Date: Fri, 17 Nov 2006 09:55:55 +0300 From: Yar Tikhiy <yar@comp.chem.msu.su> To: "M. Warner Losh" <imp@bsdimp.com> Cc: src-committers@FreeBSD.ORG, bde@zeta.org.au, jkoshy@FreeBSD.ORG, cvs-all@FreeBSD.ORG, phk@phk.freebsd.dk, cvs-src@FreeBSD.ORG Subject: Re: cvs commit: src/include ar.h Message-ID: <20061117065555.GE49602@comp.chem.msu.su> In-Reply-To: <20061116.165207.1661914048.imp@bsdimp.com> References: <20061113214928.P76443@delplex.bde.org> <20061113.101958.-861030824.imp@bsdimp.com> <20061116090412.GB37133@comp.chem.msu.su> <20061116.165207.1661914048.imp@bsdimp.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Thu, Nov 16, 2006 at 04:52:07PM -0700, M. Warner Losh wrote: > In message: <20061116090412.GB37133@comp.chem.msu.su> > Yar Tikhiy <yar@comp.chem.msu.su> writes: > : On Mon, Nov 13, 2006 at 10:19:58AM -0700, M. Warner Losh wrote: > : > : > : > : BTW, you are responsible for the __packed in <netinet/ip.h>. Please remove > : > : it. The __CTASSERT() is enough to detect if heroic packing is ever needed. > : > : The only danger is if something has grown to depend on __packed reducing > : > : alignment as a side effect. E.g., suppose we had a byte string containing > : > : a bytewise copy of a struct ip. If the copy might be misaligned, then it > : > : should be copied to an actual struct ip before accessing it as a struct, > : > : but code that accesses it directly using ((struct ip *)&bs[N]) would work > : > : now due to the reduced alignment. Places that really need __packed should > : > : probably use __aligned() to restore the natural alignment. > : > > : > DO NOT REMOVE IT. IT IS ABSOLUTELY REQUIRED FOR ARM TO WORK RIGHT. > : > If you want to remove it, then you must make sure arm works right > : > after it because I'll add it back. > : > : Many years ago I was taught that comments in code could help to > : avoid such clashes in software development. Is this true no more? ;-) > > You don't add comments like: > > i++; // Add one to i. > > This is a similar class. It is for any compiler that has differing > alignment requirements than i386. This is an oversimplification of the case. If it were so simple, no doubts about it would be raised. That's why I suggested adding a comment explaining that historically struct ip was lucky to be packed/aligned properly, but that wasn't backed by the C standard in fact, and eventually architectures appeared, e.g., ARM, which broke the false assumption. It's a rather edifying case. Then you'll have a smaller chance of having to yell in capital letters again, "DO NOT REMOVE IT. IT IS ABSOLUTELY REQUIRED FOR ARM TO WORK RIGHT." -- hopefully, not only regarding struct ip. -- Yar
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20061117065555.GE49602>