[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

RE: [Xen-devel] [xen 3.4.2][2.6.31.6-pvops] Problems with cpufreq=xen power management on amd k8 cpu



Ok, then we can narrow down the issue to Powernow driver. CC Mark who is the 
author of Powernow.

Regards
Ke

-----Original Message-----
From: Sander Eikelenboom [mailto:linux@xxxxxxxxxxxxxx] 
Sent: Saturday, January 09, 2010 1:24 AM
To: Yu, Ke
Cc: xen-devel@xxxxxxxxxxxxxxxxxxx
Subject: Re: [Xen-devel] [xen 3.4.2][2.6.31.6-pvops] Problems with cpufreq=xen 
power management on amd k8 cpu

Hello Ke,

With xen-3.4.2 with your patch and 2.6.31.6-pvops it returns

But when i look at the current linux kernels powernow-k8.c your check isn't 
valid for family 15 (CPUID_XFAM_K8) but only for CPUID_XFAM_10H cpu's.
I have a family 15 cpu.
I don't know what this means for the rest of the code in de powernow-k8 version 
in xen. Should it support the family 15 cpu's ?

Regards,

Sander


static void check_supported_cpu(void *_rc)
512 {
513         u32 eax, ebx, ecx, edx;
514         int *rc = _rc;
515
516         *rc = -ENODEV;
517
518         if (current_cpu_data.x86_vendor != X86_VENDOR_AMD)
519                 return;
520
521         eax = cpuid_eax(CPUID_PROCESSOR_SIGNATURE);
522         if (((eax & CPUID_XFAM) != CPUID_XFAM_K8) &&
523             ((eax & CPUID_XFAM) < CPUID_XFAM_10H))
524                 return;
525
526         if ((eax & CPUID_XFAM) == CPUID_XFAM_K8) {
527                 if (((eax & CPUID_USE_XFAM_XMOD) != CPUID_USE_XFAM_XMOD) ||
528                     ((eax & CPUID_XMOD) > CPUID_XMOD_REV_MASK)) {
529                         printk(KERN_INFO PFX
530                                 "Processor cpuid %x not supported\n", eax);
531                         return;
532                 }
533
534                 eax = cpuid_eax(CPUID_GET_MAX_CAPABILITIES);
535                 if (eax < CPUID_FREQ_VOLT_CAPABILITIES) {
536                         printk(KERN_INFO PFX
537                                "No frequency change capabilities 
detected\n");
538                         return;
539                 }
540
541                 cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
542                 if ((edx & P_STATE_TRANSITION_CAPABLE)
543                         != P_STATE_TRANSITION_CAPABLE) {
544                         printk(KERN_INFO PFX
545                                 "Power state transitions not supported\n");
546                         return;
547                 }
548         } else { /* must be a HW Pstate capable processor */
549                 cpuid(CPUID_FREQ_VOLT_CAPABILITIES, &eax, &ebx, &ecx, &edx);
550                 if ((edx & USE_HW_PSTATE) == USE_HW_PSTATE)
551                         cpu_family = CPU_HW_PSTATE;
552                 else
553                         return;
554         }
555
556         *rc = 0;
557 }


Friday, January 8, 2010, 3:32:18 PM, you wrote:

> Hi Sander,

> For option1 "cpufreq=dom0-kernel", it is not working currently, the reason is 
> a bit complex, and let me explain. For every processor, there is 
> corresponding acpi processor object in ACPI DSDT table. Kernel use struct 
> acpi_processor to represent the acpi processor object. There is a field "id" 
> in acpi_processor representing the cpu id (the same id as 
> smp_processor_id()). Kernel cpufreq use acpi_processor->id to do the 
> frequency scaling. Unfortunately, in pv_ops domain0, the acpi_processor->id 
> value is not correct, because it is get by acpi_id->apic_id->id mapping, and 
> in pv_ops dom0, the apic_id of vcpu is not correctly initialized. Please see 
> drivers/acpi/processor_core.c:get_cpu_id() for detail code, where 
> cpu_physical_id() is not initialized in pv_ops dom0 case. I am still thinking 
> if there is better approach to do the acpi_id->apic_id->id mapping in pv_ops 
> dom0 case. 

> BTW, "cpufreq=dom0-kernel" is not recommend, due to that it require 
> #pCPU=#vCPU, and all vCPU in dom0 must be pined.

> For option2 "cpufreq=xen", from the log, Xen can get all the P states info, 
> but it failed to do cpufreq cpu initialization 
> (xen/arch/x86/acpi/cpufreq/cpufreq.c: cpufreq_cpu_init()), so the cpufreq is 
> not started finally, and xenpm cannot set cpufreq parameter. To get more 
> debug info, I write a simple patch as attached, could you please have a try 
> and send the debug info?

> Another thing is, does "cpufreq=xen" work with linux-2.6.18-xen dom0? 

> Best Regards
> Ke

> -----Original Message-----
> From: xen-devel-bounces@xxxxxxxxxxxxxxxxxxx 
> [mailto:xen-devel-bounces@xxxxxxxxxxxxxxxxxxx] On Behalf Of Sander Eikelenboom
> Sent: Thursday, January 07, 2010 11:08 PM
> To: xen-devel@xxxxxxxxxxxxxxxxxxx
> Subject: [Xen-devel] [xen 3.4.2][2.6.31.6-pvops] Problems with cpufreq=xen 
> power management on amd k8 cpu

> Hi All,

> I'm trying to get powermanagement working with xen and a 2.6.31.6-pvops 
> kernel on xen 3.4.2
> From what I have read, I have 2 options:
> 1) Let the xen hypervisor do the powermanagement (cpufreq=xen)
> 2) Let the dom0 kernel do the powermanagement    (cpufreq=dom0-kernel)

> When booting the 2.6.31.6-pvops kernel on bare metal without hypervisor the 
> powernow-k8 module works.
> When booting the 2.6.31.6-pvops kernel as dom0 on the xen 3.4.2 hypervisor 
> the powernow-k8 module doesn't load and complains:
>      Jan  7 15:25:37 localhost kernel: [   12.797914] powernow-k8: Found 1 
> AMD Athlon(tm) Dual Core Processor 4850e processors (2 cpu cores) (version 
> 2.20.00)
>      Jan  7 15:25:37 localhost kernel: [   12.797926] powernow-k8: register 
> performance failed: bad ACPI data
>      Jan  7 15:25:37 localhost kernel: [   12.797929] [Firmware Bug]: 
> powernow-k8: No compatible ACPI _PSS objects found.
>      Jan  7 15:25:37 localhost kernel: [   12.797931] [Firmware Bug]: 
> powernow-k8: Try again with latest BIOS.
>      Jan  7 15:25:37 localhost kernel: [   12.797941] powernow-k8: register 
> performance failed: bad ACPI data
>      Jan  7 15:25:37 localhost kernel: [   12.822315] acpi-cpufreq: 
> acpi_cpufreq_init
>      Jan  7 15:25:37 localhost kernel: [   12.822319] acpi-cpufreq: 
> acpi_cpufreq_early_init
>      Jan  7 15:25:37 localhost kernel: [   12.822333] acpi-cpufreq: 
> acpi_cpufreq_cpu_init
>      Jan  7 15:25:37 localhost kernel: [   12.822338] acpi-cpufreq: 
> acpi_cpufreq_cpu_init

> It could be that additional changes are needed to the pvops kernel that have 
> been made to the xen 2.6.18.8 kernel ?

> When trying the first option, and let the hypervisor do the powermanagement 
> (booting with cpufreq=xen cpufreq.debug=2 loglvl=all)
> it shows the following in xm dmesg (complete xm dmesg attachted):

> (XEN) Set CPU acpi_id(0) cpuid(0) Px State info:
> (XEN)   _PPC: 0
> (XEN) Set CPU acpi_id(0) cpuid(0) Px State info:
> (XEN)   _PCT: descriptor=130, length=12, space_id=127, bit_width=0, 
> bit_offset=0, reserved=0, address=0
> (XEN)   _PCT: descriptor=130, length=12, space_id=127, bit_width=0, 
> bit_offset=0, reserved=0, address=0
> (XEN)   _PSS: state_count=6
> (XEN)   State0: 2500MHz 50000mW 100us 9us 0xe8201391 0x391
> (XEN)   State1: 2400MHz 46020mW 100us 9us 0xe82013d0 0x3d0
> (XEN)   State2: 2200MHz 38671mW 100us 9us 0xe820144e 0x44e
> (XEN)   State3: 2000MHz 32100mW 100us 9us 0xe82014cc 0x4cc
> (XEN)   State4: 1800MHz 26265mW 100us 9us 0xe820154a 0x54a
> (XEN)   State5: 1000MHz 13888mW 100us 9us 0xe8201582 0x582
> (XEN)   _PSD: num_entries=5 rev=0 domain=0 coord_type=253 num_processors=2
> (XEN)   _PPC: 0
> (XEN) Set CPU acpi_id(1) cpuid(1) Px State info:
> (XEN)   _PPC: 0
> (XEN) Set CPU acpi_id(1) cpuid(1) Px State info:
> (XEN)   _PCT: descriptor=130, length=12, space_id=127, bit_width=0, 
> bit_offset=0, reserved=0, address=0
> (XEN)   _PCT: descriptor=130, length=12, space_id=127, bit_width=0, 
> bit_offset=0, reserved=0, address=0
> (XEN)   _PSS: state_count=6
> (XEN)   State0: 2500MHz 50000mW 100us 9us 0xe8201391 0x391
> (XEN)   State1: 2400MHz 46020mW 100us 9us 0xe82013d0 0x3d0
> (XEN)   State2: 2200MHz 38671mW 100us 9us 0xe820144e 0x44e
> (XEN)   State3: 2000MHz 32100mW 100us 9us 0xe82014cc 0x4cc
> (XEN)   State4: 1800MHz 26265mW 100us 9us 0xe820154a 0x54a
> (XEN)   State5: 1000MHz 13888mW 100us 9us 0xe8201582 0x582
> (XEN)   _PSD: num_entries=5 rev=0 domain=0 coord_type=253 num_processors=2
> (XEN)   _PPC: 0

> So it seems to find all available powerstates, but using xenpm doesn't seem 
> to work:

> serveerstertje:~# xenpm  get-cpu-topology
> CPU     core    socket
> CPU0     0       0
> CPU1     1       0
> serveerstertje:~# xenpm  set-scaling-governor userspace
> [CPU0] failed to set governor name
> [CPU1] failed to set governor name
> serveerstertje:~# xenpm  set-scaling-governor ondemand
> [CPU0] failed to set governor name
> [CPU1] failed to set governor name
> serveerstertje:~# xenpm  set-scaling-governor performance
> [CPU0] failed to set governor name
> [CPU1] failed to set governor name
> serveerstertje:~# xenpm  set-scaling-governor powersave
> [CPU0] failed to set governor name
> [CPU1] failed to set governor name
> serveerstertje:~# xenpm  get-cpufreq-para
> [CPU0] failed to get cpufreq parameter
> [CPU1] failed to get cpufreq parameter





-- 
Best regards,
 Sander                            mailto:linux@xxxxxxxxxxxxxx


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


 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.