Re: [PATCH 0/6] perf trace: Add filter support

[Frederic Weisbecker <fweisbec@gmail.com>]

On Mon, Sep 07, 2009 at 04:11:57PM +0800, Li Zefan wrote:
> This patchset adds filter support for perf counter, so not all
> profile events are recorded but only those match the filters
> we set.


Thanks a lot for this work!



> An example:
> 
>  #./perf record -f -e irq:irq_handler_entry:irq==18:record
>  or
>  #./perf record -f -e irq:irq_handler_entry:irq==18 -R
>  ^C



We may want to write complex filters. While looking at this patchset,
that seems possible this way, but the raw line may become unreadable:

perf record -f -e irq:irq_handler_entry:"irq==18 && (name == 'foo' || bar == 'blah')":record

May be we should add an option to let one also set the filters seperately then
we could do:

perf record -f -e -R irq:irq_handler_entry --filter "irq==18 && (name == 'foo' || bar == 'blah')"

Hmm.


>  # ./perf trace
>  version = 0.5
>             perf-4303  ... irq_handler_entry: irq=18 handler=eth0
>             init-0     ... irq_handler_entry: irq=18 handler=eth0
>             init-0     ... irq_handler_entry: irq=18 handler=eth0
>             init-0     ... irq_handler_entry: irq=18 handler=eth0
>             init-0     ... irq_handler_entry: irq=18 handler=eth0
> 
> ---
>  include/linux/ftrace_event.h       |   19 +++-
>  include/linux/perf_counter.h       |    1 +
>  include/linux/syscalls.h           |   14 ++-
>  include/trace/ftrace.h             |   10 +-
>  kernel/perf_counter.c              |   40 ++++++-
>  kernel/trace/trace.h               |    9 +-
>  kernel/trace/trace_event_profile.c |   18 +++
>  kernel/trace/trace_events_filter.c |  247 +++++++++++++++++++++---------------
>  kernel/trace/trace_syscalls.c      |    9 +-
>  tools/perf/builtin-record.c        |   12 ++
>  tools/perf/util/parse-events.c     |   48 ++++++-
>  tools/perf/util/parse-events.h     |    1 +
>  12 files changed, 303 insertions(+), 125 deletions(-)
>