Date:   Mon, 6 Feb 2023 09:51:19 +1100
From:   Dave Chinner <david@fromorbit.com>
To:     "Gustavo A. R. Silva" <gustavoars@kernel.org>
Cc:     "Darrick J. Wong" <djwong@kernel.org>, linux-xfs@vger.kernel.org,
        linux-kernel@vger.kernel.org, linux-hardening@vger.kernel.org
Subject: Re: [PATCH][next] xfs: Replace one-element arrays with
 flexible-array members
Message-ID: <20230205225119.GU360264@dread.disaster.area>
References: <Y9xiYmVLRIKdpJcC@work>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <Y9xiYmVLRIKdpJcC@work>
Precedence: bulk

On Thu, Feb 02, 2023 at 07:24:50PM -0600, Gustavo A. R. Silva wrote:
> One-element arrays are deprecated, and we are replacing them with flexible
> array members instead. So, replace one-element arrays with flexible-array
> members in structures xfs_attr_leaf_name_local and
> xfs_attr_leaf_name_remote.
> 
> The only binary differences reported after the changes are all like
> these:
> 
> fs/xfs/libxfs/xfs_attr_leaf.o
> _@@ -435,7 +435,7 @@
>       3b8:      movzbl 0x2(%rbx),%eax
>       3bc:      rol    $0x8,%bp
>       3c0:      movzwl %bp,%ebp
> -     3c3:      lea    0x2(%rax,%rbp,1),%ebx
> +     3c3:      lea    0x3(%rax,%rbp,1),%ebx
>       3c7:      call   3cc <xfs_attr_leaf_entsize+0x8c>
>                         3c8: R_X86_64_PLT32     __tsan_func_exit-0x4
>       3cc:      or     $0x3,%ebx
> _@@ -454,7 +454,7 @@
>       3ea:      movzbl 0x8(%rbx),%ebx
>       3ee:      call   3f3 <xfs_attr_leaf_entsize+0xb3>
>                         3ef: R_X86_64_PLT32     __tsan_func_exit-0x4
> -     3f3:      add    $0xa,%ebx
> +     3f3:      add    $0xb,%ebx
>       3f6:      or     $0x3,%ebx
>       3f9:      add    $0x1,%ebx
>       3fc:      mov    %ebx,%eax
> 
> similar changes in fs/xfs/scrub/attr.o and fs/xfs/xfs.o object files.

That seems like a red flag to me - an off-by-one change in the
compiled code that calculates of the on-disk size of a structure as
a result of an in-memory structure change just smells like a bug.

How did you test this change?

> And the reason for this is because of the round_up() macro called in
> functions xfs_attr_leaf_entsize_remote() and xfs_attr_leaf_entsize_local(),
> which is compensanting for the one-byte reduction in size (due to the
> flex-array transformation) of structures xfs_attr_leaf_name_remote and
> xfs_attr_leaf_name_local. So, sizes remain the same before and after
> changes.

I'm not sure that is true. Before this change:

sizeof(xfs_attr_leaf_name_local_t) = 4
sizeof(xfs_attr_leaf_name_remote_t) = 12

After this change:

sizeof(xfs_attr_leaf_name_local_t) = 4
sizeof(xfs_attr_leaf_name_remote_t) = 12

i.e. no change because the structures aren't defined as packed
structures.  Hence the compiler pads them to out to 4 byte alignment
naturally regardless of the flex array definition. pahole on x86-64
also confirms that the (padded) size of the structure is not
changed.

However, the on-disk structure it is being used to decode is packed,
and we're only using pointer arithmetic to pull the location of the
name/value pairs out of the buffer to copy them - it's the structure
size calculations that actually define the size of the structures
for a given name length, not the sizeof() value or the flex array
definitions...

> This helps with the ongoing efforts to tighten the FORTIFY_SOURCE
> routines on memcpy() and help us make progress towards globally
> enabling -fstrict-flex-arrays=3 [1].
> 
> Link: https://github.com/KSPP/linux/issues/79
> Link: https://github.com/KSPP/linux/issues/251
> Link: https://gcc.gnu.org/pipermail/gcc-patches/2022-October/602902.html [1]
> Signed-off-by: Gustavo A. R. Silva <gustavoars@kernel.org>
> ---
>  fs/xfs/libxfs/xfs_da_format.h | 8 ++++----
>  1 file changed, 4 insertions(+), 4 deletions(-)
> 
> diff --git a/fs/xfs/libxfs/xfs_da_format.h b/fs/xfs/libxfs/xfs_da_format.h
> index 25e2841084e1..e1e62ebb0c44 100644
> --- a/fs/xfs/libxfs/xfs_da_format.h
> +++ b/fs/xfs/libxfs/xfs_da_format.h
> @@ -620,14 +620,14 @@ typedef struct xfs_attr_leaf_entry {	/* sorted on key, not name */
>  typedef struct xfs_attr_leaf_name_local {
>  	__be16	valuelen;		/* number of bytes in value */
>  	__u8	namelen;		/* length of name bytes */
> -	__u8	nameval[1];		/* name/value bytes */
> +	__u8	nameval[];		/* name/value bytes */
>  } xfs_attr_leaf_name_local_t;
>  
>  typedef struct xfs_attr_leaf_name_remote {
>  	__be32	valueblk;		/* block number of value bytes */
>  	__be32	valuelen;		/* number of bytes in value */
>  	__u8	namelen;		/* length of name bytes */
> -	__u8	name[1];		/* name bytes */
> +	__u8	name[];			/* name bytes */
>  } xfs_attr_leaf_name_remote_t;
>  
>  typedef struct xfs_attr_leafblock {
> @@ -747,13 +747,13 @@ xfs_attr3_leaf_name_local(xfs_attr_leafblock_t *leafp, int idx)
>   */
>  static inline int xfs_attr_leaf_entsize_remote(int nlen)
>  {
> -	return round_up(sizeof(struct xfs_attr_leaf_name_remote) - 1 +
> +	return round_up(sizeof(struct xfs_attr_leaf_name_remote) +
>  			nlen, XFS_ATTR_LEAF_NAME_ALIGN);
>  }

To be honest, the actual padding and alignment calculations are
kinda whacky because that's the way they were defined back in 1995.
And, well, once set in the on-disk format, it can't easily be
changed. FYI, here's the original definition from 1995:

#define XFS_ATTR_LEAF_ENTSIZE_REMOTE(nlen)	/* space for remote struct */ \
	(((sizeof(xfs_attr_leaf_name_remote_t)-1 + (nlen)) +3)&~0x3)

So apart using round_up and defines instead of magic numbers, the
current calculation is unchanged from the original definition.

AFAICT, the modification you are proposing above breaks this because the
sizeof(xfs_attr_leaf_name_remote) result has not changed with the
change of the structure definition.

e.g. if namelen = 17, before we had:

	size	= round_up(12 - 1 + 17, 4)
		= round_up(28, 4)
		= 28

Which is correct because the on-disk format is packed:

        0   4   89  12      20   26 28
	+---+---++--+-------+-----+-+-----....
                  |---------------| 17 bytes of name. 
		                  |-| 2 bytes of padding
				    |-----.... Next attr record.

We end up with 2 bytes of padded between the end of the name and the
start of the next attribute record in the block.

But after this patch, now we calculate the size as:

	size	= round_up(12 + 17, 4)
		= round_up(29, 4)
		= 32

Which is a different result, and would result in incorrect parsing
of the attribute records in the buffer. Hence I don't think it is
valid to be changing the entsize calculations like this if sizeof()
is not changing results.

Which comes back to my original question: how did you test this?

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com