Re: [PATCH v6] numa: make node_to_cpumask_map() NUMA_NO_NODE aware

From: Greg KH
Date: Mon Oct 14 2019 - 05:25:15 EST


On Mon, Oct 14, 2019 at 04:00:46PM +0800, Yunsheng Lin wrote:
> On 2019/10/12 18:47, Greg KH wrote:
> > On Sat, Oct 12, 2019 at 12:40:01PM +0200, Greg KH wrote:
> >> On Sat, Oct 12, 2019 at 05:47:56PM +0800, Yunsheng Lin wrote:
> >>> On 2019/10/12 15:40, Greg KH wrote:
> >>>> On Sat, Oct 12, 2019 at 02:17:26PM +0800, Yunsheng Lin wrote:
> >>>>> add pci and acpi maintainer
> >>>>> cc linux-pci@xxxxxxxxxxxxxxx and linux-acpi@xxxxxxxxxxxxxxx
> >>>>>
> >>>>> On 2019/10/11 19:15, Peter Zijlstra wrote:
> >>>>>> On Fri, Oct 11, 2019 at 11:27:54AM +0800, Yunsheng Lin wrote:
> >>>>>>> But I failed to see why the above is related to making node_to_cpumask_map()
> >>>>>>> NUMA_NO_NODE aware?
> >>>>>>
> >>>>>> Your initial bug is for hns3, which is a PCI device, which really _MUST_
> >>>>>> have a node assigned.
> >>>>>>
> >>>>>> It not having one, is a straight up bug. We must not silently accept
> >>>>>> NO_NODE there, ever.
> >>>>>>
> >>>>>
> >>>>> I suppose you mean reporting a lack of affinity when the node of a pcie
> >>>>> device is not set by "not silently accept NO_NODE".
> >>>>
> >>>> If the firmware of a pci device does not provide the node information,
> >>>> then yes, warn about that.
> >>>>
> >>>>> As Greg has asked about in [1]:
> >>>>> what is a user to do when the user sees the kernel reporting that?
> >>>>>
> >>>>> We may tell user to contact their vendor for info or updates about
> >>>>> that when they do not know about their system well enough, but their
> >>>>> vendor may get away with this by quoting ACPI spec as the spec
> >>>>> considering this optional. Should the user believe this is indeed a
> >>>>> fw bug or a misreport from the kernel?
> >>>>
> >>>> Say it is a firmware bug, if it is a firmware bug, that's simple.
> >>>>
> >>>>> If this kind of reporting is common pratice and will not cause any
> >>>>> misunderstanding, then maybe we can report that.
> >>>>
> >>>> Yes, please do so, that's the only way those boxes are ever going to get
> >>>> fixed. And go add the test to the "firmware testing" tool that is based
> >>>> on Linux that Intel has somewhere, to give vendors a chance to fix this
> >>>> before they ship hardware.
> >>>>
> >>>> This shouldn't be a big deal, we warn of other hardware bugs all the
> >>>> time.
> >>>
> >>> Ok, thanks for clarifying.
> >>>
> >>> Will send a patch to catch the case when a pcie device without numa node
> >>> being set and warn about it.
> >>>
> >>> Maybe use dev->bus to verify if it is a pci device?
> >>
> >> No, do that in the pci bus core code itself, when creating the devices
> >> as that is when you know, or do not know, the numa node, right?
> >>
> >> This can't be in the driver core only, as each bus type will have a
> >> different way of determining what the node the device is on. For some
> >> reason, I thought the PCI core code already does this, right?
> >
> > Yes, pci_irq_get_node(), which NO ONE CALLS! I should go delete that
> > thing...
> >
> > Anyway, it looks like the pci core code does call set_dev_node() based
> > on the PCI bridge, so if that is set up properly, all should be fine.
> >
> > If not, well, you have buggy firmware and you need to warn about that at
> > the time you are creating the bridge. Look at the call to
> > pcibus_to_node() in pci_register_host_bridge().
>
> Thanks for pointing out the specific function.
> Maybe we do not need to warn about the case when the device has a parent,
> because we must have warned about the parent if the device has a parent
> and the parent also has a node of NO_NODE, so do not need to warn the child
> device anymore? like blew:
>
> @@ -932,6 +932,10 @@ static int pci_register_host_bridge(struct pci_host_bridge *bridge)
> list_add_tail(&bus->node, &pci_root_buses);
> up_write(&pci_bus_sem);
>
> + if (nr_node_ids > 1 && !parent &&

Why do you need to check this? If you have a parent, it's your node
should be set, if not, that's an error, right?

> + dev_to_node(bus->bridge) == NUMA_NO_NODE)
> + dev_err(bus->bridge, FW_BUG "No node assigned on NUMA capable HW. Please contact your vendor for updates.\n");
> +
> return 0;

Who set that bus->bridge node to NUMA_NO_NODE?
If that is set, the firmware is broken, as you say, but you need to tell
the user what firmware is broken.

Try something like this out and see what happens on your machine that
had things "broken". What does it say?

> Also, we do not need to warn about that in pci_device_add(), Right?
> Because we must have warned about the pci host bridge of the pci device.

That should be true, yes.

> I may be wrong about above because I am not so familiar with the pci.
>
> >
> > And yes, you need to do this all on a per-bus-type basis, as has been
> > pointed out. It's up to the bus to create the device and set this up
> > properly.
>
> Thanks.
> Will do that on per-bus-type basis.

Good luck, I don't really think that most, if any, of this is needed,
but hey, it's nice to clean it up where it can be :)

greg k-h