[syslinux] Domain-specific HLT when idle

[Geert Stappers]

Full quote for Debian Bug Tracking System bugnumber 504495
 ( http://bugs.debian.org/504495 )

Op 20090503 om 00:58 schreef Peter Iannucci:
> H. Peter Anvin wrote:
>> Peter Iannucci wrote:
>>   
>>> Hello hpa,
>>>
>>> We are developing a Xen-based virtualization service, and we are trying
>>> to reduce excessive CPU utilization at isolinux boot prompts.  When
>>> users boot up a virtual machine on an isolinux boot CD, the CPU starts
>>> spinning at 100%.  These machines frequently get left in this state for
>>> days or weeks, degrading the service for everyone else.
>>>
>>> You replied to the work-around discussed in the earlier thread at
>>>
>>> http://syslinux.zytor.com/archives/2005-June/005291.html
>>>
>>>     
>>>> Breaks serial console.
>>>>     -hpa
>>>>       
>>> In our case, the serial console is not an issue.  We would like to
>>> submit a patch to isolinux which HLTs on getchar if isolinux is not
>>> running with the serial console option.  Do you think this is the right
>>> solution to this problem?  Can we work with you to resolve this?
>>>
>>>     
>>
>> It would *probably* be okay to do when serial console is not enabled,
>> but it seems a bit crude... and also it's somewhat confusing that one
>> should enable serial console and get drastically different behavior.
>>
>> What is much more confusing to me is why do you have VMs left like this
>> for "days or weeks"?

It is about wasted resources.
That used to be a non-issue, but now on virtualisation we have to pay ...

>> 	-hpa
>>
> One improvement might be enabling interrupts for the serial port.  On  
> line 1277 of 2.6.29/drivers/serial/8250.c, this is accomplished with
>
> serial_outp(up, UART_IER, 0x0f);
>
> where the UART Interrupt Enable Register is defined on line 23 of  
> include/linux/serial_reg.h, and serial_outp is responsible for  
> special-casing unusual devices.  The protocol for enabling interrupts  
> via this register is common to all standard PC serial ports (and can  
> easily be scraped from 8250.c for exceptions as desired).
>
> On scanning through syslinux/core/parseconfig.inc, it looks like the  
> infrastructure for enabling serial IRQ's is already present, but is not  
> being used.  Once these interrupts are enabled, the HLT should take care  
> of both cases fine.

Proposal:

Introduce a configuration keyword named like 'halt_on_idle'.

'halt_on_idle' is optional
and when set is HLT used in the "idle loop" that waits for a key stroke.

> By way of explanation, when our users boot a VM with a live CD and  
> reboot from within the guest, the domain still has the CD still  
> attached.  This leaves the machine spinning at the boot prompt.  Users  
> forget and leave their machines in this state frequently enough that it  
> has become a problem.  Similarly, users occasionally create a new VM and  
> specify a first-time boot CD, but never log in to complete the install.   
> It only takes a few VMs using 100% of a CPU to degrade the performance  
> of a host.

Probably should 'halt_on_idle' be default
and is a 'no_halt_on_idle' or a 'active_poll' needed.

'active_poll' has to be mentioned in de section 'serial port' of the docs.


> -Peter Iannucci


Cheers
Geert Stappers