On Sun, Dec 15, 2002 at 03:29:09PM -0600, Steve Langasek wrote: > On Sun, Dec 15, 2002 at 03:19:04PM -0500, Daniel Jacobowitz wrote: > > On Sun, Dec 15, 2002 at 02:02:38PM -0600, Steve Langasek wrote: > > > On Sun, Dec 15, 2002 at 11:50:23AM -0800, Jeff Bailey wrote: > > > > On Sun, Dec 15, 2002 at 12:49:56PM -0600, Steve Langasek wrote: > > > > > > I've made some progress into figuring out the source of the glibc 2.3, > > > > > apache, php4, php4-imap segfaults. I suspect that php4 and openssl > > > > > are > > > > > ultimately at fault (at least, a change with regards to these two > > > > > packages seems to correct the error), but I'm not certain, so I'm > > > > > sending a follow-up to the bug report to get feedback before > > > > > reassigning > > > > > back to PHP4. > > > > > Wow. If it does turn out to be a php4 problem, aj has told me that it's > > > > okay to fill our conflicts lines with other packages so whoever winds up > > > > with this bug, please make sure that [email protected] knows > > > > which version to conflict against. > > > > If no one on the glibc side can point to a bug in glibc 2.3 that would > > > have caused the libcrypto dependency to suddenly become a problem, then > > > I imagine this is what we'll do. I'm still curious to know what changed > > > between 2.2 and 2.3, so if anyone has any ideas... > > > I'm exceedingly suspicious here; it would be nice to figure out if this > > is a bug of the dynamic linker, since that seems likely from your > > description. > > I agree, it does sound like a bug in the dynamic linker. In order for > the main Apache binary to have referenced a symbol in libcrypto.so > (let's assume it's crypt()), the linker would have had to either search > libcrypto.so before searching libcrypt.so, despite the latter having > been loaded first; or not find anything within libcrypt.so that met its > criteria. > > The only possible way I can imagine it makes sense for this to happen is > if php4 is linked with -Bsymbolic (which it does not appear to be); PHP4 > calls crypt() while loaded the first time, causing the symbol name to be > resolved to the libcrypto.so version; the linker caches this because > Apache needs crypt() as well; and things go downhill. Does that sound > about right?
Except that I don't think -Bsymbolic is actually necessary for that. I'd need to brush up on my ELF specs to figure out what's really supposed to happen here... -- Daniel Jacobowitz MontaVista Software Debian GNU/Linux Developer
