Re: kvm-78 - kernel panic after using system_reset except when using -no-kvm-irqchip

[Jan Kiszka <jan.kiszka@web.de>]

[-- Attachment #1: Type: text/plain, Size: 2353 bytes --]

Charles Duffy wrote:
> This happens every other use of system_reset -- ie. resetting the system
>  using system_reset while in this panic'd state results in correct
> functionality. Following the suggestion and booting the guest with
> noapic results in other IRQ-related errors.
> 
> The guest kernel is CentOS 5 2.6.18-53.el5 on x86_64. The host is
> running 2.6.27.5.
> 
> Kernel command line: ro root=/dev/VolGroup01/LogVol00
> Initializing CPU#0
> PID hash table entries: 4096 (order: 12, 32768 bytes)
> Console: colour VGA+ 80x25
> Dentry cache hash table entries: 131072 (order: 8, 1048576 bytes)
> Inode-cache hash table entries: 65536 (order: 7, 524288 bytes)
> Checking aperture...
> Memory: 746368k/76936k available (2434k kernel code, 21180k reserved,
> 1235k data, 192k init)
> Calibrating delay using timer specific routine.. 4793.50 BogoMIPS
> (lpj=2396750)
> Security Framework v1.0.0 initialized
> SELinux:  Initializing.
> selinux_register_security:  Registering secondary module capability
> Capability LSM initialized as secondary
> Mount-cache hash table entries: 256
> CPU: L1 I cache: 32K, L1 D cache: 32K
> CPU: L2 cache: 2048K
> SMP alternatives: switching to UP code
> ACPI: Core revision 20060707
> irq 25, desc: ffffffff803afc80, depth: 1, count: 0, unhandled: 0
> ->handle_irq():  ffffffff800b54e3, handle_bad_irq+0x0/0x1f6
> ->chip(): ffffffff802ea700, 0xffffffff802ea700
> ->action(): 0000000000000000
>   IRQ_DISABLED set
> unexpected IRQ trap at vector 19
> ..MP-BIOS bug: 8254 timer not connected to IO-APIC
> timer doesn't work through the IO-APIC - disabling NMI Watchdog!
> Kernel panic - not syncing: IO-APIC + timer doesn't work! Try using the
> 'noapic' kernel parameter

Still too early for me, so I didn't get yet if you can trigger this
guest panic reliably or only sporadically (like I can). In the former
case I would be very interested in the how!

So far my theory on this is that the guest happens to loose too many PIC
timer ticks during the test window where it checks the PIC IRQ routing
(it does so with 10 ticks, out of which at least 4 must be delivered).
The theory also says that this is hard to avoid, maybe just less likely
with Gleb's timer drift compensation patches.

Jan

PS: Booting with noapic leaves many real boxes useless as well.


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 257 bytes --]