Counting CPU_CLK_UNHALTED on an idle system

Counting CPU_CLK_UNHALTED on an idle system

[John Sigler]

[ Note: cross-posted to linux-rt-users and oprofile-list ]

Hello,

The x86 Performance Monitoring Event "CPU_CLK_UNHALTED" is defined as:
Number of cycles during which the processor is not halted.

IIUC, when the CPU is in the HALT state, CPU_CLK_UNHALTED should not 
increment, right? When Linux has nothing to do, it executes HLT.

http://lxr.linux.no/source/arch/i386/kernel/process.c#L101

I set up OProfile to sample EIP every 200000 CPU_CLK_UNHALTED cycles on 
an idle system. I was expecting a very small number of samples, since 
CPU_CLK_UNHALTED should not increment (most of the time).

I ran the daemon 5 minutes on a idle system

# opreport -l
CPU: PIII, speed 1266.7 MHz (estimated)
Counted CPU_CLK_UNHALTED events (clocks processor is not halted) with a 
unit mask of 0x00 (No unit mask) count 200000
samples    %      symbol name
1868484  99.4914  cpu_idle
    4560   0.2428  (oprofiled)
     727   0.0387  ip_route_input
     457   0.0243  acpi_ev_asynch_execute_gpe_method
     454   0.0242  ip_route_output_slow
     286   0.0152  dma_free_coherent
     280   0.0149  ip_route_input_slow
     273   0.0145  acpi_ex_opcode_6A_0T_1R
     220   0.0117  __acpi_nmi_disable
     130   0.0069  Dta1xxRxProcessStatusFlags
      94   0.0050  Dta1xxTxProcessStatusFlags
      93   0.0050  (bash)
      85   0.0045  Dta1xxIsr
      79   0.0042  intel_machine_check
      56   0.0030  sys_olduname
      51   0.0027  Dta1xxPeriodicIntHandler

cpu_idle breakdown:

   c0100e70 1        5.4e-05
   c0100e71 1        5.4e-05
   c0100e78 1        5.4e-05
   c0100e88 1        5.4e-05
   c0100e8a 1        5.4e-05
   c0100e97 4        2.1e-04
   c0100ea8 1868475  99.9995

cpu_idle objdump:

c0100ea6:       fb                      sti
c0100ea7:       f4                      hlt
c0100ea8:       eb 0b                   jmp    c0100eb5

(If an interrupt is used to resume execution after a HLT instruction, 
the saved instruction pointer points to the instruction following the 
HLT instruction.)

oprofiled collected 1878300 samples. In other words, the system received 
1878300 NMIs. (5 * 60 * 1.2667e9) / 1878300 = 202316 cycles between NMI

Seems like CPU_CLK_UNHALTED increments every cycle.

Perhaps HLT does not put my CPU in the HALT state?

Did I miss something?

Regards.


-------------------------------------------------------------------------
This SF.net email is sponsored by DB2 Express
Download DB2 Express C - the FREE version of DB2 express and take
control of your XML. No limits. Just data. Click to get it now.
http://sourceforge.net/powerbar/db2/

Re: Counting CPU_CLK_UNHALTED on an idle system

[Stephane Eranian]

John,

On PIII, CPU_CLK_UNHALTED is broken and keeps on counting while in HALT state.

On Fri, Jul 13, 2007 at 02:33:09PM +0200, John Sigler wrote:
> [ Note: cross-posted to linux-rt-users and oprofile-list ]
> 
> Hello,
> 
> The x86 Performance Monitoring Event "CPU_CLK_UNHALTED" is defined as:
> Number of cycles during which the processor is not halted.
> 
> IIUC, when the CPU is in the HALT state, CPU_CLK_UNHALTED should not 
> increment, right? When Linux has nothing to do, it executes HLT.
> 
> http://lxr.linux.no/source/arch/i386/kernel/process.c#L101
> 
> I set up OProfile to sample EIP every 200000 CPU_CLK_UNHALTED cycles on 
> an idle system. I was expecting a very small number of samples, since 
> CPU_CLK_UNHALTED should not increment (most of the time).
> 
> I ran the daemon 5 minutes on a idle system
> 
> # opreport -l
> CPU: PIII, speed 1266.7 MHz (estimated)
> Counted CPU_CLK_UNHALTED events (clocks processor is not halted) with a 
> unit mask of 0x00 (No unit mask) count 200000
> samples    %      symbol name
> 1868484  99.4914  cpu_idle
>     4560   0.2428  (oprofiled)
>      727   0.0387  ip_route_input
>      457   0.0243  acpi_ev_asynch_execute_gpe_method
>      454   0.0242  ip_route_output_slow
>      286   0.0152  dma_free_coherent
>      280   0.0149  ip_route_input_slow
>      273   0.0145  acpi_ex_opcode_6A_0T_1R
>      220   0.0117  __acpi_nmi_disable
>      130   0.0069  Dta1xxRxProcessStatusFlags
>       94   0.0050  Dta1xxTxProcessStatusFlags
>       93   0.0050  (bash)
>       85   0.0045  Dta1xxIsr
>       79   0.0042  intel_machine_check
>       56   0.0030  sys_olduname
>       51   0.0027  Dta1xxPeriodicIntHandler
> 
> cpu_idle breakdown:
> 
>    c0100e70 1        5.4e-05
>    c0100e71 1        5.4e-05
>    c0100e78 1        5.4e-05
>    c0100e88 1        5.4e-05
>    c0100e8a 1        5.4e-05
>    c0100e97 4        2.1e-04
>    c0100ea8 1868475  99.9995
> 
> cpu_idle objdump:
> 
> c0100ea6:       fb                      sti
> c0100ea7:       f4                      hlt
> c0100ea8:       eb 0b                   jmp    c0100eb5
> 
> (If an interrupt is used to resume execution after a HLT instruction, 
> the saved instruction pointer points to the instruction following the 
> HLT instruction.)
> 
> oprofiled collected 1878300 samples. In other words, the system received 
> 1878300 NMIs. (5 * 60 * 1.2667e9) / 1878300 = 202316 cycles between NMI
> 
> Seems like CPU_CLK_UNHALTED increments every cycle.
> 
> Perhaps HLT does not put my CPU in the HALT state?
> 
> Did I miss something?
> 
> Regards.
> 
> 
> -------------------------------------------------------------------------
> This SF.net email is sponsored by DB2 Express
> Download DB2 Express C - the FREE version of DB2 express and take
> control of your XML. No limits. Just data. Click to get it now.
> http://sourceforge.net/powerbar/db2/
> _______________________________________________
> oprofile-list mailing list
> oprofile-list@lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/oprofile-list

-- 

-Stephane

-------------------------------------------------------------------------
This SF.net email is sponsored by DB2 Express
Download DB2 Express C - the FREE version of DB2 express and take
control of your XML. No limits. Just data. Click to get it now.
http://sourceforge.net/powerbar/db2/

Re: Counting CPU_CLK_UNHALTED on an idle system

[John Sigler]

Stephane Eranian wrote:

> On PIII, CPU_CLK_UNHALTED is broken and keeps on counting while in HALT state.

Stephane,

Would you have a reference?

There is no such mention in the latest Specification Update (May 2007).

http://developer.intel.com/design/intarch/specupdt/244453.htm

Regards.

Re: Counting CPU_CLK_UNHALTED on an idle system

[Stephane Eranian]

John,


On Fri, Jul 13, 2007 at 03:02:12PM +0200, John Sigler wrote:
> Stephane Eranian wrote:
> 
> > On PIII, CPU_CLK_UNHALTED is broken and keeps on counting while in HALT state.
> 
> Stephane,
> 
> Would you have a reference?
> 
No references, just my presonal experience. You'd have to ask Intel for a confirmation
or just file a bug report. I don't know where, unfortunately.

Note that the event works on Pentium M.

> There is no such mention in the latest Specification Update (May 2007).
> 
> http://developer.intel.com/design/intarch/specupdt/244453.htm
> 
> Regards.
> 
> -------------------------------------------------------------------------
> This SF.net email is sponsored by DB2 Express
> Download DB2 Express C - the FREE version of DB2 express and take
> control of your XML. No limits. Just data. Click to get it now.
> http://sourceforge.net/powerbar/db2/
> _______________________________________________
> oprofile-list mailing list
> oprofile-list@lists.sourceforge.net
> https://lists.sourceforge.net/lists/listinfo/oprofile-list

-- 

-Stephane