Re: [Bug 219984] New: [BISECTED] High power usage since 'PCI/ASPM: Correct LTR_L1.2_THRESHOLD computation'

[Bjorn Helgaas <helgaas@kernel.org>]

On Mon, Apr 07, 2025 at 10:57:44AM -0500, Bjorn Helgaas wrote:
> [bugzilla reporter bcc'd]
> 
> Bisected to https://git.kernel.org/linus/7afeb84d14ea
> 
> I'll take a look; including linux-pci for completeness and in case
> others are interested.

Sergey, would you mind collecting the output of "sudo lspci -vv" for
both v6.14 and v6.14 with 7afeb84d14ea ("PCI/ASPM: Correct
LTR_L1.2_THRESHOLD computation") reverted?

You can include the output here via email or (if you prefer not to
expose your email address) attach to the bugzilla.

> On Sun, Apr 06, 2025 at 03:11:54PM +0000, bugzilla-daemon@kernel.org wrote:
> > https://bugzilla.kernel.org/show_bug.cgi?id=219984
> > 
> >             Bug ID: 219984
> >            Summary: [BISECTED] High power usage since 'PCI/ASPM: Correct
> >                     LTR_L1.2_THRESHOLD computation'
> >           Reporter: sergey.v.dolgov@gmail.com
> >         Regression: No
> > 
> > Created attachment 307924
> >   --> https://bugzilla.kernel.org/attachment.cgi?id=307924&action=edit
> > Bisection log
> > 
> > I have been observing increased power consumption on my HP Spectre x360
> > Convertible 15-df1xxx (CoffeeLake) since kernel 6.1. Bisection revealed a
> > regression in commit 7afeb84d14eaaebb71f5c558ed57ca858e4304e7 (PCI/ASPM:
> > Correct LTR_L1.2_THRESHOLD computation). This still affects the mainline. 
> > 
> > With the original kernel 6.14, turbostat shows that the CPU sticks to Pkg%pc3
> > more than 50% of the time, resulting in a discharge rate of 3.52 W reported by
> > powertop during an idle empty plasma6 session with screen off.
> > 
> > After reverting 7afeb84d14eaaebb71f5c558ed57ca858e4304e7 and recompiling the
> > kernel with the same config, Pk%pc10, CPU%LPI and SYS%LPI residencies are all
> > above 80% during the same session, and "The battery reports a discharge rate of
> > 1.35 W."
> > 
> > The two kernels above had no nvidia drivers to avoid tainting. However,
> > compiling also the proprietary nvidia modules reduces the idle power
> > consumption further to 796 mW.
> > 
> > Could you please revert 7afeb84d14eaaebb71f5c558ed57ca858e4304e7 upstream? Or,
> > at least, to switch between the two encode_l12_threshold algorithms selectively
> > depending on the system.