From mboxrd@z Thu Jan 1 00:00:00 1970 From: Lorenzo Pieralisi Subject: Re: [PATCH RESEND] PM / sleep: Fix racing timers Date: Mon, 12 Jan 2015 16:14:53 +0000 Message-ID: <20150112161453.GA25281@red-moon> References: <1411405623-7869-1-git-send-email-soren.brinkmann@xilinx.com> <1025ebd0a9984037bf15adf3ae4889d9@BN1BFFO11FD013.protection.gbl> <99d5bcca1d6541a49ff8a9f41f9b18af@BL2FFO11FD039.protection.gbl> <12360948.gSvRbI0dui@vostro.rjw.lan> <20150112154347.GA23297@red-moon> <2313f65696c04c48867a3e8ce65ceb0b@BL2FFO11FD015.protection.gbl> Mime-Version: 1.0 Content-Type: text/plain; charset=iso-8859-1 Content-Transfer-Encoding: QUOTED-PRINTABLE Return-path: Content-Disposition: inline In-Reply-To: <2313f65696c04c48867a3e8ce65ceb0b@BL2FFO11FD015.protection.gbl> Sender: linux-kernel-owner@vger.kernel.org To: =?iso-8859-1?Q?S=F6ren?= Brinkmann Cc: "Rafael J. Wysocki" , Thomas Gleixner , John Stultz , Pavel Machek , Len Brown , "linux-kernel@vger.kernel.org" , "linux-pm@vger.kernel.org" List-Id: linux-pm@vger.kernel.org On Mon, Jan 12, 2015 at 03:55:05PM +0000, S=F6ren Brinkmann wrote: > On Mon, 2015-01-12 at 03:43PM +0000, Lorenzo Pieralisi wrote: > > Hi Rafael, Soren, > > > > On Sun, Jan 11, 2015 at 11:20:36PM +0000, Rafael J. Wysocki wrote: > > > On Friday, January 09, 2015 01:50:59 PM S=F6ren Brinkmann wrote: > > > > On Sat, 2014-11-08 at 03:56PM -0800, S=F6ren Brinkmann wrote: > > > > > Hi Rafael, > > > > > > > > > > On Thu, 2014-11-06 at 01:33AM +0100, Rafael J. Wysocki wrote: > > > > > > On Thursday, October 02, 2014 09:01:15 AM S=F6ren Brinkmann= wrote: > > > > > > > Hi Rafael, > > > > > > > > > > > > Hi, > > > > > > > > > > > > Sorry for the huge delay. > > > > > > > > > > > > > On Tue, 2014-09-23 at 01:01AM +0200, Rafael J. Wysocki wr= ote: > > > > > > > > On Monday, September 22, 2014 10:07:03 AM Soren Brinkma= nn wrote: > > > > > > > > > On platforms that do not power off during suspend, su= ccessfully entering > > > > > > > > > suspend races with timers. > > > > > > > > > > > > > > > > > > The race happening in a couple of location is: > > > > > > > > > > > > > > > > > > 1. disable IRQs (e.g. arch_suspend_di= sable_irqs()) > > > > > > > > > ... > > > > > > > > > 2. syscore_suspend() > > > > > > > > > -> timekeeping_suspend() > > > > > > > > > -> clockevents_notify(SUSPEND) > > > > > > > > > -> tick_suspend() (timers are turned of= f here) > > > > > > > > > ... > > > > > > > > > 3. wfi (wait for wake-IRQ he= re) > > > > > > > > > > > > > > > > > > Between steps 1 and 2 the timers can still generate i= nterrupts that are > > > > > > > > > not handled and stay pending until step 3. That pendi= ng IRQ causes an > > > > > > > > > immediate - spurious - wake. > > > > > > > > > > > > > > > > > > The solution is to move the clockevents suspend/resum= e notification > > > > > > > > > out of the syscore_suspend step and explictly call th= em at the appropriate > > > > > > > > > time in the suspend/hibernation paths. I.e. timers ar= e suspend _before_ > > > > > > > > > IRQs get disabled. And accordingly in the resume path= =2E > > > > > > > > > > > > > > > > > > Signed-off-by: Soren Brinkmann > > > > > > > > > --- > > > > > > > > > Hi, > > > > > > > > > > > > > > > > > > there was not a lot of discussion on the last submiss= ion. Just one comment from > > > > > > > > > Rafael (https://lkml.org/lkml/2014/8/26/780), which -= as I outlined in my > > > > > > > > > response, does not apply, IMHO, since the platform do= es not re-enable > > > > > > > > > interrupts. > > > > > > > > > > > > > > > > Well, you just don't agree with it. > > > > > > > > > > > > > > > > The problem with your approach is that timer interrupts= aren't actually as > > > > > > > > special as you think and any other IRQF_NO_SUSPEND inte= rrupts would have caused > > > > > > > > similar issues to appear under specific conditions. > > > > > > > > > > > > > > > > The solution I would suggest and that actually covers a= ll IRQF_NO_SUSPEND > > > > > > > > interrupts would be to use a wait_event() loop like the= one in freeze_enter() > > > > > > > > (on top of the current linux-next or the pm-genirq bran= ch of linux-pm.git), > > > > > > > > but wait for pm_abort_suspend to become true, to implem= ent system suspend. > > > > > > > > > > > > > > sorry, it took me a while since I needed to get some depe= ndencies ported > > > > > > > to the pm-genirq base. Once I had that, it reproduced my = original issue. > > > > > > > So far so good. I then looked into finding a solution fol= lowing your > > > > > > > guidance. I'm not sure I really found what you had in min= d, but below is > > > > > > > what I came up with, which seems to do it. > > > > > > > Please let me know how far off I am. > > > > > > > > > > > > > > Thanks, > > > > > > > S=F6ren > > > > > > > > > > > > > > -------8<------------------8<----------------8<----------= ------8<--------------- > > > > > > > diff --git a/drivers/base/power/wakeup.c b/drivers/base/p= ower/wakeup.c > > > > > > > index c2744b30d5d9..a4f9914571f1 100644 > > > > > > > --- a/drivers/base/power/wakeup.c > > > > > > > +++ b/drivers/base/power/wakeup.c > > > > > > > @@ -25,7 +25,7 @@ > > > > > > > bool events_check_enabled __read_mostly; > > > > > > > > > > > > > > /* If set and the system is suspending, terminate the su= spend. */ > > > > > > > -static bool pm_abort_suspend __read_mostly; > > > > > > > +bool pm_abort_suspend __read_mostly; > > > > > > > > > > > > > > /* > > > > > > > * Combined counters of registered wakeup events and wak= eup events in progress. > > > > > > > diff --git a/kernel/power/suspend.c b/kernel/power/suspen= d.c > > > > > > > index 6dadb25cb0d8..e6a6de8f76d0 100644 > > > > > > > --- a/kernel/power/suspend.c > > > > > > > +++ b/kernel/power/suspend.c > > > > > > > @@ -33,6 +33,7 @@ > > > > > > > > > > > > > > static const char *pm_labels[] =3D { "mem", "standby", "= freeze", }; > > > > > > > const char *pm_states[PM_SUSPEND_MAX]; > > > > > > > +extern bool pm_abort_suspend; > > > > > > > > > > > > > > static const struct platform_suspend_ops *suspend_ops; > > > > > > > static const struct platform_freeze_ops *freeze_ops; > > > > > > > @@ -294,25 +295,27 @@ static int suspend_enter(suspend_st= ate_t state, bool *wakeup) > > > > > > > if (error || suspend_test(TEST_CPUS)) > > > > > > > goto Enable_cpus; > > > > > > > > > > > > > > - arch_suspend_disable_irqs(); > > > > > > > - BUG_ON(!irqs_disabled()); > > > > > > > - > > > > > > > - error =3D syscore_suspend(); > > > > > > > - if (!error) { > > > > > > > - *wakeup =3D pm_wakeup_pending(); > > > > > > > - if (!(suspend_test(TEST_CORE) || *wakeup)) { > > > > > > > - trace_suspend_resume(TPS("machine_sus= pend"), > > > > > > > - state, true); > > > > > > > - error =3D suspend_ops->enter(state); > > > > > > > - trace_suspend_resume(TPS("machine_sus= pend"), > > > > > > > - state, false); > > > > > > > - events_check_enabled =3D false; > > > > > > > + while (!pm_abort_suspend) { > > > > > > > > > > > > That won't work in general, because pm_abort_suspend may no= t be set on some > > > > > > platforms on wakeup. It is only set if a wakeup interrupt = triggers which > > > > > > may not be the case on ACPI systems if the BIOS has woken u= p the system. > > > > > > > > > > > > But that could be addressed by making those platforms simpl= y set pm_wakeup_pending > > > > > > in their BIOS exit path. > > > > > > > > > > > > > + arch_suspend_disable_irqs(); > > > > > > > + BUG_ON(!irqs_disabled()); > > > > > > > + > > > > > > > + error =3D syscore_suspend(); > > > > > > > > > > > > Also it shouldn't be necessary to do syscore_suspend()/sysc= ore_resume() in > > > > > > every iteration of the loop. > > > > > > > > > > > > > + if (!error) { > > > > > > > + *wakeup =3D pm_wakeup_pending(); > > > > > > > > > > > > Plus pm_wakeup_pending() returns true if pm_abort_suspend i= s set > > > > > > > > > > > > > + if (!(suspend_test(TEST_CORE) || *wak= eup)) { > > > > > > > + trace_suspend_resume(TPS("mac= hine_suspend"), > > > > > > > + state, true); > > > > > > > > > > > > Did you try to add the loop here instead of above? Like: > > > > > > > > > > > > for (;;) { > > > > > > *wakeup =3D pm_wakeup_pending= (); > > > > > > if (*wakeup) > > > > > > break; > > > > > > > > > > I think, that doesn't work. I chose the start/end points of t= he loop > > > > > to include the IRQ enable/disable calls. AFAICT, pm_abort_sus= pend is > > > > > set in an ISR. Without enabling interrupts the abort conditio= n of > > > > > this loop never becomes true. > > > > > > > > Any further ideas how to resolve this? > > > > > > Sorry about the delay, lost track of this. > > > > > > You're right, the IRQ disabling/enabling needs to happen in the l= oop. > > > > > > So the direction of the patch looks OK, but it needs to ensure th= at pm_wakeup_pending > > > is set properly by all platforms. Also it should be sufficient t= o check > > > pm_wakeup_pending() to detect wakeup. > > > > > > Have you tested the patch? > > > > Before considering this patch a solution, can I ask you to rewind > > the discussion a bit since I have a question. > > > > I thought that "suspending" the tick through syscore meant shutting= down > > the respective clock_event_device, and that's how it is implemented= in the > > kernel. > > > > Now, do we expect a shutdown clock_event_device to still signal pen= ding > > IRQs ? I do not think that should be the case, at least that's not = what > > happens for eg arm arch timers - ie disabling them implicitly gates > > the signal connected to the IRQ line. > > > > So the question is more related to the zynq platform and how their = clock > > event device (which is ?) is shutdown, and what's the correct behav= iour we > > are expecting. >=20 > As outlined in the commit message, there is a race condition in the c= ore > code. Looking at the timers is just fighting the symptoms. I gathered there is a race condition between 1 and 2 in your code path. What I am asking you is why are we getting a pending IRQ at step 3 when= the clock event device is supposed to be shutdown. My question is: Should a clock event device in shutdown mode (ie disabled) still signal IRQs to the interrupt controller (and consequently to the core) ? It is for me to understand if that's the behaviour we are expecting. > > FWIW, the problem here is not related to the simple wfi state on zy= nq, > > even some other ARM platforms with power management capabilities wo= uld wake > > up from the state entered by executing wfi (ie possibly through res= et) if > > there is a pending timer IRQ, the question is more "should the IRQ = be > > allowed to be there" instead IMHO. > > > > I still think that Stephen's query related to what timer is causing > > the wake-up is worth investigating. >=20 > As I reported earlier, I see these spurious wakes with the cadence_tt= c > as well as the ARM twd timers. I thought that a shutdown clock event device explicitly disables IRQ assertion, that's why I am inquiring, I do not understand how this can happen - how can you have a pending timer IRQ at step 3 above. Thanks, Lorenzo