lkml.org 
[lkml]   [2008]   [Nov]   [21]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    /
    Date
    From
    SubjectRe: debugctl msr
    Markus,

    On Fri, Nov 21, 2008 at 9:22 AM, Metzger, Markus T
    <markus.t.metzger@intel.com> wrote:
    >>-----Original Message-----
    >>From: stephane eranian [mailto:eranian@googlemail.com]
    >>Sent: Donnerstag, 20. November 2008 22:19
    >>To: Markus Metzger
    >
    >>- in ds_put_context(), you need to mark task->thread.ds_ctx = NULL or
    >>this_system_context = NULL when you are
    >> done, otherwise a subsequent session on the same task or CPU will
    >>think there is already a context allocated but
    >> the pointer will be stale.
    >
    > This should happen in:
    > *(context->this) = NULL;
    >
    Ah, I missed that. I admit this code is hard to follow. Maybe it would easier to
    explicitly use the actual fields.
    >
    >>- in ptrace.c:ptrace_disable(), you systematically invoke ds_release()
    >>without checking if TIF_BTS_TRACE_TS
    >> is set. That causes extraneous calls to ds_release() which messes up
    >>the reference counting if PEBS is in use.
    >
    > That TIF only triggers the recording of scheduling timestamps.
    >
    > That ds_release_bts() call should:
    > - get the context
    > - fail the ownership validation in ds_validate()
    > - put the context
    >
    > If ds_validate() is disabled, it would put the context twice.
    >
    Yes, that was my case.


    >
    > I am currently working on a patch to replace the validation
    > mechanism by using a handle returned from ds_request().
    >
    Yes, that's the solution.

    > I also added the interrupt threshold to the parameters of ds_request().
    > It will be fix for one tracing session, if that's OK with you.
    >
    Yes.

    I found the reason why in system-wide my kernel was crashing.
    It was due to the way you are writing DS_AREA_MSR in
    ds_allocate_context():

    if (!task || (task == current))
    wrmsr(MSR_IA32_DS_AREA, (unsigned long)context->ds, 0);


    This does not work in 64-bit, you truncate the DS address. You must use wrmsrl()
    and let the macro break it down in two halves.

    I am still seeing an issue whereby I have a buffer with > 100 entries
    but I get an interrupt
    or each sample, which is wrong. I am investigating why this is. There
    should only be one
    interrupt per buffer full. This use to work with my old ways of handling PEBS.

    Thanks.


    \
     
     \ /
      Last update: 2008-11-21 09:51    [W:0.026 / U:0.056 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site