Date: Thu, 12 Feb 2009 22:11:01 +0200 From: Kostik Belousov <kostikbel@gmail.com> To: Doug Ambrisko <ambrisko@ambrisko.com> Cc: arch@freebsd.org Subject: Re: rtld enhancement to add osversion sub-directory search Message-ID: <20090212201101.GI2723@deviant.kiev.zoral.com.ua> In-Reply-To: <200902121859.n1CIxaJd083445@ambrisko.com> References: <200902121859.n1CIxaJd083445@ambrisko.com>
next in thread | previous in thread | raw e-mail | index | archive | help
--LiQwW4YX+w4axhAx Content-Type: text/plain; charset=us-ascii Content-Disposition: inline Content-Transfer-Encoding: quoted-printable On Thu, Feb 12, 2009 at 10:59:36AM -0800, Doug Ambrisko wrote: > Hi folks, >=20 > I'd like to discuss the idea of adding an automatic directory search > feature to rtld. At work we need to run 3rd party code that we get > as objects or even binaries that are compiled by other groups. This > works okay when our base OS is the same as the 3rd party but not so > good when we upgrade our base OS and we don't want to force th 3rd > party code to upgrade at the same time. Sometimes they can't since > legacy systems that needs to run the 3rd party code run the older > OS. Now WRT to FreeBSD's base libraries this isn't much of a problem > since the OS's lib's general bump version between releases so just > via the name we have unique lib's and the loader can get the correct > one. However, once you start to link to stuff in ports (ie.=20 > /usr/local/lib) then these names are no longer unique and something > built for FreeBSD 4 through 7 (as part of FreeBSD's release, ie=20 > pre-built packages) end up with the same name. This is a problem since > sticking the old version in /usr/lib/compat is a problem since it > looks based on name (ie. /usr/local/lib/libiconv.so.3). Things can > get interesting when the /usr/local/lib pulls in libc. So now your > FreeBSD 4 binary could pull in a libiconv.so.3 for FreeBSD 7 and > libc from FreeBSD 7. Interesting things start to happen! >=20 > What I've done for work is to teach the rtld to look for the > .note.ABI-tag and extract the osversion. I then put that and the=20 > osversion major number (ie. 6 instead of 603000) into a "try" list. > Then whenever an object/lib is attemped to be accessed I look in the=20 > osversion sub-directory then the osversion major and finally the > standard location. I even extended this to LD_PRELOAD and=20 > insert this try into the path just before the object/lib if fully > qualified. Actually, I do it for all fully qualified things. > Now I can stick most things into /usr/lib/compat/6 and it just > works. For fully qualified LD_PRELOAD I "mv" things on boot > into the sub-directory for that OS. Also I had to change LD_PRELOAD > to accept failure if the thing wasn't found since children in herit > that and if it a different version that LD_PRELOAD might not be > applicable. >=20 > I don't really see a maintain for rtld since jdp when I worked with > him to add LD_PRELOAD. I'd like to get feed back on this idea and > get it into FreeBSD. It would help solve some of these issues for > other people as well. It definitely, made my life easier at work. > It would also make things easier on my FreeBSD machines, when I=20 > upgrade the OS and ports but have legacy compiled things that I > don't want to recompile again. I could move all of my /usr/local/libs > into /usr/local/lib/compat/<osversion major), upgrade packages and=20 > still run my old stuff. I've had issues before when OO was cross-linked > with libc from 6 and 7 due to this. Rebuilding OO was not fun and > I shouldn't have needed to. >=20 > There is one glitch in atleast FreeBSD 6.1 didn't put the .note.ABI-tag > into binaries so for now if not specified I assume FreeBSD 6. That > is good enough for work but I could add an env. variable to hold > the default. If not set then it wouldn't do the sub-directory thing > on binaries that are not known. >=20 > Loading of objects will slow done a bit since it adds extra=20 > searches. It might be good enough to just do the os version major=20 > and not for the complete osversion. >=20 > BTW, I also added a feature to look at LD32_ first then LD_ > when running 32bit on 64bit. Legacy SW doesn't know about > LD32_ and just sets LD_. This way that stuff doesn't need > to be taught about LD32_ when it shouldn't really need to. There is a popular feature, unfortunately, not supported by FreeBSD ld.so, called Dynamic String Tokens, see http://docs.sun.com/app/docs/doc/817-1984/appendixc-4?l=3Den&a=3Dview I have almost abandoned patch that adds support for $ORIGIN, $OSREL, $OSNAME, and $PLATFORM. Quite amazingly, it merged with today CURRENT without serious conflicts. http://people.freebsd.org/~kib/misc/rtld_locks.4.patch --LiQwW4YX+w4axhAx Content-Type: application/pgp-signature Content-Disposition: inline -----BEGIN PGP SIGNATURE----- Version: GnuPG v1.4.9 (FreeBSD) iEYEARECAAYFAkmUglQACgkQC3+MBN1Mb4jXKwCg7pWUsPFfaSLafOs4TLiyEMfN WTwAoNnOoF1fOh0jmEMLbEajxM+fwIZ1 =h7ot -----END PGP SIGNATURE----- --LiQwW4YX+w4axhAx--
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20090212201101.GI2723>