On Tue, May 02, 2017 at 12:02:53PM +0000, Patel, Mayurkumar wrote:

quoted

On Fri, Apr 21, 2017 at 2:46 AM, Patel, Mayurkumar
[off-list ref] wrote:

quoted

quoted

On 4/17/2017 12:38 PM, Bjorn Helgaas wrote:

quoted

quoted

Like you said, what do we do by default is the question. Should we opt
for safe like we are doing, or try to save some power.

I think safety is paramount.  Every user should be able to boot safely
without any kernel parameters.  We don't want users to have a problem
booting and then have to search for a workaround like booting with
"pcie_aspm=off".  Most users will never do that.

OK, no problem with leaving the behavior as it is.

My initial approach was #2. We knew this way that user had full control
over the ASPM policy by changing the BIOS option. Then, Mayurkumar
complained that ASPM is not enabled following a hotplug insertion to an
empty slot. That's when I switched to #3 as it sounded like a good thing
to have for us.

quoted

Here's a long-term strawman proposal, see what you think:

  - Deprecate CONFIG_PCIEASPM_DEFAULT, CONFIG_PCIEASPM_POWERSAVE, etc.
  - Default aspm_policy is POLICY_DEFAULT always.
  - POLICY_DEFAULT means Linux doesn't touch anything: if BIOS enabled
ASPM, we leave it that way; we leave ASPM disabled on hot-added
devices.

I am also ok with leaving the same behavior as now.  But still
following is something open I feel besides, Which may be there in
your comments redundantly.  The current problem is,
pcie_aspm_exit_link_state() disables the ASPM configuration even
if POLICY_DEFAULT was set.

We call pcie_aspm_exit_link_state() when removing an endpoint.  When
we remove an endpoint, I think disabling ASPM is the right thing to
do.  The spec (PCIe r3.1, sec 5.4.1.3) says "Software must not enable
L0s in either direction on a given Link unless components on both
sides of the Link each support L0s; otherwise, the result is
undefined."

Yes, you are right and per spec also it makes sense that ASPM needs
to be disabled.  But, if POLICY_DEFAULT is set then, shouldn't BIOS
take care of disabling ASPM?

No, I don't think so.  POLICY_DEFAULT is a Linux thing and BIOS
doesn't know anything about it.

ASPM can be configured by BIOS before handoff to Linux, but after
handoff it should be managed either entirely by BIOS or entirely by
Linux.  If BIOS wants to retain ASPM control, it would have to tell
the OS *not* to use ASPM, and it would have to use ACPI hotplug.  In
this case, POLICY_DEFAULT is irrelevant because Linux shouldn't do
anything with ASPM.

But normally BIOS allows Linux to control ASPM, and we would use
native PCIe hotplug (pciehp) instead of ACPI hotplug, and BIOS has no
opportunity to enable or disable ASPM on hotplug events.

quoted

quoted

I am seeing already following problem(or may be influence) with
it. The Endpoint I have does not have does not have "Presence
detect change" mechanism. Hot plug is working with Link status
events.  When link is in L1 or L1SS and if EP is powered off, no
Link status change event are triggered (It might be the expected
behavior in L1 or L1SS).  When next time EP is powered on there
are link down and link up events coming one after other. BIOS
enables ASPM on Root port and Endpoint, but while processing link
status down, pcie_aspm_exit_link_state() clears the ASPM already
which were enabled by BIOS.  If we want to follow above approach
then shall we consider having something similar as following?

The proposal was to leave ASPM disabled on hot-added devices.  If
the endpoint was powered off and powered back on again, I think
that device looks like a hot-added device, doesn't it?

Yes, it is hot-added device. Also, I understand, for POLICY_DEFAULT,
OS would/should not touch ASPM(enable/disable), but BIOS could still
(enable/disable), right?

Maybe I'm misunderstanding your question.  There are two questions
here:

  1) Does the BIOS allow Linux to manage ASPM?

  2) If Linux does manage ASPM, what policy does it use?

If BIOS doesn't allow Linux to manage ASPM, POLICY_DEFAULT is
irrelevant.  If BIOS does allow Linux to manage ASPM, POLICY_DEFAULT
means Linux should use the settings made by BIOS.  The user could
select a different policy, and then Linux would change the ASPM
configuration accordingly.

Currently, what happens in my system is as following, (each 2nd
power cycle/hotplug of Endpoint disables ASPM):


First Power cycle (When ASPM L1 is already enabled): device gets
powered off -> there are no Link status events, so no pcie hotplug
interrupt and pcie_aspm_exit_link_state() triggered.  

If the Downstream Port leading to your Endpoint is hotplug capable,
doesn't the spec require that it can report link state changes (PCIe
r3.1, sec 7.8.6, 7.8.10, 7.8.11)?

When the device gets powered on again -> Link down/Link up events
are coming back to back.  First Link down is served. (BIOS checks
for the Link status and enables ASPM already, as the device is
actually powered back). OS calls pcie_aspm_exit_link_state() and
ASPM gets disabled by OS.

Second Power cycle (When ASPM L1 is disabled after above): device
gets powered off -> there are link status events, pcie hotplug
interrupt is triggered and pcie_aspm_exit_link_state() triggered.
OS disables ASPM. BIOS checks Link status and disables ASPM too.
When the device gets powered on -> BIOS enables ASPM and as this is
pcie hotplug insertion, OS does not interfere and we have ASPM
enabled.

I don't understand this sequence.  If we're using native PCIe hotplug,
BIOS should not be involved to enable ASPM when the device is powered
on.

The above sequence happens each 2nd power cycle of the hotplug
device.

So One could still argue if POLICY_DEFAULT is set, then why OS
disables ASPM if it is not meant to touch configuration.  This is
why I proposed following kind of change, so that OS would not touch
ASPM, if POLICY_DEFAULT is set.  Also, With the below change,
everything relies on BIOS for ASPM when POLICY_DEFAULT is set and I
see above problem gets resolved. Also, the existing ASPM behavior
does not have impact, unless specific BIOS does not disable ASPM on
Root Port when device gets removed.