lkml.org 
[lkml]   [2012]   [Apr]   [26]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    SubjectRe: [BUG] perf stat: useless output for raw events with new event parser
    From
    Date
    On Thu, 2012-04-26 at 15:12 +0200, Robert Richter wrote:
    > On 26.04.12 12:27:11, Peter Zijlstra wrote:
    > > Furthermore, once we have a common format, we could even ask Intel/AMD
    > > (and other vendors) to provide their data in this format.
    >
    > I don't think that can be done with a reasonable effort.

    I'm thinking you mis-understand, all we're talking about is a copy of
    your event list (BKDG Fam 10h Rev 3.48, section 3.14) in a usable
    format.

    > Why not simply pass an identifier for each kind of pmu and then only
    > add pmu specific code in userland? Much easier than all this sysfs
    > format thing, where the kernel tries to tell userland what to do,
    > which the kernel never can do exactly. And even if we can describe
    > everything with sysfs, kernel and userland code becomes bloated, it
    > actually is already, looking at the recent perf tool and kernel
    > updates.

    The format is only about encoding rules, it doesn't do constraints. All
    it wants to convey is if you have an event-code of: 0x4E2 where to stick
    it in perf_event_attr::config*. It doesn't want to tell you if that
    event exists and if there's a particular umask that ought to go with it.

    Its an aid to simplify constructing raw events, nothing more.

    When I want to use funny events I'm staring at the Intel-SDM/AMD-BKDG
    anyway and I find writing:

    cpu/event=0x4e2,umask=0xf8/

    A lot easier than:

    r40000f8e2

    because while I like numbers I cannot actually count and would get that
    4 wrong half the time, furthermore I'd have to actually remember the
    encoding rules; or something like:

    cpu/event=L3_fills_caused_by_L2_evictions.modified_any_core/

    Then again, some people are scared of numbers and prefer to wear down
    their finger-tips writing silly names.

    That said, a unique identifier in there might make sense, esp on things
    like ARM that don't have CPUID and even if they do it doesn't actually
    correlate to the PMU :-)

    So I'd be perfectly ok with adding something
    like /sys/bus/events/device/*/name or so.


    \
     
     \ /
      Last update: 2012-04-26 16:31    [W:5.322 / U:0.316 seconds]
    ©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site