kexec on panic

[Denys Fedoryshchenko]

On 2017-02-18 09:42, Jon Masters wrote:
> Hi Denys,
> 
> On 02/10/2017 03:14 AM, Denys Fedoryshchenko wrote:
> 
>> After years of using kexec and recent unpleasant experience with 
>> modern (supposed to be blazing fast to boot) hardware that need 5-10 
>> minutes just to pass POST tests,
>> one question came up to me:
>> Is it possible anyhow to execute regular (not special "panic" one to 
>> capture crash data) kexec on panic to reduce reboot time?
> 
> Generally, you don't want to do this, because various platform hardware
> might be in non-quiescent states (still doing DMA to random memory, 
> etc.)
> and other nastiness that means you don't want to do more than the 
> minimal
> amount in a kexec on panic (crash). We've seen no end of fun and games
> even with just regular crash dumps while hardware is busily writing to
> memory that it shouldn't be. An IOMMU helps, but isn't a cure-all.
> 
> Jon.
Well, i have to try, even sometimes i am facing issues with non-booting 
hardware even on regular kexec, but having at small customer HP server 
that need almost 6 minutes to boot,
no hot-spare(and hard to do by many reasons, no spare 10G ports, cost of 
hardware and etc) and some nasty bugs that is not resolved yet - forcing 
me to search way to reduce reboot time.
If i will find way to save backtrace and reboot fast, it will help a lot 
to debug kernels with minimal downtime, if bug is reproducible only on 
live system.

What i did now, might be insanely wrong, but:
diff -Naur linux-4.9.9-vanilla/kernel/kexec_core.c 
linux-4.9.9/kernel/kexec_core.c
--- linux-4.9.9-vanilla/kernel/kexec_core.c	2017-02-09 
07:08:40.000000000 +0000
+++ linux-4.9.9/kernel/kexec_core.c	2017-02-17 12:54:49.000000000 +0000
@@ -897,6 +897,10 @@
  			machine_crash_shutdown(&fixed_regs);
  			machine_kexec(kexec_crash_image);
  		}
+		if (kexec_image) {
+		    machine_shutdown();
+		    machine_kexec(kexec_image);
+		}
  		mutex_unlock(&kexec_mutex);
  	}
  }

Then

kexec -l /mnt/flash/kernel --append="intel_idle.max_cstate=0 
processor.max_cstate=1"

and
echo c >/proc/sysrq-trigger
worked even on busy network router, but i'm not sure it will be same on 
real networking stack crash.