From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1753485AbcIBMGl (ORCPT ); Fri, 2 Sep 2016 08:06:41 -0400 Received: from mx1.redhat.com ([209.132.183.28]:58834 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752073AbcIBMGk (ORCPT ); Fri, 2 Sep 2016 08:06:40 -0400 Date: Fri, 2 Sep 2016 14:06:02 +0200 From: Oleg Nesterov To: Peter Zijlstra Cc: Ingo Molnar , Al Viro , Bart Van Assche , Johannes Weiner , Linus Torvalds , Neil Brown , linux-kernel@vger.kernel.org Subject: Re: [PATCH 2/2] sched/wait: avoid abort_exclusive_wait() in __wait_on_bit_lock() Message-ID: <20160902120601.GA26495@redhat.com> References: <20160826124453.GA28894@redhat.com> <20160826124552.GB28904@redhat.com> <20160901190141.GJ10138@twins.programming.kicks-ass.net> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20160901190141.GJ10138@twins.programming.kicks-ass.net> User-Agent: Mutt/1.5.18 (2008-05-17) X-Greylist: Sender IP whitelisted, not delayed by milter-greylist-4.5.16 (mx1.redhat.com [10.5.110.38]); Fri, 02 Sep 2016 12:06:39 +0000 (UTC) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On 09/01, Peter Zijlstra wrote: > > On Fri, Aug 26, 2016 at 02:45:52PM +0200, Oleg Nesterov wrote: > > > We do not need anything tricky to avoid the race, > > The race being: > > CPU0 CPU1 CPU2 > > __wait_on_bit_lock() > bit_wait_io() > io_schedule() > > clear_bit_unlock() > __wake_up_common(.nr_exclusive=1) > list_for_each_entry() > if (curr->func() && --nr_exclusive) > break > > signal() > > if (signal_pending_state()) == TRUE > return -EINTR > > And no progress because CPU1 exits without acquiring the lock and CPU0 > thinks its done because it woke someone. Yes, > > we can just call finish_wait() if action() fails. > > That would be bit_wait*() returning -EINTR because sigpending. Hmm. Not sure I understand... Let me reply just in case, even if I am sure you get it right. Yes, in the likely case we are going to fail with -EINTR, but only if test-and-set after thar fails. > Sure, you can always call that, first thing through the loop does > prepare again, so no harm. That however does not connect to your > condition,.. /me puzzled If ->action() fails we will abort the loop in any case, prepare won't be called. So in this case finish_wait() does the right thing. > > test_and_set_bit() implies mb() so > > the lockless list_empty_careful() case is fine, we can not miss the > > condition if we race with unlock_page(). > > You're talking about this ordering?: > > finish_wait() clear_bit_unlock(); > list_empty_careful() > > /* MB implied */ smp_mb__after_atomic(); > test_and_set_bit() wake_up_page() > ... > autoremove_wake_function() > list_del_init(); > > > That could do with spelling out I feel.. :-) Yes, yes. > > __wait_on_bit_lock(wait_queue_head_t *wq, struct wait_bit_queue *q, > > wait_bit_action_f *action, unsigned mode) > > { > > + int ret = 0; > > > > + for (;;) { > > prepare_to_wait_exclusive(wq, &q->wait, mode); > > + if (test_bit(q->key.bit_nr, q->key.flags)) { > > + ret = action(&q->key, mode); > > + /* > > + * Ensure that clear_bit() + wake_up() right after > > + * test_and_set_bit() below can't see us; it should > > + * wake up another exclusive waiter if we fail. > > + */ > > + if (ret) > > + finish_wait(wq, &q->wait); > > + } > > + if (!test_and_set_bit(q->key.bit_nr, q->key.flags)) { > > So this is the actual difference, instead of failing the lock and > aborting on signal, we acquire the lock if possible. If its not > possible, someone else has it, which guarantees that someone else will > do an unlock which implies another wakeup and life goes on. Yes. This way we eliminate the need for the additional wake_up. Oleg.