lkml.org 
[lkml]   [2009]   [Apr]   [29]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [PATCH 0/29] x86/perfcounters: x86 and AMD cpu updates

    * Robert Richter <robert.richter@amd.com> wrote:

    > This patch series updates the perfcounters implementation mainly
    > for the x86 architecture.

    Wow, very nice series! Still havent looked through all of them, but
    wanted to give some quick feedback that the splitup and direction
    looks all good.

    > Also, it introduces a data structure (struct pmu) describing a
    > generic performance monitoring unit (pmu). This structure is a
    > replacement for struct hw_perf_counter_ops. Similiar, I introduced
    > struct x86_pmu for the x86 architecture (as a replacement for
    > struct pmc_x86_ops).

    Looks sensible. There will eventually be PMU features that dont fit
    the hw-counter abstraction but which can still be expressed at the
    general counter level.

    > There are patches for x86 with some fixes and cleanups, a change
    > in the model specific split and a complete rework of AMD pmu code.
    > The result is simplified model specific code and more generalized
    > and unified code. Features that are only supported by AMD or Intel
    > are now implemented in vendor specific functions.

    Nice!

    > The AMD pmu differs to Intel, especially there is no status
    > register and also there are no fixed counters. This makes a
    > separate interrupt handler for AMD cpus necessary. Also, a global
    > disable/enable of the performance counters (e.g. to avoid NMIs to
    > protect the modification of a list) is expensive on AMD cpus
    > leading to up to 4 msr reads/writes per counter. There is still
    > some more work to do here to avoid this.

    Yeah. The previous code was really just a first-level approximation
    to show that it can be done.

    > This patch series bases on the tip/percounters/core branch.
    >
    > I developed this patches based on 03ced43 and later rebased to
    > 1b88991. The latest tip/percounters/core branch seems to be
    > broken, no nmis are delivered, only perfcounter interrupts with no
    > results on kerneltop. I am still debugging this. However, I could
    > test successfully the patch series based on 03ced43 and want to
    > release the patches anyway.

    hm, it works all fine for me. This is "perf top" output from an
    AMD/Barcelona box:

    ------------------------------------------------------------------------------
    KernelTop: 139908 irqs/sec kernel: 9.5% [NMI, 100000 CPU cycles], (all, 16 CPUs)
    ------------------------------------------------------------------------------

    events pcnt RIP kernel function
    ______ ______ _____ ________________ _______________

    11038.00 - 22.2% - ffffffff8037a090 : clear_page_c
    5842.00 - 11.7% - ffffffff804c6e02 : acpi_pm_read
    2235.00 - 4.5% - ffffffff80579530 : page_fault
    1518.00 - 3.0% - ffffffff8037a300 : copy_user_generic_string!
    1184.00 - 2.4% - ffffffff80291598 : get_page_from_freelist
    899.00 - 1.8% - ffffffff8057919a : _spin_lock
    824.00 - 1.7% - ffffffff802a0c0a : unmap_vmas
    739.00 - 1.5% - ffffffff8029d8a4 : __dec_zone_state
    696.00 - 1.4% - ffffffff8028aafe : perf_swcounter_event
    672.00 - 1.3% - ffffffff802a1b2e : handle_mm_fault

    that's NMIs delivered to 16 cores. No lockups and no stuck IRQ
    handling.

    Would be nice to fix this...

    Ingo


    \
     
     \ /
      Last update: 2009-04-29 13:13    [W:0.027 / U:31.208 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site