Re: [Bug 215533] [BISECTED][REGRESSION] UI becomes unresponsive every couple of seconds

[Bjorn Helgaas]

[+cc linux-kernel for visibility]

On Wed, Jan 26, 2022 at 06:12:50AM -0600, Bjorn Helgaas wrote:
> On Wed, Jan 26, 2022 at 08:18:12AM +0000, bugzilla-daemon@xxxxxxxxxxxxxxxxxxx wrote:
> > https://bugzilla.kernel.org/show_bug.cgi?id=215533
> > 
> > --- Comment #1 from joey.corleone@xxxxxxx ---
> > I accidentally sent the report prematurely. So here come my findings:
> > 
> > Since 5.16
> > (1) my system becomes unresponsive every couple of seconds (micro lags), which
> > makes it more or less unusable.
> > (2) wrong(?) CPU frequencies are reported. 
> > 
> > - 5.15 works fine.
> > - Starting from some commit in 5.17, it seems (1) is fixed (unsure), but
> > definitely not (2).
> > 
> > I have bisected the kernel between 5.15 and 5.16, and found that the offending
> > commit is 0e8ae5a6ff5952253cd7cc0260df838ab4c21009 ("PCI/portdrv: Do not setup
> > up IRQs if there are no users"). Bisection log attached.
> > 
> > Reverting this commit on linux-git[1] fixes both (1) and (2).
> > 
> > Important notes:
> > - This regression was reported on a DELL XPS 9550 laptop by two users [2], so
> > it might be related strictly to that model. 
> > - According to user mallocman, the issue can also be fixed by reverting the
> > BIOS version of the laptop to v1.12.
> > - The issue ONLY occurs when AC is plugged in (and stays there even when I
> > unplug it).
> > - When booting on battery power, there is no issue at all.
> > 
> > You can easily observe the regression via: 
> > 
> > watch cat /sys/devices/system/cpu/cpu[0-9]*/cpufreq/scaling_cur_fre
> > 
> > As soon as I plug in AC, all frequencies go up to values around 3248338 and
> > stay there even if I unplug AC. This does not happen at all when booted on
> > battery power. 
> > 
> > Also note: 
> > - the laptop's fans are not really affected by the high frequencies.
> > - setting the governor to "powersave" has no effect on the frequencies (as
> > compared to when on battery power).
> > - lowering the maximum frequency manually works, but does not fix (1).
> > 
> > [1] https://aur.archlinux.org/pkgbase/linux-git/ (pulled commits up to
> > 0280e3c58f92b2fe0e8fbbdf8d386449168de4a8).
> > [2] https://bbs.archlinux.org/viewtopic.php?id=273330

I hope we can find a better solution, but since the responsiveness
issue is a significant regression, I queued up a revert of
0e8ae5a6ff59 ("PCI/portdrv: Do not setup up IRQs if there are no
users") in case we don't find one.

If/when we get to the bottom of this, I'll replace the revert with the
solution.  0e8ae5a6ff59 appeared in v5.16, so we'll have to make sure
we fix that as well.

Bjorn