Re: S3 resume regression [1cf4f629d9d2 ("cpu/hotplug: Move online calls to hotplugged cpu")]

From: Ville Syrjälä
Date: Wed Jul 13 2016 - 10:55:58 EST


On Tue, May 31, 2016 at 10:26:50AM +0300, Ville Syrjälä wrote:
> On Mon, May 30, 2016 at 10:43:51PM +0200, Rafael J. Wysocki wrote:
> > On Thu, May 26, 2016 at 8:32 PM, Ville Syrjälä
> > <ville.syrjala@xxxxxxxxxxxxxxx> wrote:
> > > On Wed, May 18, 2016 at 10:24:24AM +0300, Ville Syrjälä wrote:
> > >> On Wed, May 18, 2016 at 01:14:42AM +0200, Rafael J. Wysocki wrote:
> > >> > On 5/16/2016 9:39 PM, Ville Syrjälä wrote:
> > >> > > On Wed, May 11, 2016 at 04:34:06PM +0300, Ville Syrjälä wrote:
> > >> > >> On Wed, May 11, 2016 at 08:44:45AM -0400, Steven Rostedt wrote:
> > >> > >>> On Wed, 11 May 2016 15:21:16 +0300
> > >> > >>> Ville Syrjälä <ville.syrjala@xxxxxxxxxxxxxxx> wrote:
> > >> > >>>
> > >> > >>>> Yeah can't get anything from the machine at that point. netconsole
> > >> > >>>> didn't help either, and no serial on this machine. And IIRC I've
> > >> > >>>> tried ramoops on this thing in the past but unfortunately the memory
> > >> > >>>> got cleared on reboot.
> > >> > >>>>
> > >> > >>> Can you look at the documentation in the kernel code at
> > >> > >>>
> > >> > >>> Documentation/power/basic-pm-debugging.txt And follow the procedures
> > >> > >>> for testing suspend to RAM (although it requires mostly running the
> > >> > >>> same tests as for hibernation suspending).
> > >> > >>>
> > >> > >>> You can also use the tool s2ram for this as well.
> > >> > >>>
> > >> > >>> See Documentation/power/s2ram.txt
> > >> > >>>
> > >> > >>> Perhaps this can give us a bit more light onto the problem.
> > >> > >>>
> > >> > >>> Basically the above does partial suspend and resume, and can pinpoint
> > >> > >>> problem areas down to a more select location.
> > >> > >> All the pm_test modes work fine. The only difference between them was
> > >> > >> that 'platform' required me to manually wake up the machine (hitting a
> > >> > >> key was sufficient), whereas the others woke up without help.
> > >> > >>
> > >> > >> pm_trace gave me
> > >> > >> [ 1.306633] Magic number: 0:185:178
> > >> > >> [ 1.322880] hash matches ../drivers/base/power/main.c:1070
> > >> > >> [ 1.339270] acpi device:0e: hash matches
> > >> > >> [ 1.355414] platform: hash matches
> > >> > >>
> > >> > >> which is the TRACE_SUSPEND in __device_suspend_noirq(), so no help
> > >> > >> there.
> > >> > >>
> > >> > >> I guess I could try to sprinkle more TRACE_RESUMEs around into some
> > >> > >> early resume code. If anyone has good ideas where to put them it
> > >> > >> might speed things up a bit.
> > >> > > So I did a bunch of that and found that it gets stuck somewhere
> > >> > > around executing the _WAK method:
> > >> > > platform_resume_noirq
> > >> > > acpi_pm_finish
> > >> > > acpi_leave_sleep_state
> > >> > > acpi_hw_sleep_dispatch
> > >> > > acpi_hw_legacy_wake
> > >> > > acpi_hw_execute_sleep_method
> > >> > > acpi_evaluate_object
> > >> > > acpi_ns_evaluate
> > >> > > acpi_ps_execute_method
> > >> > > acpi_ps_parse_aml
> > >> > >
> > >> > > It also seesm that adding a few TRACE_RESUME()s or an msleep() right
> > >> > > after enable_nonboot_cpus() can avoid the hang, sometimes.
> > >> > >
> > >> > > I've attached the DSDT in case anyone is interested in looking at it.
> > >> > >
> > >> >
> > >> > What if you comment out the execution of _WAK (line 318 of
> > >> > drivers/acpi/acpica/hwsleep.c in 4.6)? Does that make any difference?
> > >>
> > >> Indeed it does. Tried with acpi_idle and intel_idle, and both appear to
> > >> resume just fine with that hack.
> > >>
> > >> - acpi_hw_execute_sleep_method(METHOD_PATHNAME__WAK, sleep_state);
> > >> + //acpi_hw_execute_sleep_method(METHOD_PATHNAME__WAK, sleep_state);
> > >> + printk(KERN_CRIT "skipping _WAK\n");
> > >
> > > Continuing with my detective work a bit, I decided to hack the DSDT a
> > > bit to see if I can narrow the it down further, and looks like I found
> > > it on the first guess. The following change stops it from hanging.
> > >
> > > @ -5056,7 +5056,7 @@
> > > If (LEqual (Arg0, 0x03))
> > > {
> > > Store (0x01, \SPNF)
> > > - TRAP (0x46)
> > > + //TRAP (0x46)
> > > P8XH (0x00, 0x03)
> > > }
> > >
> > > So what does that do? Let's see:
> > >
> > > OperationRegion (IO_T, SystemIO, 0x0800, 0x10)
> > > Field (IO_T, ByteAcc, NoLock, Preserve)
> > > {
> > > Offset (0x08),
> > > TRP0, 8
> > > }
> > >
> > > OperationRegion (GNVS, SystemMemory, 0x3F5E0C7C, 0x0200)
> > > Field (GNVS, AnyAcc, Lock, Preserve)
> > > {
> > > OSYS, 16,
> > > SMIF, 8,
> > > ...
> > >
> > > Method (TRAP, 1, Serialized)
> > > {
> > > Store (Arg0, SMIF) /* \SMIF */
> > > Store (0x00, TRP0) /* \TRP0 */
> > > Return (SMIF) /* \SMIF */
> > > }
> > >
> > > and a dump of the IOTR registers shows:
> > >
> > > 0x1e80: 0x0000fe01
> > > 0x1e84: 0x00020001
> > > 0x1e98: 0x000c0801
> > > 0x1e9c: 0x000200f0
> > >
> > > which seems to be telling me that ports 0x800-0x80f and
> > > 0xfe00-0xfe03 would trigger an SMI.
> >
> > Well, the name of the method kind of suggests that it triggers an SMM trap. :-)
>
> Which is why I wanted confirm that by looking at the IOTR regs ;)
>
> >
> > > So the next question is how do the idle drivers and cpu hotplug
> > > fit into this picture. Do we need to force the second HT into
> > > a specific C state before the SMI or something?
> >
> > Or you can ask why exactly someone put that SMM trap into _WAK.
> >
> > Apparently, it was regarded as necessary or no one would have
> > bothered. The only reason I can see why it might be regarded as
> > necessary was that Windows did something Linux doesn't do on that
> > platform, or, which to me is far more interesting, that Windows didn't
> > do something actually done by Linux.
> >
> > My theory would be that Windows didn't reinitialize the second HT
> > properly during resume and the trap was added to let SMM do that. If
> > that's the case, the trap may trigger by the time the second HT
> > already executes code in Linux and then it will mess up with it and
> > crash.
> >
> > Now, what do idles states have to do with that? IIRC, Windows puts
> > nonboot CPUs into idle states before suspend, so the SMM code
> > triggered by the trap may make assumptions about the CPU being in such
> > a state or similar.
>
> BTW I also tried to move the enable_nonboot_cpus() after _WAK, and I
> tried to boot with nosmp, but neither trick helped. If someone could
> throw some patches my way to force things into a specific state
> before suspend/_WAK I'd be happy to test them out.

Ping. Anyone have any ideas what to try here? Would be nice to get this
machine working again...

--
Ville Syrjälä
Intel OTC