On Thu, 2012-12-27 at 01:57 +0100, Tobias Gasser wrote: > Am 25.12.2012 10:28, schrieb Simon Geard: > > > Hmm... it occurs to me that while using FS monitoring (or your 'find' > > based approach) is neat, it's not parallel-safe. I'm guessing you don't > > install more than one package simultaneously? My current build scripts > > basically consist of a generated Makefile to deal with dependencies, and > > it copes quite happily with being run with half a dozen processes... > > i first used a 'find' aproach. > fine for lfs. but after having installed xfce it gets really slow. > > so i switched to an inotify approach. but not for long. very buggy as > files can't be monitored as long as the path is not registred. so i > missed lots of files which were installed but not logged...
For me, I think the biggest problem would be the inability to build and install multiple packages at once. With my old build scripts, I used to build individual packages with -j6, but found that just meant I was spending all my time waiting for configure scripts to run single-threaded. Especially for something like Xorg with it's many small packages, being able to build a dozen packages in parallel helps a lot... > now i use destdir for everything. > i create 2 tempfs for each package. one for the extracted sources (and > build if required), one for the destdir. as i have 16g of ram in my > build-system, this even works for libreoffice. no need to remove the > sources, just umount the tempfs. miliseconds to remove the libreoffice > source. Does the tmpfs really help that much? Last time I tried it, I didn't see much of a speedup from building on tmpfs vs on the regular disk - a few percent at most. I suspect that OS-level disk caching is keeping most of the activity in-memory anyway... > > and the destdir (tempfs mounted on pathes build with mktemp) is the only > one of these 3 which works with parallel builds. > > most packages support destdir or something equivalent. only a few > packages require some nasty patching... For that reason, I've dropped most of the bits in LFS that move stuff between / and /usr after running "make install". I'm never going to have them on separate partitions, so it's just added complexity. Actually, my current system just has /bin, /sbin and /lib symlinked into /usr now anyway... > to automate things like user-creation i use a _pre.sh, to fix up config > files i use a _post.sh. the destdir and those 2 scripts i put into an > tar.xz and have some kind of simple package management which allows to > update mulitple systems with one build. removing the 'old' version first > is peanuts: just get a list of installed files from the old tar.xz. (ok, > i have to manually save the configuration stuff first.. it's not really > package management, but makes life much easier) A little more complicated than what I do, but then, I don't care about copying stuff to other systems, only about being able to remove packages I no longer need. So stuff like user creation can be done as part of the build... > > the kernel i don't build with destdir. as i need the sources for some > packages, there is no saving in building in ram and copying to disk > after install - and the kernel differs on each machine so i can't > distribute updates. Same here... that one's a special case, keeping multiple versions installed, and not removing the source tree after build, etc. Plus, uninstalling a kernel is easier than most most packages, since everything is in either /usr/src, /lib/modules, or /boot, and is clearly tagged with the version... Simon.
signature.asc
Description: This is a digitally signed message part
-- http://linuxfromscratch.org/mailman/listinfo/blfs-support FAQ: http://www.linuxfromscratch.org/blfs/faq.html Unsubscribe: See the above information page
