Re: [PATCH][1/7] perfctr-2.7.2 for 2.6.6-mm2: core

From: Mikael Pettersson (mikpe_at_csd.uu.se)
Date: 05/16/04

  • Next message: tmp: "Kernel OOPS" 
    Date:	Sun, 16 May 2004 13:52:43 +0200 (MEST)
To: akpm@osdl.org

    On Sat, 15 May 2004 22:39:37 -0700, Andrew Morton wrote:
    >Mikael Pettersson <mikpe@csd.uu.se> wrote:
    >>
    >> The per-process perfctrs used to be accessed via /proc/pid/perfctr,
    >> but the /proc/pid/-now-denotes-that-posixy-process-grop-thingy
    >> change in 2.6 broke that, so I went away from /proc/pid/ last year.
    >>
    >> The per-process perfctrs would need their own file system mount point,
    >> with files or directories named by actual kernel task id. readdir()
    >> won't be fun to implement. The top-level access point can certainly
    >> be in a special fs, the question is whether I must go further and
    >> do that also for the individual control data fields?
    >>
    >> The global-mode perfctrs could be accessed via /dev/cpu/$cpu/gperfctr
    >> for per-cpu operations, and /dev/cpu/gperfctr/$file for global
    >> operations (like start and stop). However, global-mode perfctrs
    >> are considerably less important than per-process perfctrs, and
    >> I'd rather remove them until the per-process stuff is done.
    >
    >Well standing back and squinting at the problem:
    >
    >As it collects samples globally, oprofile is a system-wide thing. And a
    >filesytem is a system-wide thing too, so one maps onto the other nicely.
    >
    >But perfctr is a *per process* thing, and that doesn't map onto a
    >filesystem abstraction very well at all.
    >
    >So unless someone comes up with a cunning way of getting your square peg
    >into a filesystem's round hole, I'd be inclined to stick with a syscall
    >interface. Six syscalls would be preferable to
    >one-which-contains-a-switch-statement, please.

    If I drop the global-mode counters I'll still need seven calls:
    six for the per-process counters, and one get-information call.
    There is information that user-space needs which no other
    kernel interface provides (AFAIK): the timebase-to-core multiplier
    on PowerPC, and the set of forbidden(*) CPUs on x86/x86_64.
    I also export several CPU feature flags. One of them, the
    "can use overflow interrupt counters" flag, can't be detected
    by user-space from the CPU type alone since it also depends
    on local APIC availability, which in turn depends on kernel
    .config, DMI scan, and kernel boot options.

    I think I can eliminate the structure marshalling code, but
    it will require padding structures with dummy fields for
    future hardware extensions.

    So seven syscalls, sys_vperfctr+0,...,sys_vperfctr+6, no
    global-mode counters, and no marshalling code. Sounds Ok?

    /Mikael

    (*) You can thank Intel's HT P4 for that. Hyperthreaded P4s are
    _asymmetric_ wrt the availability of the performance counters.
    The solution is to restrict processes to thread #0 in each physical
    CPU, but users must be told about this so they don't try to change
    affinity to one of the forbidden (non-thread-#0) CPUs.
    There are safety checks in place, so if they do so anyway their
    counters are terminated before any damage is done.
    -
    To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
    the body of a message to majordomo@vger.kernel.org
    More majordomo info at http://vger.kernel.org/majordomo-info.html
    Please read the FAQ at http://www.tux.org/lkml/

  • Next message: tmp: "Kernel OOPS"

    Relevant Pages