From mboxrd@z Thu Jan 1 00:00:00 1970 From: "Paul E. McKenney" Subject: Re: [RFC PATCH] Fix abnormal rcu dynticks_nesting values related to async page fault Date: Tue, 27 Nov 2012 07:25:50 -0800 Message-ID: <20121127152550.GV2474@linux.vnet.ibm.com> References: <1353993325.14050.49.camel@ThinkPad-T5421.cn.ibm.com> <20121127130703.GW25516@redhat.com> <50B4C7BD.90807@oracle.com> Reply-To: paulmck@linux.vnet.ibm.com Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Return-path: Received: from e36.co.us.ibm.com ([32.97.110.154]:41277 "EHLO e36.co.us.ibm.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1755373Ab2K0P0I (ORCPT ); Tue, 27 Nov 2012 10:26:08 -0500 Received: from /spool/local by e36.co.us.ibm.com with IBM ESMTP SMTP Gateway: Authorized Use Only! Violators will be prosecuted for from ; Tue, 27 Nov 2012 08:26:07 -0700 Content-Disposition: inline In-Reply-To: <50B4C7BD.90807@oracle.com> Sender: linux-next-owner@vger.kernel.org List-ID: To: Sasha Levin Cc: Gleb Natapov , Li Zhong , linux-next list , LKML , avi@redhat.com, fweisbec@gmail.com On Tue, Nov 27, 2012 at 09:01:33AM -0500, Sasha Levin wrote: > On 11/27/2012 08:07 AM, Gleb Natapov wrote: > > Those rcu_irq_enter()/rcu_irq_exit() were introduced by commit > > c5e015d4949aa665 "KVM guest: exit idleness when handling > > KVM_PV_REASON_PAGE_NOT_PRESENT", but now I am starting to question this > > commit. KVM_PV_REASON_PAGE_NOT_PRESENT should not kick cpu out of > > idleness. kvm_async_pf_task_wait() checks that cpu is idle and calls > > halt if it is. After that commit schedule() may be called between > > rcu_irq_enter()/rcu_irq_exit() which is probably illegal. Paul? It is legal to call rcu_irq_enter() and then schedule(). In fact, it turns out that it -has- to be legal, due to some architectures' quaint habit of entering interrupt/exception handlers that they never leave, and possibly vice versa. > otoh, calling schedule() apparently kicks cpu out of idleness now. But if you call rcu_irq_enter() and then schedule(), and if schedule() switches to the idle thread, and if execution proceeds to the point where rcu_idle_enter() is called, then, RCU will quite naturally decide that it is fully idle. At that point, it is illegal to invoke rcu_irq_exit() unless/until you have either: (1) exited the idle loop (as in called rcu_idle_exit()) or (2) taken an interrupt, which will call rcu_irq_enter(). And to think that when I started coding RCU's dyntick-idle funtionality, I was thinking in terms of a simple nesting counter. Silly me! ;-) Thanx, Paul