On Mon, May 22, 2023 at 01:45:55PM +0200, Linux regression tracking (Thorsten Leemhuis) wrote: > On 05.05.23 08:56, Koba Ko wrote: > > On Thu, May 4, 2023 at 5:23 PM Bjorn Helgaas <helgaas@xxxxxxxxxx> wrote: > >> [+cc Koba, Ajay, Tasev, Mark, Thomas, regressions list] > >> On Tue, Apr 11, 2023 at 03:42:29PM -0500, Bjorn Helgaas wrote: > >>> On Tue, Apr 11, 2023 at 08:32:04AM +0000, bugzilla-daemon@xxxxxxxxxx wrote: > >>>> https://bugzilla.kernel.org/show_bug.cgi?id=217321 > >>>> ... > >>>> Regression: No > >>>> > >>>> [Symptom] > >>>> Intel cpu can't sleep deeper than pcˇ during long idle > >>>> ~~~ > >>>> Pkg%pc2 Pkg%pc3 Pkg%pc6 Pkg%pc7 Pkg%pc8 Pkg%pc9 Pk%pc10 > >>>> 15.08 75.02 0.00 0.00 0.00 0.00 0.00 > >>>> 15.09 75.02 0.00 0.00 0.00 0.00 0.00 > >>>> ^CPkg%pc2 Pkg%pc3 Pkg%pc6 Pkg%pc7 Pkg%pc8 Pkg%pc9 Pk%pc10 > >>>> 15.38 68.97 0.00 0.00 0.00 0.00 0.00 > >>>> 15.38 68.96 0.00 0.00 0.00 0.00 0.00 > >>>> ~~~ > >>>> [How to Reproduce] > >>>> 1. run turbostat to monitor > >>>> 2. leave machine idle > >>>> 3. turbostat show cpu only go into pc2~pc3. > >>>> > >>>> [Misc] > >>>> The culprit are this > >>>> a7152be79b62) Revert "PCI/ASPM: Save L1 PM Substates Capability for > >>>> suspend/resume” > >>>> > >>>> if revert a7152be79b62, the issue is gone > >>> > >>> Relevant commits: > >>> > >>> 4ff116d0d5fd ("PCI/ASPM: Save L1 PM Substates Capability for suspend/resume") > >>> a7152be79b62 ("Revert "PCI/ASPM: Save L1 PM Substates Capability for suspend/resume"") > >>> > >>> 4ff116d0d5fd appeared in v6.1-rc1. Prior to 4ff116d0d5fd, ASPM L1 PM > >>> Substates configuration was not preserved across suspend/resume, so > >>> the system *worked* after resume, but used more power than expected. > >>> > >>> But 4ff116d0d5fd caused resume to fail completely on some systems, so > >>> a7152be79b62 reverted it. With a7152be79b62 reverted, ASPM L1 PM > >>> Substates configuration is likely not preserved across suspend/resume. > >>> a7152be79b62 appeared in v6.2-rc8 and was backported to the v6.1 > >>> stable series starting with v6.1.12. > >>> > >>> KobaKo, you don't mention any suspend/resume in this bug report, but > >>> neither patch should make any difference unless suspend/resume is > >>> involved. Does the platform sleep as expected *before* suspend, but > >>> fail to sleep after resume? > >>> > >>> Or maybe some individual device was suspended via runtime power > >>> management, and that device lost its L1 PM Substates config? I don't > >>> know if there's a way to disable runtime PM easily. > >> > >> Koba, per your bugzilla update, the issue happens even without > >> suspend/resume. And we don't know whether some particular device is > >> responsible. > >> > >> But if we save/restore L1SS state, we can sleep deeper than PC3. If > >> we don't preserve L1SS state, we can't. > >> > >> We definitely want to preserve the L1SS state, but we can't simply > >> apply 4ff116d0d5fd ("PCI/ASPM: Save L1 PM Substates Capability for > >> suspend/resume") again because it caused its own regressions [1,2,3] > >> > >> So somebody needs to figure out what was wrong with 4ff116d0d5fd, fix > >> it, verify that it doesn't cause the issues reported by Tasev, Thomas, > >> and Mark, and then we can apply it. > > > > Good days, discussed with Kai-Heng and he mentioned the GPU may not > > be pulled off the power. > > then the GPU needs L1ss to get into power saving. > > > > I will investigate further on this way. > > Did anything come our of this? > > FWIW, I'm considering to drop this from the list of tracked regressions. > Yes, this is a regression, but it's caused by fix for another (worse) > regression -- so there is nothing we can do for now anyway (and Koba > seems motivated already to look properly into all of this). Or does > anyone consider this to be a problem? I would drop this from the regression list. Yes, bz 217321 is a bug, and yes, 4ff116d0d5fd is a partial fix for it, but 4ff116d0d5fd causes worse problems (it breaks resume from suspend) than just living with bz 217321, which is a "mere" power consumption issue. I updated bz 217321 to drop the "regression" label there. Bjorn
