for the case here the existing test can allow rcu-walk to continue > > > > for > > > > a > > > > dentry that would attempt to trigger an automount so it's also a bug in > > > > the > > > > existing code. > > > > > > I don't think the existing code is buggy. As I read __follow_mount_rcu > > > if DCACHE_NEED_AUTOMOUNT is set on the dentry after return from d_manage > > > the code will break out of the rcu walk. > > > > That's right. > > > > I'm working on follow up patches now. > > > > > > > > > Any thoughts on how we can handle this Neil, I'm having a bit of trouble > > > > working > > > > out how to resolve this one. > > > > > > I believe in this case d_mountpoint is enough for the rcu walk case. If > > > the mountpoint turns out not to be local __follow_mount_rcu will kick > > > out of the rcu walk as it will return false. Because: > > > return !mounted && !(path->dentry->d_flags & DCACHE_NEED_AUTOMOUNT); > > > > > > I am not quite certain about the non-rcu case. That can't be > > > is_local_mountpoint as that is inside a spinlock and is_local_mountpoint > > > can sleep. Perhaps d_mountpoint is simply correct for autofs4_d_manage. > > > > That's right, I think I have other mistakes there too. > > > > I'm checking those too. > > It may be that is_local_mountpoint() isn't the right thing to use for this. > > When I originally set out to do this I used a lookup_mnt() type check which > should be sufficient in this case so I might need to go back to that.
Eric, Mateusz, I appreciate your spending time on this and particularly pointing out my embarrassingly stupid is_local_mountpoint() usage mistake. Please accept my apology for the inconvenience. If all goes well (in testing) I'll have follow up patches to correct this fairly soon. Ian
