Re: [PATCH][RFC] Use pipe() to simulate signalfd()

[Anthony Liguori <aliguori@us.ibm.com>]

Marcelo Tosatti wrote:
> True.
>
> Either way, this starvation due to signals can't happen with the current
> scheme because signals are blocked. It seems a significant drawback.
>   

Practically speaking, I don't see signal starving being a problem.  Part 
of the benefit of this approach is that we're reducing the overall 
number of signals received.  With the in-kernel PIT, the number of 
userspace signals is even further reduced.

> Moving the signal handling + pipe write to a separate thread should get
> rid of it.
>   

Yeah, but then you just introduce buffering problems since if you're 
getting that many signals, the pipe will get full.

No point in designing for something that isn't likely to happen in practice.

Regards,

Anthony Liguori

>> No, they're independent of the patch.  The symptom is that the guest 
>> tends to hang during boot for prolonged periods of time.  It tends to be 
>> random whether it will hang at all, how long it will hang, and at what 
>> point it will hang.  It tends to hang more often in particular places.  
>> In my ubuntu server guest, for instance, it tends to hang right after 
>> partition probing, right after "Loading hardware drivers", and right 
>> before spawning a getty on /dev/tty0.
>>
>> Normally, it will only hang for a few seconds.  However, with -smp N 
>> where N > number of cpus in the host, it tends to happen more 
>> frequently.  Using clock=jiffies in the guest makes the problem go away.
>>     
>
> I'll try to reproduce that, thanks.
>
>   
>> -no-kvm-pit doesn't make a difference (the guest is normally using the 
>> pm-timer).
>>
>> Regards,
>>
>> Anthony Liguori
>>     


-------------------------------------------------------------------------
This SF.net email is sponsored by the 2008 JavaOne(SM) Conference 
Don't miss this year's exciting event. There's still time to save $100. 
Use priority code J8TL2D2. 
http://ad.doubleclick.net/clk;198757673;13503038;p?http://java.sun.com/javaone