[PATCH v6 0/2] vfs: fix dedupe permission check

[PATCH v6 0/2] vfs: fix dedupe permission check

[Mark Fasheh]

Hi Andrew/Al,

Could I please have these patches put in a tree for more public testing?
They've hit fsdevel a few times now, I have links to the discussions in the
change log below.


The following patches fix a couple of issues with the permission check
we do in vfs_dedupe_file_range().

The first patch expands our check to allow dedupe of a file if the
user owns it or otherwise would be allowed to write to it.

Current behavior is that we'll allow dedupe only if:

- the user is an admin (root)
- the user has the file open for write

This makes it impossible for a user to dedupe their own file set
unless they do it as root, or ensure that all files have write
permission. There's a couple of duperemove bugs open for this:

https://github.com/markfasheh/duperemove/issues/129
https://github.com/markfasheh/duperemove/issues/86

The other problem we have is also related to forcing the user to open
target files for write - A process trying to exec a file currently
being deduped gets ETXTBUSY. The answer (as above) is to allow them to
open the targets ro - root can already do this. There was a patch from
Adam Borowski to fix this back in 2016:

https://lkml.org/lkml/2016/7/17/130

which I have incorporated into my changes.


The 2nd patch fixes our return code for permission denied to be
EPERM. For some reason we're returning EINVAL - I think that's
probably my fault. At any rate, we need to be returning something
descriptive of the actual problem, otherwise callers see EINVAL and
can't really make a valid determination of what's gone wrong.

This has also popped up in duperemove, mostly in the form of cryptic
error messages. Because this is a code returned to userspace, I did
check the other users of extent-same that I could find. Both 'bees'
and 'rust-btrfs' do the same as duperemove and simply report the error
(as they should).


One way I tested these patches was to make non-root owned files with
read-only permissions and see if I could dedupe them as the owning user. For
example, the following script fails on an unpatched kernel and succeeds with
the patches applied.

  TESTDIR=/btrfs
  USER=mfasheh

  rm -f $TESTDIR/file*

  dd if=/dev/zero of=$TESTDIR/file1 count=1024 bs=1024
  dd if=/dev/zero of=$TESTDIR/file2 count=1024 bs=1024

  chown $USER $TESTDIR/file*
  chmod 444 $TESTDIR/file*

  # open file* for read and dedupe
  sudo -u $USER duperemove -Ad $TESTDIR/file*


Lastly, I have an update to the fi_deduperange manpage to reflect these
changes. That patch is attached below.


git pull https://github.com/markfasheh/linux dedupe-perms

Thanks,
  --Mark

Changes from V5 to V6:
- Rebase and retest on 4.19-rc3
- Add a note on testing

Changes from V4 to V5:
- Rebase and retest on 4.18-rc8
- Place updated manpage patch below, CC linux-api
- V4 discussion: https://patchwork.kernel.org/patch/10530365/

Changes from V3 to V4:
- Add a patch (below) to ioctl_fideduperange.2 explaining our
  changes. I will send this patch once the kernel update is
  accepted. Thanks to Darrick Wong for this suggestion.
- V3 discussion: https://www.spinics.net/lists/linux-btrfs/msg79135.html

Changes from V2 to V3:
- Return bool from allow_file_dedupe
- V2 discussion: https://www.spinics.net/lists/linux-btrfs/msg78421.html

Changes from V1 to V2:
- Add inode_permission check as suggested by Adam Borowski
- V1 discussion: https://marc.info/?l=linux-xfs&m=152606684017965&w=2


From: Mark Fasheh <mfasheh@suse.de>

[PATCH] ioctl_fideduperange.2: clarify permission requirements

dedupe permission checks were recently relaxed - update our man page to
reflect those changes.

Signed-off-by: Mark Fasheh <mfasheh@suse.de>
---
 man2/ioctl_fideduperange.2 | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)

diff --git a/man2/ioctl_fideduperange.2 b/man2/ioctl_fideduperange.2
index 84d20a276..4040ee064 100644
--- a/man2/ioctl_fideduperange.2
+++ b/man2/ioctl_fideduperange.2
@@ -105,9 +105,12 @@ The field
 must be zero.
 During the call,
 .IR src_fd
-must be open for reading and
+must be open for reading.
 .IR dest_fd
-must be open for writing.
+can be open for writing, or reading.
+If
+.IR dest_fd
+is open for reading, the user must have write access to the file.
 The combined size of the struct
 .IR file_dedupe_range
 and the struct
@@ -185,8 +188,8 @@ This can appear if the filesystem does not support deduplicating either file
 descriptor, or if either file descriptor refers to special inodes.
 .TP
 .B EPERM
-.IR dest_fd
-is immutable.
+This will be returned if the user lacks permission to dedupe the file referenced by
+.IR dest_fd .
 .TP
 .B ETXTBSY
 One of the files is a swap file.
-- 
2.15.1

[PATCH 1/2] vfs: allow dedupe of user owned read-only files

[Mark Fasheh]

The permission check in vfs_dedupe_file_range_one() is too coarse - We only
allow dedupe of the destination file if the user is root, or they have the
file open for write.

This effectively limits a non-root user from deduping their own read-only
files. In addition, the write file descriptor that the user is forced to
hold open can prevent execution of files. As file data during a dedupe
does not change, the behavior is unexpected and this has caused a number of
issue reports. For an example, see:

https://github.com/markfasheh/duperemove/issues/129

So change the check so we allow dedupe on the target if:

- the root or admin is asking for it
- the process has write access
- the owner of the file is asking for the dedupe
- the process could get write access

That way users can open read-only and still get dedupe.

Signed-off-by: Mark Fasheh <mfasheh@suse.de>
---
 fs/read_write.c | 16 +++++++++++++++-
 1 file changed, 15 insertions(+), 1 deletion(-)

diff --git a/fs/read_write.c b/fs/read_write.c
index 39b4a21dd933..be0e8723a049 100644
--- a/fs/read_write.c
+++ b/fs/read_write.c
@@ -1964,6 +1964,20 @@ int vfs_dedupe_file_range_compare(struct inode *src, loff_t srcoff,
 }
 EXPORT_SYMBOL(vfs_dedupe_file_range_compare);
 
+/* Check whether we are allowed to dedupe the destination file */
+static bool allow_file_dedupe(struct file *file)
+{
+	if (capable(CAP_SYS_ADMIN))
+		return true;
+	if (file->f_mode & FMODE_WRITE)
+		return true;
+	if (uid_eq(current_fsuid(), file_inode(file)->i_uid))
+		return true;
+	if (!inode_permission(file_inode(file), MAY_WRITE))
+		return true;
+	return false;
+}
+
 int vfs_dedupe_file_range_one(struct file *src_file, loff_t src_pos,
 			      struct file *dst_file, loff_t dst_pos, u64 len)
 {
@@ -1978,7 +1992,7 @@ int vfs_dedupe_file_range_one(struct file *src_file, loff_t src_pos,
 		goto out_drop_write;
 
 	ret = -EINVAL;
-	if (!(capable(CAP_SYS_ADMIN) || (dst_file->f_mode & FMODE_WRITE)))
+	if (!allow_file_dedupe(dst_file))
 		goto out_drop_write;
 
 	ret = -EXDEV;
-- 
2.15.1

[PATCH 2/2] vfs: dedupe should return EPERM if permission is not granted

[Mark Fasheh]

Right now we return EINVAL if a process does not have permission to dedupe a
file. This was an oversight on my part. EPERM gives a true description of
the nature of our error, and EINVAL is already used for the case that the
filesystem does not support dedupe.

Signed-off-by: Mark Fasheh <mfasheh@suse.de>
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>
Acked-by: David Sterba <dsterba@suse.com>
---
 fs/read_write.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/fs/read_write.c b/fs/read_write.c
index be0e8723a049..c734bc2880a5 100644
--- a/fs/read_write.c
+++ b/fs/read_write.c
@@ -1991,7 +1991,7 @@ int vfs_dedupe_file_range_one(struct file *src_file, loff_t src_pos,
 	if (ret < 0)
 		goto out_drop_write;
 
-	ret = -EINVAL;
+	ret = -EPERM;
 	if (!allow_file_dedupe(dst_file))
 		goto out_drop_write;
 
-- 
2.15.1

Re: [PATCH 1/2] vfs: allow dedupe of user owned read-only files

[Mark Fasheh]

On Sun, May 13, 2018 at 10:50:25PM +0200, Adam Borowski wrote:
> On Sun, May 13, 2018 at 06:16:53PM +0000, Mark Fasheh wrote:
> > On Sat, May 12, 2018 at 04:49:20AM +0200, Adam Borowski wrote:
> > > On Fri, May 11, 2018 at 12:26:50PM -0700, Mark Fasheh wrote:
> > > > The permission check in vfs_dedupe_file_range() is too coarse - We
> > > > only allow dedupe of the destination file if the user is root, or
> > > > they have the file open for write.
> > > > 
> > > > This effectively limits a non-root user from deduping their own
> > > > read-only files. As file data during a dedupe does not change,
> > > > this is unexpected behavior and this has caused a number of issue
> > > > reports.
> [...]
> > > > So change the check so we allow dedupe on the target if:
> > > > 
> > > > - the root or admin is asking for it
> > > > - the owner of the file is asking for the dedupe
> > > > - the process has write access
> > > 
> > > I submitted a similar patch in May 2016, yet it has never been applied
> > > despite multiple pings, with no NAK.  My version allowed dedupe if:
> > > - the root or admin is asking for it
> > > - the file has w permission (on the inode -- ie, could have been opened rw)
> > 
> > Ahh, yes I see that now. I did wind up acking it too :)
> > > 
> > > I like this new version better than mine: "root or owner or w" is more
> > > Unixy than "could have been opened w".
> > 
> > I agree, IMHO the behavior in this patch is intuitive. What we had before
> > would surprise users.
> 
> Actually, there's one reason to still consider "could have been opened w":
> with it, deduplication programs can simply open the file r and not care
> about ETXTBSY at all.  Otherwise, every program needs to stat() and have
> logic to pick the proper argument to the open() call (r if owner/root,
> rw or w if not).

That makes sense. The goal after all is to be able to just open r and not
worry about it. I hadn't considered the other possibilities that
inode_permission() covers. I'll make that change for my next series.

Thanks,
  --Mark

Re: [PATCH 1/2] vfs: allow dedupe of user owned read-only files

[Adam Borowski]

On Sun, May 13, 2018 at 06:16:53PM +0000, Mark Fasheh wrote:
> On Sat, May 12, 2018 at 04:49:20AM +0200, Adam Borowski wrote:
> > On Fri, May 11, 2018 at 12:26:50PM -0700, Mark Fasheh wrote:
> > > The permission check in vfs_dedupe_file_range() is too coarse - We
> > > only allow dedupe of the destination file if the user is root, or
> > > they have the file open for write.
> > > 
> > > This effectively limits a non-root user from deduping their own
> > > read-only files. As file data during a dedupe does not change,
> > > this is unexpected behavior and this has caused a number of issue
> > > reports.
[...]
> > > So change the check so we allow dedupe on the target if:
> > > 
> > > - the root or admin is asking for it
> > > - the owner of the file is asking for the dedupe
> > > - the process has write access
> > 
> > I submitted a similar patch in May 2016, yet it has never been applied
> > despite multiple pings, with no NAK.  My version allowed dedupe if:
> > - the root or admin is asking for it
> > - the file has w permission (on the inode -- ie, could have been opened rw)
> 
> Ahh, yes I see that now. I did wind up acking it too :)
> > 
> > I like this new version better than mine: "root or owner or w" is more
> > Unixy than "could have been opened w".
> 
> I agree, IMHO the behavior in this patch is intuitive. What we had before
> would surprise users.

Actually, there's one reason to still consider "could have been opened w":
with it, deduplication programs can simply open the file r and not care
about ETXTBSY at all.  Otherwise, every program needs to stat() and have
logic to pick the proper argument to the open() call (r if owner/root,
rw or w if not).

I also have a sister patch: btrfs_ioctl_defrag wants the same change, for
the very same reason.  But, let's discuss dedupe first to avoid unnecessary
round trips.


Meow!
-- 
⢀⣴⠾⠻⢶⣦⠀ 
⣾⠁⢰⠒⠀⣿⡁ 
⢿⡄⠘⠷⠚⠋⠀ Certified airhead; got the CT scan to prove that!
⠈⠳⣄⠀⠀⠀⠀ 

Re: [PATCH 1/2] vfs: allow dedupe of user owned read-only files

[Mark Fasheh]

Hey Adam,

On Sat, May 12, 2018 at 04:49:20AM +0200, Adam Borowski wrote:
> On Fri, May 11, 2018 at 12:26:50PM -0700, Mark Fasheh wrote:
> > The permission check in vfs_dedupe_file_range() is too coarse - We
> > only allow dedupe of the destination file if the user is root, or
> > they have the file open for write.
> > 
> > This effectively limits a non-root user from deduping their own
> > read-only files. As file data during a dedupe does not change,
> > this is unexpected behavior and this has caused a number of issue
> > reports. For an example, see:
> > 
> > https://github.com/markfasheh/duperemove/issues/129
> > 
> > So change the check so we allow dedupe on the target if:
> > 
> > - the root or admin is asking for it
> > - the owner of the file is asking for the dedupe
> > - the process has write access
> 
> I submitted a similar patch in May 2016, yet it has never been applied
> despite multiple pings, with no NAK.  My version allowed dedupe if:
> - the root or admin is asking for it
> - the file has w permission (on the inode -- ie, could have been opened rw)

Ahh, yes I see that now. I did wind up acking it too :)

> There was a request to include in xfstests a test case for the ETXTBSY race
> this patch fixes, but there's no reasonable way to make such a test case:
> the race condition is not a bug, it's write-xor-exec working as designed.
> 
> Another idea discussed was about possibly just allowing everyone who can
> open the file to deduplicate it, as the file contents are not modified in
> any way.  Zygo Blaxell expressed a concern that it could be used by an
> unprivileged user who can trigger a crash to abuse writeout bugs.
> 
> I like this new version better than mine: "root or owner or w" is more
> Unixy than "could have been opened w".

I agree, IMHO the behavior in this patch is intuitive. What we had before
would surprise users.

Yeah we've been careful about not letting a user dedupe some other users
file. Data-wise it technically might be ok but I can imagine several
scenarios where you might not want some other user deduping your files.

Thanks for the review,
	--Mark

Re: [PATCH 1/2] vfs: allow dedupe of user owned read-only files

[Adam Borowski]

On Fri, May 11, 2018 at 12:26:50PM -0700, Mark Fasheh wrote:
> The permission check in vfs_dedupe_file_range() is too coarse - We
> only allow dedupe of the destination file if the user is root, or
> they have the file open for write.
> 
> This effectively limits a non-root user from deduping their own
> read-only files. As file data during a dedupe does not change,
> this is unexpected behavior and this has caused a number of issue
> reports. For an example, see:
> 
> https://github.com/markfasheh/duperemove/issues/129
> 
> So change the check so we allow dedupe on the target if:
> 
> - the root or admin is asking for it
> - the owner of the file is asking for the dedupe
> - the process has write access

I submitted a similar patch in May 2016, yet it has never been applied
despite multiple pings, with no NAK.  My version allowed dedupe if:
- the root or admin is asking for it
- the file has w permission (on the inode -- ie, could have been opened rw)

There was a request to include in xfstests a test case for the ETXTBSY race
this patch fixes, but there's no reasonable way to make such a test case:
the race condition is not a bug, it's write-xor-exec working as designed.

Another idea discussed was about possibly just allowing everyone who can
open the file to deduplicate it, as the file contents are not modified in
any way.  Zygo Blaxell expressed a concern that it could be used by an
unprivileged user who can trigger a crash to abuse writeout bugs.

I like this new version better than mine: "root or owner or w" is more
Unixy than "could have been opened w".

> Signed-off-by: Mark Fasheh <mfasheh@suse.de>
> ---
>  fs/read_write.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/read_write.c b/fs/read_write.c
> index c4eabbfc90df..77986a2e2a3b 100644
> --- a/fs/read_write.c
> +++ b/fs/read_write.c
> @@ -2036,7 +2036,8 @@ int vfs_dedupe_file_range(struct file *file, struct file_dedupe_range *same)
>  
>  		if (info->reserved) {
>  			info->status = -EINVAL;
> -		} else if (!(is_admin || (dst_file->f_mode & FMODE_WRITE))) {
> +		} else if (!(is_admin || (dst_file->f_mode & FMODE_WRITE) ||
> +			     uid_eq(current_fsuid(), dst->i_uid))) {
I had:
  +		} else if (!(is_admin || !inode_permission(dst, MAY_WRITE))) {
>  			info->status = -EINVAL;
>  		} else if (file->f_path.mnt != dst_file->f_path.mnt) {
>  			info->status = -EXDEV;
> -- 

Meow!
-- 
⢀⣴⠾⠻⢶⣦⠀ 
⣾⠁⢰⠒⠀⣿⡁ 
⢿⡄⠘⠷⠚⠋⠀ Certified airhead; got the CT scan to prove that!
⠈⠳⣄⠀⠀⠀⠀ 

Re: [PATCH 1/2] vfs: allow dedupe of user owned read-only files

[Darrick J. Wong]

On Fri, May 11, 2018 at 12:26:50PM -0700, Mark Fasheh wrote:
> The permission check in vfs_dedupe_file_range() is too coarse - We
> only allow dedupe of the destination file if the user is root, or
> they have the file open for write.
> 
> This effectively limits a non-root user from deduping their own
> read-only files. As file data during a dedupe does not change,
> this is unexpected behavior and this has caused a number of issue
> reports. For an example, see:
> 
> https://github.com/markfasheh/duperemove/issues/129
> 
> So change the check so we allow dedupe on the target if:
> 
> - the root or admin is asking for it
> - the owner of the file is asking for the dedupe
> - the process has write access
> 
> Signed-off-by: Mark Fasheh <mfasheh@suse.de>

Sounds fine I guess....
Acked-by: Darrick J. Wong <darrick.wong@oracle.com>

--D

> ---
>  fs/read_write.c | 3 ++-
>  1 file changed, 2 insertions(+), 1 deletion(-)
> 
> diff --git a/fs/read_write.c b/fs/read_write.c
> index c4eabbfc90df..77986a2e2a3b 100644
> --- a/fs/read_write.c
> +++ b/fs/read_write.c
> @@ -2036,7 +2036,8 @@ int vfs_dedupe_file_range(struct file *file, struct file_dedupe_range *same)
>  
>  		if (info->reserved) {
>  			info->status = -EINVAL;
> -		} else if (!(is_admin || (dst_file->f_mode & FMODE_WRITE))) {
> +		} else if (!(is_admin || (dst_file->f_mode & FMODE_WRITE) ||
> +			     uid_eq(current_fsuid(), dst->i_uid))) {
>  			info->status = -EINVAL;
>  		} else if (file->f_path.mnt != dst_file->f_path.mnt) {
>  			info->status = -EXDEV;
> -- 
> 2.15.1
> 

[PATCH 1/2] vfs: allow dedupe of user owned read-only files

[Mark Fasheh]

The permission check in vfs_dedupe_file_range() is too coarse - We
only allow dedupe of the destination file if the user is root, or
they have the file open for write.

This effectively limits a non-root user from deduping their own
read-only files. As file data during a dedupe does not change,
this is unexpected behavior and this has caused a number of issue
reports. For an example, see:

https://github.com/markfasheh/duperemove/issues/129

So change the check so we allow dedupe on the target if:

- the root or admin is asking for it
- the owner of the file is asking for the dedupe
- the process has write access

Signed-off-by: Mark Fasheh <mfasheh@suse.de>
---
 fs/read_write.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/fs/read_write.c b/fs/read_write.c
index c4eabbfc90df..77986a2e2a3b 100644
--- a/fs/read_write.c
+++ b/fs/read_write.c
@@ -2036,7 +2036,8 @@ int vfs_dedupe_file_range(struct file *file, struct file_dedupe_range *same)
 
 		if (info->reserved) {
 			info->status = -EINVAL;
-		} else if (!(is_admin || (dst_file->f_mode & FMODE_WRITE))) {
+		} else if (!(is_admin || (dst_file->f_mode & FMODE_WRITE) ||
+			     uid_eq(current_fsuid(), dst->i_uid))) {
 			info->status = -EINVAL;
 		} else if (file->f_path.mnt != dst_file->f_path.mnt) {
 			info->status = -EXDEV;
-- 
2.15.1