Skip to contentSkip to conversation

Some fixes for overlayfs remove privs

PATCH In review Amir Goldstein 3 participants 12 messages Oct 3, 2022 – Oct 5, 2022

Subsystems (4)
lore
Share & tools
lore

Amir Goldstein

author maintainer
#0 lore raw

cover Some fixes for overlayfs remove privs

Miklos,

While running latest fstests on overlayfs-next, I noticed these failures:
generic/673 generic/683 generic/684 generic/685 generic/686 generic/687

Christian has also reported those failures earlier.

Those are not regressions, those are 5 new tests added to fstests and one test whose expected result was modified by fstests commit b3a59bb6 ("generic/673: fix golden output to reflect vfs setgid behavior").

The following two patches aim to fix those test failures, but they are incomplete - without those patches, the tests fail miserably in all test cases, because no privs are stripped.

With those two patches, only two test cases are failing, which are the two test cases whose expectation was changed by fstests commit b3a59bb6. The reason was explained in [1] and the issue was fixed for xfs by kernel commit e014f37db1a2 ("xfs: use setattr_copy to set vfs inode attributes").

Trying to figure out how to fix this hurts my brain, so I'll need suggestions how to proceed.

Thanks,
Amir.

[1] https://lore.kernel.org/linux-xfs/CAL3q7H47iNQ=Wmk83WcGB-KBJVOEtR9+qGczzCeXJ9Y2KCV25Q@mail.gmail.com/

Amir Goldstein (2):
ovl: remove privs in ovl_copyfile()
ovl: remove privs in ovl_fallocate()

 fs/overlayfs/file.c | 28 +++++++++++++++++++++++++---
 1 file changed, 25 insertions(+), 3 deletions(-)
Signature · 1 line
2.25.1
author maintainer
#1 lore raw

PATCH 1/2 ovl: remove privs in ovl_copyfile()

Underlying fs doesn't remove privs because copy_range/remap_range are called with privileged mounter credentials.

This fixes some failures in fstest generic/673.

Fixes 8ede205541ff ("ovl: add reflink/copyfile/dedup support")
---
 fs/overlayfs/file.c | 16 ++++++++++++++--
 1 file changed, 14 insertions(+), 2 deletions(-)
fs/overlayfs/file.c +14 -2
index d17faeb014e5..c8308da8909a 100644
@@ -567,14 +567,23 @@ static loff_t ovl_copyfile(struct file *file_in, loff_t pos_in,
 	const struct cred *old_cred;
 	loff_t ret;
 
+	inode_lock(inode_out);
+	if (op != OVL_DEDUPE) {
+		/* Update mode */
+		ovl_copyattr(inode_out);
+		ret = file_remove_privs(file_out);
+		if (ret)
+			goto out_unlock;
+	}
+
 	ret = ovl_real_fdget(file_out, &real_out);
 	if (ret)
-		return ret;
+		goto out_unlock;
 
 	ret = ovl_real_fdget(file_in, &real_in);
 	if (ret) {
 		fdput(real_out);
-		return ret;
+		goto out_unlock;
 	}
 
 	old_cred = ovl_override_creds(file_inode(file_out)->i_sb);
@@ -603,6 +612,9 @@ static loff_t ovl_copyfile(struct file *file_in, loff_t pos_in,
 	fdput(real_in);
 	fdput(real_out);
 
+out_unlock:
+	inode_unlock(inode_out);
+
 	return ret;
 }
 
Signature · 1 line
2.25.1
maintainer Acked-by on v1 1/2
#2 lore raw

Re: [PATCH 1/2] ovl: remove privs in ovl_copyfile()

Quoted 8 lines · Amir Goldstein

Underlying fs doesn't remove privs because copy_range/remap_range are called with privileged mounter credentials.

This fixes some failures in fstest generic/673.

Fixes 8ede205541ff ("ovl: add reflink/copyfile/dedup support")
author maintainer
#3 lore raw

PATCH 2/2 ovl: remove privs in ovl_fallocate()

In reply to Amir Goldstein

Underlying fs doesn't remove privs because fallocate is called with privileged mounter credentials.

This fixes some failure in fstests generic/683..687.

Fixes aab8848cee5e ("ovl: add ovl_fallocate()")
---
 fs/overlayfs/file.c | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)
fs/overlayfs/file.c +11 -1
index c8308da8909a..e90ac5376456 100644
@@ -517,9 +517,16 @@ static long ovl_fallocate(struct file *file, int mode, loff_t offset, loff_t len
 	const struct cred *old_cred;
 	int ret;
 
+	inode_lock(inode);
+	/* Update mode */
+	ovl_copyattr(inode);
+	ret = file_remove_privs(file);
+	if (ret)
+		goto out_unlock;
+
 	ret = ovl_real_fdget(file, &real);
 	if (ret)
-		return ret;
+		goto out_unlock;
 
 	old_cred = ovl_override_creds(file_inode(file)->i_sb);
 	ret = vfs_fallocate(real.file, mode, offset, len);
@@ -530,6 +537,9 @@ static long ovl_fallocate(struct file *file, int mode, loff_t offset, loff_t len
 
 	fdput(real);
 
+out_unlock:
+	inode_unlock(inode);
+
 	return ret;
 }
 
Signature · 1 line
2.25.1
#4 lore raw

Re: [PATCH 2/2] ovl: remove privs in ovl_fallocate()

Quoted 16 lines · Amir Goldstein

Underlying fs doesn't remove privs because fallocate is called with privileged mounter credentials.

This fixes some failure in fstests generic/683..687.

Fixes aab8848cee5e ("ovl: add ovl_fallocate()")
---
 fs/overlayfs/file.c | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)
fs/overlayfs/file.c +0 -0
index c8308da8909a..e90ac5376456 100644

fs/overlayfs/file.c static long ovl_fallocate(struct file *file, int mode, loff_t offset, loff_t len · line 523

Quoted 7 lines
fs/overlayfs/file.c +4 -0
@@ -517,9 +517,16 @@ static long ovl_fallocate(struct file *file, int mode, loff_t offset, loff_t len
 	const struct cred *old_cred;
 	int ret;
 
+	inode_lock(inode);
+	/* Update mode */
+	ovl_copyattr(inode);
+	ret = file_remove_privs(file);

First, thank you for picking this up!

Let me analyze generic/683 failure of Test1 to see why you still see failures in this test:

echo "Test 1 - qa_user, non-exec file $verb"
setup_testfile
chmod a+rws $junk_file
commit_and_check "$qa_user" "$verb" 64k 64k

So this creates a file with 6666 permissions. While the file has the S_ISUID and S_ISGID bits set it does not have the S_IXGRP set. This is important in a little bit.

On a regular filesystem like xfs what will happen is:

sys_fallocate()
-> vfs_fallocate()
-> xfs_file_fallocate()

      -> file_modified()
         -> __file_remove_privs()
            -> dentry_needs_remove_privs()
               -> should_remove_suid()
            -> __remove_privs()
               newattrs.ia_valid = ATTR_FORCE | kill;
               -> notify_change()

In should_remove_suid() we can see that ATTR_KILL_SUID is raised unconditionally because the file in the test has S_ISUID set.

But we also see that ATTR_KILL_SGID won't be set because while the file is S_ISGID it is not S_IXGRP (see above) which is a condition for ATTR_KILL_SGID being raised.

So by the time we call notify_change() we have attr->ia_valid set to ATTR_KILL_SUID | ATTR_FORCE. Now notify_change() sees that ATTR_KILL_SUID is set and does:

ia_valid = attr->ia_valid |= ATTR_MODE
attr->ia_mode = (inode->i_mode & ~S_ISUID);

which means that when we call setattr_copy() later we will definitely update inode->i_mode. Note that attr->ia_mode still contain S_ISGID.

Now we call into the filesystem's ->setattr() inode operation which will end up calling setattr_copy(). Since ATTR_MODE is set we will hit:

if (ia_valid & ATTR_MODE) {
        umode_t mode = attr->ia_mode;
        vfsgid_t vfsgid = i_gid_into_vfsgid(mnt_userns, inode);
        if (!vfsgid_in_group_p(vfsgid) &&
            !capable_wrt_inode_uidgid(mnt_userns, inode, CAP_FSETID))
                mode &= ~S_ISGID;
        inode->i_mode = mode;
}

and since the caller in the test is neither capable nor in the group of the inode the S_ISGID bit is stripped.

But now contrast this with overlayfs even after your changes. When ovl_setattr() is hit from ovl_fallocate()'s call to file_remove_privs() and calls ovl_do_notify_change() then we are doing this under the mounter's creds and so the S_ISGID bit is retained:

sys_fallocate()
-> vfs_fallocate()
-> ovl_fallocate()

      -> file_remove_privs()
         -> dentry_needs_remove_privs()
            -> should_remove_suid()
         -> __remove_privs()
            newattrs.ia_valid = attr_force | kill;
            -> notify_change()
	       -> ovl_setattr()
	          // TAKE ON MOUNTER'S CREDS
	          -> ovl_do_notify_change()
	          // GIVE UP MOUNTER'S CREDS
     // TAKE ON MOUNTER'S CREDS
     -> vfs_fallocate()
        -> xfs_file_fallocate()
           -> file_modified()
              -> __file_remove_privs()
                 -> dentry_needs_remove_privs()
                    -> should_remove_suid()
                 -> __remove_privs()
                    newattrs.ia_valid = attr_force | kill;
                    -> notify_change()
#5 lore raw

Re: [PATCH 2/2] ovl: remove privs in ovl_fallocate()

Quoted 92 lines · Christian Brauner
Quoted 16 lines · Amir Goldstein

Underlying fs doesn't remove privs because fallocate is called with privileged mounter credentials.

This fixes some failure in fstests generic/683..687.

Fixes aab8848cee5e ("ovl: add ovl_fallocate()")
---
 fs/overlayfs/file.c | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)
fs/overlayfs/file.c +0 -0
index c8308da8909a..e90ac5376456 100644

fs/overlayfs/file.c static long ovl_fallocate(struct file *file, int mode, loff_t offset, loff_t len · line 523

Quoted 7 lines
fs/overlayfs/file.c +4 -0
@@ -517,9 +517,16 @@ static long ovl_fallocate(struct file *file, int mode, loff_t offset, loff_t len
 	const struct cred *old_cred;
 	int ret;
 
+	inode_lock(inode);
+	/* Update mode */
+	ovl_copyattr(inode);
+	ret = file_remove_privs(file);

First, thank you for picking this up!

Let me analyze generic/683 failure of Test1 to see why you still see failures in this test:

echo "Test 1 - qa_user, non-exec file $verb"
setup_testfile
chmod a+rws $junk_file
commit_and_check "$qa_user" "$verb" 64k 64k

So this creates a file with 6666 permissions. While the file has the S_ISUID and S_ISGID bits set it does not have the S_IXGRP set. This is important in a little bit.

On a regular filesystem like xfs what will happen is:

sys_fallocate()
-> vfs_fallocate()
-> xfs_file_fallocate()

      -> file_modified()
         -> __file_remove_privs()
            -> dentry_needs_remove_privs()
               -> should_remove_suid()
            -> __remove_privs()
               newattrs.ia_valid = ATTR_FORCE | kill;
               -> notify_change()

In should_remove_suid() we can see that ATTR_KILL_SUID is raised unconditionally because the file in the test has S_ISUID set.

But we also see that ATTR_KILL_SGID won't be set because while the file is S_ISGID it is not S_IXGRP (see above) which is a condition for ATTR_KILL_SGID being raised.

So by the time we call notify_change() we have attr->ia_valid set to ATTR_KILL_SUID | ATTR_FORCE. Now notify_change() sees that ATTR_KILL_SUID is set and does:

ia_valid = attr->ia_valid |= ATTR_MODE
attr->ia_mode = (inode->i_mode & ~S_ISUID);

which means that when we call setattr_copy() later we will definitely update inode->i_mode. Note that attr->ia_mode still contain S_ISGID.

Now we call into the filesystem's ->setattr() inode operation which will end up calling setattr_copy(). Since ATTR_MODE is set we will hit:

if (ia_valid & ATTR_MODE) {
        umode_t mode = attr->ia_mode;
        vfsgid_t vfsgid = i_gid_into_vfsgid(mnt_userns, inode);
        if (!vfsgid_in_group_p(vfsgid) &&
            !capable_wrt_inode_uidgid(mnt_userns, inode, CAP_FSETID))
                mode &= ~S_ISGID;
        inode->i_mode = mode;
}

and since the caller in the test is neither capable nor in the group of the inode the S_ISGID bit is stripped.

But now contrast this with overlayfs even after your changes. When ovl_setattr() is hit from ovl_fallocate()'s call to file_remove_privs() and calls ovl_do_notify_change() then we are doing this under the mounter's creds and so the S_ISGID bit is retained:

sys_fallocate()
-> vfs_fallocate()
-> ovl_fallocate()

Quoted 7 lines
      -> file_remove_privs()
         -> dentry_needs_remove_privs()
            -> should_remove_suid()
         -> __remove_privs()
            newattrs.ia_valid = attr_force | kill;
            -> notify_change()
	       -> ovl_setattr()

Note also that the setattr_prepare() call skips setgid checks because ATTR_FORCE is passed from file_remove_privs().

Quoted 13 lines
	          // TAKE ON MOUNTER'S CREDS
	          -> ovl_do_notify_change()
	          // GIVE UP MOUNTER'S CREDS
     // TAKE ON MOUNTER'S CREDS
     -> vfs_fallocate()
        -> xfs_file_fallocate()
           -> file_modified()
              -> __file_remove_privs()
                 -> dentry_needs_remove_privs()
                    -> should_remove_suid()
                 -> __remove_privs()
                    newattrs.ia_valid = attr_force | kill;
                    -> notify_change()
author maintainer
#6 lore raw

Re: [PATCH 2/2] ovl: remove privs in ovl_fallocate()

In reply to Christian Brauner
Quoted 73 lines · Christian Brauner
Quoted 16 lines · Amir Goldstein

Underlying fs doesn't remove privs because fallocate is called with privileged mounter credentials.

This fixes some failure in fstests generic/683..687.

Fixes aab8848cee5e ("ovl: add ovl_fallocate()")
---
 fs/overlayfs/file.c | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)
fs/overlayfs/file.c +0 -0
index c8308da8909a..e90ac5376456 100644

fs/overlayfs/file.c static long ovl_fallocate(struct file *file, int mode, loff_t offset, loff_t len · line 523

Quoted 7 lines
fs/overlayfs/file.c +4 -0
@@ -517,9 +517,16 @@ static long ovl_fallocate(struct file *file, int mode, loff_t offset, loff_t len
      const struct cred *old_cred;
      int ret;
 
+     inode_lock(inode);
+     /* Update mode */
+     ovl_copyattr(inode);
+     ret = file_remove_privs(file);

First, thank you for picking this up!

Let me analyze generic/683 failure of Test1 to see why you still see failures in this test:

echo "Test 1 - qa_user, non-exec file $verb"
setup_testfile
chmod a+rws $junk_file
commit_and_check "$qa_user" "$verb" 64k 64k

So this creates a file with 6666 permissions. While the file has the S_ISUID and S_ISGID bits set it does not have the S_IXGRP set. This is important in a little bit.

On a regular filesystem like xfs what will happen is:

sys_fallocate()
-> vfs_fallocate()
-> xfs_file_fallocate()

      -> file_modified()
         -> __file_remove_privs()
            -> dentry_needs_remove_privs()
               -> should_remove_suid()
            -> __remove_privs()
               newattrs.ia_valid = ATTR_FORCE | kill;
               -> notify_change()

In should_remove_suid() we can see that ATTR_KILL_SUID is raised unconditionally because the file in the test has S_ISUID set.

But we also see that ATTR_KILL_SGID won't be set because while the file is S_ISGID it is not S_IXGRP (see above) which is a condition for ATTR_KILL_SGID being raised.

So by the time we call notify_change() we have attr->ia_valid set to ATTR_KILL_SUID | ATTR_FORCE. Now notify_change() sees that ATTR_KILL_SUID is set and does:

ia_valid = attr->ia_valid |= ATTR_MODE
attr->ia_mode = (inode->i_mode & ~S_ISUID);

which means that when we call setattr_copy() later we will definitely update inode->i_mode. Note that attr->ia_mode still contain S_ISGID.

Now we call into the filesystem's ->setattr() inode operation which will end up calling setattr_copy(). Since ATTR_MODE is set we will hit:

Quoted 9 lines
if (ia_valid & ATTR_MODE) {
        umode_t mode = attr->ia_mode;
        vfsgid_t vfsgid = i_gid_into_vfsgid(mnt_userns, inode);
        if (!vfsgid_in_group_p(vfsgid) &&
            !capable_wrt_inode_uidgid(mnt_userns, inode, CAP_FSETID))
                mode &= ~S_ISGID;
        inode->i_mode = mode;
}

Can you think of a reason why the above should not be done in notify_change() before even calling to ->setattr()?

Although, it wouldn't help because ovl_setattr() does:

    if (attr->ia_valid & (ATTR_KILL_SUID|ATTR_KILL_SGID))
        attr->ia_valid &= ~ATTR_MODE;
Quoted 23 lines

and since the caller in the test is neither capable nor in the group of the inode the S_ISGID bit is stripped.

But now contrast this with overlayfs even after your changes. When ovl_setattr() is hit from ovl_fallocate()'s call to file_remove_privs() and calls ovl_do_notify_change() then we are doing this under the mounter's creds and so the S_ISGID bit is retained:

sys_fallocate()
-> vfs_fallocate()
-> ovl_fallocate()

      -> file_remove_privs()
         -> dentry_needs_remove_privs()
            -> should_remove_suid()
         -> __remove_privs()
            newattrs.ia_valid = attr_force | kill;
            -> notify_change()
               -> ovl_setattr()
                  // TAKE ON MOUNTER'S CREDS
                  -> ovl_do_notify_change()
                  // GIVE UP MOUNTER'S CREDS
     // TAKE ON MOUNTER'S CREDS
     -> vfs_fallocate()
Quoted 8 lines
        -> xfs_file_fallocate()
           -> file_modified()
              -> __file_remove_privs()
                 -> dentry_needs_remove_privs()
                    -> should_remove_suid()
                 -> __remove_privs()
                    newattrs.ia_valid = attr_force | kill;
                    -> notify_change()

The model in overlayfs is that security is checked twice once on overlay inode with caller creds and once again on xfs inode with mounter creds. Either of these checks could result in clearing SUID/SGID bits.

In the call stack above, the outer should_remove_suid() with caller creds sets ATTR_KILL_SUID and then the outer notify_change() clears SUID and sets ATTR_MODE, but ovl_setattr() clears ATTR_MODE and then the inner notify_change() re-clears SUID and sets ATTR_MODE again.

If the outer notify_change() would have checked the in_group_p() condition, clear SGID and set a flag ATTR_KILL_SGID_FORCE then the inner notify_change() would see this flag and re-clear SGID bit, just the same as it does with SUID bit in the stack stace above.

Is this making any sense?

Thanks,
Amir.

#7 lore raw

Re: [PATCH 2/2] ovl: remove privs in ovl_fallocate()

Quoted 124 lines · Amir Goldstein
Quoted 73 lines · Christian Brauner
Quoted 16 lines · Amir Goldstein

Underlying fs doesn't remove privs because fallocate is called with privileged mounter credentials.

This fixes some failure in fstests generic/683..687.

Fixes aab8848cee5e ("ovl: add ovl_fallocate()")
---
 fs/overlayfs/file.c | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)
fs/overlayfs/file.c +0 -0
index c8308da8909a..e90ac5376456 100644

fs/overlayfs/file.c static long ovl_fallocate(struct file *file, int mode, loff_t offset, loff_t len · line 523

Quoted 7 lines
fs/overlayfs/file.c +4 -0
@@ -517,9 +517,16 @@ static long ovl_fallocate(struct file *file, int mode, loff_t offset, loff_t len
      const struct cred *old_cred;
      int ret;
 
+     inode_lock(inode);
+     /* Update mode */
+     ovl_copyattr(inode);
+     ret = file_remove_privs(file);

First, thank you for picking this up!

Let me analyze generic/683 failure of Test1 to see why you still see failures in this test:

echo "Test 1 - qa_user, non-exec file $verb"
setup_testfile
chmod a+rws $junk_file
commit_and_check "$qa_user" "$verb" 64k 64k

So this creates a file with 6666 permissions. While the file has the S_ISUID and S_ISGID bits set it does not have the S_IXGRP set. This is important in a little bit.

On a regular filesystem like xfs what will happen is:

sys_fallocate()
-> vfs_fallocate()
-> xfs_file_fallocate()

      -> file_modified()
         -> __file_remove_privs()
            -> dentry_needs_remove_privs()
               -> should_remove_suid()
            -> __remove_privs()
               newattrs.ia_valid = ATTR_FORCE | kill;
               -> notify_change()

In should_remove_suid() we can see that ATTR_KILL_SUID is raised unconditionally because the file in the test has S_ISUID set.

But we also see that ATTR_KILL_SGID won't be set because while the file is S_ISGID it is not S_IXGRP (see above) which is a condition for ATTR_KILL_SGID being raised.

So by the time we call notify_change() we have attr->ia_valid set to ATTR_KILL_SUID | ATTR_FORCE. Now notify_change() sees that ATTR_KILL_SUID is set and does:

ia_valid = attr->ia_valid |= ATTR_MODE
attr->ia_mode = (inode->i_mode & ~S_ISUID);

which means that when we call setattr_copy() later we will definitely update inode->i_mode. Note that attr->ia_mode still contain S_ISGID.

Now we call into the filesystem's ->setattr() inode operation which will end up calling setattr_copy(). Since ATTR_MODE is set we will hit:

Quoted 9 lines
if (ia_valid & ATTR_MODE) {
        umode_t mode = attr->ia_mode;
        vfsgid_t vfsgid = i_gid_into_vfsgid(mnt_userns, inode);
        if (!vfsgid_in_group_p(vfsgid) &&
            !capable_wrt_inode_uidgid(mnt_userns, inode, CAP_FSETID))
                mode &= ~S_ISGID;
        inode->i_mode = mode;
}

Can you think of a reason why the above should not be done in notify_change() before even calling to ->setattr()?

Although, it wouldn't help because ovl_setattr() does:

    if (attr->ia_valid & (ATTR_KILL_SUID|ATTR_KILL_SGID))
        attr->ia_valid &= ~ATTR_MODE;
Quoted 23 lines

and since the caller in the test is neither capable nor in the group of the inode the S_ISGID bit is stripped.

But now contrast this with overlayfs even after your changes. When ovl_setattr() is hit from ovl_fallocate()'s call to file_remove_privs() and calls ovl_do_notify_change() then we are doing this under the mounter's creds and so the S_ISGID bit is retained:

sys_fallocate()
-> vfs_fallocate()
-> ovl_fallocate()

      -> file_remove_privs()
         -> dentry_needs_remove_privs()
            -> should_remove_suid()
         -> __remove_privs()
            newattrs.ia_valid = attr_force | kill;
            -> notify_change()
               -> ovl_setattr()
                  // TAKE ON MOUNTER'S CREDS
                  -> ovl_do_notify_change()
                  // GIVE UP MOUNTER'S CREDS
     // TAKE ON MOUNTER'S CREDS
     -> vfs_fallocate()
Quoted 8 lines
        -> xfs_file_fallocate()
           -> file_modified()
              -> __file_remove_privs()
                 -> dentry_needs_remove_privs()
                    -> should_remove_suid()
                 -> __remove_privs()
                    newattrs.ia_valid = attr_force | kill;
                    -> notify_change()
Quoted 4 lines

The model in overlayfs is that security is checked twice once on overlay inode with caller creds and once again on xfs inode with mounter creds. Either of these checks could result in clearing SUID/SGID bits.

Yep.

Quoted 4 lines

In the call stack above, the outer should_remove_suid() with caller creds sets ATTR_KILL_SUID and then the outer notify_change() clears SUID and sets ATTR_MODE,

Yes.

Quoted 2 lines

but ovl_setattr() clears ATTR_MODE and then the inner notify_change() re-clears SUID and sets ATTR_MODE again.

Yes.

Quoted 8 lines

If the outer notify_change() would have checked the in_group_p() condition, clear SGID and set a flag ATTR_KILL_SGID_FORCE then the inner notify_change() would see this flag and re-clear SGID bit, just the same as it does with SUID bit in the stack stace above.

Is this making any sense?

What I kept thinking was sm along the lines of:

fs/inode.c +6 -2
index ba1de23c13c1..e62a564201b7 100644
@@ -1968,8 +1968,12 @@ int should_remove_suid(struct dentry *dentry)
         * sgid without any exec bits is just a mandatory locking mark; leave
         * it alone.  If some exec bits are set, it's a real sgid; kill it.
         */
-       if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
-               kill |= ATTR_KILL_SGID;
+       if (unlikely(mode & S_ISGID)) {
+               if ((mode & S_IXGRP) ||
+                   (!vfsgid_in_group_p(vfsgid) &&
+                    !capable_wrt_inode_uidgid(mnt_userns, inode, CAP_FSETID)))
+                       kill |= ATTR_KILL_SGID;
+       }
 
        if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
                return kill;

mandatory locks have been removed as well so that remark seems pointless as well?

#8 lore raw

Re: [PATCH 2/2] ovl: remove privs in ovl_fallocate()

Quoted 175 lines · Christian Brauner
Quoted 124 lines · Amir Goldstein
Quoted 73 lines · Christian Brauner
Quoted 16 lines · Amir Goldstein

Underlying fs doesn't remove privs because fallocate is called with privileged mounter credentials.

This fixes some failure in fstests generic/683..687.

Fixes aab8848cee5e ("ovl: add ovl_fallocate()")
---
 fs/overlayfs/file.c | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)
fs/overlayfs/file.c +0 -0
index c8308da8909a..e90ac5376456 100644

fs/overlayfs/file.c static long ovl_fallocate(struct file *file, int mode, loff_t offset, loff_t len · line 523

Quoted 7 lines
fs/overlayfs/file.c +4 -0
@@ -517,9 +517,16 @@ static long ovl_fallocate(struct file *file, int mode, loff_t offset, loff_t len
      const struct cred *old_cred;
      int ret;
 
+     inode_lock(inode);
+     /* Update mode */
+     ovl_copyattr(inode);
+     ret = file_remove_privs(file);

First, thank you for picking this up!

Let me analyze generic/683 failure of Test1 to see why you still see failures in this test:

echo "Test 1 - qa_user, non-exec file $verb"
setup_testfile
chmod a+rws $junk_file
commit_and_check "$qa_user" "$verb" 64k 64k

So this creates a file with 6666 permissions. While the file has the S_ISUID and S_ISGID bits set it does not have the S_IXGRP set. This is important in a little bit.

On a regular filesystem like xfs what will happen is:

sys_fallocate()
-> vfs_fallocate()
-> xfs_file_fallocate()

      -> file_modified()
         -> __file_remove_privs()
            -> dentry_needs_remove_privs()
               -> should_remove_suid()
            -> __remove_privs()
               newattrs.ia_valid = ATTR_FORCE | kill;
               -> notify_change()

In should_remove_suid() we can see that ATTR_KILL_SUID is raised unconditionally because the file in the test has S_ISUID set.

But we also see that ATTR_KILL_SGID won't be set because while the file is S_ISGID it is not S_IXGRP (see above) which is a condition for ATTR_KILL_SGID being raised.

So by the time we call notify_change() we have attr->ia_valid set to ATTR_KILL_SUID | ATTR_FORCE. Now notify_change() sees that ATTR_KILL_SUID is set and does:

ia_valid = attr->ia_valid |= ATTR_MODE
attr->ia_mode = (inode->i_mode & ~S_ISUID);

which means that when we call setattr_copy() later we will definitely update inode->i_mode. Note that attr->ia_mode still contain S_ISGID.

Now we call into the filesystem's ->setattr() inode operation which will end up calling setattr_copy(). Since ATTR_MODE is set we will hit:

Quoted 9 lines
if (ia_valid & ATTR_MODE) {
        umode_t mode = attr->ia_mode;
        vfsgid_t vfsgid = i_gid_into_vfsgid(mnt_userns, inode);
        if (!vfsgid_in_group_p(vfsgid) &&
            !capable_wrt_inode_uidgid(mnt_userns, inode, CAP_FSETID))
                mode &= ~S_ISGID;
        inode->i_mode = mode;
}

Can you think of a reason why the above should not be done in notify_change() before even calling to ->setattr()?

Although, it wouldn't help because ovl_setattr() does:

    if (attr->ia_valid & (ATTR_KILL_SUID|ATTR_KILL_SGID))
        attr->ia_valid &= ~ATTR_MODE;
Quoted 23 lines

and since the caller in the test is neither capable nor in the group of the inode the S_ISGID bit is stripped.

But now contrast this with overlayfs even after your changes. When ovl_setattr() is hit from ovl_fallocate()'s call to file_remove_privs() and calls ovl_do_notify_change() then we are doing this under the mounter's creds and so the S_ISGID bit is retained:

sys_fallocate()
-> vfs_fallocate()
-> ovl_fallocate()

      -> file_remove_privs()
         -> dentry_needs_remove_privs()
            -> should_remove_suid()
         -> __remove_privs()
            newattrs.ia_valid = attr_force | kill;
            -> notify_change()
               -> ovl_setattr()
                  // TAKE ON MOUNTER'S CREDS
                  -> ovl_do_notify_change()
                  // GIVE UP MOUNTER'S CREDS
     // TAKE ON MOUNTER'S CREDS
     -> vfs_fallocate()
Quoted 8 lines
        -> xfs_file_fallocate()
           -> file_modified()
              -> __file_remove_privs()
                 -> dentry_needs_remove_privs()
                    -> should_remove_suid()
                 -> __remove_privs()
                    newattrs.ia_valid = attr_force | kill;
                    -> notify_change()
Quoted 4 lines

The model in overlayfs is that security is checked twice once on overlay inode with caller creds and once again on xfs inode with mounter creds. Either of these checks could result in clearing SUID/SGID bits.

Yep.

Quoted 4 lines

In the call stack above, the outer should_remove_suid() with caller creds sets ATTR_KILL_SUID and then the outer notify_change() clears SUID and sets ATTR_MODE,

Yes.

Quoted 2 lines

but ovl_setattr() clears ATTR_MODE and then the inner notify_change() re-clears SUID and sets ATTR_MODE again.

Yes.

Quoted 8 lines

If the outer notify_change() would have checked the in_group_p() condition, clear SGID and set a flag ATTR_KILL_SGID_FORCE then the inner notify_change() would see this flag and re-clear SGID bit, just the same as it does with SUID bit in the stack stace above.

Is this making any sense?

What I kept thinking was sm along the lines of:

fs/inode.c +6 -2
index ba1de23c13c1..e62a564201b7 100644
@@ -1968,8 +1968,12 @@ int should_remove_suid(struct dentry *dentry)
         * sgid without any exec bits is just a mandatory locking mark; leave
         * it alone.  If some exec bits are set, it's a real sgid; kill it.
         */
-       if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
-               kill |= ATTR_KILL_SGID;
+       if (unlikely(mode & S_ISGID)) {
+               if ((mode & S_IXGRP) ||
+                   (!vfsgid_in_group_p(vfsgid) &&
+                    !capable_wrt_inode_uidgid(mnt_userns, inode, CAP_FSETID)))
+                       kill |= ATTR_KILL_SGID;
+       }
 
        if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
                return kill;
Quoted 2 lines

mandatory locks have been removed as well so that remark seems pointless as well?

My feeling here is that both should_remove_suid() and notify_change() need to apply the same permission checks as setattr_prepare() and setattr_copy() instead of all this special casing them. I don't see a good reason to not require the same checks. So sm like (__completely untested__):

From 922f9f123ab6531c29bf05585ef88c17fe65dba3 Mon Sep 17 00:00:00 2001
From: Christian Brauner <[email protected]>
Date: Tue, 4 Oct 2022 16:13:34 +0200
Subject: [UNTESTED PATCH] attr: use consistent sgid stripping checks

Require the same permissions as setattr_prepare() and setattr_copy() have instead of all these special cases. We can probably consolidate this even more...

---
 fs/attr.c          |  2 +-
 fs/fuse/file.c     |  2 +-
 fs/inode.c         | 41 ++++++++++++++++++++++++-----------------
 fs/internal.h      |  4 +++-
 fs/ocfs2/file.c    |  4 ++--
 fs/open.c          |  2 +-
 include/linux/fs.h |  2 +-
 7 files changed, 33 insertions(+), 24 deletions(-)
fs/attr.c +1 -1
index 1552a5f23d6b..9262c6b31c26 100644
@@ -375,7 +375,7 @@ int notify_change(struct user_namespace *mnt_userns, struct dentry *dentry,
 		}
 	}
 	if (ia_valid & ATTR_KILL_SGID) {
-		if ((mode & (S_ISGID | S_IXGRP)) == (S_ISGID | S_IXGRP)) {
+		if (should_remove_sgid(mnt_userns, dentry)) {
 			if (!(ia_valid & ATTR_MODE)) {
 				ia_valid = attr->ia_valid |= ATTR_MODE;
 				attr->ia_mode = inode->i_mode;
fs/fuse/file.c +1 -1
index 1a3afd469e3a..fccc2c7e88fd 100644
@@ -1313,7 +1313,7 @@ static ssize_t fuse_cache_write_iter(struct kiocb *iocb, struct iov_iter *from)
 			return err;
 
 		if (fc->handle_killpriv_v2 &&
-		    should_remove_suid(file_dentry(file))) {
+		    should_remove_suid(&init_user_ns, file_dentry(file))) {
 			goto writethrough;
 		}
 
fs/inode.c +24 -17

72 lines not on this page · raw message

fs/internal.h +3 -1
index 1e67b4b9a4d1..ae152ded227c 100644
@@ -139,7 +139,8 @@ extern int vfs_open(const struct path *, struct file *);
  * inode.c
  */
 extern long prune_icache_sb(struct super_block *sb, struct shrink_control *sc);
-extern int dentry_needs_remove_privs(struct dentry *dentry);
+extern int dentry_needs_remove_privs(struct user_namespace *,
+				     struct dentry *dentry);
 
 /*
  * fs-writeback.c
@@ -226,3 +227,4 @@ int do_set_acl(struct user_namespace *mnt_userns, struct dentry *dentry,
 	       const char *acl_name, const void *kvalue, size_t size);
 ssize_t do_get_acl(struct user_namespace *mnt_userns, struct dentry *dentry,
 		   const char *acl_name, void *kvalue, size_t size);
+int should_remove_sgid(struct user_namespace *mnt_userns, struct dentry *dentry);
fs/ocfs2/file.c +2 -2
index af900aaa9275..0a33096a3852 100644
@@ -1991,7 +1991,7 @@ static int __ocfs2_change_file_space(struct file *file, struct inode *inode,
 		}
 	}
 
-	if (file && should_remove_suid(file->f_path.dentry)) {
+	if (file && should_remove_suid(&init_user_ns, file->f_path.dentry)) {
 		ret = __ocfs2_write_remove_suid(inode, di_bh);
 		if (ret) {
 			mlog_errno(ret);
@@ -2279,7 +2279,7 @@ static int ocfs2_prepare_inode_for_write(struct file *file,
 		 * inode. There's also the dinode i_size state which
 		 * can be lost via setattr during extending writes (we
 		 * set inode->i_size at the end of a write. */
-		if (should_remove_suid(dentry)) {
+		if (should_remove_suid(&init_user_ns, dentry)) {
 			if (meta_level == 0) {
 				ocfs2_inode_unlock_for_extent_tree(inode,
 								   &di_bh,
fs/open.c +1 -1
index 8a813fa5ca56..ecb7b8a58275 100644
@@ -54,7 +54,7 @@ int do_truncate(struct user_namespace *mnt_userns, struct dentry *dentry,
 	}
 
 	/* Remove suid, sgid, and file capabilities on truncate too */
-	ret = dentry_needs_remove_privs(dentry);
+	ret = dentry_needs_remove_privs(mnt_userns, dentry);
 	if (ret < 0)
 		return ret;
 	if (ret)
include/linux/fs.h +1 -1
index badff81b9dde..58347514421e 100644
@@ -3092,7 +3092,7 @@ extern void __destroy_inode(struct inode *);
 extern struct inode *new_inode_pseudo(struct super_block *sb);
 extern struct inode *new_inode(struct super_block *sb);
 extern void free_inode_nonrcu(struct inode *inode);
-extern int should_remove_suid(struct dentry *);
+extern int should_remove_suid(struct user_namespace *, struct dentry *);
 extern int file_remove_privs(struct file *);
 
 /*
Signature · 1 line
2.34.1
author maintainer
#9 lore raw

Re: [PATCH 2/2] ovl: remove privs in ovl_fallocate()

Quoted 177 lines · Christian Brauner
Quoted 175 lines · Christian Brauner
Quoted 124 lines · Amir Goldstein
Quoted 73 lines · Christian Brauner
Quoted 16 lines · Amir Goldstein

Underlying fs doesn't remove privs because fallocate is called with privileged mounter credentials.

This fixes some failure in fstests generic/683..687.

Fixes aab8848cee5e ("ovl: add ovl_fallocate()")
---
 fs/overlayfs/file.c | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)
fs/overlayfs/file.c +0 -0
index c8308da8909a..e90ac5376456 100644

fs/overlayfs/file.c static long ovl_fallocate(struct file *file, int mode, loff_t offset, loff_t len · line 523

Quoted 7 lines
fs/overlayfs/file.c +4 -0
@@ -517,9 +517,16 @@ static long ovl_fallocate(struct file *file, int mode, loff_t offset, loff_t len
      const struct cred *old_cred;
      int ret;
 
+     inode_lock(inode);
+     /* Update mode */
+     ovl_copyattr(inode);
+     ret = file_remove_privs(file);

First, thank you for picking this up!

Let me analyze generic/683 failure of Test1 to see why you still see failures in this test:

echo "Test 1 - qa_user, non-exec file $verb"
setup_testfile
chmod a+rws $junk_file
commit_and_check "$qa_user" "$verb" 64k 64k

So this creates a file with 6666 permissions. While the file has the S_ISUID and S_ISGID bits set it does not have the S_IXGRP set. This is important in a little bit.

On a regular filesystem like xfs what will happen is:

sys_fallocate()
-> vfs_fallocate()
-> xfs_file_fallocate()

      -> file_modified()
         -> __file_remove_privs()
            -> dentry_needs_remove_privs()
               -> should_remove_suid()
            -> __remove_privs()
               newattrs.ia_valid = ATTR_FORCE | kill;
               -> notify_change()

In should_remove_suid() we can see that ATTR_KILL_SUID is raised unconditionally because the file in the test has S_ISUID set.

But we also see that ATTR_KILL_SGID won't be set because while the file is S_ISGID it is not S_IXGRP (see above) which is a condition for ATTR_KILL_SGID being raised.

So by the time we call notify_change() we have attr->ia_valid set to ATTR_KILL_SUID | ATTR_FORCE. Now notify_change() sees that ATTR_KILL_SUID is set and does:

ia_valid = attr->ia_valid |= ATTR_MODE
attr->ia_mode = (inode->i_mode & ~S_ISUID);

which means that when we call setattr_copy() later we will definitely update inode->i_mode. Note that attr->ia_mode still contain S_ISGID.

Now we call into the filesystem's ->setattr() inode operation which will end up calling setattr_copy(). Since ATTR_MODE is set we will hit:

Quoted 9 lines
if (ia_valid & ATTR_MODE) {
        umode_t mode = attr->ia_mode;
        vfsgid_t vfsgid = i_gid_into_vfsgid(mnt_userns, inode);
        if (!vfsgid_in_group_p(vfsgid) &&
            !capable_wrt_inode_uidgid(mnt_userns, inode, CAP_FSETID))
                mode &= ~S_ISGID;
        inode->i_mode = mode;
}

Can you think of a reason why the above should not be done in notify_change() before even calling to ->setattr()?

Although, it wouldn't help because ovl_setattr() does:

    if (attr->ia_valid & (ATTR_KILL_SUID|ATTR_KILL_SGID))
        attr->ia_valid &= ~ATTR_MODE;
Quoted 23 lines

and since the caller in the test is neither capable nor in the group of the inode the S_ISGID bit is stripped.

But now contrast this with overlayfs even after your changes. When ovl_setattr() is hit from ovl_fallocate()'s call to file_remove_privs() and calls ovl_do_notify_change() then we are doing this under the mounter's creds and so the S_ISGID bit is retained:

sys_fallocate()
-> vfs_fallocate()
-> ovl_fallocate()

      -> file_remove_privs()
         -> dentry_needs_remove_privs()
            -> should_remove_suid()
         -> __remove_privs()
            newattrs.ia_valid = attr_force | kill;
            -> notify_change()
               -> ovl_setattr()
                  // TAKE ON MOUNTER'S CREDS
                  -> ovl_do_notify_change()
                  // GIVE UP MOUNTER'S CREDS
     // TAKE ON MOUNTER'S CREDS
     -> vfs_fallocate()
Quoted 8 lines
        -> xfs_file_fallocate()
           -> file_modified()
              -> __file_remove_privs()
                 -> dentry_needs_remove_privs()
                    -> should_remove_suid()
                 -> __remove_privs()
                    newattrs.ia_valid = attr_force | kill;
                    -> notify_change()
Quoted 4 lines

The model in overlayfs is that security is checked twice once on overlay inode with caller creds and once again on xfs inode with mounter creds. Either of these checks could result in clearing SUID/SGID bits.

Yep.

Quoted 4 lines

In the call stack above, the outer should_remove_suid() with caller creds sets ATTR_KILL_SUID and then the outer notify_change() clears SUID and sets ATTR_MODE,

Yes.

Quoted 2 lines

but ovl_setattr() clears ATTR_MODE and then the inner notify_change() re-clears SUID and sets ATTR_MODE again.

Yes.

Quoted 8 lines

If the outer notify_change() would have checked the in_group_p() condition, clear SGID and set a flag ATTR_KILL_SGID_FORCE then the inner notify_change() would see this flag and re-clear SGID bit, just the same as it does with SUID bit in the stack stace above.

Is this making any sense?

What I kept thinking was sm along the lines of:

fs/inode.c +6 -2
index ba1de23c13c1..e62a564201b7 100644
@@ -1968,8 +1968,12 @@ int should_remove_suid(struct dentry *dentry)
         * sgid without any exec bits is just a mandatory locking mark; leave
         * it alone.  If some exec bits are set, it's a real sgid; kill it.
         */
-       if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
-               kill |= ATTR_KILL_SGID;
+       if (unlikely(mode & S_ISGID)) {
+               if ((mode & S_IXGRP) ||
+                   (!vfsgid_in_group_p(vfsgid) &&
+                    !capable_wrt_inode_uidgid(mnt_userns, inode, CAP_FSETID)))
+                       kill |= ATTR_KILL_SGID;
+       }
 
        if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
                return kill;
Quoted 8 lines
Quoted 2 lines

mandatory locks have been removed as well so that remark seems pointless as well?

My feeling here is that both should_remove_suid() and notify_change() need to apply the same permission checks as setattr_prepare() and setattr_copy() instead of all this special casing them. I don't see a good reason to not require the same checks. So sm like (__completely untested__):

I like it :)
Few small nits. Some you may have already noticed after testing...

Quoted 26 lines

From 922f9f123ab6531c29bf05585ef88c17fe65dba3 Mon Sep 17 00:00:00 2001
From: Christian Brauner <[email protected]>
Date: Tue, 4 Oct 2022 16:13:34 +0200
Subject: [UNTESTED PATCH] attr: use consistent sgid stripping checks

Require the same permissions as setattr_prepare() and setattr_copy() have instead of all these special cases. We can probably consolidate this even more...

---
 fs/attr.c          |  2 +-
 fs/fuse/file.c     |  2 +-
 fs/inode.c         | 41 ++++++++++++++++++++++++-----------------
 fs/internal.h      |  4 +++-
 fs/ocfs2/file.c    |  4 ++--
 fs/open.c          |  2 +-
 include/linux/fs.h |  2 +-
 7 files changed, 33 insertions(+), 24 deletions(-)
fs/attr.c +0 -0
index 1552a5f23d6b..9262c6b31c26 100644

fs/attr.c int notify_change(struct user_namespace *mnt_userns, struct dentry *dentry, · line 378

Quoted 5 lines
fs/attr.c +1 -1
@@ -375,7 +375,7 @@ int notify_change(struct user_namespace *mnt_userns, struct dentry *dentry,
                }
        }
        if (ia_valid & ATTR_KILL_SGID) {
-               if ((mode & (S_ISGID | S_IXGRP)) == (S_ISGID | S_IXGRP)) {
+               if (should_remove_sgid(mnt_userns, dentry)) {

This should just be:
+ if (mode & S_ISGID) {

Just like the case with ATTR_KILL_SUID,
notify_change() should follow ATTR_KILL_* hints issued by dentry_needs_remove_privs() without re-checking the conditions. This is what makes overlayfs SUID stripping work in current code.

Quoted 55 lines
fs/attr.c +0 -0
                        if (!(ia_valid & ATTR_MODE)) {
                                ia_valid = attr->ia_valid |= ATTR_MODE;
                                attr->ia_mode = inode->i_mode;
fs/fuse/file.c +1 -1
index 1a3afd469e3a..fccc2c7e88fd 100644
@@ -1313,7 +1313,7 @@ static ssize_t fuse_cache_write_iter(struct kiocb *iocb, struct iov_iter *from)
                        return err;
 
                if (fc->handle_killpriv_v2 &&
-                   should_remove_suid(file_dentry(file))) {
+                   should_remove_suid(&init_user_ns, file_dentry(file))) {
                        goto writethrough;
                }
 
fs/inode.c +18 -8
index ba1de23c13c1..c639aefe01c3 100644
@@ -1949,27 +1949,33 @@ void touch_atime(const struct path *path)
 }
 EXPORT_SYMBOL(touch_atime);
 
-/*
- * The logic we want is
- *
- *     if suid or (sgid and xgrp)
- *             remove privs
- */
-int should_remove_suid(struct dentry *dentry)
+bool should_remove_sgid(struct user_namespace *mnt_userns, struct dentry *dentry)
+{
+       struct inode *inode = d_inode(dentry);
+       umode_t mode = inode->i_mode;
+
+       if (unlikely(mode & S_ISGID)) {
+               if ((mode & S_IXGRP) ||
+                   (!vfsgid_in_group_p(i_gid_into_vfsgid(mnt_userns, inode)) &&
+                    !capable_wrt_inode_uidgid(mnt_userns, inode, CAP_FSETID)))
+                       return true;
+       }
+
+       return false;
+}
+
+int should_remove_suid(struct user_namespace *mnt_userns, struct dentry *dentry)
 {
-       umode_t mode = d_inode(dentry)->i_mode;
+       struct inode *inode = d_inode(dentry);
+       umode_t mode = inode->i_mode;
        int kill = 0;
 
        /* suid always must be killed */
        if (unlikely(mode & S_ISUID))

fs/inode.c void touch_atime(const struct path *path) · line 1978

Quoted 10 lines
fs/inode.c +2 -6
                kill = ATTR_KILL_SUID;
 
-       /*
-        * sgid without any exec bits is just a mandatory locking mark; leave
-        * it alone.  If some exec bits are set, it's a real sgid; kill it.
-        */
-       if (unlikely((mode & S_ISGID) && (mode & S_IXGRP)))
-               kill |= ATTR_KILL_SGID;
+       if (should_remove_sgid(mnt_userns, dentry))
+               kill = ATTR_KILL_SGID;

kill |= ATTR_KILL_SGID;

Quoted 46 lines
fs/inode.c +4 -3
        if (unlikely(kill && !capable(CAP_FSETID) && S_ISREG(mode)))
                return kill;
@@ -1983,7 +1989,8 @@ EXPORT_SYMBOL(should_remove_suid);
  * response to write or truncate. Return 0 if nothing has to be changed.
  * Negative value on error (change should be denied).
  */
-int dentry_needs_remove_privs(struct dentry *dentry)
+int dentry_needs_remove_privs(struct user_namespace *mnt_userns,
+                             struct dentry *dentry)
 {
        struct inode *inode = d_inode(dentry);
        int mask = 0;
@@ -1992,7 +1999,7 @@ int dentry_needs_remove_privs(struct dentry *dentry)
        if (IS_NOSEC(inode))
                return 0;
 
-       mask = should_remove_suid(dentry);
+       mask = should_remove_suid(mnt_userns, dentry);
        ret = security_inode_need_killpriv(dentry);
        if (ret < 0)
                return ret;
@@ -2024,7 +2031,7 @@ static int __file_remove_privs(struct file *file, unsigned int flags)
        if (IS_NOSEC(inode) || !S_ISREG(inode->i_mode))
                return 0;
 
-       kill = dentry_needs_remove_privs(dentry);
+       kill = dentry_needs_remove_privs(file_mnt_user_ns(file), dentry);
        if (kill < 0)
                return kill;
 
fs/internal.h +2 -1
index 1e67b4b9a4d1..ae152ded227c 100644
@@ -139,7 +139,8 @@ extern int vfs_open(const struct path *, struct file *);
  * inode.c
  */
 extern long prune_icache_sb(struct super_block *sb, struct shrink_control *sc);
-extern int dentry_needs_remove_privs(struct dentry *dentry);
+extern int dentry_needs_remove_privs(struct user_namespace *,
+                                    struct dentry *dentry);
 
 /*
  * fs-writeback.c

fs/internal.h int do_set_acl(struct user_namespace *mnt_userns, struct dentry *dentry, · line 230

Quoted 4 lines
fs/internal.h +1 -0
@@ -226,3 +227,4 @@ int do_set_acl(struct user_namespace *mnt_userns, struct dentry *dentry,
               const char *acl_name, const void *kvalue, size_t size);
 ssize_t do_get_acl(struct user_namespace *mnt_userns, struct dentry *dentry,
                   const char *acl_name, void *kvalue, size_t size);
+int should_remove_sgid(struct user_namespace *mnt_userns, struct dentry *dentry);

Pls move this up to /* inode.c */ section

Miklos,

Do you want to ACK the ovl_fallocate() and ovl_copyfile() patches so Chritain can pick them to his tree or do you prefer to take them through your tree? They fix SUID stripping bugs regardless of Christain's extra patch for fixing vfs SGID stripping.

Thanks,
Amir.

maintainer Acked-by on v1 2/2
#10 lore raw

Re: [PATCH 2/2] ovl: remove privs in ovl_fallocate()

In reply to Amir Goldstein
Quoted 8 lines · Amir Goldstein

Underlying fs doesn't remove privs because fallocate is called with privileged mounter credentials.

This fixes some failure in fstests generic/683..687.

Fixes aab8848cee5e ("ovl: add ovl_fallocate()")

Christian, please feel free to take these if you already have a bunch of related patches.

Thanks,
Miklos

#11 lore raw

Re: [PATCH 2/2] ovl: remove privs in ovl_fallocate()

Quoted 7 lines · Miklos Szeredi
Quoted 6 lines · Amir Goldstein

Underlying fs doesn't remove privs because fallocate is called with privileged mounter credentials.

This fixes some failure in fstests generic/683..687.

Quoted 7 lines
Quoted 2 lines
Fixes aab8848cee5e ("ovl: add ovl_fallocate()")

Christian, please feel free to take these if you already have a bunch of related patches.

Thank you, Miklos! Will do. I resend the series with the vfs related change. If you have an opinion on it I'd appreciate it.A

Thanks!
Christian