On 25/07/2023 21.02, Haiyang Zhang wrote:

quoted

-----Original Message-----
From: Jesper Dangaard Brouer <redacted>
Sent: Tuesday, July 25, 2023 2:01 PM

quoted

quoted

Our driver is using NUMA 0 by default, so I implicitly assign NUMA node id
to zero during pool init.

And, if the IRQ/CPU affinity is changed, the page_pool_nid_changed()
will update the nid for the pool. Does this sound good?

Also, since our driver is getting the default node from here:
	gc->numa_node = dev_to_node(&pdev->dev);
I will update this patch to set the default node as above, instead of implicitly
assigning it to 0.

In that case, I agree that it make sense to use dev_to_node(&pdev->dev),
like:
	pprm.nid = dev_to_node(&pdev->dev);

Driver must have a reason for assigning gc->numa_node for this hardware,
which is okay. That is why page_pool API allows driver to control this.

But then I don't think you should call page_pool_nid_changed() like

	page_pool_nid_changed(rxq->page_pool, numa_mem_id());

Because then you will (at first packet processing event) revert the
dev_to_node() setting to use numa_mem_id() of processing/running CPU.
(In effect this will be the same as setting NUMA_NO_NODE).

I know, mlx5 do call page_pool_nid_changed(), but they showed benchmark
numbers that this was preferred action, even-when sysadm had
"misconfigured" the default smp_affinity RX-processing to happen on a
remote NUMA node.  AFAIK mlx5 keeps the descriptor rings on the
originally configured NUMA node that corresponds to the NIC PCIe slot.

In mana_gd_setup_irqs(), we set the default IRQ/CPU affinity to gc->numa_node
too, so it won't revert the nid initial setting.

Currently, the Azure hypervisor always indicates numa 0 as default. (In
the future, it will start to provide the accurate default dev node.) When a
user manually changes the IRQ/CPU affinity for perf tuning, we want to
allow page_pool_nid_changed() to update the pool. Is this OK?

If I were you, I would wait with the page_pool_nid_changed()
"optimization" and do a benchmark mark to see if this actually have a
benefit.  (You can do this in another patch).  (In a Azure hypervisor
environment is might not be the right choice).

This reminds me, do you have any benchmark data on the improvement this
patch (using page_pool) gave?

--Jesper