[PATCH] 2.5.72: follow_mount / follow_link

[PATCH] 2.5.72: follow_mount / follow_link

[Mike Waychison]

[-- Attachment #1: Type: text/plain, Size: 865 bytes --]

Hi Al,

When a follow_link dentry call is made, the implementation expects the 
dentry to be followed as well as a nameidata struct to be filled in. 
The received nd->mnt is expected to contain the vfsmount of the dentry 
being followed, so that a subsequent call to vfs_follow_link may 
properly pivot off that mount and onto another vfsmount as the path of 
the link is walked, thus keeping reference counts proper.

The changes made in fs/namei.c@1.42 break this behaviour iff the dentry 
being follow_link'ed is a root dentry.  This is because follow_mount 
follows down next.mnt and not nd->mnt like it used to.  So, if a root 
dentry has a follow_link op, the nd->mnt it receives is in fact the 
vfsmount of the mount it is mounted upon (which breaks reference counts).

Please apply the attached patch which was made against 2.5.72.

Thanks,

Mike Waychison

[-- Attachment #2: ndmnt.patch --]
[-- Type: text/plain, Size: 1178 bytes --]

# This is a BitKeeper generated patch for the following project:
# Project Name: Linux kernel tree
# This patch format is intended for GNU patch command version 2.5 or higher.
# This patch includes the following deltas:
#	           ChangeSet	1.1348  -> 1.1349 
#	          fs/namei.c	1.76    -> 1.77   
#
# The following is the BitKeeper ChangeSet Log
# --------------------------------------------
# 03/06/23	mikew@nisserv.test.com	1.1349
# namei.c:
#   nameidata should have the mnt we are looking at when doing a follow_link after a follow_mount
# --------------------------------------------
#
diff -Nru a/fs/namei.c b/fs/namei.c
--- a/fs/namei.c	Tue Jun 24 13:29:37 2003
+++ b/fs/namei.c	Tue Jun 24 13:29:37 2003
@@ -655,6 +655,7 @@
 
 		if (inode->i_op->follow_link) {
 			mntget(next.mnt);
+		        nd->mnt = next.mnt;
 			err = do_follow_link(next.dentry, nd);
 			dput(next.dentry);
 			mntput(next.mnt);
@@ -708,6 +709,7 @@
 		if ((lookup_flags & LOOKUP_FOLLOW)
 		    && inode && inode->i_op && inode->i_op->follow_link) {
 			mntget(next.mnt);
+		        nd->mnt = next.mnt;
 			err = do_follow_link(next.dentry, nd);
 			dput(next.dentry);
 			mntput(next.mnt);

Re: [PATCH] 2.5.72: follow_mount / follow_link

[viro]

On Tue, Jun 24, 2003 at 10:41:59AM -0400, Mike Waychison wrote:

> The changes made in fs/namei.c@1.42 break this behaviour iff the dentry 
> being follow_link'ed is a root dentry.  This is becauseo

... mixing symlinks and mounting is ripe with very ugly races and corner
cases.  Not allowed - symlink can't be a mountpoint or a mounted object.

Re: [PATCH] 2.5.72: follow_mount / follow_link

[H. Peter Anvin]

Followup to:  <20030624145418.GP6754@parcelfarce.linux.theplanet.co.uk>
By author:    viro@parcelfarce.linux.theplanet.co.uk
In newsgroup: linux.dev.kernel
>
> On Tue, Jun 24, 2003 at 10:41:59AM -0400, Mike Waychison wrote:
> 
> > The changes made in fs/namei.c@1.42 break this behaviour iff the dentry 
> > being follow_link'ed is a root dentry.  This is becauseo
> 
> ... mixing symlinks and mounting is ripe with very ugly races and corner
> cases.  Not allowed - symlink can't be a mountpoint or a mounted object.
>

Unfortunately, this is probably the only realistic way to ever get
working direct mounts, so please don't dismiss it out of hand.
follow_link on a directory has turned out to be a really useful way of
doing automounting.

	-hpa
-- 
<hpa@transmeta.com> at work, <hpa@zytor.com> in private!
"Unix gives you enough rope to shoot yourself in the foot."
Architectures needed: ia64 m68k mips64 ppc ppc64 s390 s390x sh v850 x86-64

Re: [PATCH] 2.5.72: follow_mount / follow_link

[jlnance]

On Tue, Jun 24, 2003 at 08:42:56AM -0700, H. Peter Anvin wrote:

> Unfortunately, this is probably the only realistic way to ever get
> working direct mounts, so please don't dismiss it out of hand.
> follow_link on a directory has turned out to be a really useful way of
> doing automounting.

Hi Peter,
    I have always wondered why direct mounts, as well as things like
/net/host are difficult with Linux.  If you have a couple of minutes,
would you explain the problem?  Also, do you have any idea how Solaris
does this and is it easier there?

Thanks,

Jim

Re: [PATCH] 2.5.72: follow_mount / follow_link

[H. Peter Anvin]

jlnance@unity.ncsu.edu wrote:
> On Tue, Jun 24, 2003 at 08:42:56AM -0700, H. Peter Anvin wrote:
> 
> 
>>Unfortunately, this is probably the only realistic way to ever get
>>working direct mounts, so please don't dismiss it out of hand.
>>follow_link on a directory has turned out to be a really useful way of
>>doing automounting.
> 
> 
> Hi Peter,
>     I have always wondered why direct mounts, as well as things like
> /net/host are difficult with Linux.  If you have a couple of minutes,
> would you explain the problem?  Also, do you have any idea how Solaris
> does this and is it easier there?
> 

It's a pretty long lecture, and I have had it enough time that I'm not
really keen on repeating it every time anyone asks.  Look at the
archives of the autofs mailing list on linux.kernel.org.

	-hpa

Re: [PATCH] 2.5.72: follow_mount / follow_link

[H. Peter Anvin]

jlnance@unity.ncsu.edu wrote:
> On Tue, Jun 24, 2003 at 08:42:56AM -0700, H. Peter Anvin wrote:
> 
> 
>>Unfortunately, this is probably the only realistic way to ever get
>>working direct mounts, so please don't dismiss it out of hand.
>>follow_link on a directory has turned out to be a really useful way of
>>doing automounting.
> 
> 
> Hi Peter,
>     I have always wondered why direct mounts, as well as things like
> /net/host are difficult with Linux.  If you have a couple of minutes,
> would you explain the problem?  Also, do you have any idea how Solaris
> does this and is it easier there?
> 

Oh yes... no, I don't know what Solaris does, although I suspect they
had to adjust their VFS to *make* it work (management directive: this
shalt work) even if that meant overall performance loss.

If you don't find something in the autofs archives get back to me or
corner me on IRC sometime.

	-hpa