From mboxrd@z Thu Jan 1 00:00:00 1970 From: "Rafael J. Wysocki" Subject: Re: linux-next: build failure after merge of the tip tree Date: Tue, 14 Jan 2014 23:18:16 +0100 Message-ID: <2433882.lLbOWJV5fl@vostro.rjw.lan> References: <20140114142627.537a702607e703c6eff63640@canb.auug.org.au> <1742600.ID61iX8uKd@vostro.rjw.lan> Mime-Version: 1.0 Content-Type: text/plain; charset="utf-8" Content-Transfer-Encoding: 7Bit Return-path: In-Reply-To: Sender: cpufreq-owner@vger.kernel.org To: Mikulas Patocka Cc: Peter Zijlstra , Stephen Rothwell , Thomas Gleixner , Ingo Molnar , "H. Peter Anvin" , Viresh Kumar , linux-next@vger.kernel.org, linux-kernel@vger.kernel.org, cpufreq@vger.kernel.org List-Id: linux-next.vger.kernel.org On Tuesday, January 14, 2014 04:52:20 PM Mikulas Patocka wrote: > > On Tue, 14 Jan 2014, Rafael J. Wysocki wrote: > > > On Tuesday, January 14, 2014 04:43:43 PM Mikulas Patocka wrote: > > > > > > On Tue, 14 Jan 2014, Peter Zijlstra wrote: > > > > > > > On Tue, Jan 14, 2014 at 02:06:57PM -0500, Mikulas Patocka wrote: > > > > > On Tue, 14 Jan 2014, Peter Zijlstra wrote: > > > > > > > Caused by commit 62b94a08da1b ("sched/preempt: Take away > > > > > > > preempt_enable_no_resched() from modules") > > > > > > > > Read these two lines, then note that: > > > > > > > > > Try adding #include to speedstep-lib.c. Does it help? > > > > > > > > this obviously will not work as preempt_check_resched() and > > > > preempt_enable_no_resched() are no longer available to modules. > > > > > > I see, you added commit 62b94a08da1bae9d187d49dfcd6665af393750f8 to > > > linux-next, that broke my patch. > > > > > > > > > I think that pm commit is a very good example of why the sched/preempt > > > > > > patch is a very good idea. > > > > > > > > > > > > Also that Changelog fails to explain why enabling interrupts helps. What > > > > > > interrupt is required for progress, and how does it make the progress > > > > > > happen. > > > > > > > > > > There is no explanation. It's hardware issue and I have no documentation > > > > > for the hardware. > > > > > > > > Rafael works for Intel, he ought to be able to figure out wtf the > > > > hardware does/needs. > > > > > > > > > The general problem is that if there are bus-master transfers running (or > > > > > possibly for other hardware reasons), the CPU refuses to change frequency. > > > > > You can wait a little bit and retry and maybe you succeed changing the > > > > > frequency next time. > > > > > > > > > > If you enable interrupts, wait, disable interrupts and retry, you may > > > > > succeed. If you keep interrupts disabled and retry, you never succeed, no > > > > > matter how long do you wait. I found it experimentally, I don't know > > > > > reason for that. > > > > > > > > Sounds like magic goo.. > > > > > > > > In any case, try the below, it does the same but is far less horrid. > > > > > > > > --- > > > > drivers/cpufreq/speedstep-smi.c | 4 ++++ > > > > 1 file changed, 4 insertions(+) > > > > > > > > diff --git a/drivers/cpufreq/speedstep-smi.c b/drivers/cpufreq/speedstep-smi.c > > > > index 0f5326d6f79f..57d31538c248 100644 > > > > --- a/drivers/cpufreq/speedstep-smi.c > > > > +++ b/drivers/cpufreq/speedstep-smi.c > > > > @@ -188,6 +188,7 @@ static void speedstep_set_state(unsigned int state) > > > > return; > > > > > > > > /* Disable IRQs */ > > > > + preempt_disable(); > > > > local_irq_save(flags); > > > > > > > > command = (smi_sig & 0xffffff00) | (smi_cmd & 0xff); > > > > @@ -200,7 +201,9 @@ static void speedstep_set_state(unsigned int state) > > > > if (retry) { > > > > pr_debug("retry %u, previous result %u, waiting...\n", > > > > retry, result); > > > > + local_irq_restore(flags); > > > > > > ^^^ this is wrong, because the function speedstep_set_state may already be > > > called with interrupts disabled from speedstep_get_freqs. So, you need to > > > enable interrupts unconditionally, even if they were disabled at the > > > beginning of the function speedstep_set_state. > > > > > > I know it's dirty to enable interrupts in a function that was called with > > > disabled interrupts, but here it must be so (you could rewrite > > > speedstep_get_freqs to not disable interrupts if you want to avoid this > > > dirtiness). > > > > > > > mdelay(retry * 50); > > > > + local_irq_save(flags); > > > > } > > > > retry++; > > > > __asm__ __volatile__( > > > > @@ -217,6 +220,7 @@ static void speedstep_set_state(unsigned int state) > > > > > > > > /* enable IRQs */ > > > > local_irq_restore(flags); > > > > + preempt_enable(); > > > > > > > > if (new_state == state) > > > > pr_debug("change to %u MHz succeeded after %u tries " > > > > > > You need also preempt_disable/enable in speedstep_get_freqs. > > > > > > > > > Here I'm resending the patch, to account for > > > 62b94a08da1bae9d187d49dfcd6665af393750f8. > > > > Do I think correctly that this should work regardless of whether or not > > 62b94a08da1bae9d187d49dfcd6665af393750f8 is applied? > > Yes. OK I'll replace your original patch with this version, then. Thanks! -- I speak only for myself. Rafael J. Wysocki, Intel Open Source Technology Center.