linux-headers-5.2 and proper use of SIOCGSTAMP

Commit https://github.com/torvalds/linux/commit/0768e17073dc527ccd18ed5f96ce85f9985e9115
("net: socket: implement 64-bit timestamps") caused a bit of userspace breakage
for existing programs:
- firefox: https://bugs.gentoo.org/689808
- qemu: https://lists.sr.ht/~philmd/qemu/%3C20190604071915.288045-1-borntraeger%40de.ibm.com%3E
- linux-atm: https://gitweb.gentoo.org/repo/gentoo.git/tree/net-dialup/linux-atm/files/linux-atm-2.5.2-linux-5.2-SIOCGSTAMP.patch?id=408621819a85bf67a73efd33a06ea371c20ea5a2

I have a question: how a well-behaved app should include 'SIOCGSTAMP' definition to keep being buildable against old and new linux-headers?

'man 7 socket' explicitly mentions SIOCGSTAMP and mentions only

    #include <sys/socket.h>

as needed header.

Should #include <linux/sockios.h> always be included by user app? Or should glibc tweak it's definition of '#include <sys/socket.h>' to make it available on both old and new version of linux headers?

CCing both kernel and glibc folk as I don't understand on which side issue should be fixed.

Thanks!

  Sergei

* Sergei Trofimovich:

What is the reason for dropping SIOCGSTAMP from <asm/socket.h>?

If we know that, it will be much easier to decide what to do about <sys/socket.h>.

As far as I can tell, nobody thought it would be a problem to move it from asm/sockios.h to linux/sockios.h, as the general rule is that one should use the linux/*.h version if both exist, and that the asm/*.h version only contains architecture specific definitions. The new definition is the same across all architectures, so it made sense to have it in the common file.

If the assumption was wrong, the obvious solution is to duplicate the definitions everywhere or move the common parts into asm-generic/sockios.h, but it would have been better to hear about that earlier.

Arnd

* Arnd Bergmann:

Most of the socket-related constants are not exposed in UAPI headers, although userspace is expected to use them. It seems to me that due to the lack of other options among the UAPI headers, <asm/socket.h> has been a dumping ground for various socket-related things in the past, whether actually architecture-specific or not.

<linux/socket.h> does not include <asm/socket.h>, so that's why we usually end up with including <asm/socket.h> (perhaps indirectly via <sys/socket.h>), which used to include <asm/sockios.h> on most (all?) architectures. That in turn provided some of the SIOC* constants in the past, so people didn't investigate other options.

I think we can change glibc to include <linux/sockios.h> in addition to <asm/socket.h>. <linux/sockios.h> looks reasonably clean to me, much better than <asm/socket.h>. I'm still working on the other breakage, and I'm severely limited by the machine resources I have access to.

It seems that both the missing constants and the fact that linux/socket.h doesn't include asm/socket.h and linux/sockios.h goes back to a 21 year old commit:

commit 74f513101058f7585176ea8cdf6fb026faea8a7e

Author: linus1 <[email protected]>
Date:   Wed May 20 11:00:00 1998 -0800

[tytso] include/asm-i386/posix_types.h
This quick fix eliminates a lot of warning messages when compiling e2fsprogs under glibc. This is because the glibc header files defines its own version of FD_SET, FD_ZERO, etc., and so if you need to #include the kernel include files, you get a lot of duplicate defined macro warning messages. This patch simply #ifdef's out the kernel versions of these function if the kernel is not being compiled and the glibc header files are in use.

index 08f0d281401c..35a7629b6b70 100644
@@ -1,6 +1,8 @@
 #ifndef _LINUX_SOCKET_H
 #define _LINUX_SOCKET_H
 
+#if defined(__KERNEL__) || !defined(__GLIBC__) || (__GLIBC__ < 2)
+
 #include <asm/socket.h>                        /* arch-dependent

defines */

 #include <linux/sockios.h>             /* the SIOCxxx I/O controls     */
 #include <linux/uio.h>                 /* iovec support                */

@@ -256,4 +258,5 @@ extern int move_addr_to_user(void *kaddr, int klen, void *uaddr, int *ulen);
extern int move_addr_to_kernel(void *uaddr, int ulen, void *kaddr); extern int put_cmsg(struct msghdr*, int level, int type, int len, void *data); #endif
+#endif /* not kernel and not glibc */
#endif /* _LINUX_SOCKET_H */

(the same commit did similar changes in linux/stat.h and asm/posix_types.h)

Over time, the check for glibc was removed (to allow including linux/socket.h before sys/socket.h), and all the #ifdef __KERNEL__ bits were removed from the installed header as part of the uapi header split.

That seems reasonable to me, but overall my fear is that these headers are already so broken that any change will risk breaking something in more or less unexpected ways.

Arnd