On Tuesday 09 February 2010 06:50:06 pm David Miller wrote: > [ Josip, when I hit this I thought that if it were a gcc bug it might > explain your gcc-4.4 kernel failures. That's why I haven't played > with your machines yet. Now that I know this is actually a kernel > bug effecting only user applications, I will go pursue your > issues. ] > > I originally thought this was a gcc bug: > > http://gcc.gnu.org/bugzilla/show_bug.cgi?id=43004 > > But it's actually a kernel issue. > > When we clone or create a signal stack we only 8-byte align the stack > frame, it should be 16-byte aligned on 64-bit. Use 16-byte alignment > for 32-bit too since that's harmless and it's cheaper that way. > > Distribution folks you really want this in your kernels, otherwise > 64-bit threads and signal handlers will have broken alloca() and > corrupt their stacks. I thought that maybe this would help with a localdef generation issue that is being seen while building glibc, at least for debian and fedora. however it seems not. I have built the latest Fedora 12 stable kernel with your patch applied and glibc failed to build still. patches had been submitted to glibc http://sources.redhat.com/bugzilla/show_bug.cgi?id=10855 http://sources.redhat.com/bugzilla/show_bug.cgi?id=10283 however one had been outright rejected and the sparc specific one has not been looked at. Is this something that should be fixed in the kernel? Dennis
Attachment:
signature.asc
Description: This is a digitally signed message part.
