Re: [PATCH] PM / QoS: Fix default runtime_pm device resume latency

From: Rafael J. Wysocki
Date: Wed Nov 01 2017 - 16:50:54 EST


On Wed, Nov 1, 2017 at 11:28 AM, Tero Kristo <t-kristo@xxxxxx> wrote:
> On 01/11/17 00:32, Rafael J. Wysocki wrote:
>>
>> On Tue, Oct 31, 2017 at 7:07 PM, Geert Uytterhoeven
>> <geert@xxxxxxxxxxxxxx> wrote:
>>>
>>> Hi Rafael,
>>>
>>> On Tue, Oct 31, 2017 at 6:22 PM, Rafael J. Wysocki <rafael@xxxxxxxxxx>
>>> wrote:
>>>>
>>>> On Tue, Oct 31, 2017 at 2:55 PM, Geert Uytterhoeven
>>>> <geert@xxxxxxxxxxxxxx> wrote:
>>>>>
>>>>> Hi Rafael, Tero,
>>>>>
>>>>> CC pinchartl, dri-devel
>>>>>
>>>>> On Tue, Oct 31, 2017 at 2:10 PM, Geert Uytterhoeven
>>>>> <geert@xxxxxxxxxxxxxx> wrote:
>>>>>>
>>>>>> CC linux-renesas-soc
>>>>>>
>>>>>> On Tue, Oct 31, 2017 at 2:09 PM, Geert Uytterhoeven
>>>>>> <geert@xxxxxxxxxxxxxx> wrote:
>>>>>>>
>>>>>>> On Tue, Oct 31, 2017 at 12:27 AM, Rafael J. Wysocki
>>>>>>> <rjw@xxxxxxxxxxxxx> wrote:
>>>>>>>>
>>>>>>>> On Monday, October 30, 2017 11:19:08 AM CET Rafael J. Wysocki wrote:
>>>>>>>>>
>>>>>>>>> On Mon, Oct 30, 2017 at 8:10 AM, Tero Kristo <t-kristo@xxxxxx>
>>>>>>>>> wrote:
>>>>>>>>>>
>>>>>>>>>> The recent change to the PM QoS framework to introduce a proper
>>>>>>>>>> no constraint value overlooked to handle the devices which don't
>>>>>>>>>> implement PM QoS OPS. Runtime PM is one of the more severely
>>>>>>>>>> impacted subsystems, failing every attempt to runtime suspend
>>>>>>>>>> a device. This leads into some nasty second level issues like
>>>>>>>>>> probe failures and increased power consumption among other things.
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Oh, that's bad.
>>>>>>>>>
>>>>>>>>> Sorry about breaking it and thanks for the fix!
>>>>>>>>>
>>>>>>>>>> Fix this by adding a proper return value for devices that don't
>>>>>>>>>> implement PM QoS implicitly.
>>>>>>>>>>
>>>>>>>>>> Fixes: 0cc2b4e5a020 ("PM / QoS: Fix device resume latency PM QoS")
>>>>>>>>>> Signed-off-by: Tero Kristo <t-kristo@xxxxxx>
>>>>>>>>>> Cc: Rafael J. Wysocki <rafael.j.wysocki@xxxxxxxxx>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Applied.
>>>>>>>>
>>>>>>>>
>>>>>>>> And pushed to Linus.
>>>>>>>
>>>>>>>
>>>>>>> I'm afraid it is not sufficient.
>>>>>>>
>>>>>>> Commit 0cc2b4e5a020fc7f ("PM / QoS: Fix device resume latency PM
>>>>>>> QoS")
>>>>>>> introduced two issues on Renesas platforms:
>>>>>>> 1. After boot up, many devices have changed their state from
>>>>>>> "suspended"
>>>>>>> to "active", according to
>>>>>>> /sys/kernel/debug/pm_genpd/pm_genpd_summary
>>>>>>> (comparing that file across boots is one of my standard tests).
>>>>>>> Interestingly, doing a system suspend/resume cycle restores
>>>>>>> their state
>>>>>>> to "suspended".
>>>>>>>
>>>>>>> 2. During system suspend, the following warning is printed on
>>>>>>> r8a7791/koelsch:
>>>>>>>
>>>>>>> i2c-rcar e6530000.i2c: runtime PM trying to suspend device
>>>>>>> but
>>>>>>> active child
>>>>>
>>>>>
>>>>> 3. I've just bisected a seemingly unrelated issue to the same commit.
>>>>> On Salvator-XS with R-Car H3, initialization of the rcar-du driver
>>>>> now
>>>>> takes more than 1 minute due to flip_done time outs, while it took
>>>>> 0.12s
>>>>> before:
>>>>>
>>>>> [ 3.015035] [drm] Supports vblank timestamp caching Rev 2
>>>>> (21.10.2013).
>>>>> [ 3.021721] [drm] No driver support for vblank timestamp query.
>>>>> [ 13.280738] [drm:drm_atomic_helper_wait_for_flip_done] *ERROR*
>>>>> [CRTC:58:crtc-3] flip_done timed out
>>>>> [ 23.520707] [drm:drm_atomic_helper_commit_cleanup_done] *ERROR*
>>>>> [CRTC:58:crtc-3] flip_done timed out
>>>>> [ 33.760708] [drm:drm_atomic_helper_wait_for_flip_done] *ERROR*
>>>>> [CRTC:58:crtc-3] flip_done timed out
>>>>> [ 44.000755] [drm:drm_atomic_helper_commit_cleanup_done] *ERROR*
>>>>> [CRTC:58:crtc-3] flip_done timed out
>>>>> [ 44.003597] Console: switching to colour frame buffer device
>>>>> 128x48
>>>>> [ 54.240707] [drm:drm_atomic_helper_wait_for_flip_done] *ERROR*
>>>>> [CRTC:58:crtc-3] flip_done timed out
>>>>> [ 64.480706] [drm:drm_atomic_helper_commit_cleanup_done] *ERROR*
>>>>> [CRTC:58:crtc-3] flip_done timed out
>>>>> [ 64.544876] rcar-du feb00000.display: fb0: frame buffer device
>>>>> [ 64.552013] [drm] Initialized rcar-du 1.0.0 20130110 for
>>>>> feb00000.display on minor 0
>>>>> [ 64.559873] [drm] Device feb00000.display probed
>>>>>
>>>>>>> Commit 2a9a86d5c81389cd ("PM / QoS: Fix default runtime_pm device
>>>>>>> resume
>>>>>>> latency") fixes the second issue, but not the first.
>>>>>
>>>>>
>>>>> ... nor the third.
>>>>>
>>>>>>> Reverting commits 2a9a86d5c81389cd ("PM / QoS: Fix default runtime_pm
>>>>>>> device resume latency") and 0cc2b4e5a020fc7f ("PM / QoS: Fix device
>>>>>>> resume
>>>>>>> latency PM QoS") fixes both.
>>>>>
>>>>>
>>>>> ... all three.
>>>>
>>>>
>>>> Sorry for the breakage.
>>>>
>>>> OK, I'll just push the reverts to Linus later today.
>>>>
>>>>>>> Do you have a clue?
>>>>
>>>>
>>>> Well, kind of.
>>>>
>>>> There is a change in behavior in domain_governor.c that should not
>>>> have made any difference to my eyes, but maybe that's it.
>>>>
>>>> Can you please check if the attached patch makes any difference?
>>>
>>>
>>> Thanks, but it doesn't seem to fix the issues.
>>
>>
>> Thanks for testing!
>>
>> I've just pushed the reverts, but the PM QoS still needs to be fixed,
>> so we have to get to the bottom of this.
>>
>> The current theory goes that the changes in domain_governor.c are to
>> blame. Is genpd involved in all of the issues with the PM QoS fix you
>> have seen?
>>
>> Thanks,
>> Rafael
>>
>
> It seems the default values for pm_qos have changed with the patch, and that
> breaks genpd at least. I only fixed PM runtime initially, but you could try
> this diff to fix the genpd part also:
>
> diff --git a/include/linux/pm_qos.h b/include/linux/pm_qos.h
> index d68b056..7c8f643 100644
> --- a/include/linux/pm_qos.h
> +++ b/include/linux/pm_qos.h
> @@ -34,9 +34,9 @@ enum pm_qos_flags_status {
> #define PM_QOS_NETWORK_LAT_DEFAULT_VALUE (2000 * USEC_PER_SEC)
> #define PM_QOS_NETWORK_THROUGHPUT_DEFAULT_VALUE 0
> #define PM_QOS_MEMORY_BANDWIDTH_DEFAULT_VALUE 0
> -#define PM_QOS_RESUME_LATENCY_DEFAULT_VALUE 0
> +#define PM_QOS_RESUME_LATENCY_DEFAULT_VALUE PM_QOS_LATENCY_ANY
> #define PM_QOS_RESUME_LATENCY_NO_CONSTRAINT PM_QOS_LATENCY_ANY
> -#define PM_QOS_LATENCY_TOLERANCE_DEFAULT_VALUE 0
> +#define PM_QOS_LATENCY_TOLERANCE_DEFAULT_VALUE (-1)
> #define PM_QOS_LATENCY_TOLERANCE_NO_CONSTRAINT (-1)
>
> #define PM_QOS_FLAG_NO_POWER_OFF (1 << 0)

This is the original change in pm_qos.h (up to the GMail-induced
whitespace breakage):

-#define PM_QOS_DEFAULT_VALUE -1
+#define PM_QOS_DEFAULT_VALUE (-1)
+#define PM_QOS_LATENCY_ANY S32_MAX

#define PM_QOS_CPU_DMA_LAT_DEFAULT_VALUE (2000 * USEC_PER_SEC)
#define PM_QOS_NETWORK_LAT_DEFAULT_VALUE (2000 * USEC_PER_SEC)
#define PM_QOS_NETWORK_THROUGHPUT_DEFAULT_VALUE 0
#define PM_QOS_MEMORY_BANDWIDTH_DEFAULT_VALUE 0
#define PM_QOS_RESUME_LATENCY_DEFAULT_VALUE 0
+#define PM_QOS_RESUME_LATENCY_NO_CONSTRAINT PM_QOS_LATENCY_ANY
#define PM_QOS_LATENCY_TOLERANCE_DEFAULT_VALUE 0
#define PM_QOS_LATENCY_TOLERANCE_NO_CONSTRAINT (-1)

-#define PM_QOS_LATENCY_ANY ((s32)(~(__u32)0 >> 1))

#define PM_QOS_FLAG_NO_POWER_OFF (1 << 0)
#define PM_QOS_FLAG_REMOTE_WAKEUP (1 << 1)

so the only thing that really has changed is the addition of
PM_QOS_RESUME_LATENCY_NO_CONSTRAINT, so I'm not really sure what you
mean. Care to elaborate?

There is a bug in the genpd part of the original patch (the
multiplication by NSEC_PER_USEC in dev_update_qos_constraint() should
not be applied to the effective_constraint value), but it doesn't
matter too much now that the problematic commit has been reverted.

I'll post a new version of it for testing shortly, but on top of a
genpd governor patch to make it behave consistently.

Thanks,
Rafael