Re: [V3 PATCH 1/2] x86/panic: Replace smp_send_stop() with kdump friendly version

From: 'Dave Young'
Date: Tue Jul 12 2016 - 22:04:36 EST


On 07/12/16 at 02:49am, æåèå / KAWAIïHIDEHIRO wrote:
> Hi Dave,
>
> Thanks for the comments.
>
> > From: Dave Young [mailto:dyoung@xxxxxxxxxx]
> > Sent: Monday, July 11, 2016 5:35 PM
> >
> > On 07/05/16 at 08:33pm, Hidehiro Kawai wrote:
> > > This patch fixes one of the problems reported by Daniel Walker
> > > (https://lkml.org/lkml/2015/6/24/44).
> > >
> > > If crash_kexec_post_notifiers boot option is specified, other CPUs
> > > are stopped by smp_send_stop() instead of machine_crash_shutdown()
> > > in crash_kexec() path. This behavior change leads two problems.
> > >
> > > Problem 1:
> > > octeon_generic_shutdown() for MIPS OCTEON assumes that other CPUs are
> > > still online and try to stop their watchdog timer. If
> > > smp_send_stop() is called before octeon_generic_shutdown(), stopping
> > > watchdog timer will fail because other CPUs have been offlined by
> > > smp_send_stop().
> > >
> > > panic()
> > > if crash_kexec_post_notifiers == 1
> > > smp_send_stop()
> > > atomic_notifier_call_chain()
> > > kmsg_dump()
> > > crash_kexec()
> > > machine_crash_shutdown()
> > > octeon_generic_shutdown() // shutdown watchdog for ONLINE CPUs
> > >
> > > Problem 2:
> > > Most of architectures stop other CPUs in machine_crash_shutdown()
> > > path, and they also do something needed for kdump. For example,
> > > they save registers, disable virtualization extensions, and so on.
> > > However, if smp_send_stop() stops other CPUs before
> > > machine_crash_shutdown(), we miss those operations.
> > >
> > > How do we fix these problems? In the first place, we should stop
> > > other CPUs as soon as possible when panic() was called, otherwise
> > > other CPUs may wipe out a clue to the cause of the failure. So, we
> > > replace smp_send_stop() with more suitable one for kdump.
> >
> > We have been avoiding extra things in panic path, but unfortunately
> > crash_kexec_post_notifiers were added. I tend to agree the best place
> > for this stuff is in 2nd kernel or purgatory instead of in 1st kernel.
>
> Several months ago, I posted a patch set which writes regs to SEL, generate
> an event to send SNMP message, and start/stop BMC's watchdog timer in
> purgatory. This feature requires BMC with KCS (Keyboard Controller Style)
> I/F, but the most of enterprise grade server would have it.
> (http://thread.gmane.org/gmane.linux.kernel.kexec/15382)
>
> Doing kmsg_dump things in purgatory wouldn't be suitable (should be done
> in the 2nd kernel before enabling devices and IRQs?)

In theory it is doable maybe do something like oldmem_kmsg_dump while /proc/vmcore
initializing?

>
> > As for this patch I'm not sure it is safe to replace the smp_send_stop
> > with the kdump friendly function. I'm also not sure if the kdump friendly
> > function is safe for kdump. Will glad to hear opinions from other
> > arch experts.
>
> This stuff depends on architectures, so I speak only about
> x86 (the logic doesn't change on other architectures at this time).
>
> kdump path with crash_kexec_post_notifiers disabled:
> panic()
> __crash_kexec()
> crash_setup_regs()
> crash_save_vmcoreinfo()
> machine_crash_shutdown()
> native_machine_crash_shutdown()
> panic_smp_send_stop() /* mostly same as original
> * kdump_nmi_shootdown_cpus()
> */
>
> kdump path with crash_kexec_post_notifiers enabled:
> panic()
> panic_smp_send_stop()
> __crash_kexec()
> crash_setup_regs()
> crash_save_vmcoreinfo()
> machine_crash_shutdown()
> native_machine_crash_shutdown()
> panic_smp_send_stop() // do nothing
>
> The difference is that stopping other CPUs before crash_setup_regs()
> and crash_save_vmcoreinfo() or not. Since crash_setup_regs() and
> crash_save_vmcoreinfo() just save information to some memory area,
> they wouldn't be affected by panic_smp_send_stop(). This means
> placing panic_smp_send_stop before __crash_kexec is safe.
>
> BTW, I noticed my patch breaks Xen kernel. I'll fix it in the
> next version.

But it does breaks stuff which depends on cpu not being disabled
like problem 1 you mentioned in patch log.

>
> > BTW, if one want to use crash_kexec_post_notifiers he should take the
> > risk of unreliable kdump. How about only call smp_send_stop in case no
> > crash_kexec_post_notifiers being used.
>
> Unlike panic_smp_send_stop()/kdump_nmi_shootdown_cpus(), smp_send_stop()
> for x86 tries to stop other CPUs with normal IPI before issuing NMI IPI.
> This would be because NMI IPI has a risk of deadlock. We checked if
> the kdump path has a risk of deadlock in the case of NMI panic and fixed
> it. But I'm not sure about normal panic path. I agree with that use
> smp_send_stop if crash_kexec_post_notifiers or kdump is disabled.

What I mean is like below, problem 1 will not exist in this way, but
kdump will be unreliable:
if (!crash_kexec_post_notifiers)
smp_send_stop()

>
> > > This patch solves Problem 2 by replacing smp_send_stop() in panic()
> > > with panic_smp_send_stop(). This is a weak function which calls
> > > smp_send_stop(), and architecture dependent code may override this
> > > with appropriate one. This patch only provides x86-specific version.
> >
> > It does not fix the Problem 1, it seem not possible to fix it?
>
> Problem 1 depends on architectures, and at least it doesn't happen
> on x86. I can try to fix the Problem 1 for MIPS, but I can't test it.
> Possible solution will be to use an smp_send_stop variant which stop
> the CPU without offlining.
>
> Best regards,
>
> Hidehiro Kawai
> Hitachi, Ltd. Research & Development Group
>
[snip]
Thanks
Dave