On 01/18/2013 05:44 AM, Pedro Alves wrote: > On 01/18/2013 04:22 AM, Carlos O'Donell wrote: >> On Thu, Jan 17, 2013 at 11:20 PM, Mike Frysinger <vapier@xxxxxxxxxx> wrote: >>> On Wednesday 16 January 2013 22:15:38 David Miller wrote: >>>> From: Carlos O'Donell <carlos@xxxxxxxxxxxxxxxx> >>>> Date: Wed, 16 Jan 2013 21:15:03 -0500 >>>> >>>>> +/* If a glibc-based userspace has already included in.h, then we will >>>>> not + * define in6_addr (nor the defines), sockaddr_in6, or ipv6_mreq. >>>>> The + * ABI used by the kernel and by glibc match exactly. Neither the >>>>> kernel + * nor glibc should break this ABI without coordination. >>>>> + */ >>>>> +#ifndef _NETINET_IN_H >>>>> + >>>> >>>> I think we should shoot for a non-glibc-centric solution. >>>> >>>> I can't imagine that other libc's won't have the same exact problem >>>> with their netinet/in.h conflicting with the kernel's, redefining >>>> structures like in6_addr, that we'd want to provide a protection >>>> scheme for here as well. >>> >>> yes, the kernel's use of __GLIBC__ in exported headers has already caused >>> problems in the past. fortunately, it's been reduced down to just one case >>> now (stat.h). let's not balloon it back up. >>> -mike >> >> I also see coda.h has grown a __GLIBC__ usage. >> >> In the next revision of the patch I created a single libc-compat.h header >> which encompasses the logic for any libc that wants to coordinate with >> the kernel headers. > > >> It's simple enough to move all of the __GLIBC__ uses into libc-compat.h, >> then you control userspace libc coordination from one file. > > How about just deciding on a single macro/symbol both the > kernel and libc (any libc that needs this) define? Something > like both the kernel and userland doing: > > #ifndef __IPV6_BITS_DEFINED > #define __IPV6_BITS_DEFINED > ... > define in6_addr, sockaddr_in6, ipv6_mreq, whatnot > #endif > > So whichever the application includes first, wins. > Too naive? I didn't see this option being discarded, so > not sure it was considered. Too naive, but *close* to what my patch does :-) The kernel definitions when included first, and in a glibc userspace, must try to mimic the glibc userspace headers and we need more than one guard macro to do that effectively. The reason I jumped into the code is because this kind of problem is easy to talk about, but the devil is in the details. There are certainly some compromises on both sides, but the solution promises to solve this problem. Honestly without UAPI this would have been an impossible task. Cheers, Carlos. -- libvir-list mailing list libvir-list@xxxxxxxxxx https://www.redhat.com/mailman/listinfo/libvir-list