Date: Thu, 18 Sep 2008 11:38:02 +0200 From: Marc =?iso-8859-1?q?L=F6rner?= <marc.loerner@hob.de> To: Marcel Moolenaar <xcllnt@mac.com> Cc: freebsd-ia64@freebsd.org Subject: Re: Montecito running and little SMP-bug Message-ID: <200809181138.02217.marc.loerner@hob.de> In-Reply-To: <9A984C7A-66FE-4E5C-9B36-D5B0E458AADD@mac.com> References: <200809111023.31516.marc.loerner@hob.de> <9A984C7A-66FE-4E5C-9B36-D5B0E458AADD@mac.com>
index | next in thread | previous in thread | raw e-mail
On Wednesday 17 September 2008 18:28, Marcel Moolenaar wrote: > On Sep 11, 2008, at 1:23 AM, Marc Lörner wrote: > > Hello! > > I've got good news: > > Finally GENERIC-kernel without SMP-support is running now on my > > montecito > > box :-) > > Is the 7-STABLE? > It's 7.0-RELEASE-p3 > I've been running 8-CURRENT on a Montecito for a while now. > > > When turning on SMP, I always get an Page-Not-Present in function vcnt > > in vm/vm_meter.c when it iterates through found CPUs. > > There are optimization bugs, which pretty much forces us > to run with INVARIANTS for now. I haven't had the time to > dig into it. > > > Following is the output of the APIC-table: > > > > Table 'APIC' at 0xe0000000ff1b6620 > > Local APIC address=0xfee00000 > > Local SAPIC entry > > ProcessorId=0x0, Id=0x0, Eid=0x0 > > Local SAPIC entry > > ProcessorId=0x1, Id=0x1, Eid=0x0 (disabled) > > Local SAPIC entry > > ProcessorId=0x2, Id=0x2, Eid=0x0 > > Local SAPIC entry > > ProcessorId=0x3, Id=0x3, Eid=0x0 (disabled) > > > > > > I think the case of running/enabled CPUs being not contiguous is not > > thought > > of when initializing pcpus. > > I don't think that's the problem. My box has disables CPUs > as well and there are no panics there. In any case, if we > want to support hot-pluggable CPUs, we cannot assume that > if there are N CPUs, index [0..N> are in use... > > FYI, I did following little patch in sys/ia64/acpica/madt.c in function ia64_probe_sapics: - added local variable (initialized to zero) for counting cpus - in local-sapic case: use local variable as first parameter in cpu_mp_add and increment it After that I sometimes went into an Page-Not-Present fault (see discussion of asymmetric mtx_[un]lock_sleep on hackers mailinglist). After some more debugging I recognized that curthread and access to pcpu in machine-dependent code (sys/ia64/include/pcpu.h) is not preemption-ready. So I made some little adaption from sys/amd64/include/pcpu.h. => Since today, I'm testing stability of SMP-kernel (7.0-RELEASE-p3) with these two bugfixes and without invariant and witness support. Regards, Marc Loernerhelp
Want to link to this message? Use this
URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?200809181138.02217.marc.loerner>