Date:   Sat, 7 Sep 2019 10:10:34 +0800
From:   Yunsheng Lin <linyunsheng@...wei.com>
To:     Greg KH <gregkh@...uxfoundation.org>
CC:     <rafael@...nel.org>, <linux-kernel@...r.kernel.org>,
        <peterz@...radead.org>, <mingo@...nel.org>, <mhocko@...nel.org>,
        <linuxarm@...wei.com>
Subject: Re: [PATCH RFC] driver core: ensure a device has valid node id in
 device_add()

On 2019/9/6 22:00, Greg KH wrote:
> 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@...wei.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.

The above isn't related to my problem really.

It just that there may be three fields that can indicate
the node id of a cpu:
per_cpu(numa_node), cpu->node_id and cpu->dev.numa_node

The per_cpu(numa_node) may be used for the fast path, and
cpu->node_id does not seems to be used, so it can be removed
when cpu->dev.numa_node is there.

Anyway, this is different problem here, maybe a separate patch
to "fix" it or clean it up if the above makes sense to you.

Sorry for the confusion.

> 
> 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?

Yes

Powered by blists - more mailing lists