[Xen-ia64-devel] soft lockup warning when domain creation

[Isaku Yamahata <yamahata@xxxxxxxxxxxxx>]

Hi.
Soft lockup warning pops up when creating a domain with much memory
(e.g. over 100GB). I'm guessing that populate physmap hypercall causes it.

There is the continuation mechanism for hypercalls which might take long
time. (The destroy domain hypercall was a exception.)
The continuation gives a guest kernel a chance to handle interruptions
during a long hypercall. However the soft lockup detector uses
kernel thread and the xenLinux/ia64 kernel with the default configuration
isn't kernel preemptive. So a continuable hypercall may cause the warning
message.

There are several ways to address it.
What do you think? My preference (at this moment) is the choice C.

A. Change hypercall continuation ABI.
   Currently the continuation automatically restart the hypercall.
   Change the ABI such that the hypercall returns -EAGAIN and the guest
   kernel is responsible for reissuing hypercall.
   Then the non-preemptable guest kernel can switch thread context.
   Pros
        This would be the most generic way and would solves
        the potential future similar issues.
   Cons
        This will break the hypercall ABI
        It would be difficult to convince other arch developers so that
        this would be ia64 specific.

B. Enable kernel preemption of the default configuration.
   Pros
        This would be a logical result based on the hypercall continuation
        mechanism.
   Cons
        We would have to debug the kernel preemption probably.
        People may not want to enable the option.
        
C. modify the caller of the hypercall so that it repeats the hypercall.
   Probably at this moment, we need to modify only the callers of
   populate physmap hypercall.
   There were two choices to modify. libxc or privcmd driver.
   I think that any user processes shouldn't cause soft lockup, so
   privcmd driver is the place to modify.
   Pros
        When we want to create a domain with much more momery
        in the future, such a repetition would be necessary anyway.
        Probably we need to address only populate physmap hypercall.
   Cons
        Potentially we may have to modify more another hypercall callers.
        (This would be very ironic to the xencomm generalization patch.)

Thanks,
-- 
yamahata

_______________________________________________
Xen-ia64-devel mailing list
Xen-ia64-devel@xxxxxxxxxxxxxxxxxxx
http://lists.xensource.com/xen-ia64-devel