[PATCH] ovl: make private mounts longterm

[PATCH] ovl: make private mounts longterm

[Miklos Szeredi]

Overlayfs is using clone_private_mount() to create internal mounts for
underlying layers.  These are used for operations requiring a path, such as
dentry_open().

Since these private mounts are not in any namespace they are treated as
short term, "detached" mounts and mntput() involves taking the global
mount_lock, which can result in serious cacheline pingpong.

Make these private mounts longterm instead, which trade the penalty on
mntput() for a slightly longer shutdown time due to an added RCU grace
period when putting these mounts.

Introduce a new helper kern_unmount_many() that can take care of multiple
longterm mounts with a single RCU grace period.

Cc: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Miklos Szeredi <mszeredi@redhat.com>
---
 fs/namespace.c        | 16 ++++++++++++++++
 fs/overlayfs/super.c  | 19 ++++++++++++++-----
 include/linux/mount.h |  2 ++
 3 files changed, 32 insertions(+), 5 deletions(-)

diff --git a/fs/namespace.c b/fs/namespace.c
index a28e4db075ed..5d16d87b6b8b 100644
--- a/fs/namespace.c
+++ b/fs/namespace.c
@@ -1879,6 +1879,9 @@ struct vfsmount *clone_private_mount(const struct path *path)
 	if (IS_ERR(new_mnt))
 		return ERR_CAST(new_mnt);
 
+	/* Longterm mount to be removed by kern_unmount*() */
+	new_mnt->mnt_ns = MNT_NS_INTERNAL;
+
 	return &new_mnt->mnt;
 }
 EXPORT_SYMBOL_GPL(clone_private_mount);
@@ -3804,6 +3807,19 @@ void kern_unmount(struct vfsmount *mnt)
 }
 EXPORT_SYMBOL(kern_unmount);
 
+void kern_unmount_many(struct vfsmount *mnt[], unsigned int num)
+{
+	unsigned int i;
+
+	for (i = 0; i < num; i++)
+		if (mnt[i])
+			real_mount(mnt[i])->mnt_ns = NULL;
+	synchronize_rcu_expedited();
+	for (i = 0; i < num; i++)
+		mntput(mnt[i]);
+}
+EXPORT_SYMBOL(kern_unmount_many);
+
 bool our_mnt(struct vfsmount *mnt)
 {
 	return check_mnt(real_mount(mnt));
diff --git a/fs/overlayfs/super.c b/fs/overlayfs/super.c
index 60dfb27bc12b..a938dd2521b2 100644
--- a/fs/overlayfs/super.c
+++ b/fs/overlayfs/super.c
@@ -225,12 +225,21 @@ static void ovl_free_fs(struct ovl_fs *ofs)
 	dput(ofs->workbasedir);
 	if (ofs->upperdir_locked)
 		ovl_inuse_unlock(ofs->upper_mnt->mnt_root);
-	mntput(ofs->upper_mnt);
-	for (i = 1; i < ofs->numlayer; i++) {
-		iput(ofs->layers[i].trap);
-		mntput(ofs->layers[i].mnt);
+
+	if (!ofs->layers) {
+		/* Deal with partial setup */
+		kern_unmount(ofs->upper_mnt);
+	} else {
+		/* Hack!  Reuse ofs->layers as a mounts array */
+		struct vfsmount **mounts = (struct vfsmount **) ofs->layers;
+
+		for (i = 0; i < ofs->numlayer; i++) {
+			iput(ofs->layers[i].trap);
+			mounts[i] = ofs->layers[i].mnt;
+		}
+		kern_unmount_many(mounts, ofs->numlayer);
+		kfree(ofs->layers);
 	}
-	kfree(ofs->layers);
 	for (i = 0; i < ofs->numfs; i++)
 		free_anon_bdev(ofs->fs[i].pseudo_dev);
 	kfree(ofs->fs);
diff --git a/include/linux/mount.h b/include/linux/mount.h
index bf8cc4108b8f..e3e994bfcecb 100644
--- a/include/linux/mount.h
+++ b/include/linux/mount.h
@@ -109,4 +109,6 @@ extern unsigned int sysctl_mount_max;
 
 extern bool path_is_mountpoint(const struct path *path);
 
+extern void kern_unmount_many(struct vfsmount *mnt[], unsigned int num);
+
 #endif /* _LINUX_MOUNT_H */
-- 
2.21.1

Re: [PATCH] ovl: make private mounts longterm

[Al Viro]

On Fri, May 22, 2020 at 10:57:23AM +0200, Miklos Szeredi wrote:
> Overlayfs is using clone_private_mount() to create internal mounts for
> underlying layers.  These are used for operations requiring a path, such as
> dentry_open().
> 
> Since these private mounts are not in any namespace they are treated as
> short term, "detached" mounts and mntput() involves taking the global
> mount_lock, which can result in serious cacheline pingpong.
> 
> Make these private mounts longterm instead, which trade the penalty on
> mntput() for a slightly longer shutdown time due to an added RCU grace
> period when putting these mounts.
> 
> Introduce a new helper kern_unmount_many() that can take care of multiple
> longterm mounts with a single RCU grace period.

Umm...

1) Documentation/filesystems/porting - something along the lines
of "clone_private_mount() returns a longterm mount now, so the proper
destructor of its result is kern_unmount()"

2) the name kern_unmount_many() has an unfortunate clash with
fput_many(), with arguments that look similar and mean something
entirely different.  How about kern_unmount_array()?

3)
> -	mntput(ofs->upper_mnt);
> -	for (i = 1; i < ofs->numlayer; i++) {
> -		iput(ofs->layers[i].trap);
> -		mntput(ofs->layers[i].mnt);
> +
> +	if (!ofs->layers) {
> +		/* Deal with partial setup */
> +		kern_unmount(ofs->upper_mnt);
> +	} else {
> +		/* Hack!  Reuse ofs->layers as a mounts array */
> +		struct vfsmount **mounts = (struct vfsmount **) ofs->layers;
> +
> +		for (i = 0; i < ofs->numlayer; i++) {
> +			iput(ofs->layers[i].trap);
> +			mounts[i] = ofs->layers[i].mnt;
> +		}
> +		kern_unmount_many(mounts, ofs->numlayer);
> +		kfree(ofs->layers);

That's _way_ too subtle.  AFAICS, you rely upon ->upper_mnt == ->layers[0].mnt,
->layers[0].trap == NULL, without even mentioning that.  And the hack you do
mention...  Yecchhh...  How many layers are possible, again?

Re: [PATCH] ovl: make private mounts longterm

[Miklos Szeredi]

On Fri, May 22, 2020 at 6:08 PM Al Viro <viro@zeniv.linux.org.uk> wrote:
>
> On Fri, May 22, 2020 at 10:57:23AM +0200, Miklos Szeredi wrote:
> > Overlayfs is using clone_private_mount() to create internal mounts for
> > underlying layers.  These are used for operations requiring a path, such as
> > dentry_open().
> >
> > Since these private mounts are not in any namespace they are treated as
> > short term, "detached" mounts and mntput() involves taking the global
> > mount_lock, which can result in serious cacheline pingpong.
> >
> > Make these private mounts longterm instead, which trade the penalty on
> > mntput() for a slightly longer shutdown time due to an added RCU grace
> > period when putting these mounts.
> >
> > Introduce a new helper kern_unmount_many() that can take care of multiple
> > longterm mounts with a single RCU grace period.
>
> Umm...
>
> 1) Documentation/filesystems/porting - something along the lines
> of "clone_private_mount() returns a longterm mount now, so the proper
> destructor of its result is kern_unmount()"
>
> 2) the name kern_unmount_many() has an unfortunate clash with
> fput_many(), with arguments that look similar and mean something
> entirely different.  How about kern_unmount_array()?
>
> 3)
> > -     mntput(ofs->upper_mnt);
> > -     for (i = 1; i < ofs->numlayer; i++) {
> > -             iput(ofs->layers[i].trap);
> > -             mntput(ofs->layers[i].mnt);
> > +
> > +     if (!ofs->layers) {
> > +             /* Deal with partial setup */
> > +             kern_unmount(ofs->upper_mnt);
> > +     } else {
> > +             /* Hack!  Reuse ofs->layers as a mounts array */
> > +             struct vfsmount **mounts = (struct vfsmount **) ofs->layers;
> > +
> > +             for (i = 0; i < ofs->numlayer; i++) {
> > +                     iput(ofs->layers[i].trap);
> > +                     mounts[i] = ofs->layers[i].mnt;
> > +             }
> > +             kern_unmount_many(mounts, ofs->numlayer);
> > +             kfree(ofs->layers);
>
> That's _way_ too subtle.  AFAICS, you rely upon ->upper_mnt == ->layers[0].mnt,
> ->layers[0].trap == NULL, without even mentioning that.  And the hack you do
> mention...  Yecchhh...  How many layers are possible, again?

500, mounts array would fit inside a page and a page can be allocated
with __GFP_NOFAIL. But why bother?  It's not all that bad, is it?

Thanks,
Miklos

Re: [PATCH] ovl: make private mounts longterm

[Amir Goldstein]

> > > -     mntput(ofs->upper_mnt);
> > > -     for (i = 1; i < ofs->numlayer; i++) {
> > > -             iput(ofs->layers[i].trap);
> > > -             mntput(ofs->layers[i].mnt);
> > > +
> > > +     if (!ofs->layers) {
> > > +             /* Deal with partial setup */
> > > +             kern_unmount(ofs->upper_mnt);
> > > +     } else {
> > > +             /* Hack!  Reuse ofs->layers as a mounts array */
> > > +             struct vfsmount **mounts = (struct vfsmount **) ofs->layers;
> > > +
> > > +             for (i = 0; i < ofs->numlayer; i++) {
> > > +                     iput(ofs->layers[i].trap);
> > > +                     mounts[i] = ofs->layers[i].mnt;
> > > +             }
> > > +             kern_unmount_many(mounts, ofs->numlayer);
> > > +             kfree(ofs->layers);
> >
> > That's _way_ too subtle.  AFAICS, you rely upon ->upper_mnt == ->layers[0].mnt,
> > ->layers[0].trap == NULL, without even mentioning that.  And the hack you do
> > mention...  Yecchhh...  How many layers are possible, again?
>
> 500, mounts array would fit inside a page and a page can be allocated
> with __GFP_NOFAIL. But why bother?  It's not all that bad, is it?

FWIW, it seems fine to me.
We can transfer the reference from upperdir_trap to layers[0].trap
when initializing layers[0] for the sake of clarity.

Thanks,
Amir.

Re: [PATCH] ovl: make private mounts longterm

[Miklos Szeredi]

On Fri, May 22, 2020 at 7:02 PM Amir Goldstein <amir73il@gmail.com> wrote:
>
> > > > -     mntput(ofs->upper_mnt);
> > > > -     for (i = 1; i < ofs->numlayer; i++) {
> > > > -             iput(ofs->layers[i].trap);
> > > > -             mntput(ofs->layers[i].mnt);
> > > > +
> > > > +     if (!ofs->layers) {
> > > > +             /* Deal with partial setup */
> > > > +             kern_unmount(ofs->upper_mnt);
> > > > +     } else {
> > > > +             /* Hack!  Reuse ofs->layers as a mounts array */
> > > > +             struct vfsmount **mounts = (struct vfsmount **) ofs->layers;
> > > > +
> > > > +             for (i = 0; i < ofs->numlayer; i++) {
> > > > +                     iput(ofs->layers[i].trap);
> > > > +                     mounts[i] = ofs->layers[i].mnt;
> > > > +             }
> > > > +             kern_unmount_many(mounts, ofs->numlayer);
> > > > +             kfree(ofs->layers);
> > >
> > > That's _way_ too subtle.  AFAICS, you rely upon ->upper_mnt == ->layers[0].mnt,
> > > ->layers[0].trap == NULL, without even mentioning that.  And the hack you do
> > > mention...  Yecchhh...  How many layers are possible, again?
> >
> > 500, mounts array would fit inside a page and a page can be allocated
> > with __GFP_NOFAIL. But why bother?  It's not all that bad, is it?
>
> FWIW, it seems fine to me.
> We can transfer the reference from upperdir_trap to layers[0].trap
> when initializing layers[0] for the sake of clarity.

Right, we should just get rid of ofs->upper_mnt and ofs->upperdir_trap
and use ofs->layers[0] to store those.

Thanks,
Miklos

Re: [PATCH] ovl: make private mounts longterm

[Al Viro]

On Fri, May 22, 2020 at 08:53:49PM +0200, Miklos Szeredi wrote:
> On Fri, May 22, 2020 at 7:02 PM Amir Goldstein <amir73il@gmail.com> wrote:
> >
> > > > > -     mntput(ofs->upper_mnt);
> > > > > -     for (i = 1; i < ofs->numlayer; i++) {
> > > > > -             iput(ofs->layers[i].trap);
> > > > > -             mntput(ofs->layers[i].mnt);
> > > > > +
> > > > > +     if (!ofs->layers) {
> > > > > +             /* Deal with partial setup */
> > > > > +             kern_unmount(ofs->upper_mnt);
> > > > > +     } else {
> > > > > +             /* Hack!  Reuse ofs->layers as a mounts array */
> > > > > +             struct vfsmount **mounts = (struct vfsmount **) ofs->layers;
> > > > > +
> > > > > +             for (i = 0; i < ofs->numlayer; i++) {
> > > > > +                     iput(ofs->layers[i].trap);
> > > > > +                     mounts[i] = ofs->layers[i].mnt;
> > > > > +             }
> > > > > +             kern_unmount_many(mounts, ofs->numlayer);
> > > > > +             kfree(ofs->layers);
> > > >
> > > > That's _way_ too subtle.  AFAICS, you rely upon ->upper_mnt == ->layers[0].mnt,
> > > > ->layers[0].trap == NULL, without even mentioning that.  And the hack you do
> > > > mention...  Yecchhh...  How many layers are possible, again?
> > >
> > > 500, mounts array would fit inside a page and a page can be allocated
> > > with __GFP_NOFAIL. But why bother?  It's not all that bad, is it?
> >
> > FWIW, it seems fine to me.
> > We can transfer the reference from upperdir_trap to layers[0].trap
> > when initializing layers[0] for the sake of clarity.
> 
> Right, we should just get rid of ofs->upper_mnt and ofs->upperdir_trap
> and use ofs->layers[0] to store those.

For that you'd need to allocate ->layers before you get to ovl_get_upper(),
though.  I'm not saying it's a bad idea - doing plain memory allocations
before anything else tends to make failure exits cleaner; it's just that
it'll take some massage.  Basically, do ovl_split_lowerdirs() early,
then allocate everything you need, then do lookups, etc., filling that
stuff.

Regarding this series - the points regarding the name choice and the
need to document the calling conventions change still remain.

Re: [PATCH] ovl: make private mounts longterm

[Miklos Szeredi]

On Fri, May 22, 2020 at 9:56 PM Al Viro <viro@zeniv.linux.org.uk> wrote:
>
> On Fri, May 22, 2020 at 08:53:49PM +0200, Miklos Szeredi wrote:

> > Right, we should just get rid of ofs->upper_mnt and ofs->upperdir_trap
> > and use ofs->layers[0] to store those.
>
> For that you'd need to allocate ->layers before you get to ovl_get_upper(),
> though.  I'm not saying it's a bad idea - doing plain memory allocations
> before anything else tends to make failure exits cleaner; it's just that
> it'll take some massage.  Basically, do ovl_split_lowerdirs() early,
> then allocate everything you need, then do lookups, etc., filling that
> stuff.

That was exactly the plan I set out.

> Regarding this series - the points regarding the name choice and the
> need to document the calling conventions change still remain.

Agreed.

Thanks,
Miklos