lkml.org 
[lkml]   [2010]   [Aug]   [6]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
    Patch in this message
    /
    From
    Subject[RFC PATCH 00/9] RAS daemon prototype, v2
    Date
    From: Borislav Petkov <borislav.petkov@amd.com>

    Hi,

    I finally found some time for a second take at the ras daemon.
    This one is based on Steven's trace-cmd integration into perf:

    git://git.kernel.org/pub/scm/linux/kernel/git/rostedt/linux-2.6-trace.git, branch tip/perf/parse-events

    which is still alpha, as are those patches too, btw. I'm sending them
    now as a RFC/FYI type-thing only, work on them continues.

    I've tried to incorporate and accomodate all comments from the last
    round. Using debugfs files for mmaping the perf buffers makes the code
    much simpler and introduces a lot less changes to the perf tool wrt to
    exporting stuff into the library for external use.

    The are still unresolved issues like the VM_SHARED check in perf_mmap()
    which fails for read-only files in debugfs. Peter, what is your take at
    this, do we want to relax that for persistent read-only events?

    Also, I need this hunk otherwise I'm oopsing on exit:

    ---
    diff --git a/kernel/perf_event.c b/kernel/perf_event.c
    index 8ff5292..8edb400 100644
    --- a/kernel/perf_event.c
    +++ b/kernel/perf_event.c
    @@ -2665,7 +2665,9 @@ static void perf_mmap_close(struct vm_area_struct *vma)
    struct user_struct *user = event->mmap_user;
    struct perf_buffer *buffer = event->buffer;

    - atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
    + if (user)
    + atomic_long_sub((size >> PAGE_SHIFT) + 1, &user->locked_vm);
    +
    vma->vm_mm->locked_vm -= event->mmap_locked;
    rcu_assign_pointer(event->buffer, NULL);
    mutex_unlock(&event->mmap_mutex);
    --
    and I'll add it as an additional patch to the series if there are no
    objections. See, we add the current user when we mmap but on already
    allocated event buffers we skip that by exiting early so this change
    should be fine for events with preallocated buffers, right?

    @Arnaldo: I think this version should take care of the build issues you
    had last time.

    To the individual patches:

    #1-2: added for completeness here, already upstream
    #3: export perf funcs for mce
    #4: initialize persistence mce event and allocate buffers
    #5: move Steven's trace-cmd stuff to tools/lib/trace
    #6-8: needed by the ras daemon and others maybe
    #9: rough ras daemon skeleton, will flesh out later

    Thanks for all your comments last time.

    --

    Changelog:

    V1:

    here's the first rough version of all the jerky code that attempts to
    implement a RAS daemon listening for MCEs using perf. This is a preview
    code only, I still need to figure out how to do the sample parsing
    the easiest and flesh out the daemon functionality a bit more. Also,
    I wanted to reuse as much code as possible therefore a lot had to be
    reengineered with the perf tool and all its library-like compilation
    units.

    With this, you can do

    make perf

    or

    make ras

    and build the respective tool.

    Even though I tried to split the patchset for easier review, please bear
    in mind that there are some fat guys there (241K is the biggest one).
    However, they don't do anything special except moving code around. As
    such, they might not appear on lkml due to vger size constraints so I've
    upped the whole patchset also at

    git://git.kernel.org/pub/scm/linux/kernel/git/bp/bp.git perf-v1

    The patchset is based on tip/perf/core from last week. Here are some
    more details to some of the patches individually:

    2: enables the mce tracepoint unconditionally. I had a problem with
    perf_event_attr.sample_period which is checked in perf_swevent_add().
    Currently, I'm setting it to ULLONG_MAX but this is icky. I'd much
    rather have the check do something like

    if (event->type != PERF_EVENT_TYPE_PERSISTENT)
    if (!hwc->sample_period)
    return;
    or similar.

    4: sys_perf_event_open needs to know about already allocated and
    enabled events.

    5-10: carves out a bunch of generic perf compilation units into a common
    lib. Split into 5 patches for easier review.

    12-14: those are pulled in when exporting parse_events.c for external use.

    16: lib/perf/misc.c contains functions and global variables like
    pager_in_use() or perf_guest or usage_with_options() which are used in
    generic utilities but are strictly perf-specific. Long term we should
    strive in making the library self-contained and getting rid of those.

    19-20: those are only needed for testing, they'll go in over the edac
    tree in the next merge window. Added for completeness here.

    21: only bare-bones implementation, needs a lot of fleshing out

    Anyways, please have a look and let me know either way :)

    Thanks,
    Boris.


    \
     
     \ /
      Last update: 2010-08-06 16:29    [W:0.049 / U:1.948 seconds]
    ©2003-2016 Jasper Spaans. hosted at Digital OceanAdvertise on this site