Re: [PM] bfcc1e67ff: kernel-selftests.breakpoints.step_after_suspend_test.fail

["Rafael J. Wysocki" <rafael@kernel.org>]

  On Wed, Oct 20, 2021 at 8:17 PM Florian Fainelli <f.fainelli@gmail.com> wrote:
>
> On 10/20/21 9:00 AM, Rafael J. Wysocki wrote:
> > On Wed, Oct 20, 2021 at 5:34 PM Florian Fainelli <f.fainelli@gmail.com> wrote:
> >>
> >>
> >>
> >> On 10/20/2021 6:49 AM, Rafael J. Wysocki wrote:
> >>> On Tue, Oct 19, 2021 at 9:04 PM Florian Fainelli <f.fainelli@gmail.com> wrote:
> >>>>
> >>>> On 10/19/21 11:53 AM, Rafael J. Wysocki wrote:
> >>>>> On 10/15/2021 9:40 PM, Florian Fainelli wrote:
> >>>>>> On 10/15/21 11:45 AM, Rafael J. Wysocki wrote:
> >>>>>>> On 10/14/2021 11:55 PM, Florian Fainelli wrote:
> >>>>>>>> On 10/14/21 12:23 PM, Rafael J. Wysocki wrote:
> >>>>>>>>> On 10/14/2021 6:26 PM, Florian Fainelli wrote:
> >>>>>>>>>> On 10/14/21 12:57 AM, kernel test robot wrote:
> >>>>>>>>>>> Greeting,
> >>>>>>>>>>>
> >>>>>>>>>>> FYI, we noticed the following commit (built with gcc-9):
> >>>>>>>>>>>
> >>>>>>>>>>> commit: bfcc1e67ff1e4aa8bfe2ca57f99390fc284c799d ("PM: sleep: Do not
> >>>>>>>>>>> assume that "mem" is always present")
> >>>>>>>>>>> https://git.kernel.org/cgit/linux/kernel/git/next/linux-next.git
> >>>>>>>>>>> master
> >>>>>>>>>>>
> >>>>>>>>>>>
> >>>>>>>>>>> in testcase: kernel-selftests
> >>>>>>>>>>> version: kernel-selftests-x86_64-c8c9111a-1_20210929
> >>>>>>>>>>> with following parameters:
> >>>>>>>>>>>
> >>>>>>>>>>>        group: group-00
> >>>>>>>>>>>        ucode: 0x11
> >>>>>>>>>>>
> >>>>>>>>>>> test-description: The kernel contains a set of "self tests" under
> >>>>>>>>>>> the
> >>>>>>>>>>> tools/testing/selftests/ directory. These are intended to be small
> >>>>>>>>>>> unit tests to exercise individual code paths in the kernel.
> >>>>>>>>>>> test-url: https://www.kernel.org/doc/Documentation/kselftest.txt
> >>>>>>>>>>>
> >>>>>>>>>>>
> >>>>>>>>>>> on test machine: 288 threads 2 sockets Intel(R) Xeon Phi(TM) CPU
> >>>>>>>>>>> 7295
> >>>>>>>>>>> @ 1.50GHz with 80G memory
> >>>>>>>>>>>
> >>>>>>>>>>> caused below changes (please refer to attached dmesg/kmsg for entire
> >>>>>>>>>>> log/backtrace):
> >>>>>>>>>>>
> >>>>>>>>>>>
> >>>>>>>>>>>
> >>>>>>>>>>>
> >>>>>>>>>>> If you fix the issue, kindly add following tag
> >>>>>>>>>>> Reported-by: kernel test robot <oliver.sang@intel.com>
> >>>>>>>>>> Thanks for your report. Assuming that the code responsible for
> >>>>>>>>>> registering the suspend operations is drivers/acpi/sleep.c for your
> >>>>>>>>>> platform, and that acpi_sleep_suspend_setup() iterated over all
> >>>>>>>>>> possible
> >>>>>>>>>> sleep states, your platform must somehow be returning that
> >>>>>>>>>> ACPI_STATE_S3
> >>>>>>>>>> is not a supported state somehow?
> >>>>>>>>>>
> >>>>>>>>>> Rafael have you ever encountered something like that?
> >>>>>>>>> Yes, there are systems with ACPI that don't support S3.
> >>>>>>>> OK and do you know what happens when we enter suspend with "mem" in
> >>>>>>>> those cases? Do we immediately return because ultimately the firmware
> >>>>>>>> does not support ACPI S3?
> >>>>>>> "mem" should not be present in the list of available strings then, so it
> >>>>>>> should be rejected right away.
> >>>>>> Well yes, that was the purpose of the patch I submitted, but assuming
> >>>>>> that we did provide "mem" as one of the possible standby modes even
> >>>>>> though that was wrong (before patch), and the test was trying to enter
> >>>>>> ACPI S3 standby, what would have happened, would the ACPI firmware honor
> >>>>>> the request but return an error, or would it actually enter ACPI S3?
> >>>>>>
> >>>>>> In any case, I will change the test to check that this is a supported
> >>>>>> standby mode before trying it.
> >>>>>
> >>>>> Unfortunately, I will need to revert bfcc1e67ff1e4aa8bfe2, because it
> >>>>> breaks user space compatibility and that's got caught properly by the test.
> >>>>
> >>>> Reverting my commit will break powerpc and other ARM/ARM64 platforms
> >>>> where mem is not supported (via PSCI),
> >>>
> >>> It won't break anything, although the things that didn't work before
> >>> will still not work after it.
> >>>
> >>> And "mem" is always supported even if there are no suspend_ops at all,
> >>> in which case it becomes an alternative way to trigger s2idle.
> >>>
> >>> So, on the affected systems, what's there in /sys/power/?  Is
> >>> mem_sleep present?  If so, what's in it?
> >>
> >> With 4.9 which is what I used initially:
> >>
> >> # cat /sys/power/state
> >> freeze standby
> >> # cat /sys/power/
> >> pm_async           pm_print_times     pm_wakeup_irq      wakeup_count
> >> pm_freeze_timeout  pm_test            state
> >>
> >> With a newer kernel without my patch:
> >>
> >> # cat /sys/power/state
> >> freeze standby mem
> >> # cat /sys/power/mem_sleep
> >> s2idle shallow [deep]
> >
> > OK, so the "deep" and "shallow" suspend variants appear to be
> > supported.  What's the problem with advertising "mem" then?
>
> s2idle and shallow are, but deep is not.

Why is it there in mem_sleep, then?  It should not be there if
valid_state() returns 'false' for it.

mem_sleep_states[PM_SUSPEND_MEM] is only set by suspend_set_ops() if
valid_state(PM_SUSPEND_MEM) is 'true'.

> >
> >> # cat /sys/power/
> >> mem_sleep          pm_freeze_timeout  pm_wakeup_irq      wakeup_count
> >> pm_async           pm_print_times     state
> >> pm_debug_messages  pm_test            suspend_stats/
> >>
> >>
> >>>
> >>>> I have a change pending for PSCI
> >>>> that will actually check that SYSTEM_SUSPEND is supported before
> >>>> unconditionally making use of it.
> >>>>
> >>>>>
> >>>>> What happens is that "mem" is a "pointer" to a secondary list of
> >>>>> possible states and that generally is "s2idle shallow deep" and if
> >>>>> s2idle is the only available option, it will be just "s2idle".
> >>>>>
> >>>>> This list is there in /sys/power/mem_sleep.
> >>>>>
> >>>>> It was done this way, because some variants of user space expect "mem"
> >>>>> to be always present and don't recognize "freeze" properly.
> >>>>>
> >>>>> Sorry for the confusion.
> >>>>
> >>>> So how do we all get our cookie here? Should we just slap an #ifndef
> >>>> CONFIG_ACPI in order to allow platforms that do not have "mem" to not
> >>>> have it?
> >>>
> >>> Certainly not.
> >>>
> >>> I've just hacked my test-bed system with ACPI so it does not register
> >>> any suspend_ops at all and I have "freeze mem disk" in
> >>> /sys/power/state and "s2idle" in /sys/power/mem_sleep.  Writing "mem"
> >>> to /sys/power/state causes s2idle to be carried out.
> >>>
> >>> Since this is the expected behavior, I'm not sure what the problem is.
> >>
> >> The problem is advertising "mem" in /sys/power/state when the state is
> >> not actually supported by the platform firmware here, whether that
> >> translates into the form of s2idle or not. It is not supported, and it
> >> should not be there IMHO.
> >
> > Well, it is there, because some user space expects it to be there on
> > systems supporting any kind of system-wide suspend, including s2idle.
> > Like it or not.
>
> But that was not the case before 406e79385f32 ("PM / sleep: System sleep
> state selection interface rework") and clearly nobody complained about
> that, did they?

Yes, it was and yes, they did.  Changes like that are not made without a reason.

> >
> > If it is not there, the utilities in question assume that system-wide
> > suspend is not supported at all.
>
> What utilities do depend on that? That selftest that does not even check
> that "mem" is actually present in /sys/power/state and just fails its
> test if it is not, yes it's not great, but that can be fixed.

Various GUI-based things like KDE, GNOME and similar plus the Chrome
user space IIRC.

> >
> >> I was late to the game in identifying that,
> >> but the 4.9 kernel makes sense to me.
> >>
> >> Similarly, if you take arch/powerpc/sysdev/fsl_pmc.c only
> >> PM_SUSPEND_STANDBY is valid, so advertising mem would be wrong if we
> >> don't look at what ->valid tells us.
> >
> > Again: "mem" appears in /sys/power/state if the system supports any
> > kind of system-wide suspend (because of the expectations of user space
> > mentioned above) and mem_sleep decides what it really means.
> >
> > And this is documented too (see Documentation/admin-guide/pm/sleep-states.html).
>
> The documentation just states that if the kernel supports *any* suspend
> state, then /sys/power/state will be present and likewise for
> /sys/power/mem_sleep, it does not say what the contents will be and that
> "mem" would always be present in there.

It doesn't say so directly, but it kind of wouldn't make sense to have
"mem_sleep" without "mem" in "state" and it implies that "mem_sleep"
is not empty if it is present.  Ergo "mem" is present in "state" if
"mem_sleep" is present which is the case if (at least) s2idle is
supported.  That is always the case if CONFIG_SUSPEND is set which
follows from the suspend-to-idle description.

Anyway, I'm still not sure what the problem really is.  Commit
406e79385f32 still allows user space to only trigger transitions to
s2idle and other states explicitly reported as valid by the platform.