From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1754962AbZIJCG2 (ORCPT ); Wed, 9 Sep 2009 22:06:28 -0400 Received: (majordomo@vger.kernel.org) by vger.kernel.org id S1754919AbZIJCG2 (ORCPT ); Wed, 9 Sep 2009 22:06:28 -0400 Received: from mail-ew0-f217.google.com ([209.85.219.217]:34129 "EHLO mail-ew0-f217.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1754881AbZIJCG1 (ORCPT ); Wed, 9 Sep 2009 22:06:27 -0400 DomainKey-Signature: a=rsa-sha1; c=nofws; d=gmail.com; s=gamma; h=date:from:to:cc:subject:message-id:references:mime-version :content-type:content-disposition:in-reply-to:user-agent; b=j4fCLvGdXjYEye8GjcfkNUijs7KN7bJ9gdhcS9UoXDutBx3dbrz1CL4Z6P8Yb4gdsL t+wBG5ZmFxUUCJ3X57SI1JU6jug03Eq0Sa3I1mhzkWY7kHA2n52JecH27Su1m2ukKFQT yMVorFf137ovO5X9t2HdM2/7iAfLM55wPjYtE= Date: Thu, 10 Sep 2009 04:06:26 +0200 From: Frederic Weisbecker To: Steven Rostedt Cc: Lai Jiangshan , Ingo Molnar , LKML Subject: Re: [PATCH 2/3] tracing: block-able ring_buffer consumer Message-ID: <20090910020624.GD8450@nowhere> References: <4A95F768.7070701@cn.fujitsu.com> <20090829102100.GB6418@nowhere> <1252530602.27001.71.camel@gandalf.stny.rr.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <1252530602.27001.71.camel@gandalf.stny.rr.com> User-Agent: Mutt/1.5.18 (2008-05-17) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Wed, Sep 09, 2009 at 05:10:02PM -0400, Steven Rostedt wrote: > On Sat, 2009-08-29 at 12:21 +0200, Frederic Weisbecker wrote: > > On Thu, Aug 27, 2009 at 11:03:04AM +0800, Lai Jiangshan wrote: > > > > > > makes consumer side(per_cpu/cpu#/trace_pipe_raw) block-able, > > > which is a TODO in trace.c > > > > > > Signed-off-by: Lai Jiangshan > > > --- > > > > > > int ring_buffer_print_entry_header(struct trace_seq *s); > > > diff --git a/kernel/timer.c b/kernel/timer.c > > > index 6e712df..79f5596 100644 > > > --- a/kernel/timer.c > > > +++ b/kernel/timer.c > > > @@ -39,6 +39,7 @@ > > > #include > > > #include > > > #include > > > +#include > > > > > > #include > > > #include > > > @@ -1178,6 +1179,7 @@ void update_process_times(int user_tick) > > > printk_tick(); > > > scheduler_tick(); > > > run_posix_cpu_timers(p); > > > + tracing_notify(); > > > > > > > > Hmm, that looks really not a good idea. The tracing shouldn't ever impact > > the system when it is inactive. > > Especially in such a fast path like the timer interrupt. > > > > Perhaps we should put a trace point there instead. Then we could add a > probe to it (doesn't need to be an event). > > trace_update_process_times() ? Yeah that would do the trick although I still doubt about the need to do this check at every tick. > > > > > > > > } > > > > > > /* > > > diff --git a/kernel/trace/ring_buffer.c b/kernel/trace/ring_buffer.c > > > index f1e1533..db82b38 100644 > > > --- a/kernel/trace/ring_buffer.c > > > +++ b/kernel/trace/ring_buffer.c > > > @@ -443,6 +443,7 @@ struct ring_buffer_per_cpu { > > > u64 write_stamp; > > > u64 read_stamp; > > > atomic_t record_disabled; > > > + wait_queue_head_t sleepers; > > > > > > That seems a too generic name. May be consumer_queue? > > "waiters" is what is usually used. Yeah. > > > > > > > > }; > > > > > > struct ring_buffer { > > > @@ -999,6 +999,7 @@ rb_allocate_cpu_buffer(struct ring_buffer *buffer, int cpu) > > > spin_lock_init(&cpu_buffer->reader_lock); > > > lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key); > > > cpu_buffer->lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED; > > > + init_waitqueue_head(&cpu_buffer->sleepers); > > > > > > bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()), > > > GFP_KERNEL, cpu_to_node(cpu)); > > > @@ -3318,6 +3319,77 @@ ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts) > > > EXPORT_SYMBOL_GPL(ring_buffer_read); > > > > > > /** > > > + * ring_buffer_notify - notify the sleepers when there is any available page > > > + * @buffer: The ring buffer. > > > + */ > > > +void ring_buffer_notify(struct ring_buffer *buffer) > > > +{ > > > + unsigned long flags; > > > + struct ring_buffer_per_cpu *cpu_buffer; > > > + > > > + cpu_buffer = buffer->buffers[smp_processor_id()]; > > > + > > > + if (!spin_trylock_irqsave(&cpu_buffer->reader_lock, flags)) > > > + return; > > > + > > > + if (waitqueue_active(&cpu_buffer->sleepers)) { > > > + struct buffer_page *reader_page; > > > + struct buffer_page *commit_page; > > > + > > > + reader_page = cpu_buffer->reader_page; > > > + commit_page = ACCESS_ONCE(cpu_buffer->commit_page); > > > > > > ACCESS_ONCE makes sense if you loop, to ensure the value > > is not cached through iteration, but there I'm not sure this is > > useful. > > ACCESS_ONCE is fine. Otherwise we may read it again on the check below. > Of course the worse that will happen is we don't wake up on this tick. Yeah, I haven't seen the fact we may check more than once there. > > > > > > > > > + > > > + /* > > > + * ring_buffer_notify() is fast path, so we don't use the slow > > > + * rb_get_reader_page(cpu_buffer, 1) to detect available pages. > > > + */ > > > + if (reader_page == commit_page) > > > + goto out; > > > + > > > + if (reader_page->read < rb_page_commit(reader_page) > > > + || rb_set_head_page(cpu_buffer) != commit_page) > > > > > > > > This may need a small comment to explain you are checking that the reader > > is not completely consumed. > > Heh, it was obvious for me ;-) For you, of course ;-) > > -- Steve > >