lkml.org 
[lkml]   [2010]   [Mar]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: [PATCH] x86,perf: Unmask LVTPC only if we have APIC supported

    * Cyrill Gorcunov <gorcunov@gmail.com> wrote:

    > On Sat, Mar 13, 2010 at 01:40:36PM +0100, Ingo Molnar wrote:
    > >
    > > * Cyrill Gorcunov <gorcunov@gmail.com> wrote:
    > >
    > > > On Sat, Mar 13, 2010 at 01:24:32PM +0100, Ingo Molnar wrote:
    > > > >
    > > > > * Cyrill Gorcunov <gorcunov@openvz.org> wrote:
    > > > >
    > > > > > Ingo reported
    > > > > > |
    > > > > > | There's a build failure on -tip with the P4 driver, on UP 32-bit, if
    > > > > > | PERF_EVENTS is enabled but UP_APIC is disabled:
    > > > > > |
    > > > > > | arch/x86/built-in.o: In function `p4_pmu_handle_irq':
    > > > > > | perf_event.c:(.text+0xa756): undefined reference to `apic'
    > > > > > | perf_event.c:(.text+0xa76e): undefined reference to `apic'
    > > > > > |
    > > > > >
    > > > > > So we have to unmask LVTPC only if we're configured to have one.
    > > > > >
    > > > > > Reported-by: Ingo Molnar <mingo@elte.hu>
    > > > > > CC: Lin Ming <ming.m.lin@intel.com>
    > > > > > CC: Peter Zijlstra <peterz@infradead.org>
    > > > > > Signed-off-by: Cyrill Gorcunov <gorcunov@openvz.org>
    > > > > > ---
    > > > > > arch/x86/kernel/cpu/perf_event_p4.c | 2 ++
    > > > > > 1 file changed, 2 insertions(+)
    > > > > >
    > > > > > Index: linux-2.6.git/arch/x86/kernel/cpu/perf_event_p4.c
    > > > > > =====================================================================
    > > > > > --- linux-2.6.git.orig/arch/x86/kernel/cpu/perf_event_p4.c
    > > > > > +++ linux-2.6.git/arch/x86/kernel/cpu/perf_event_p4.c
    > > > > > @@ -365,8 +365,10 @@ static int p4_pmu_handle_irq(struct pt_r
    > > > > > }
    > > > > >
    > > > > > if (handled) {
    > > > > > +#ifdef CONFIG_X86_LOCAL_APIC
    > > > > > /* p4 quirk: unmask it again */
    > > > > > apic_write(APIC_LVTPC, apic_read(APIC_LVTPC) & ~APIC_LVT_MASKED);
    > > > > > +#endif
    > > > > > inc_irq_stat(apic_perf_irqs);
    > > > >
    > > > > This ugly #ifdef looks like a workaround though. Why doesnt apic_write() map
    > > > > to nothing in that case?
    > > > >
    > > > > Ingo
    > > > >
    > > >
    > > > It is. I mean -- it maps to nothing if apic is disabled. But the scenario is
    > > > that no apic configured at all. Actually I wonder how this code is supposed
    > > > to work without apic support.
    > > >
    > > > Pehpaps better to make a p4 quirk helper here, since #ifdef at this point
    > > > looks ugly indeed.
    > > >
    > > > Don't apply it then. Will back with other solution.
    > >
    > > apic_write() is really just equivalent to a spin_lock() on UP without
    > > UP_IOAPIC set - it should do nothing. So if it does something and fails the
    > > build, then that should be fixed - not the P4 PMU code.
    > >
    > > Ingo
    > >
    >
    > Looking at code a bit and config deps I think the former proposal with
    > #ifdef is minimal (in amount of changes) and sufficient. perf_event.c
    > uses #ifdef CONFIG_X86_LOCAL_APIC for the very same reason.
    >
    > The former issue with config dependencies is that we may need to compile
    > perf_event.c without CONFIG_LOCAL_APIC support at all (and this is a case
    > for which you've posted the config). CONFIG_LOCAL_APIC deps on X86_UP_APIC,
    > the config has no X86_UP_APIC support and as result -- no CONFIG_LOCAL_APIC
    > and no apic.o compiled.
    >
    > So, as expected, no apic_write/read and friends there. We may introduce
    > apic_write/read weak(s) but this would only mess the code more and would
    > smell unpleasant I think :) .
    >
    > All-in-once: unresolved external symbol here, which could be fixed either by
    > introducing dummy symbol, or conditional compilation. I think the second is
    > preferred if the issue is just one line code.
    >
    > Or you mean something different and I took a wrong mind-path?

    Well it's not just one line of code as (like you mentioned) perf_event.c is
    affected as well.

    Introducing a dummy (NOP) placeholder method is what we are doing in all the
    other cases (such as spin_lock()), we dont pollute the kernel with #ifdefs.

    Ingo


    \
     
     \ /
      Last update: 2010-03-13 14:55    [W:0.028 / U:120.348 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site