Prior to commit 7c03e2cda4a5 ("vfs: move cap_convert_nscap() call into
vfs_setxattr()") the translation of nscap->rootid did not take stacked
filesystems (overlayfs and ecryptfs) into account.
That patch fixed the overlay case, but made the ecryptfs case worse.
Restore old the behavior for ecryptfs that existed before the overlayfs
fix. This does not fix ecryptfs's handling of complex user namespace
setups, but it does make sure existing setups don't regress.
Reported-by: Eric W. Biederman <[email protected]>
Cc: Tyler Hicks <[email protected]>
Fixes: 7c03e2cda4a5 ("vfs: move cap_convert_nscap() call into vfs_setxattr()")
Signed-off-by: Miklos Szeredi <[email protected]>
---
fs/ecryptfs/inode.c | 10 +++++++---
1 file changed, 7 insertions(+), 3 deletions(-)
diff --git a/fs/ecryptfs/inode.c b/fs/ecryptfs/inode.c
index e23752d9a79f..58d0f7187997 100644
--- a/fs/ecryptfs/inode.c
+++ b/fs/ecryptfs/inode.c
@@ -1016,15 +1016,19 @@ ecryptfs_setxattr(struct dentry *dentry, struct inode *inode,
{
int rc;
struct dentry *lower_dentry;
+ struct inode *lower_inode;
lower_dentry = ecryptfs_dentry_to_lower(dentry);
- if (!(d_inode(lower_dentry)->i_opflags & IOP_XATTR)) {
+ lower_inode = d_inode(lower_dentry);
+ if (!(lower_inode->i_opflags & IOP_XATTR)) {
rc = -EOPNOTSUPP;
goto out;
}
- rc = vfs_setxattr(lower_dentry, name, value, size, flags);
+ inode_lock(lower_inode);
+ rc = __vfs_setxattr_locked(lower_dentry, name, value, size, flags, NULL);
+ inode_unlock(lower_inode);
if (!rc && inode)
- fsstack_copy_attr_all(inode, d_inode(lower_dentry));
+ fsstack_copy_attr_all(inode, lower_inode);
out:
return rc;
}
--
2.26.2
Miklos Szeredi <[email protected]> writes:
> Prior to commit 7c03e2cda4a5 ("vfs: move cap_convert_nscap() call into
> vfs_setxattr()") the translation of nscap->rootid did not take stacked
> filesystems (overlayfs and ecryptfs) into account.
>
> That patch fixed the overlay case, but made the ecryptfs case worse.
>
> Restore old the behavior for ecryptfs that existed before the overlayfs
> fix. This does not fix ecryptfs's handling of complex user namespace
> setups, but it does make sure existing setups don't regress.
Today vfs_setxattr handles handles a delegated_inode and breaking
leases. Code that is enabled with CONFIG_FILE_LOCKING. So unless
I am missing something this introduces a different regression into
ecryptfs.
>
> Reported-by: Eric W. Biederman <[email protected]>
> Cc: Tyler Hicks <[email protected]>
> Fixes: 7c03e2cda4a5 ("vfs: move cap_convert_nscap() call into vfs_setxattr()")
> Signed-off-by: Miklos Szeredi <[email protected]>
> ---
> fs/ecryptfs/inode.c | 10 +++++++---
> 1 file changed, 7 insertions(+), 3 deletions(-)
>
> diff --git a/fs/ecryptfs/inode.c b/fs/ecryptfs/inode.c
> index e23752d9a79f..58d0f7187997 100644
> --- a/fs/ecryptfs/inode.c
> +++ b/fs/ecryptfs/inode.c
> @@ -1016,15 +1016,19 @@ ecryptfs_setxattr(struct dentry *dentry, struct inode *inode,
> {
> int rc;
> struct dentry *lower_dentry;
> + struct inode *lower_inode;
>
> lower_dentry = ecryptfs_dentry_to_lower(dentry);
> - if (!(d_inode(lower_dentry)->i_opflags & IOP_XATTR)) {
> + lower_inode = d_inode(lower_dentry);
> + if (!(lower_inode->i_opflags & IOP_XATTR)) {
> rc = -EOPNOTSUPP;
> goto out;
> }
> - rc = vfs_setxattr(lower_dentry, name, value, size, flags);
> + inode_lock(lower_inode);
> + rc = __vfs_setxattr_locked(lower_dentry, name, value, size, flags, NULL);
> + inode_unlock(lower_inode);
> if (!rc && inode)
> - fsstack_copy_attr_all(inode, d_inode(lower_dentry));
> + fsstack_copy_attr_all(inode, lower_inode);
> out:
> return rc;
> }
Eric
On Tue, Jan 19, 2021 at 10:11 PM Eric W. Biederman
<[email protected]> wrote:
>
> Miklos Szeredi <[email protected]> writes:
>
> > Prior to commit 7c03e2cda4a5 ("vfs: move cap_convert_nscap() call into
> > vfs_setxattr()") the translation of nscap->rootid did not take stacked
> > filesystems (overlayfs and ecryptfs) into account.
> >
> > That patch fixed the overlay case, but made the ecryptfs case worse.
> >
> > Restore old the behavior for ecryptfs that existed before the overlayfs
> > fix. This does not fix ecryptfs's handling of complex user namespace
> > setups, but it does make sure existing setups don't regress.
>
> Today vfs_setxattr handles handles a delegated_inode and breaking
> leases. Code that is enabled with CONFIG_FILE_LOCKING. So unless
> I am missing something this introduces a different regression into
> ecryptfs.
This is in line with all the other cases of ecryptfs passing NULL as
delegated inode.
I'll defer this to the maintainer of ecryptfs.
Thanks,
Miklos
On 2021-01-20 08:52:27, Miklos Szeredi wrote:
> On Tue, Jan 19, 2021 at 10:11 PM Eric W. Biederman
> <[email protected]> wrote:
> >
> > Miklos Szeredi <[email protected]> writes:
> >
> > > Prior to commit 7c03e2cda4a5 ("vfs: move cap_convert_nscap() call into
> > > vfs_setxattr()") the translation of nscap->rootid did not take stacked
> > > filesystems (overlayfs and ecryptfs) into account.
> > >
> > > That patch fixed the overlay case, but made the ecryptfs case worse.
> > >
> > > Restore old the behavior for ecryptfs that existed before the overlayfs
> > > fix. This does not fix ecryptfs's handling of complex user namespace
> > > setups, but it does make sure existing setups don't regress.
> >
> > Today vfs_setxattr handles handles a delegated_inode and breaking
> > leases. Code that is enabled with CONFIG_FILE_LOCKING. So unless
> > I am missing something this introduces a different regression into
> > ecryptfs.
>
> This is in line with all the other cases of ecryptfs passing NULL as
> delegated inode.
>
> I'll defer this to the maintainer of ecryptfs.
eCryptfs cannot be exported so I do not think this proposed fix to
ecryptfs_setxattr() creates a new regression wrt inode delegation.
Tyler
>
> Thanks,
> Miklos
On 2021-01-19 17:22:03, Miklos Szeredi wrote:
> Prior to commit 7c03e2cda4a5 ("vfs: move cap_convert_nscap() call into
> vfs_setxattr()") the translation of nscap->rootid did not take stacked
> filesystems (overlayfs and ecryptfs) into account.
>
> That patch fixed the overlay case, but made the ecryptfs case worse.
Thanks for sending a fix!
I know that you don't have an eCryptfs setup to test with but I'm at a
loss about how to test this from the userns/fscaps side of things. Do
you have a sequence of unshare/setcap/getcap commands that I can run on
a file inside of an eCryptfs mount to verify that the bug exists after
7c03e2cda4a5 and then again to verify that this patch fixes the bug?
Tyler
>
> Restore old the behavior for ecryptfs that existed before the overlayfs
> fix. This does not fix ecryptfs's handling of complex user namespace
> setups, but it does make sure existing setups don't regress.
>
> Reported-by: Eric W. Biederman <[email protected]>
> Cc: Tyler Hicks <[email protected]>
> Fixes: 7c03e2cda4a5 ("vfs: move cap_convert_nscap() call into vfs_setxattr()")
> Signed-off-by: Miklos Szeredi <[email protected]>
> ---
> fs/ecryptfs/inode.c | 10 +++++++---
> 1 file changed, 7 insertions(+), 3 deletions(-)
>
> diff --git a/fs/ecryptfs/inode.c b/fs/ecryptfs/inode.c
> index e23752d9a79f..58d0f7187997 100644
> --- a/fs/ecryptfs/inode.c
> +++ b/fs/ecryptfs/inode.c
> @@ -1016,15 +1016,19 @@ ecryptfs_setxattr(struct dentry *dentry, struct inode *inode,
> {
> int rc;
> struct dentry *lower_dentry;
> + struct inode *lower_inode;
>
> lower_dentry = ecryptfs_dentry_to_lower(dentry);
> - if (!(d_inode(lower_dentry)->i_opflags & IOP_XATTR)) {
> + lower_inode = d_inode(lower_dentry);
> + if (!(lower_inode->i_opflags & IOP_XATTR)) {
> rc = -EOPNOTSUPP;
> goto out;
> }
> - rc = vfs_setxattr(lower_dentry, name, value, size, flags);
> + inode_lock(lower_inode);
> + rc = __vfs_setxattr_locked(lower_dentry, name, value, size, flags, NULL);
> + inode_unlock(lower_inode);
> if (!rc && inode)
> - fsstack_copy_attr_all(inode, d_inode(lower_dentry));
> + fsstack_copy_attr_all(inode, lower_inode);
> out:
> return rc;
> }
> --
> 2.26.2
>
On 2021-01-25 14:25:38, Miklos Szeredi wrote:
> On Fri, Jan 22, 2021 at 7:31 PM Tyler Hicks <[email protected]> wrote:
> >
> > On 2021-01-19 17:22:03, Miklos Szeredi wrote:
> > > Prior to commit 7c03e2cda4a5 ("vfs: move cap_convert_nscap() call into
> > > vfs_setxattr()") the translation of nscap->rootid did not take stacked
> > > filesystems (overlayfs and ecryptfs) into account.
> > >
> > > That patch fixed the overlay case, but made the ecryptfs case worse.
> >
> > Thanks for sending a fix!
> >
> > I know that you don't have an eCryptfs setup to test with but I'm at a
> > loss about how to test this from the userns/fscaps side of things. Do
> > you have a sequence of unshare/setcap/getcap commands that I can run on
> > a file inside of an eCryptfs mount to verify that the bug exists after
> > 7c03e2cda4a5 and then again to verify that this patch fixes the bug?
>
> You need two terminals:
> $ = <USER>
> # = root
>
> $ unshare -Um
> $ echo $$
> <PID>
> # echo "0 1000 1" > uid_map
> # cp uid_map gid_map
> # echo 1000 2000 1 >> uid_map
> # echo 2000 3000 1 >> uid_map
> # cat uid_map > /proc/<PID>/uid_map
> # cat gid_map > /proc/<PID>/gid_map
> $ mkdir ~/tmp ~/mnt
> $ mount -t tmpfs tmpfs ~/tmp
> $ pwd
> /home/<USER>
> # nsenter -t <PID> -m
> # [setup ecryptfs on /home/<USER>/mnt using /home/<USER>/tmp]
> $ cd ~/mnt
> $ touch test
> $ /sbin/setcap -n 1000 cap_dac_override+eip test
> $ /sbin/getcap -n test
> test = cap_dac_override+eip [rootid=1000]
>
> Without the patch, I'm thinking that it will do a double translate and
> end up with rootid=2000 in the user namespace, but I might well have
> messed it up...
>
> Let me know how this goes.
Spot-on instructions. Thank you for taking the time to provide the
steps.
I was able to repro the bug and verify the fix. The change visually
looks good to me and it passed through the eCryptfs regression tests.
I've pushed it to the eCryptfs next branch and I plan to submit it to
Linus on Thursday. Thanks again!
Tyler
>
> Thanks,
> Miklos
>
On Fri, Jan 22, 2021 at 7:31 PM Tyler Hicks <[email protected]> wrote:
>
> On 2021-01-19 17:22:03, Miklos Szeredi wrote:
> > Prior to commit 7c03e2cda4a5 ("vfs: move cap_convert_nscap() call into
> > vfs_setxattr()") the translation of nscap->rootid did not take stacked
> > filesystems (overlayfs and ecryptfs) into account.
> >
> > That patch fixed the overlay case, but made the ecryptfs case worse.
>
> Thanks for sending a fix!
>
> I know that you don't have an eCryptfs setup to test with but I'm at a
> loss about how to test this from the userns/fscaps side of things. Do
> you have a sequence of unshare/setcap/getcap commands that I can run on
> a file inside of an eCryptfs mount to verify that the bug exists after
> 7c03e2cda4a5 and then again to verify that this patch fixes the bug?
You need two terminals:
$ = <USER>
# = root
$ unshare -Um
$ echo $$
<PID>
# echo "0 1000 1" > uid_map
# cp uid_map gid_map
# echo 1000 2000 1 >> uid_map
# echo 2000 3000 1 >> uid_map
# cat uid_map > /proc/<PID>/uid_map
# cat gid_map > /proc/<PID>/gid_map
$ mkdir ~/tmp ~/mnt
$ mount -t tmpfs tmpfs ~/tmp
$ pwd
/home/<USER>
# nsenter -t <PID> -m
# [setup ecryptfs on /home/<USER>/mnt using /home/<USER>/tmp]
$ cd ~/mnt
$ touch test
$ /sbin/setcap -n 1000 cap_dac_override+eip test
$ /sbin/getcap -n test
test = cap_dac_override+eip [rootid=1000]
Without the patch, I'm thinking that it will do a double translate and
end up with rootid=2000 in the user namespace, but I might well have
messed it up...
Let me know how this goes.
Thanks,
Miklos
On Mon, Jan 25, 2021 at 2:25 PM Miklos Szeredi <[email protected]> wrote:
>
> On Fri, Jan 22, 2021 at 7:31 PM Tyler Hicks <[email protected]> wrote:
> >
> > On 2021-01-19 17:22:03, Miklos Szeredi wrote:
> > > Prior to commit 7c03e2cda4a5 ("vfs: move cap_convert_nscap() call into
> > > vfs_setxattr()") the translation of nscap->rootid did not take stacked
> > > filesystems (overlayfs and ecryptfs) into account.
> > >
> > > That patch fixed the overlay case, but made the ecryptfs case worse.
> >
> > Thanks for sending a fix!
> >
> > I know that you don't have an eCryptfs setup to test with but I'm at a
> > loss about how to test this from the userns/fscaps side of things. Do
> > you have a sequence of unshare/setcap/getcap commands that I can run on
> > a file inside of an eCryptfs mount to verify that the bug exists after
> > 7c03e2cda4a5 and then again to verify that this patch fixes the bug?
>
> You need two terminals:
> $ = <USER>
> # = root
>
> $ unshare -Um
> $ echo $$
> <PID>
> # echo "0 1000 1" > uid_map
NOTE: <USER> is assumed to have uid=1000, so this and following
"1000" values need to be fixed up if it's not the case.
Thanks,
Miklos