Re: [BUG] [linux-pm] Commit "Hibernation: Enter platform hibernation state in a consistent way)" makes my system to resume instantly from S4

["Rafael J. Wysocki" <rjw@sisk.pl>]

On Thursday, 1 November 2007 16:25, Maxim Levitsky wrote:
> On Tuesday 30 October 2007 01:31:13 Rafael J. Wysocki wrote:
> > On Monday, 29 October 2007 23:36, Maxim Levitsky wrote:
> > > On Monday 29 October 2007 22:37:37 Rafael J. Wysocki wrote:
> > > > On Monday, 29 October 2007 07:45, Maxim Levitsky wrote:
> > > > > On Sunday 28 October 2007 22:23:15 Rafael J. Wysocki wrote:
> > > > > > On Sunday, 28 October 2007 21:00, Maxim Levitsky wrote:
> > > > > > > On Saturday 27 October 2007 23:46:45 Rafael J. Wysocki wrote:
> > > > > > > > On Saturday, 27 October 2007 14:05, Maxim Levitsky wrote:
> > > > > > > > > Hi,
> > > > > > > > > 
> > > > > > > > > Recently I noticed that my system resumes just after suspend to disk.
> > > > > > > > > 
> > > > > > > > > I traced this to commit 9cd9a0058dd35268b24fa16795a92c800f4086d4.
> > > > > > > > > 
> > > > > > > > > Note:
> > > > > > > > > 
> > > > > > > > > This happens only if I enable WOL using /proc/acpi/wakeup
> > > > > > > > > (echo "ILAN" > /proc/acpi/wakeup)
> > > > > > > > 
> > > > > > > > What happens after a suspend to RAM?
> > > > > > > I do more testing tomorrow, but what I remember is that,
> > > > > > > I had no issues with suspend-to-ram.
> > > > > > > (And take into account the fact that _all_ wakeup devices were enabled always,
> > > > > > > the rc.local does that for me, and still I did quite a lot of suspends, but I
> > > > > > > got the immediate resume, only and always after suspend to disk.
> > > > > > 
> > > > > > Hm, can you please retest with the above commit applied, but with
> > > > > > device_suspend() in hibernation_platform_enter() replaced by device_shutdown()?
> > > > > Bingo! doing the above fixes this problem.
> > > > > And... strange, since I did test the system with e1000 unloaded, but still got same bug,
> > > > > back when I chased that bug, thus it is probably not related to e1000 and WOL.
> > > > > 
> > > > > the ACPI wake-up source (I have read the DSDT, and ICH datasheets, thus I know that this is a set of bits in southbridge)
> > > > > of e1000 is shared with EHCI and HDA sound codec.
> > > > > (But usb on my system only able to wake from S3, and the soundchip which can theoretically wake up from S3 too, doesn't have this ability)
> > > > > 
> > > > > (But disabling this wakeup source, fixes this bug too)
> > > > 
> > > > OK
> > > > 
> > > > Please get back to device_suspend() in hibernation_platform_enter() and try
> > > > to unload ehci_hcd before hibernation with the wakeup source enabled.
> > > Well, ... bingo! again!
> > > 
> > > Replacing back the device_suspend() , and unloading the ehci_hcd fixes the problem!
> > > Even more, loading this module back causes this problem to reappear, and vise-verso.
> > > All wakeup sources are enabled always.
> > > 
> > > Note that I don't have any USB 2.0 devices yet.
> > > The usb devices I have are: keyboard,mouse,and joystick.
> > > 
> > > So now this problem looks fixable, I take a look at ehci driver.
> > 
> > This is very similar to what I've been observing for some time on one of my
> > test boxes wrt suspending to RAM and in that case the problem is also related
> > to EHCI (this is a pretty old Celeron-based box with an Intel chipset).
> > 
> > It looks like in some situations, on some platforms the EHCI controller
> > triggers a resume right after the suspend (or hibernation) if the driver's
> > .suspend() routine is executed before invoking the ACPI firmware to put the
> > system into the sleep state.
> > 
> > [The "ehci_hcd causes box to resume immediately after suspend to RAM" thread on
> > linux-pm was about this problem.]
> > 
> > Greetings,
> > Rafael
> > 
> 
> Sorry, for a bit late response,
> 
> Do you meant that EHCI driver triggers the resume regardless of the above commit?

Rather, the hardware does it, but the driver is at least partially involved (ie.
unloading ehci-hcd before a suspend to RAM makes the box behave as expected,
which is to stay suspended).

> Now I have checked the source and understand that the above commit is 150% right.
> (Actually I always assumed that it was always there, and some developers did that too)
> 
> 
> I assumed that on hibernation the sequence is:
> 
> .suspend() - disable device + enable wake
> snapshot
> 
> .resume() - enabled device + disable wake
> snapshot writeout
> .suspend() - disable device + enable wake
> 
> 
> But before the above commit was introduced, the last .suspend()
> didn't happen, and thus all the code that enables wake was disabled.
> For the same reason rtc-cmos didn't work, since its .resume disables the alarm.

Yes.
 
> The mystery why my system did react to WOL in S4 is resolved by the fact that
> e1000 .shutdown() calls .suspend() to make WOL work in S5 too.
> (Btw, this is wrong that rtc-cmos disables alarm in .shutdown()  since it can be used in S5 too)
> 
> 
> So, this commit caused the EHCI .suspend to activate the wakeup, and this code probably contains a bug
> (Or a lack of hardware workaround)

I bet on the lack of hardware workaround in there.

Also, the problem seems to be related to the sharing of resources between EHCI
and some devices (sharing interrupts, wakeup sources).

> Thus, since the suspend to ram isn't affected by the above commit, I want to know
> whenever it affects the system you have seen resuming from ram instantly with ehci driver.

No it doesn't, AFAICS.

I tested this patch on the affected system before posting it.

Greetings,
Rafael
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/