Re: [PATCH RFC] driver core: ensure a device has valid node id in device_add()

[Greg KH <gregkh@linuxfoundation.org>]

On Fri, Sep 06, 2019 at 04:21:47PM +0800, Yunsheng Lin wrote:
> On 2019/9/6 14:52, Greg KH wrote:
> > On Fri, Sep 06, 2019 at 02:41:36PM +0800, Yunsheng Lin wrote:
> >> On 2019/9/5 15:33, Greg KH wrote:
> >>> On Thu, Sep 05, 2019 at 02:48:24PM +0800, Yunsheng Lin wrote:
> >>>> On 2019/9/5 13:57, Greg KH wrote:
> >>>>> On Thu, Sep 05, 2019 at 09:33:50AM +0800, Yunsheng Lin wrote:
> >>>>>> Currently a device does not belong to any of the numa nodes
> >>>>>> (dev->numa_node is NUMA_NO_NODE) when the FW does not provide
> >>>>>> the node id and the device has not no parent device.
> >>>>>>
> >>>>>> According to discussion in [1]:
> >>>>>> Even if a device's numa node is not set by fw, the device
> >>>>>> really does belong to a node.
> >>>>>>
> >>>>>> This patch sets the device node to node 0 in device_add() if
> >>>>>> the fw has not specified the node id and it either has no
> >>>>>> parent device, or the parent device also does not have a valid
> >>>>>> node id.
> >>>>>>
> >>>>>> There may be explicit handling out there relying on NUMA_NO_NODE,
> >>>>>> like in nvme_probe().
> >>>>>>
> >>>>>> [1] https://lkml.org/lkml/2019/9/2/466
> >>>>>>
> >>>>>> Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
> >>>>>> ---
> >>>>>>  drivers/base/core.c  | 17 ++++++++++++++---
> >>>>>>  include/linux/numa.h |  2 ++
> >>>>>>  2 files changed, 16 insertions(+), 3 deletions(-)
> >>>>>>
> >>>>>> diff --git a/drivers/base/core.c b/drivers/base/core.c
> >>>>>> index 1669d41..466b8ff 100644
> >>>>>> --- a/drivers/base/core.c
> >>>>>> +++ b/drivers/base/core.c
> >>>>>> @@ -2107,9 +2107,20 @@ int device_add(struct device *dev)
> >>>>>>  	if (kobj)
> >>>>>>  		dev->kobj.parent = kobj;
> >>>>>>  
> >>>>>> -	/* use parent numa_node */
> >>>>>> -	if (parent && (dev_to_node(dev) == NUMA_NO_NODE))
> >>>>>> -		set_dev_node(dev, dev_to_node(parent));
> >>>>>> +	/* use parent numa_node or default node 0 */
> >>>>>> +	if (!numa_node_valid(dev_to_node(dev))) {
> >>>>>> +		int nid = parent ? dev_to_node(parent) : NUMA_NO_NODE;
> >>>>>
> >>>>> Can you expand this to be a "real" if statement please?
> >>>>
> >>>> Sure. May I ask why "? :" is not appropriate here?
> >>>
> >>> Because it is a pain to read, just spell it out and make it obvious what
> >>> is happening.  You write code for developers first, and the compiler
> >>> second, and in this case, either way is identical to the compiler.
> >>>
> >>>>>> +
> >>>>>> +		if (numa_node_valid(nid)) {
> >>>>>> +			set_dev_node(dev, nid);
> >>>>>> +		} else {
> >>>>>> +			if (nr_node_ids > 1U)
> >>>>>> +				pr_err("device: '%s': has invalid NUMA node(%d)\n",
> >>>>>> +				       dev_name(dev), dev_to_node(dev));
> >>>>>
> >>>>> dev_err() will show you the exact device properly, instead of having to
> >>>>> rely on dev_name().
> >>>>>
> >>>>> And what is a user to do if this message happens?  How do they fix this?
> >>>>> If they can not, what good is this error message?
> >>>>
> >>>> If user know about their system's topology well enough and node 0
> >>>> is not the nearest node to the device, maybe user can readjust that by
> >>>> writing the nearest node to /sys/class/pci_bus/XXXX/device/numa_node,
> >>>> if not, then maybe user need to contact the vendor for info or updates.
> >>>>
> >>>> Maybe print error message as below:
> >>>>
> >>>> dev_err(dev, FW_BUG "has invalid NUMA node(%d). Readjust it by writing to sysfs numa_node or contact your vendor for updates.\n",
> >>>> 	dev_to_node(dev));
> >>>
> >>> FW_BUG?
> >>>
> >>> Anyway, if you make this change, how many machines start reporting this
> >>> error? 
> >>
> >> Any machines with more than one numa node will start reporting this error.
> >>
> >> 1) many virtual deivces maybe do not set the node id before calling
> >>    device_register(), such as vfio, tun, etc.
> >>
> >> 2) struct cpu has a dev, but does not set the dev' node according to
> >>    cpu_to_node().
> >>
> >> 3) Many platform Device also do not have a node id provided by FW.
> > 
> > Then this patch is not ok, as you are flooding the kernel log saying the
> > system is "broken" when this is just what it always has been like.  How
> > is anyone going to "fix" things?
> 
> cpu->node_id does not seem to be used, maybe we can fix the cpu device:
> 
> diff --git a/drivers/base/cpu.c b/drivers/base/cpu.c
> index cc37511d..ad0a841 100644
> --- a/drivers/base/cpu.c
> +++ b/drivers/base/cpu.c
> @@ -41,7 +41,7 @@ static void change_cpu_under_node(struct cpu *cpu,
>         int cpuid = cpu->dev.id;
>         unregister_cpu_under_node(cpuid, from_nid);
>         register_cpu_under_node(cpuid, to_nid);
> -       cpu->node_id = to_nid;
> +       set_dev_node(&cpu->dev, to_nid);
>  }
> 
>  static int cpu_subsys_online(struct device *dev)
> @@ -367,7 +367,7 @@ int register_cpu(struct cpu *cpu, int num)
>  {
>         int error;
> 
> -       cpu->node_id = cpu_to_node(num);
> +       set_dev_node(&cpu->dev, cpu_to_node(num));
>         memset(&cpu->dev, 0x00, sizeof(struct device));
>         cpu->dev.id = num;
>         cpu->dev.bus = &cpu_subsys;
> diff --git a/include/linux/cpu.h b/include/linux/cpu.h
> index fcb1386..9a6fc51 100644
> --- a/include/linux/cpu.h
> +++ b/include/linux/cpu.h
> @@ -24,7 +24,6 @@ struct device_node;
>  struct attribute_group;
> 
>  struct cpu {
> -       int node_id;            /* The node which contains the CPU */
>         int hotpluggable;       /* creates sysfs control file if hotpluggable */
>         struct device dev;
>  };

I have no idea what you are trying to do here, it feels like you are
flailing around trying to set something that almost no bios/firmware
sets or cares about.

If setting the proper node is a requirement, make sure your firmware
does this and then all should be fine.  Otherwise just use the default
node like what happens today, right?

thanks,

greg k-h