Re: [PATCH] PCI/ASPM: Enable ASPM for links under VMD domain

From: Bjorn Helgaas
Date: Thu Sep 17 2020 - 14:39:50 EST


On Thu, Sep 10, 2020 at 07:51:05PM +0000, Derrick, Jonathan wrote:
> On Thu, 2020-09-10 at 14:17 -0500, Bjorn Helgaas wrote:
> > On Thu, Sep 10, 2020 at 06:52:48PM +0000, Derrick, Jonathan wrote:
> > > On Thu, 2020-09-10 at 12:38 -0500, Bjorn Helgaas wrote:
> > > > On Thu, Sep 10, 2020 at 04:33:39PM +0000, Derrick, Jonathan wrote:
> > > > > On Wed, 2020-09-09 at 20:55 -0500, Bjorn Helgaas wrote:
> > > > > > On Fri, Aug 21, 2020 at 08:32:20PM +0800, Kai-Heng Feng wrote:
> > > > > > > New Intel laptops with VMD cannot reach deeper power saving state,
> > > > > > > renders very short battery time.
> > > > > > >
> > > > > > > As BIOS may not be able to program the config space for devices under
> > > > > > > VMD domain, ASPM needs to be programmed manually by software. This is
> > > > > > > also the case under Windows.
> > > > > > >
> > > > > > > The VMD controller itself is a root complex integrated endpoint that
> > > > > > > doesn't have ASPM capability, so we can't propagate the ASPM settings to
> > > > > > > devices under it. Hence, simply apply ASPM_STATE_ALL to the links under
> > > > > > > VMD domain, unsupported states will be cleared out anyway.
> > > > > > >
> > > > > > > Signed-off-by: Kai-Heng Feng <kai.heng.feng@xxxxxxxxxxxxx>
> > > > > > > ---
> > > > > > > drivers/pci/pcie/aspm.c | 3 ++-
> > > > > > > drivers/pci/quirks.c | 11 +++++++++++
> > > > > > > include/linux/pci.h | 2 ++
> > > > > > > 3 files changed, 15 insertions(+), 1 deletion(-)
> > > > > > >
> > > > > > > diff --git a/drivers/pci/pcie/aspm.c b/drivers/pci/pcie/aspm.c
> > > > > > > index 253c30cc1967..dcc002dbca19 100644
> > > > > > > --- a/drivers/pci/pcie/aspm.c
> > > > > > > +++ b/drivers/pci/pcie/aspm.c
> > > > > > > @@ -624,7 +624,8 @@ static void pcie_aspm_cap_init(struct pcie_link_state *link, int blacklist)
> > > > > > > aspm_calc_l1ss_info(link, &upreg, &dwreg);
> > > > > > >
> > > > > > > /* Save default state */
> > > > > > > - link->aspm_default = link->aspm_enabled;
> > > > > > > + link->aspm_default = parent->dev_flags & PCI_DEV_FLAGS_ENABLE_ASPM ?
> > > > > > > + ASPM_STATE_ALL : link->aspm_enabled;
> > > > > >
> > > > > > This function is ridiculously complicated already, and I really don't
> > > > > > want to make it worse.
> > > > > >
> > > > > > What exactly is the PCIe topology here? Apparently the VMD controller
> > > > > > is a Root Complex Integrated Endpoint, so it's a Type 0 (non-bridge)
> > > > > > device. And it has no Link, hence no Link Capabilities or Control and
> > > > > > hence no ASPM-related bits. Right?
> > > > >
> > > > > That's correct. VMD is the Type 0 device providing config/mmio
> > > > > apertures to another segment and MSI/X remapping. No link and no ASPM
> > > > > related bits.
> > > > >
> > > > > Hierarchy is usually something like:
> > > > >
> > > > > Segment 0 | VMD segment
> > > > > Root Complex -> VMD | Type 0 (RP/Bridge; physical slot) - Type 1
> > > > > | Type 0 (RP/Bridge; physical slot) - Type 1
> > > > >
> > > > > > And the devices under the VMD controller? I guess they are regular
> > > > > > PCIe Endpoints, Switch Ports, etc? Obviously there's a Link involved
> > > > > > somewhere. Does the VMD controller have some magic, non-architected
> > > > > > Port on the downstream side?
> > > > >
> > > > > Correct: Type 0 and Type 1 devices, and any number of Switch ports as
> > > > > it's usually pinned out to physical slot.
> > > > >
> > > > > > Does this patch enable ASPM on this magic Link between VMD and the
> > > > > > next device? Configuring ASPM correctly requires knowledge and knobs
> > > > > > from both ends of the Link, and apparently we don't have those for the
> > > > > > VMD end.
> > > > >
> > > > > VMD itself doesn't have the link to it's domain. It's really just the
> > > > > config/mmio aperture and MSI/X remapper. The PCIe link is between the
> > > > > Type 0 and Type 1 devices on the VMD domain. So fortunately the VMD
> > > > > itself is not the upstream part of the link.
> > > > >
> > > > > > Or is it for Links deeper in the hierarchy? I assume those should
> > > > > > just work already, although there might be issues with latency
> > > > > > computation, etc., because we may not be able to account for the part
> > > > > > of the path above VMD.
> > > > >
> > > > > That's correct. This is for the links within the domain itself, such as
> > > > > between a type 0 and NVMe device.
> > > >
> > > > OK, great. So IIUC, below the VMD, there is a Root Port, and the Root
> > > > Port has a link to some Endpoint or Switch, e.g., an NVMe device. And
> > > > we just want to enable ASPM on that link.
> > > >
> > > > That should not be a special case; we should be able to make this so
> > > > it Just Works. Based on this patch, I guess the reason it doesn't
> > > > work is because link->aspm_enabled for that link isn't set correctly.
> > > >
> > > > So is this just a consequence of us depending on the initial Link
> > > > Control value from BIOS? That seems like something we shouldn't
> > > > really depend on.
> Seems like a good idea, that it should instead be quirked if ASPM is
> found unusable on a link. Though I'm not aware of how many platforms
> would require a quirk..
>
> > > >
> > > That's the crux. There's always pcie_aspm=force.
> > > Something I've wondered is if there is a way we could 'discover' if the
> > > link is ASPM safe?
> >
> > Sure. Link Capabilities is supposed to tell us that. If aspm.c
> > depends on the BIOS settings, I think that's a design mistake.
> >
> > But what CONFIG_PCIEASPM_* setting are you using? The default
> > is CONFIG_PCIEASPM_DEFAULT, which literally means "Use the BIOS
> > defaults". If you're using that, and BIOS doesn't enable ASPM below
> > VMD, I guess aspm.c will leave it disabled, and that seems like it
> > would be the expected behavior.
> >
> > Does "pcie_aspm=force" really help you? I don't see any uses of it
> > that should apply to your situation.
> >
> > Bjorn
>
> No you're right. I don't think we need pcie_aspm=force, just the policy
> configuration.

I'm not sure where we're at here.

If the kernel is built with CONFIG_PCIEASPM_DEFAULT=y (which means
"use the BIOS defaults"), and the BIOS doesn't enable ASPM on these
links below VMD, then Linux will leave things alone. I think that's
working as intended.

If desired, we should be able to enable ASPM using sysfs in that case.

We have a pci_disable_link_state() kernel interface that drivers can
use to *disable* ASPM for their device. But I don't think there's any
corresponding interface for drivers to *enable* ASPM. Maybe that's an
avenue to explore?

Bjorn