[Xen-devel] [QEMU-traditional] ACPI AML code races with QEMU updating the vCPU count when hotplugging more than ~14 VCPUs.

[Konrad Rzeszutek Wilk <konrad.wilk@xxxxxxxxxx>]

This is a race so the amount varies but on a 4PCPU box
I seem to get only ~14 out of 16 vCPUs I want to online.

The issue at hand is that QEMU xenstore.c hotplug code
does to the command: xl vcpu-set latest 16
(guest config has vcpus=1, maxvcpus=32) this:


QEMU:                                           Guest OS:
-xenstore_process_vcpu_set_event
 -> Gets an XenBus notification for CPU1
 -> Updates the gpe_state.cpus_state bitfield.       
        -> Pulses the ACPI SCI 
                                                - ACPI SCI kicks in

 -> Gets an XenBus notification for CPU2
 -> Updates the gpe_state.cpus_state bitfield.
        -> Pulses the ACPI SCI 

 -> Gets an XenBus notification for CPU3
 -> Updates the gpe_state.cpus_state bitfield.
        -> Pulses the ACPI SCI 
   ...
                                                 - Method(PRST) invoked
                                                          
 -> Gets an XenBus notification for CPU12
 -> Updates the gpe_state.cpus_state bitfield.
        -> Pulses the ACPI SCI 
                                                  - reads AF00 for CPU state
                                                    [gets 0xff]
                                                  - reads AF02 [gets 0x7f]

                                                        
 -> Gets an XenBus notification for CPU13
 -> Updates the gpe_state.cpus_state bitfield.
        -> Pulses the ACPI SCI 

        .. until VCPU 16
                                                 - Method PRST updates
                                                   PR01 through 13 FLG
                                                   entry.
                                                 - PR01->PR13 _MAD
                                                   invoked.

                                                 - Brings up 13 CPUs.

While QEMU updates the rest of the cpus_state bitfields the ACPI AML
only does the CPU hotplug on those it had read. For reference
please see the debug patch and also the QEMU log. Look for
'gpe_cpus_readb'.

My thinking of how to fix this is just to add in
xenstore_process_vcpu_set_event
 - a scan for all of the other availability/cpu states.
 - for each of the cpu availability states query the
   gpe_state.cpus_state and if different modify them (and set
   a bool that one of them was modified).
 - When done scanning and if the bool was set, then kick of the
   qemu_irq_pulse.

Then if the other events are triggered we can just check the
gpe_state.cpus_state against what XenBus thinks and if they
are the same just return without doing the qemu_irq_pulse.

Thoughts?

Attachment: qemu-debug.patch
Description: Text document

Attachment: qemu-dm-latest.log
Description: Text document

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel