[libcamera-devel] [PATCH v2] libcamera: media_device: prevent sign extension on casts

[Laurent Pinchart]

Hi Kieran,

On Tue, Feb 18, 2020 at 05:07:41PM +0000, Kieran Bingham wrote:
> On 18/02/2020 15:17, Laurent Pinchart wrote:
> > On Tue, Feb 18, 2020 at 02:18:39PM +0000, Kieran Bingham wrote:
> >> The conversion of pointers to integers is implementation defined and
> >> differs between g++ and clang++ when utilising a uint64_t type.
> >>
> >>     #include <iostream>
> >>
> >>     int main(int argc, char **argv)
> >>     {
> >>         void *ptr = reinterpret_cast<void *>(0xf1234567);
> >>         uint64_t u64 = reinterpret_cast<uint64_t>(ptr);
> >> 	uint64_t uint64 = reinterpret_cast<uintptr_t>(ptr);
> >>
> >>         std::cout << "ptr " << ptr
> >> 		  << " u64 0x" << std::hex << u64
> >> 		  << " uint64 0x" << std::hex << uint64
> > 
> > Maybe " ptr -> u64 0x" and " ptr -> uintptr_t -> u64 0x" ?
> > 
> >> 		  << std::endl;
> >>
> >> 	return 0;
> >>     }
> > 
> > That's a weird mix of tabs and spaces :-)
> 
> Agh, I indented the left hand items by 4 to indent the whole code block
> for the commit message.
> 
> Indenting the whole lot by a tab fits, but it looks a bit weird and too
> far to the right ... but I don't think it's important.

No big deal, no. I was however referring to lines starting with
"uint64_t uint64" and "return" that seem to have a different combination
of spaces and tabs than the other ones.

> >> When compiled with g++ produces the following unexpected output:
> > 
> > s/with g++/with g++ for a 32-bit platform/
> 
> No, I don't think that's true.
> This issue affects a 64-bit platform just the same.
> 
> If a pointer is passed through here which happens to be
> 0x00000000f1234567, it will be incorrect when it gets to the kernel.
> 
> (or any pointer with leading zeros up to a set bit 32...)

Really ? Testing the above code on x86-64 without -m32 prints

ptr 0xf1234567 u64 0xf1234567 uint64 0xf1234567

for me.

> >>   ptr 0xf1234567 u64 0xfffffffff1234567 uint64 0xf1234567
> >>
> >> The standards states:
> >>
> >> "A pointer can be explicitly converted to any integral type large enough
> >> to hold all values of its type. The mapping function is
> >> implementation-defined. [Note: It is intended to be unsurprising to
> >> those who know the addressing structure of the underlying machine. — end
> >> note]"
> >>
> >> And as such the g++ implementation appears to be little more surprising
> >> than expected in this situation.
> >>
> >> The MediaDevice passes pointers to the kernel via the struct
> >> media_v2_topology in which pointers are cast using a uint64 type (__u64),
> >> which is affected by the sign extension described above when BIT(32) is
> >> set and causes an invalid address to be given to the kernel.
> >>
> >> Ensure that we cast using uintptr_t which is not affected by the sign
> >> extension issue.
> >>
> >> Signed-off-by: Kieran Bingham <kieran.bingham at ideasonboard.com>
> > 
> > Reviewed-by: Laurent Pinchart <laurent.pinchart at ideasonboard.com>
> > 
> >> ---
> >> v2:
> >>  - Expand commit message to explain the underlying issue.
> >>  - include stdint.h
> >>  - use uintptr_t instead of std::uintptr_t
> >>
> >>  src/libcamera/media_device.cpp | 9 +++++----
> >>  1 file changed, 5 insertions(+), 4 deletions(-)
> >>
> >> diff --git a/src/libcamera/media_device.cpp b/src/libcamera/media_device.cpp
> >> index fad475b9ac76..0d6b5efd9e7a 100644
> >> --- a/src/libcamera/media_device.cpp
> >> +++ b/src/libcamera/media_device.cpp
> >> @@ -9,6 +9,7 @@
> >>  
> >>  #include <errno.h>
> >>  #include <fcntl.h>
> >> +#include <stdint.h>
> >>  #include <string>
> >>  #include <string.h>
> >>  #include <sys/ioctl.h>
> >> @@ -236,10 +237,10 @@ int MediaDevice::populate()
> >>  	 */
> >>  	while (true) {
> >>  		topology.topology_version = 0;
> >> -		topology.ptr_entities = reinterpret_cast<__u64>(ents);
> >> -		topology.ptr_interfaces = reinterpret_cast<__u64>(interfaces);
> >> -		topology.ptr_links = reinterpret_cast<__u64>(links);
> >> -		topology.ptr_pads = reinterpret_cast<__u64>(pads);
> >> +		topology.ptr_entities = reinterpret_cast<uintptr_t>(ents);
> >> +		topology.ptr_interfaces = reinterpret_cast<uintptr_t>(interfaces);
> >> +		topology.ptr_links = reinterpret_cast<uintptr_t>(links);
> >> +		topology.ptr_pads = reinterpret_cast<uintptr_t>(pads);
> >>  
> >>  		ret = ioctl(fd_, MEDIA_IOC_G_TOPOLOGY, &topology);
> >>  		if (ret < 0) {

-- 
Regards,

Laurent Pinchart