Re: [PATCH] driver core: Fix test_async_driver_probe if NUMA is disabled

[Guenter Roeck <linux@roeck-us.net>]

On 11/27/19 1:24 PM, Alexander Duyck wrote:
> On Wed, 2019-11-27 at 12:24 -0800, Guenter Roeck wrote:
>> Since commit 57ea974fb871 ("driver core: Rewrite test_async_driver_probe
>> to cover serialization and NUMA affinity"), running the test with NUMA
>> disabled results in warning messages similar to the following.
>>
>> test_async_driver test_async_driver.12: NUMA node mismatch -1 != 0
>>
>> If CONFIG_NUMA=n, dev_to_node(dev) returns -1, and numa_node_id()
>> returns 0. Both are widely used, so it appears risky to change return
>> values. Augment the check with IS_ENABLED(CONFIG_NUMA) instead
>> to fix the problem.
>>
>> Cc: Alexander Duyck <alexander.h.duyck@linux.intel.com>
>> Fixes: 57ea974fb871 ("driver core: Rewrite test_async_driver_probe to cover serialization and NUMA affinity")
>> Signed-off-by: Guenter Roeck <linux@roeck-us.net>
>> ---
>>   drivers/base/test/test_async_driver_probe.c | 3 ++-
>>   1 file changed, 2 insertions(+), 1 deletion(-)
>>
>> diff --git a/drivers/base/test/test_async_driver_probe.c b/drivers/base/test/test_async_driver_probe.c
>> index f4b1d8e54daf..3bb7beb127a9 100644
>> --- a/drivers/base/test/test_async_driver_probe.c
>> +++ b/drivers/base/test/test_async_driver_probe.c
>> @@ -44,7 +44,8 @@ static int test_probe(struct platform_device *pdev)
>>   	 * performing an async init on that node.
>>   	 */
>>   	if (dev->driver->probe_type == PROBE_PREFER_ASYNCHRONOUS) {
>> -		if (dev_to_node(dev) != numa_node_id()) {
>> +		if (IS_ENABLED(CONFIG_NUMA) &&
>> +		    dev_to_node(dev) != numa_node_id()) {
>>   			dev_warn(dev, "NUMA node mismatch %d != %d\n",
>>   				 dev_to_node(dev), numa_node_id());
>>   			atomic_inc(&warnings);
> 
> I'm not sure that is really the correct fix. It might be better to test it
> against NUMA_NO_NODE and then if it is not that make sure that it matches
> the node ID. Adding the check against NUMA_NO_NODE would resolve the issue
> for cases where the device might be assigned to multiple NUMA nodes.
> 
I think you are suggesting that dev_to_node(dev) might return NUMA_NO_NODE
even on systems with CONFIG_NUMA enabled. I have no idea if that can happen.
The code in test_async_probe_init() seems to suggest that the node is set
to a valid node id for all asynchronous nodes, so I don't immediately see
how that could be the case. I may be missing something, of course.

Thanks,
Guenter