Matthew Garrett responds to the ASPM power regression

ASPM's an interesting technology. PCIe is clocked at at least 2.5GHz, 
and later versions run at 5GHz. Running that clock takes a bunch of 
power, and a lot of the time the link will be idle. ASPM lets this be 
powered down at runtime, with the only (theoretical!) cost being some 
additional latency in bringing the link back up.

The implementation of ASPM is down to the hardware (the PCIe root and 
the endpoint handshake the transition between states), but policy is 
determined by setting bits in PCI configuration space. This can be done 
by either the OS or the BIOS. The PCIe specification defines how ASPM 
setup should occur, provides an algorithm for determining whether the 
latency will be greater than the maximum permitted and gives absolutely 
no indication under what circumstances ASPM should be enabled.

Because this is a complicated feature that will only be tested in 
hardware if something explicitly enables it, it obviously didn't work in 
the majority of early PCIe hardware. As a result, there's several 
heuristics and mechanisms for indicating the presence or absence of 
working ASPM support. Microsoft test for the presence of one of the bits 
that was only added late in the PCIe 1.1 specification, and if that's 
not present don't enable ASPM. We implemented the same policy. Microsoft 
will also not touch any PCIe configuration bits (including ASPM) if the 
ACPI/OS handshake for PCIe control doesn't grant full control of PCIe 
features to the OS. We've also implemented this. Drivers can also 
implement blacklisting, and there's support for this in Linux in the 
form of pcie_disable_link_state.

The final mechanism is a bit in the ACPI tables. If this bit is set, the 
platform is indicating to the OS that it doesn't support ASPM. In the 
past we took that to mean that we simply shouldn't touch the ASPM bits. 
However, it turns out that there's some systems where the BIOS has 
enabled ASPM itself, set the "ASPM unsupported" bit and then the 
hardware falls over when an ASPM transition occurs. The most 
straightforward thing to assume was that the BIOS was stupid (which 
is, to be fair, my default assumption) and shouldn't have enabled ASPM. 
So, since that patch, we clear the ASPM state when the BIOS indicates 
that the platform doesn't support ASPM.

So, is this the right thing to do? We don't know. The ACPI spec doesn't 
say anything about whether BIOS state should be retained or cleared, and 
the PCIe spec doesn't refer to the ACPI spec at all. Microsoft don't 
document how they behave in the presence of this bit. No BIOS vendors 
have yet seen fit to share their knowledge with us. THe patch fixed some 
hardware. However, it also disabled ASPM on some other machines where it 
appeared to work fine.

Of course, "Appeared" may be the operative word here. ASPM failures in 
hardware may be triggered by specific timing issues and may seem like 
random lockups that nobody ever tracks down. Some of the machines where 
we're seeing increased power consumption may also now be more stable. Or 
they may just be chewing more power. It's hard to tell.

What alternatives are there? We could keep the status quo and add driver 
whitelisting for hardware setups that are known to work. The problem is 
that even where we have specifications for the hardware, we often don't 
have the errata lists. We don't know for sure whether it works or not. 
We could revert this patch and add more driver blacklisting. But then we 
need to track down every device that doesn't work. Or, it's possible 
that the original code was correct and Linux simply programs the 
hardware differently, triggering ASPM issues that aren't seen elsewhere.

Right now the kernel takes the conservative approach. Users can override 
this at boot time with pcie_aspm=force. The only safe approach is to 
have more feedback from the vendors who set this bit as to what their 
expectations are. And, so far, that's not something we've had a great 
deal of luck obtaining.

-- 
Matthew Garrett | mjg59@srcf.ucam.org

Note: you can avoid this step in the future by logging into your LWN account.