lkml.org 
[lkml]   [2017]   [Dec]   [13]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
Date
SubjectRe: Linux 4.15-rc2: Regression in resume from ACPI S3
On Wed, Dec 13, 2017 at 11:39 PM, Rafael J. Wysocki <rjw@rjwysocki.net> wrote:
> On Wednesday, December 13, 2017 7:19:17 PM CET Thomas Gleixner wrote:
>> On Wed, 13 Dec 2017, Linus Torvalds wrote:
>>
>> > On Wed, Dec 13, 2017 at 8:41 AM, Thomas Gleixner <tglx@linutronix.de> wrote:
>> > >
>> > > Definitely. That was fragile forever but puzzles me is that I can't figure
>> > > out what now causes that spurious interrupt to surface out of the blue.
>> >
>> > Perhaps just timing?
>>
>> That's what I'm trying to figure out right now, because that is the only
>> sensible explanation left. The whole machinery of suspend is exactly the
>> same with and without the vector changes. I instrumented all functions
>> involved and the picture is the same. I even do not see any fundamental
>> timing differences where one would say: That's it.
>>
>> What puzzles me even more is that in the range of commits I'm fiddling with
>> there is no other change than the vector management stuff and the point
>> where it breaks makes no sense at all. The point Maarten bisected it to
>> works nicely here, so that might just point to a very subtle timing issue.
>>
>> > How hard would it be to change the ordering to just redirect irqs first?
>>
>> The whole interrupt redirection happens when the non boot CPUs are brought
>> down, which is the very last step before the actual suspend happens.
>>
>> We could probably do that earlier, but that's something Rafael needs to
>> answer ultimately.
>
> Well, that's both flattering and concerning. ;-)
>
> Anyway, yes, we can do that earlier AFAICS. Action handlers are not going to
> run after we've called suspend_device_irqs() which happens before the final
> stage of PCI devices suspend (suspend_noirq) and it doesn't matter which CPU
> gets the interrupt from that point on (it is either wakeup or unwanted then).

There is a catch that we don't and likely should not do that for
suspend-to-idle, but since we have pm_suspend_target_state now, that
case can be distinguished from the "full suspend" one readily.

Thanks,
Rafael

\
 
 \ /
  Last update: 2017-12-14 00:26    [W:0.126 / U:0.340 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site