jfs: Fix fortify moan in symlink

  From: "Dr. David Alan Gilbert" <linux@treblig.org>

JFS has in jfs_incore.h:

      /* _inline may overflow into _inline_ea when needed */
      /* _inline_ea may overlay the last part of
       * file._xtroot if maxentry = XTROOTINITSLOT
       */
      union {
        struct {
          /* 128: inline symlink */
          unchar _inline[128];
          /* 128: inline extended attr */
          unchar _inline_ea[128];
        };
        unchar _inline_all[256];

and currently the symlink code copies into _inline;
if this is larger than 128 bytes it triggers a fortify warning of the
form:

  memcpy: detected field-spanning write (size 132) of single field
     "ip->i_link" at fs/jfs/namei.c:950 (size 18446744073709551615)

when it's actually OK.

Copy it into _inline_all instead.

Reported-by: syzbot+5fc38b2ddbbca7f5c680@syzkaller.appspotmail.com
Signed-off-by: Dr. David Alan Gilbert <linux@treblig.org>
---
 fs/jfs/namei.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
  

On Sat, Oct 22, 2022 at 09:39:14PM +0100, linux@treblig.org wrote:
> From: "Dr. David Alan Gilbert" <linux@treblig.org>
> 
> JFS has in jfs_incore.h:
> 
>       /* _inline may overflow into _inline_ea when needed */
>       /* _inline_ea may overlay the last part of
>        * file._xtroot if maxentry = XTROOTINITSLOT
>        */
>       union {
>         struct {
>           /* 128: inline symlink */
>           unchar _inline[128];
>           /* 128: inline extended attr */
>           unchar _inline_ea[128];
>         };
>         unchar _inline_all[256];
> 
> and currently the symlink code copies into _inline;
> if this is larger than 128 bytes it triggers a fortify warning of the
> form:
> 
>   memcpy: detected field-spanning write (size 132) of single field
>      "ip->i_link" at fs/jfs/namei.c:950 (size 18446744073709551615)

Which compiler are you using for this build? This size report (SIZE_MAX)
should be impossible to reach. But also, the size is just wrong --
i_inline is 128 bytes, not SIZE_MAX. So, the detection is working
(132 > 128), but the report is broken, and I can't see how...

> 
> when it's actually OK.
> 
> Copy it into _inline_all instead.
> 
> Reported-by: syzbot+5fc38b2ddbbca7f5c680@syzkaller.appspotmail.com
> Signed-off-by: Dr. David Alan Gilbert <linux@treblig.org>
> ---
>  fs/jfs/namei.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/fs/jfs/namei.c b/fs/jfs/namei.c
> index 9db4f5789c0ec..4fbbf88435e69 100644
> --- a/fs/jfs/namei.c
> +++ b/fs/jfs/namei.c
> @@ -946,7 +946,7 @@ static int jfs_symlink(struct user_namespace *mnt_userns, struct inode *dip,
>  	if (ssize <= IDATASIZE) {
>  		ip->i_op = &jfs_fast_symlink_inode_operations;
>  
> -		ip->i_link = JFS_IP(ip)->i_inline;
> +		ip->i_link = JFS_IP(ip)->i_inline_all;
>  		memcpy(ip->i_link, name, ssize);
>  		ip->i_size = ssize - 1;
>  

Regardless, the fix looks correct to me!

Reviewed-by: Kees Cook <keescook@chromium.org>
  

* Kees Cook (keescook@chromium.org) wrote:
> On Sat, Oct 22, 2022 at 09:39:14PM +0100, linux@treblig.org wrote:
> > From: "Dr. David Alan Gilbert" <linux@treblig.org>
> > 
> > JFS has in jfs_incore.h:
> > 
> >       /* _inline may overflow into _inline_ea when needed */
> >       /* _inline_ea may overlay the last part of
> >        * file._xtroot if maxentry = XTROOTINITSLOT
> >        */
> >       union {
> >         struct {
> >           /* 128: inline symlink */
> >           unchar _inline[128];
> >           /* 128: inline extended attr */
> >           unchar _inline_ea[128];
> >         };
> >         unchar _inline_all[256];
> > 
> > and currently the symlink code copies into _inline;
> > if this is larger than 128 bytes it triggers a fortify warning of the
> > form:
> > 
> >   memcpy: detected field-spanning write (size 132) of single field
> >      "ip->i_link" at fs/jfs/namei.c:950 (size 18446744073709551615)
> 
> Which compiler are you using for this build?

I think that report was the same on gcc on Fedora 37 and whatever
syzkaller was running.

> This size report (SIZE_MAX)
> should be impossible to reach. But also, the size is just wrong --
> i_inline is 128 bytes, not SIZE_MAX. So, the detection is working
> (132 > 128), but the report is broken, and I can't see how...

Yeh, and led me down a blind alley for a while thinking something had
really managed to screwup the strlen somehow.

> 
> > 
> > when it's actually OK.
> > 
> > Copy it into _inline_all instead.
> > 
> > Reported-by: syzbot+5fc38b2ddbbca7f5c680@syzkaller.appspotmail.com
> > Signed-off-by: Dr. David Alan Gilbert <linux@treblig.org>
> > ---
> >  fs/jfs/namei.c | 2 +-
> >  1 file changed, 1 insertion(+), 1 deletion(-)
> > 
> > diff --git a/fs/jfs/namei.c b/fs/jfs/namei.c
> > index 9db4f5789c0ec..4fbbf88435e69 100644
> > --- a/fs/jfs/namei.c
> > +++ b/fs/jfs/namei.c
> > @@ -946,7 +946,7 @@ static int jfs_symlink(struct user_namespace *mnt_userns, struct inode *dip,
> >  	if (ssize <= IDATASIZE) {
> >  		ip->i_op = &jfs_fast_symlink_inode_operations;
> >  
> > -		ip->i_link = JFS_IP(ip)->i_inline;
> > +		ip->i_link = JFS_IP(ip)->i_inline_all;
> >  		memcpy(ip->i_link, name, ssize);
> >  		ip->i_size = ssize - 1;
> >  
> 
> Regardless, the fix looks correct to me!
> 
> Reviewed-by: Kees Cook <keescook@chromium.org>

Thanks!

Dave

> -- 
> Kees Cook
  

Applied.

Thanks,
Shaggy

On 10/24/22 1:49PM, Dr. David Alan Gilbert wrote:
> * Kees Cook (keescook@chromium.org) wrote:
>> On Sat, Oct 22, 2022 at 09:39:14PM +0100, linux@treblig.org wrote:
>>> From: "Dr. David Alan Gilbert" <linux@treblig.org>
>>>
>>> JFS has in jfs_incore.h:
>>>
>>>        /* _inline may overflow into _inline_ea when needed */
>>>        /* _inline_ea may overlay the last part of
>>>         * file._xtroot if maxentry = XTROOTINITSLOT
>>>         */
>>>        union {
>>>          struct {
>>>            /* 128: inline symlink */
>>>            unchar _inline[128];
>>>            /* 128: inline extended attr */
>>>            unchar _inline_ea[128];
>>>          };
>>>          unchar _inline_all[256];
>>>
>>> and currently the symlink code copies into _inline;
>>> if this is larger than 128 bytes it triggers a fortify warning of the
>>> form:
>>>
>>>    memcpy: detected field-spanning write (size 132) of single field
>>>       "ip->i_link" at fs/jfs/namei.c:950 (size 18446744073709551615)
>>
>> Which compiler are you using for this build?
> 
> I think that report was the same on gcc on Fedora 37 and whatever
> syzkaller was running.
> 
>> This size report (SIZE_MAX)
>> should be impossible to reach. But also, the size is just wrong --
>> i_inline is 128 bytes, not SIZE_MAX. So, the detection is working
>> (132 > 128), but the report is broken, and I can't see how...
> 
> Yeh, and led me down a blind alley for a while thinking something had
> really managed to screwup the strlen somehow.
> 
>>
>>>
>>> when it's actually OK.
>>>
>>> Copy it into _inline_all instead.
>>>
>>> Reported-by: syzbot+5fc38b2ddbbca7f5c680@syzkaller.appspotmail.com
>>> Signed-off-by: Dr. David Alan Gilbert <linux@treblig.org>
>>> ---
>>>   fs/jfs/namei.c | 2 +-
>>>   1 file changed, 1 insertion(+), 1 deletion(-)
>>>
>>> diff --git a/fs/jfs/namei.c b/fs/jfs/namei.c
>>> index 9db4f5789c0ec..4fbbf88435e69 100644
>>> --- a/fs/jfs/namei.c
>>> +++ b/fs/jfs/namei.c
>>> @@ -946,7 +946,7 @@ static int jfs_symlink(struct user_namespace *mnt_userns, struct inode *dip,
>>>   	if (ssize <= IDATASIZE) {
>>>   		ip->i_op = &jfs_fast_symlink_inode_operations;
>>>   
>>> -		ip->i_link = JFS_IP(ip)->i_inline;
>>> +		ip->i_link = JFS_IP(ip)->i_inline_all;
>>>   		memcpy(ip->i_link, name, ssize);
>>>   		ip->i_size = ssize - 1;
>>>   
>>
>> Regardless, the fix looks correct to me!
>>
>> Reviewed-by: Kees Cook <keescook@chromium.org>
> 
> Thanks!
> 
> Dave
> 
>> -- 
>> Kees Cook
  

On Mon, Oct 24, 2022 at 07:49:17PM +0100, Dr. David Alan Gilbert wrote:
> * Kees Cook (keescook@chromium.org) wrote:
> > On Sat, Oct 22, 2022 at 09:39:14PM +0100, linux@treblig.org wrote:
> > > From: "Dr. David Alan Gilbert" <linux@treblig.org>
> > > 
> > > JFS has in jfs_incore.h:
> > > 
> > >       /* _inline may overflow into _inline_ea when needed */
> > >       /* _inline_ea may overlay the last part of
> > >        * file._xtroot if maxentry = XTROOTINITSLOT
> > >        */
> > >       union {
> > >         struct {
> > >           /* 128: inline symlink */
> > >           unchar _inline[128];
> > >           /* 128: inline extended attr */
> > >           unchar _inline_ea[128];
> > >         };
> > >         unchar _inline_all[256];
> > > 
> > > and currently the symlink code copies into _inline;
> > > if this is larger than 128 bytes it triggers a fortify warning of the
> > > form:
> > > 
> > >   memcpy: detected field-spanning write (size 132) of single field
> > >      "ip->i_link" at fs/jfs/namei.c:950 (size 18446744073709551615)
> > 
> > Which compiler are you using for this build?
> 
> I think that report was the same on gcc on Fedora 37 and whatever
> syzkaller was running.
> 
> > This size report (SIZE_MAX)
> > should be impossible to reach. But also, the size is just wrong --
> > i_inline is 128 bytes, not SIZE_MAX. So, the detection is working
> > (132 > 128), but the report is broken, and I can't see how...
> 
> Yeh, and led me down a blind alley for a while thinking something had
> really managed to screwup the strlen somehow.

This looks like a GCC bug (going at least back to GCC 10.2)[1], but some
extra care around the macro appears to make it go away, so the reporting
variable doesn't get confused/re-evaluated:

diff --git a/include/linux/fortify-string.h b/include/linux/fortify-string.h
index 09a032f6ce6b..9e2d96993c30 100644
--- a/include/linux/fortify-string.h
+++ b/include/linux/fortify-string.h
@@ -550,13 +550,18 @@ __FORTIFY_INLINE bool fortify_memcpy_chk(__kernel_size_t size,
 
 #define __fortify_memcpy_chk(p, q, size, p_size, q_size,		\
 			     p_size_field, q_size_field, op) ({		\
-	size_t __fortify_size = (size_t)(size);				\
-	WARN_ONCE(fortify_memcpy_chk(__fortify_size, p_size, q_size,	\
-				     p_size_field, q_size_field, #op),	\
+	const size_t __fortify_size = (size_t)(size);			\
+	const size_t __p_size = (p_size);				\
+	const size_t __q_size = (q_size);				\
+	const size_t __p_size_field = (p_size_field);			\
+	const size_t __q_size_field = (q_size_field);			\
+	WARN_ONCE(fortify_memcpy_chk(__fortify_size, __p_size,		\
+				     __q_size, __p_size_field,		\
+				     __q_size_field, #op),		\
 		  #op ": detected field-spanning write (size %zu) of single %s (size %zu)\n", \
 		  __fortify_size,					\
 		  "field \"" #p "\" at " __FILE__ ":" __stringify(__LINE__), \
-		  p_size_field);					\
+		  __p_size_field);					\
 	__underlying_##op(p, q, __fortify_size);			\
 })
 


[1] https://syzkaller.appspot.com/bug?id=23d613df5259b977dac1696bec77f61a85890e3d
  

* Kees Cook (keescook@chromium.org) wrote:
> On Mon, Oct 24, 2022 at 07:49:17PM +0100, Dr. David Alan Gilbert wrote:
> > * Kees Cook (keescook@chromium.org) wrote:
> > > On Sat, Oct 22, 2022 at 09:39:14PM +0100, linux@treblig.org wrote:
> > > > From: "Dr. David Alan Gilbert" <linux@treblig.org>
> > > > 
> > > > JFS has in jfs_incore.h:
> > > > 
> > > >       /* _inline may overflow into _inline_ea when needed */
> > > >       /* _inline_ea may overlay the last part of
> > > >        * file._xtroot if maxentry = XTROOTINITSLOT
> > > >        */
> > > >       union {
> > > >         struct {
> > > >           /* 128: inline symlink */
> > > >           unchar _inline[128];
> > > >           /* 128: inline extended attr */
> > > >           unchar _inline_ea[128];
> > > >         };
> > > >         unchar _inline_all[256];
> > > > 
> > > > and currently the symlink code copies into _inline;
> > > > if this is larger than 128 bytes it triggers a fortify warning of the
> > > > form:
> > > > 
> > > >   memcpy: detected field-spanning write (size 132) of single field
> > > >      "ip->i_link" at fs/jfs/namei.c:950 (size 18446744073709551615)
> > > 
> > > Which compiler are you using for this build?
> > 
> > I think that report was the same on gcc on Fedora 37 and whatever
> > syzkaller was running.
> > 
> > > This size report (SIZE_MAX)
> > > should be impossible to reach. But also, the size is just wrong --
> > > i_inline is 128 bytes, not SIZE_MAX. So, the detection is working
> > > (132 > 128), but the report is broken, and I can't see how...
> > 
> > Yeh, and led me down a blind alley for a while thinking something had
> > really managed to screwup the strlen somehow.
> 
> This looks like a GCC bug (going at least back to GCC 10.2)[1], but some
> extra care around the macro appears to make it go away, so the reporting
> variable doesn't get confused/re-evaluated:

Thanks for chasing that; are you also going to file a gcc bug?

Dave

> diff --git a/include/linux/fortify-string.h b/include/linux/fortify-string.h
> index 09a032f6ce6b..9e2d96993c30 100644
> --- a/include/linux/fortify-string.h
> +++ b/include/linux/fortify-string.h
> @@ -550,13 +550,18 @@ __FORTIFY_INLINE bool fortify_memcpy_chk(__kernel_size_t size,
>  
>  #define __fortify_memcpy_chk(p, q, size, p_size, q_size,		\
>  			     p_size_field, q_size_field, op) ({		\
> -	size_t __fortify_size = (size_t)(size);				\
> -	WARN_ONCE(fortify_memcpy_chk(__fortify_size, p_size, q_size,	\
> -				     p_size_field, q_size_field, #op),	\
> +	const size_t __fortify_size = (size_t)(size);			\
> +	const size_t __p_size = (p_size);				\
> +	const size_t __q_size = (q_size);				\
> +	const size_t __p_size_field = (p_size_field);			\
> +	const size_t __q_size_field = (q_size_field);			\
> +	WARN_ONCE(fortify_memcpy_chk(__fortify_size, __p_size,		\
> +				     __q_size, __p_size_field,		\
> +				     __q_size_field, #op),		\
>  		  #op ": detected field-spanning write (size %zu) of single %s (size %zu)\n", \
>  		  __fortify_size,					\
>  		  "field \"" #p "\" at " __FILE__ ":" __stringify(__LINE__), \
> -		  p_size_field);					\
> +		  __p_size_field);					\
>  	__underlying_##op(p, q, __fortify_size);			\
>  })
>  
> 
> 
> [1] https://syzkaller.appspot.com/bug?id=23d613df5259b977dac1696bec77f61a85890e3d
> 
> -- 
> Kees Cook
  

On Sat, Oct 29, 2022 at 01:48:05PM +0100, Dr. David Alan Gilbert wrote:
> * Kees Cook (keescook@chromium.org) wrote:
> > This looks like a GCC bug (going at least back to GCC 10.2)[1], but some
> > extra care around the macro appears to make it go away, so the reporting
> > variable doesn't get confused/re-evaluated:
> 
> Thanks for chasing that; are you also going to file a gcc bug?

No, I haven't had the time to minimize a PoC, so I'm not sure it would
make a very good bug report.