Re: [PATCH] pci, add sysfs numa_node write function

From: Alexander Duyck
Date: Thu Oct 16 2014 - 13:04:16 EST



On 10/16/2014 09:07 AM, Prarit Bhargava wrote:
On 10/16/2014 10:44 AM, Alexander Duyck wrote:
On 10/16/2014 05:32 AM, Prarit Bhargava wrote:
On 10/15/2014 05:20 PM, Bjorn Helgaas wrote:
On Wed, Oct 15, 2014 at 1:47 PM, Prarit Bhargava <prarit@xxxxxxxxxx> wrote:
On 10/15/2014 03:23 PM, Bjorn Helgaas wrote:
Hi Prarit,

On Wed, Oct 15, 2014 at 1:05 PM, Prarit Bhargava <prarit@xxxxxxxxxx> wrote:
Consider a multi-node, multiple pci root bridge system which can be
configured into one large node or one node/socket. When configuring the
system the numa_node value for each PCI root bridge is always set
incorrectly to -1, or NUMA_NO_NODE, rather than to the node value of each
socket. Each PCI device inherits the numa value directly from it's parent
device, so that the NUMA_NO_NODE value is passed through the entire PCI
tree.

Some new drivers, such as the Intel QAT driver, drivers/crypto/qat,
require that a specific node be assigned to the device in order to
achieve maximum performance for the device, and will fail to load if the
device has NUMA_NO_NODE.
It seems ... unfriendly for a driver to fail to load just because it
can't guarantee maximum performance. Out of curiosity, where does
this actually happen? I had a quick look for NUMA_NO_NODE and
module_init() functions in drivers/crypto/qat, and I didn't see the
spot.
The whole point of the Intel QAT driver is to guarantee max performance. If
that is not possible the driver should not load (according to the thread
mentioned below)
This is just short-sighted thinking. The fact that the PCI device advertises -1
means that either the BIOS isn't configured, or the PCI slots are shared as was
the case on some Nehalem systems where the IOH was shared between two sockets.
Yes ...

I suspect that this driver doesn't even take that into account as it was likely
only written for Sandy Bridge architectures.

Nope. New hardware. The issue is that there is only a performance impact if
local node memory is used, o/w the claim is that the performance drops to that
of doing software crypto.

Yes, but the problem is what is considered a local node? In the case of the Nehalem architecture it wasn't uncommon to have two sockets share an IOH. As a result each node would have its own memory, but the PCI devices were shared. That platform would fail this test, but each node has its own memory and I suspect the QPI overhead wouldn't be that bad since the link is dedicated to the IOH connection.

What I am getting at is that this driver was likely intended for specifically the current and follow-on generations of Xeon processors. As such it deciding to reject all NUMA_NO_NODE configuration may be acceptable for it since they have a very specific architecture and CPU in mind. For the workaround if you are going to try and do something as generic, you will need to be mindful of how the other architectures deal with NUMA and PCI.

To use this, one can do

echo 3 > /sys/devices/pci0000:ff/0000:ff:1f.3/numa_node

to set the numa node for PCI device 0000:ff:1f.3.
It definitely seems wrong that we don't set the node number correctly.
pci_acpi_scan_root() sets the node number by looking for a _PXM method
that applies to the host bridge. Why does that not work in this case?
Does the BIOS not supply _PXM?
Yeah ... unfortunately the BIOS is broken in this case. And I know what you're
thinking ;) -- why not get the BIOS fixed? I'm through relying on BIOS fixes
which can take six months to a year to appear in a production version... I've
been bitten too many times by promises of BIOS fixes that never materialize.
Yep, I understand. The question is how we implement a workaround so
it doesn't become the accepted way to do things. Obviously we don't
want people manually grubbing through numactl/lspci output or writing
shell scripts to do things that *should* happen automatically.
I'd say if nothing else we should flag the system as tainted as soon as we start
overwriting BIOS/ACPI configured values with sysfs. This is one of the reasons
for the TAINT_FIRMWARE_WORKAROUND even existing.

I was thinking that I could modify the patch to do this, but I'd rather
investigate Bjorn's suggestion first. I think his approach has some merits, but
I will definitely TAINT if I go with that approach too.

If we are doing any sort of workaround we should mark the system as being tainted. Otherwise we are just going to let whatever workaround you implement be accepted and nobody will ever bother to fix the BIOS in this generation or the next.

Somewhere in the picture there needs to be a feedback loop that
encourages the vendor to fix the problem. I don't see that happening
yet. Having QAT fail because the platform didn't supply the
information required to make it work would be a nice loop. I don't
want to completely paper over the problem without providing some other
kind of feedback at the same time.
Okay -- I see what you're after here and I completely agree with it. But
sometimes I feel like I banging on a silent drum with some of these companies
about this stuff :( My frustration with these companies is starting to show I
guess...
Just how visible is the QAT driver load failure? I has a similar issue with DCA
not being configured in a number of BIOSes and it wasn't until I made the issue
painfully visible with TAINT_FIRMWARE_WORKAROUND that I started to see any
traction on getting this fixed in the BIOSes.

Just a couple of printks to the screen.

Yeah, taint would definitely be preferred. Then if something has a panic later because we didn't work around this correctly we should be able to see that this is a tainted system in the panic dump. I did the same thing for DSA.

We would need to sort out the systems that actually have bad BIOSes versus just
being configured without PCI slots directly associated with any given NUMA node
since there are systems where that is a valid configuration.
The problem is NUMA_NO_NODE, which as you point out can be a valid
configuration. So in some cases system designers may have intentionally done
this (Can anyone think of a valid reason to leave off the _PXM, or have it
assigned to NUMA_NO_NODE?), so the previous statement about having an opt-in,
then attempting to calculate the node location, and now TAINTING might be a good
direction to move in.

That is my thought. We would just need to come up with a good way of sorting these platforms out and that can be the challenge. From past experience I have even seen some systems where the memory layout wasn't configured correctly in the BIOS/ACPI tables so this is pretty much just a crapshoot as to if we can figure out the actual layout or not via software. Maybe this could be identified with TAINT_OVERRIDDEN_ACPI_TABLE since we are having to hand code an ACPI table.

OTOH ... what do we do about older unsupported hardware that won't have new BIOS
releases? Those would basically say "Go fix your BIOS" and there's nothing that
could be done :/. All those users see is a loud warning...

I'm not saying we cannot provide a workaround for it, but at the same time as you have pointed out this is mostly just a performance issue due to the BIOS and we could be introducing new issues by trying to fix it in the kernel.

The loud warning is both for the users and those of us that are stuck debugging the issues they find when their platform doesn't run as expected and triggers a panic. You have to keep in mind that by working around things we run the risk of introducing new issues and we have to know when we have entered one of those cases.

You're probably aware of [1], which was the same problem. Apparently
it was originally reported to RedHat as [2] (which is private, so I
can't read it). That led to a workaround hack for some AMD systems
[3, 4].
Yeah ... part of me was thinking that maybe I should do something like
the above but I didn't know how you'd feel about expanding that hack. I'll look
into it. I'd prefer it to be opt-in with a kernel parameter.

P.
Are you thinking something like a "pci=assign-numa"? The problem is there
doesn't seem to be a good way to currently determine the NUMA layout without the
information being provided by the BIOS/ACPI tables, and we probably don't want
to be creating a definition of the NUMA layout per platform.
Well ... let me think about this for a bit. The big issue is what happens
during socket hot-add events and the PCI root bridges that are added at that
time. It may not be possible to come up with a correct calculation :( but let
me give it a shot. IIRC the node-to-socket map should be static...

P.

I don't know if there is a good way to resolve all this programmatically since I have seen a number of issues where even the NUMA memory layout is messed up. The sysfs might be a good approach however I would suggest tainting the kernel at that point as you are overwritting pre-defined BIOS/ACPI table values and that may have certain side effects if things don't get configured correctly.

Thanks,

Alex
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/