RE: [RFC PATCH 1/2] sched/topology: Get rid of NUMA overlapping groups

[Valentin Schneider]

Hi Barry,

On 08/02/21 10:04, Song Bao Hua (Barry Song) wrote:
>> -----Original Message-----
>> From: Valentin Schneider [mailto:valentin.schneider@xxxxxxx]

>
> Hi Valentin,
>
> While I like your approach, this will require more time
> to evaluate possible influence as the approach also affects
> all machines without 3-hops issue. So x86 platforms need to
> be tested and benchmark is required.
>
> What about we firstly finish the review of "grandchild" approach
> v2 and have a solution for kunpeng920 and Sun Fire X4600-M2
> while not impacting other machines which haven't 3-hops issues
> first?
>

I figured I'd toss this out while the iron was hot (and I had the topology
crud paged in), but I ultimately agree that it's better to first go with
something that fixes the diameter > 2 topologies and leaves the other ones
untouched, which is exactly what you have.

> I would appreciate very much if you could comment on v2:
> https://lore.kernel.org/lkml/20210203111201.20720-1-song.bao.hua@xxxxxxxxxxxxx/
>

See my comment below on domain degeneration; with that taken care of I
would say it's good to go. Have a look at what patch1+patch3 squashed
together looks like, passing the right sd to init_overlap_sched_group()
looks a bit neater IMO.

>> +static struct sched_domain *find_node_domain(struct sched_domain *sd)
>> +{
>> +	struct sched_domain *parent;
>> +
>> +	BUG_ON(!(sd->flags & SD_NUMA));
>> +
>> +	/* Get to the level above NODE */
>> +	while (sd && sd->child) {
>> +		parent = sd;
>> +		sd = sd->child;
>> +
>> +		if (!(sd->flags & SD_NUMA))
>> +			break;
>> +	}
>> +	/*
>> +	 * We're going to create cross topology level sched_group_capacity
>> +	 * references. This can only work if the domains resulting from said
>> +	 * levels won't be degenerated, as we need said sgc to be periodically
>> +	 * updated: it needs to be attached to the local group of a domain
>> +	 * that didn't get degenerated.
>> +	 *
>> +	 * Of course, groups aren't available yet, so we can't call the usual
>> +	 * sd_degenerate(). Checking domain spans is the closest we get.
>> +	 * Start from NODE's parent, and keep going up until we get a domain
>> +	 * we're sure won't be degenerated.
>> +	 */
>> +	while (sd->parent &&
>> +	       cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) {
>> +		sd = parent;
>> +		parent = sd->parent;
>> +	}
>
> So this is because the sched_domain which doesn't contribute to scheduler
> will be destroyed during cpu_attach_domain() since sd and parent span
> the seam mask?
>

Yes; let's take your topology for instance:

node   0   1   2   3
    0:  10  12  20  22
    1:  12  10  22  24
    2:  20  22  10  12
    3:  22  24  12  10

      2       10      2
  0 <---> 1 <---> 2 <---> 3


Domains for node1 will look like (before any fixes are applied):

NUMA<=10: span=1   groups=(1)
NUMA<=12: span=0-1 groups=(1)->(0)
NUMA<=20: span=0-1 groups=(0,1)
NUMA<=22: span=0-2 groups=(0,1)->(0,2-3)
NUMA<=24: span=0-3 groups=(0-2)->(0,2-3)

As you can see, the domain representing distance <= 20 will be degenerated
(it has a single group). If we were to e.g. add some more nodes to the left
of node0, then we would trigger the "grandchildren logic" for node1 and
would end up creating a reference to node1 NUMA<=20's sgc, which is a
mistake: that domain will be degenerated, and that sgc will never be
updated. The right thing to do here would be reference node1 NUMA<=12's
sgc, which the above snippet does.

>> +
>> +	return parent;
>> +}
>> +