Since I upgraded my kernel from 2.6.10-gentoo-r4, my system tends to crash under high load. A good way to reproduce it is compiling koffice in an emerge -uD world. This started to happen with some version after 2.6.10 but it wasn't reproducable till now and I didn't have the chance to get my hands on the kernel ring buffer (dmesg over xinetd rocks :) The kernel complains that it has a "double free detected in cache 'vm_area_struct'". For some reason it seems like most the times (but not always) it starts killing apache first. I'll attach the dmesg output and my kernel config.
Created attachment 76024 [details] dmesg output
Created attachment 76026 [details] kernel config
Google sent me to this [1] mail from Andrew Morton: | slab presently goes BUG if someone tries to register an already-registered | cache. | | But this can happen if the user accidentally loads a module which is already | statically linked into the kernel. Nuking the kernel is rather a harsh | reaction. | | Change it into a warning, and just fail the kmem_cache_alloc() attempt. If | he module is well-behaved, the modprobe will fail and all is well. I actually compiled the 8250 module into the kernel when I upgraded and the startup scripts still try to load the module. It seems like there should be already some patch against this problem in the current kernel and my box doesn't crash on boot, but I'll have a try without anyway... [1] http://lkml.org/lkml/2005/10/28/7
Hmmm... end of this monologue for today: The init scripts still try to load 8250 but I never compiled it as a module for 2.6.14, so this can't really be the problem.
Please test with the latest development kernel (currently 2.6.16-rc1)
I already tested gentoo-sources-2.6.15 which shows the same behaviour. I wanted to try the vanilla kernel, too, but didn't have the time to do so till now. I'll try to check the 2.6.16 one, too, but it might take a while.
Created attachment 77690 [details] dmesg from vanilla-sources-2.6.15 vanilla-sources-2.6.15 crash too. I'll try 2.6.16-rc1 next.
vanilla-sources-2.6.16-rc1 behave fine, at least when I compile KOffice which crashed all the prevoius kernels. 2.6.16-rc1 has some annoying IPv6 bug though, see http://lists.thekelleys.org.uk/pipermail/dnsmasq-discuss/2006q1/000572.html
Created attachment 77737 [details] it still crashes with 2.6.16-rc1 I was wrong, 2.6.16-rc1 crashes, too. This time when I was compiling php.
I'm stumped. Please file this as an upstream bug against 2.6.16-rc1 at http://bugzilla.kernel.org and post the new URL here.
Created attachment 78093 [details, diff] differences between crashing and current config Hmmm... when I built that 2.6.14 kernel I actually changed some options. Now I built a new one with almost the same options as 2.6.10 and it doesn't seem to crash. I currently suspect either the CONFIG_HZ_1000 option (wasn't 1kHz the default in the old 2.6 without the config option?) or the CONFIG_SND_USB_AUDIO which I didn't add this time. Especially because the dmesg ouput shows a lot of USB hickups after the crashes and I actually tend to switch off my screen, which has the USB audio included, on long compiles. I'll try around and when I found the culprit and I am sure that its not caused by the Gentoo patchset, I'll file it upstream.
*grmf* it still crashes with that config, it just took a bit longer. I filed it upstream, see <http://bugzilla.kernel.org/show_bug.cgi?id=5964>.
