Re: kexec crash kernel running with watchdog enabled

[dwalker@fifo99.com]

On Fri, Jun 26, 2015 at 10:02:00AM +0900, Hidehiro Kawai wrote:
> Hi,
> 
> (2015/06/25 1:31), dwalker@fifo99.com wrote:
> > On Wed, Jun 24, 2015 at 03:52:48PM +0900, Masami Hiramatsu wrote:
> >> Hi,
> >>
> >> On 2015/06/23 23:05, dwalker@fifo99.com wrote:
> >>>
> >>> Hi,
> >>>
> >>> There was a commit in kernel/panic.c which altered when the kexec crash kernel is executed,
> >>>
> >>> commit f06e5153f4ae2e2f3b0300f0e260e40cb7fefd45
> >>> Author: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com>
> >>> Date:   Fri Jun 6 14:37:07 2014 -0700
> >>>
> >>>     kernel/panic.c: add "crash_kexec_post_notifiers" option for kdump after panic_notifers
> >>>
> >>>
> >>> This made it possible for smp_send_stop() to be executed prior to calling the kexec crash
> >>> kernel.
> >>>
> >>> The issue is that smp_send_stop() offlines the cores, and other code depend on the cores being online.
> >>>
> >>> In my case on Octeon here's an example,
> >>>
> >>> panic()
> >>>  crash_kexec()
> >>>   machine_crash_shutdown()
> >>>    octeon_generic_shutdown()
> >>>
> >>> Inside octeon_generic_shutdown() the Octeon watchdog is shutdown for_each_online_cpu(), but since
> >>> most of the cpu's already got offlined in smp_send_stop() it means the watchdog is still alive on
> >>> those cores. This results in a reboot during the crash kernel execution.
> >>
> >> Ah, I see.
> >>
> >>> Another example seem to be in default_machine_crash_shutdown() where crash_kexec_prepare_cpus() depends
> >>> on an IPI for saving the registers on different cores. However, the cpu's are all offlined with
> >>> interrupts disabled so they won't be running those IPI's in this case.
> >>>
> >>> I'm looking for any advice on how this should be fixed, or if it's already fixed. I'm not going to be
> >>> submitting a patch so if anyone wants to submit one feel free to do so.
> >>
> >> Hmm, IMHO, when the cpu goes to offline in appropriate way(smp_send_stop), it should stop
> >> watchdog timer on the offlined cpu too.
> >> Or, you can also register crash handler which stops all watchdogs, but it's a bit tricky.
> >>
> > 
> > That doesn't really fix all the issue tho. As I was explaining generic MIPS code depends on the cpu's
> > effectively being online for crash data collection (with an IPI). This issue may effect other architectures also,
> > because smp_send_stop() offlines the cpu on other architectures also. I haven't surveyed the other architectures
> > enough to know what issue could happen from this tho.
> > 
> > Is it possible to move the smp_send_stop() below the notifiers ? I'm just throwing out ideas.
> 
> No, that doesn't works.  Some notifiers assume that they run in
> single core mode.
> 
> Another possible solution is to add notifiers just after
> machine_crash_shutdown() like this:
> 
> void panic(const char *fmt, ...)
> ...
> -	if (!crash_kexec_post_notifiers)
> -		crash_kexec(NULL);
> +	crash_kexec(NULL, buf);
> 
>   and
> 
> -void crash_kexec(struct pt_regs *regs)
> +void crash_kexec(struct pt_regs *regs, char *msg)
>  ...
>  		if (kexec_crash_image) {
>  			struct pt_regs fixed_regs;
>  
>  			crash_setup_regs(&fixed_regs, regs);
>  			crash_save_vmcoreinfo();
>  			machine_crash_shutdown(&fixed_regs);
> +			if (crash_kexec_post_notifiers) {
> +				kmsg_dump(KMSG_DUMP_PANIC);
> +				atomic_notifier_call_chain(&panic_notifier_list, 0, msg);
> +			}
>                         machine_kexec(kexec_crash_image);
> 
> Most of archs stop other cores in machine_crash_shutdown(),
> so it will work well.  Furthermore, it simplifies the special
> case where crash_kexec() is called without entering panic().
> 
> However, we need some tweaks for sh and s390 cases.  As for sh,
> it seems not to stop other cores in the crash_kexec() sequence
> (kdump support is incompleted?).  For s390, smp_send_stop() is
> called in machine_kexec() but not machine_crash_shutdown().

You could add an ifdef into the __setup() to filter out s390 and sh, until we figure out what
to do there. So the "crash_kexec_post_notifiers" wouldn't be available for those platforms.

Daniel