Return-Path: Message-ID: <5B725A12.8050409@rock-chips.com> Date: Tue, 14 Aug 2018 12:26:58 +0800 From: JeffyChen MIME-Version: 1.0 To: Brian Norris , Andrea Parri CC: Marcel Holtmann , Johan Hedberg , "David S. Miller" , linux-bluetooth@vger.kernel.org, netdev@vger.kernel.org, linux-kernel@vger.kernel.org, Brian Norris , AL Yu-Chen Cho Subject: Re: [Question] bluetooth/{bnep,cmtp,hidp}: memory barriers References: <20180730031030.GA9430@andrea> <20180813231854.GA173912@ban.mtv.corp.google.com> In-Reply-To: <20180813231854.GA173912@ban.mtv.corp.google.com> Content-Type: text/plain; charset=UTF-8; format=flowed List-ID: Hi guys, Thanks for your mails, and sorry for the late response.. On 08/14/2018 07:18 AM, Brian Norris wrote: > > commit 5da8e47d849d3d37b14129f038782a095b9ad049 > Author: Jeffy Chen > Date: Tue Jun 27 17:34:44 2017 +0800 > > Bluetooth: hidp: fix possible might sleep error in hidp_session_thread > > that*some* kind of barrier was stuck in there simply as a response to > comments like this, that were going away: > > - * > - * Note: set_current_state() performs any necessary > - * memory-barriers for us. > */ > - set_current_state(TASK_INTERRUPTIBLE); > > + /* Ensure session->terminate is updated */ > + smp_mb__before_atomic(); > > > It was probably an attempt to fill in the gap for the > set_current_state() (and comment) which was being removed. I believe > Jeffy originally added more barriers in other places, but I convinced > him not to. right, i was trying to avoid losing memory-barriers when removing set_current_state and changing wake_up_process to wake_up_interruptible. and checking these code again, it's true the smp_mb__before_atomic before atomic_read is not needed, the smp_mb after atomic_inc(&session->terminate) should be enough. and as Brian point out, there's already an smp_store_mb at the end of wait_woken, i agree we can remove all the smp_mb__{before,after}_atomic() i wrongly added :) > > I have to say, I'm not really up-to-speed on the use of manual barriers > in Linux (it's much preferable when they're wrapped into higher-level > data structures already), but I believe the main intention here is to > ensure that any change to 'terminate' that happened during the previous > "wait_woken()" would be visible to our atomic_read(). > > Looking into wait_woken(), I'm feeling like none of these additional > barriers are necessary at all. I believe wait_woken() handles the > visibility issues we care about (that if we were woken for termination, > we'll see the terminating condition).