svn commit: r275679 - head/sys/x86/xen

[Roger Pau Monné]

El 10/12/14 a les 21.17, Justin T. Gibbs ha escrit:
> On Dec 10, 2014, at 6:25 AM, Roger Pau MonnXX <royger at FreeBSD.org> wrote:
>>
>> Author: royger
>> Date: Wed Dec 10 13:25:21 2014
>> New Revision: 275679
>> URL: https://svnweb.freebsd.org/changeset/base/275679
>>
>> Log:
>>  xen/intr: balance dynamic interrupts across available vCPUs
>>
>>  By default Xen binds all event channels to vCPU#0, and FreeBSD only shuffles
>>  the interrupt sources once, at the end of the boot process. Since new event
>>  channels might be created after this point (because new devices or backends
>>  are added), try to automatically shuffle them at creation time. 
>>
>>  This does not affect VIRQ or IPI event channels, that are already bound to a
>>  specific vCPU as requested by the caller.
>>
>>  Sponsored by: Citrix Systems R&D
>>
>> Modified:
>>  head/sys/x86/xen/xen_intr.c
>>
>> Modified: head/sys/x86/xen/xen_intr.c
>> ==============================================================================
>> --- head/sys/x86/xen/xen_intr.c	Wed Dec 10 11:42:02 2014	(r275678)
>> +++ head/sys/x86/xen/xen_intr.c	Wed Dec 10 13:25:21 2014	(r275679)
>> @@ -425,6 +425,17 @@ xen_intr_bind_isrc(struct xenisrc **isrc
>> 	/* Assign the opaque handler (the event channel port) */
>> 	*port_handlep = &isrc->xi_port;
>>
>> +#ifdef SMP
>> +	if (type == EVTCHN_TYPE_PORT) {
>> +		/*
>> +		 * By default all interrupts are assigned to vCPU#0
>> +		 * unless specified otherwise, so shuffle them to balance
>> +		 * the interrupt load.
>> +		 */
>> +		xen_intr_assign_cpu(&isrc->xi_intsrc, intr_next_cpu());
>> +	}
>> +#endif
>> +
> 
> Shouldn't this only happen for new interrupt source objects, not those that are being reclaimed/reused?

No, because the vCPU binding is done based on the event channel port,
which changes when the event channel is allocated/setup.

What I realize now is that I should probably have used intr_event_bind
instead of directly calling xen_intr_assign_cpu, so that the ithread cpu
mask is also correctly set.

Roger.