Skip to contentSkip to conversation

iommufd: Add top-level bounds check on kernel buffer size

PATCH In review Kees Cook 2 participants 5 messages Jan 27, 2023 – Feb 1, 2023

Subsystems (2)
lore
Share & tools
lore

Kees Cook

author
#0 lore raw

PATCH 1/1 iommufd: Add top-level bounds check on kernel buffer size

While the op->size assignments are already bounds-checked at static initializer time, these limits aren't aggregated and tracked when doing later variable range checking under -Warray-bounds. Help the compiler see that we know what we're talking about, and we'll never ask to write more that sizeof(ucmd.cmd) bytes during the memset() inside copy_struct_from_user(). Seen under GCC 13:

In function 'copy_struct_from_user',
inlined from 'iommufd_fops_ioctl' at ../drivers/iommu/iommufd/main.c:333:8:
../include/linux/fortify-string.h:59:33: warning: '__builtin_memset' offset [57, 4294967294] is out of the bounds [0, 56] of object 'buf' with type 'union ucmd_buffer' [-Warray-bounds=]

   59 | #define __underlying_memset     __builtin_memset
      |                                 ^
../include/linux/fortify-string.h:453:9: note: in expansion of macro '__underlying_memset'
  453 |         __underlying_memset(p, c, __fortify_size); \
      |         ^~~~~~~~~~~~~~~~~~~
../include/linux/fortify-string.h:461:25: note: in expansion of macro '__fortify_memset_chk'
  461 | #define memset(p, c, s) __fortify_memset_chk(p, c, s, \
      |                         ^~~~~~~~~~~~~~~~~~~~
../include/linux/uaccess.h:334:17: note: in expansion of macro 'memset'
  334 |                 memset(dst + size, 0, rest);
      |                 ^~~~~~

../drivers/iommu/iommufd/main.c: In function 'iommufd_fops_ioctl':
../drivers/iommu/iommufd/main.c:311:27: note: 'buf' declared here
311 | union ucmd_buffer buf;
| ^~~

Cc Jason Gunthorpe <[email protected]>
Cc Kevin Tian <[email protected]>
Cc Joerg Roedel <[email protected]>
Cc Will Deacon <[email protected]>
Cc Robin Murphy <[email protected]>
---
 drivers/iommu/iommufd/main.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)
drivers/iommu/iommufd/main.c +3 -2
index 3fbe636c3d8a..34a1785da33a 100644
@@ -330,8 +330,9 @@ static long iommufd_fops_ioctl(struct file *filp, unsigned int cmd,
 		return -EINVAL;
 
 	ucmd.cmd = &buf;
-	ret = copy_struct_from_user(ucmd.cmd, op->size, ucmd.ubuffer,
-				    ucmd.user_size);
+	ret = copy_struct_from_user(ucmd.cmd,
+				    min_t(size_t, op->size, sizeof(ucmd.cmd)),
+				    ucmd.ubuffer, ucmd.user_size);
 	if (ret)
 		return ret;
 	ret = op->execute(&ucmd);
Signature · 1 line
2.34.1
#1 lore raw
Quoted 11 lines · Kees Cook

While the op->size assignments are already bounds-checked at static initializer time, these limits aren't aggregated and tracked when doing later variable range checking under -Warray-bounds. Help the compiler see that we know what we're talking about, and we'll never ask to write more that sizeof(ucmd.cmd) bytes during the memset() inside copy_struct_from_user(). Seen under GCC 13:

In function 'copy_struct_from_user',
inlined from 'iommufd_fops_ioctl' at ../drivers/iommu/iommufd/main.c:333:8:
../include/linux/fortify-string.h:59:33: warning: '__builtin_memset' offset [57, 4294967294] is out of the bounds [0, 56] of object 'buf' with type 'union ucmd_buffer' [-Warray-bounds=]
59 | #define __underlying_memset __builtin_memset

This seems strange to me

I thought the way gcc handled this was if it knew the value must be in a certain range then it would check it

If it couldn't figure out any ranges it would not make a warning.

So why did it decide "rest" was in that really weird range?

Is this just a compiler bug?

Jason

author
#2 lore raw
Quoted 13 lines · Jason Gunthorpe
Quoted 11 lines · Kees Cook

While the op->size assignments are already bounds-checked at static initializer time, these limits aren't aggregated and tracked when doing later variable range checking under -Warray-bounds. Help the compiler see that we know what we're talking about, and we'll never ask to write more that sizeof(ucmd.cmd) bytes during the memset() inside copy_struct_from_user(). Seen under GCC 13:

In function 'copy_struct_from_user',
inlined from 'iommufd_fops_ioctl' at ../drivers/iommu/iommufd/main.c:333:8:
../include/linux/fortify-string.h:59:33: warning: '__builtin_memset' offset [57, 4294967294] is out of the bounds [0, 56] of object 'buf' with type 'union ucmd_buffer' [-Warray-bounds=]
59 | #define __underlying_memset __builtin_memset

Quoted 8 lines

This seems strange to me

I thought the way gcc handled this was if it knew the value must be in a certain range then it would check it

If it couldn't figure out any ranges it would not make a warning.

So why did it decide "rest" was in that really weird range?

It's because it got bounds-checked at the lower end (for the minimum size test).

Quoted 2 lines

Is this just a compiler bug?

I don't think so. This just keeps the bounds within the buffer size now.

Signature · 1 line
Kees Cook
#3 lore raw
Quoted 19 lines · Kees Cook
Quoted 18 lines · Jason Gunthorpe
Quoted 11 lines · Kees Cook

While the op->size assignments are already bounds-checked at static initializer time, these limits aren't aggregated and tracked when doing later variable range checking under -Warray-bounds. Help the compiler see that we know what we're talking about, and we'll never ask to write more that sizeof(ucmd.cmd) bytes during the memset() inside copy_struct_from_user(). Seen under GCC 13:

In function 'copy_struct_from_user',
inlined from 'iommufd_fops_ioctl' at ../drivers/iommu/iommufd/main.c:333:8:
../include/linux/fortify-string.h:59:33: warning: '__builtin_memset' offset [57, 4294967294] is out of the bounds [0, 56] of object 'buf' with type 'union ucmd_buffer' [-Warray-bounds=]
59 | #define __underlying_memset __builtin_memset

This seems strange to me

I thought the way gcc handled this was if it knew the value must be in a certain range then it would check it

Quoted 6 lines
Quoted 3 lines

If it couldn't figure out any ranges it would not make a warning.

So why did it decide "rest" was in that really weird range?

It's because it got bounds-checked at the lower end (for the minimum size test).

Where? There is no sizeof(ucmd.ubuffer) in this code.

There are no statically computable constants at all.

The minimum size test loads from a struct:

	if (ucmd.user_size < op->min_size)
		return -EINVAL;

So, either gcc can't see through that and thus has no idea what the bound check is

Or, gcc has figured out that struct iommufd_ioctl_op::min_size has a finite set of values

If the latter, why doesn't it also know that iommufd_ioctl_op::size has finite set too?

Combined with the weird report that the upper end of that range is -2 (not UINT_MAX), something very strange is going on inside gcc.

Jason

author
#4 lore raw
Quoted 20 lines · Jason Gunthorpe
Quoted 19 lines · Kees Cook
Quoted 18 lines · Jason Gunthorpe
Quoted 11 lines · Kees Cook

While the op->size assignments are already bounds-checked at static initializer time, these limits aren't aggregated and tracked when doing later variable range checking under -Warray-bounds. Help the compiler see that we know what we're talking about, and we'll never ask to write more that sizeof(ucmd.cmd) bytes during the memset() inside copy_struct_from_user(). Seen under GCC 13:

In function 'copy_struct_from_user',
inlined from 'iommufd_fops_ioctl' at ../drivers/iommu/iommufd/main.c:333:8:
../include/linux/fortify-string.h:59:33: warning: '__builtin_memset' offset [57, 4294967294] is out of the bounds [0, 56] of object 'buf' with type 'union ucmd_buffer' [-Warray-bounds=]
59 | #define __underlying_memset __builtin_memset

This seems strange to me

I thought the way gcc handled this was if it knew the value must be in a certain range then it would check it

Quoted 8 lines
Quoted 6 lines
Quoted 3 lines

If it couldn't figure out any ranges it would not make a warning.

So why did it decide "rest" was in that really weird range?

It's because it got bounds-checked at the lower end (for the minimum size test).

Where? There is no sizeof(ucmd.ubuffer) in this code.

memset() is internally doing that via __builtin_object_size(dst + size, 1).

Quoted 1 line

There are no statically computable constants at all.

I think it's some logic that excludes a range based on ucmd.user_size internally to the additional checks in copy_struct_from_user().

Regardless, I think the correct fix should be with copy_struct_from_user(), so please disregard this patch. :)

Signature · 1 line
Kees Cook