Message-ID: <a5779c78-eb82-16ae-3f05-16f132f29a67@huawei.com>
Date:   Mon, 9 May 2022 20:20:23 +0800
From:   Yicong Yang <yangyicong@...wei.com>
To:     Bjorn Helgaas <helgaas@...nel.org>
CC:     <yangyicong@...ilicon.com>, <bhelgaas@...gle.com>,
        <rafael@...nel.org>, <linux-pci@...r.kernel.org>,
        <linux-acpi@...r.kernel.org>, <lenb@...nel.org>,
        <linux-kernel@...r.kernel.org>, <linuxarm@...wei.com>
Subject: Re: [PATCH] PCI/ACPI: Always advertise ASPM support if
 CONFIG_PCIEASPM=y

On 2022/5/6 2:41, Bjorn Helgaas wrote:
> On Thu, May 05, 2022 at 08:36:42PM +0800, Yicong Yang wrote:
>> On 2022/5/4 6:38, Bjorn Helgaas wrote:
>>> On Mon, Apr 25, 2022 at 03:06:34PM +0800, Yicong Yang wrote:
>>>> When we have CONFIG_PCIEASPM enabled it means OS can always support ASPM no
>>>> matter user have disabled it through pcie_aspm=off or not. But currently we
>>>> won't advertise ASPM support in _OSC negotiation if user disables it, which
>>>> doesn't match the fact. This will also have side effects that other PCIe
>>>> services like AER and hotplug will be disabled as ASPM support is required
>>>> and we won't negotiate other services if ASPM support is absent.
>>>>
>>>> So this patch makes OS always advertising ASPM support if CONFIG_PCIEASPM=y.
>>>> It intends no functional change to pcie_aspm=off as it will still mark
>>>> aspm_disabled=1 and aspm_support_enabled=false, driver will check these
>>>> status before configuring ASPM.
>>>>
>>>> Tested this patch with pcie_aspm=off:
>>>> estuary:/$ dmesg | egrep -i "aspm|osc"
>>>> [    0.000000] PCIe ASPM is disabled
>>>> [    8.706961] acpi PNP0A08:00: _OSC: OS supports [ExtendedConfig ASPM
>>>> ClockPM Segments MSI EDR HPX-Type3]
>>>> [    8.726032] acpi PNP0A08:00: _OSC: platform does not support [LTR]
>>>> [    8.742818] acpi PNP0A08:00: _OSC: OS now controls [PCIeHotplug PME
>>>> AER PCIeCapability DPC]
>>>> estuary:/sys/module/pcie_aspm/parameters$ cat policy
>>>> [default] performance powersave powersupersave
>>>> estuary:/sys/module/pcie_aspm/parameters$ echo powersave > policy
>>>> bash: echo: write error: Operation not permitted
>>>>
>>>> Cc: Rafael J. Wysocki <rafael@...nel.org>
>>>> Suggested-by: Bjorn Helgaas <bhelgaas@...gle.com>
>>>> [https://lore.kernel.org/linux-pci/20220407154257.GA235990@bhelgaas/]
>>>> Signed-off-by: Yicong Yang <yangyicong@...ilicon.com>
>>>> ---
>>>>  drivers/acpi/pci_root.c | 2 +-
>>>>  1 file changed, 1 insertion(+), 1 deletion(-)
>>>>
>>>> diff --git a/drivers/acpi/pci_root.c b/drivers/acpi/pci_root.c
>>>> index 6f9e75d14808..17e78582e633 100644
>>>> --- a/drivers/acpi/pci_root.c
>>>> +++ b/drivers/acpi/pci_root.c
>>>> @@ -393,7 +393,7 @@ static u32 calculate_support(void)
>>>>  	support |= OSC_PCI_HPX_TYPE_3_SUPPORT;
>>>>  	if (pci_ext_cfg_avail())
>>>>  		support |= OSC_PCI_EXT_CONFIG_SUPPORT;
>>>> -	if (pcie_aspm_support_enabled())
>>>> +	if (IS_ENABLED(CONFIG_PCIEASPM))
>>>
>>> Is there any way firmware could tell the difference between
>>> "CONFIG_PCIEASPM not set" and "CONFIG_PCIEASPM=y and booted with
>>> 'pcie_aspm=off'"?
>>>
>>> If not, why would we even check whether CONFIG_PCIEASPM is set?
>>
>> If we announce ASPM support when CONFIG_PCIEASPM=n it'll work as well
>> but negotiation and the log don't match the fact. We'll get misleading
>> messages that ASPM is supported by OS by it cannot be enable as there's
>> no driver.
>>
>> As mentioned by the PCIe Firmware Spec r3.3,
>> "ASPM Optionality supported
>>  The operating system sets this bit to 1 if it properly recognizes
>>  and manages ASPM support on PCI Express components which report
>>  support for ASPM L1 only in the ASPM Support field within the Link
>>  Capabilities Register. Otherwise, the operating system sets this
>>  bit to 0"
> 
> Yes.  I don't completely understand this bit, but I think it's related
> to the fact that L0s support was originally required for all links, so
> the only defined ASPM Support encodings were these:
> 
>   01b - L0s supported
>   11b - L0s and L1 supported
> 
> The "ASPM Optionality" ECN [1] of June 19, 2009, added these new
> encodings:
> 
>   00b - No ASPM support
>   10b - L1 supported
> 
> So I think the _OSC "ASPM Optionality Supported" bit tells the
> firmware that the OS supports this new possibility of devices that
> support L1 but not L0s.
> 
> Linux currently never sets the "ASPM Optionality Supported" bit, but
> it probably should, because I think we *do* support L1 even if L0s
> isn't supported.
> 

Yes, it sounds sensible to me. Actually I intended to refer BIT[1] which we're
currently using for advertising ASPM support in _OSC, but I copied the wrong
field...apologize.

"Active State Power Management supported
The operating system sets this bit to 1 if it natively supports configuration
of Active State Power Management registers in PCI Express devices. Otherwise,
the operating system sets this bit to 0."

IIUC, CONFIG_PCIEASPM=y means the OS *natively* support ASPM configuration so
we should set this bit to 1 even if we boot with pcie_aspm=off; otherwise the
OS has no native support of ASPM the bit should be 0. Currently the _OSC
negotiation seems to violent the spec a bit when booting with pcie_aspm=off
on a OS with CONFIG_PCIASPM=y.

>> When CONFIG_PCIEASPM=n we have no aspm driver and apparently cannot
>> support any ASPM features so we should set the bit to 0 to match the spec.
> 
> I think you're saying that firmware could not tell the difference, but
> the Linux log messages might be slightly misleading.  I assume you
> mean this message:
> 
>   acpi PNP0A08:00: _OSC: OS supports [ExtendedConfig ASPM ClockPM Segments MSI EDR HPX-Type3]
> 
> where we would claim that we support ASPM even when CONFIG_PCIEASPM is
> unset.
> 

yes. That's what I mean misleading.

> The purpose of that message is to expose what Linux is telling the
> platform via _OSC.  If we're telling the platform we support ASPM, I
> think the message should reflect that.
> 

agree.

> But I'm actually not sure there's real value in advertising ASPM
> support to the platform when CONFIG_PCIEASPM=y but we're booted with
> "pcie_aspm=off".  It sounds like this was found by using that option
> (even though it wasn't *needed*) and finding that Linux didn't request
> control of other PCIe services.  I don't know if that's worth
> changing.
> 

It's found in one of our test scenes that the AER is not worked. The issue
is implicit as AER is influenced by the ASPM which it shouldn't be. And
the implementation of pcie_aspm=off seems don't follow the spec. This patch
intends to make the code follow the spec in this corner case and by the way
fixes the issue I met. In the general cases there intends no change.

For the usage of pcie_aspm=off there may be cases of turning off ASPM when
the firmware grant the control to the OS. On some platform user may disable
ASPM through firmware by ACPI FADT, but on other platform OS may always get
the control of ASPM so this provide a way of disabling it. But I think it's
not proper to assume user doesn't want other services like AER either.

Since we haven't met any realistic issue on this boot option, I'd really
appreciate your suggestions on this.

Regards,
Yicong

Powered by blists - more mailing lists