Re: [PATCH] cxl/test: Map mock device nodes to an online node

[Alison Schofield <alison.schofield@intel.com>]

On Thu, Jul 23, 2026 at 11:06:30AM -0700, Davidlohr Bueso wrote:
> cxl_mem_init() assigns alternating NUMA nodes to the mock memdevs
> to model multi-node topologies. On a single-node host node 1 has
> no pgdat, and the first node-aware allocation in an odd device's
> probe faults on a NULL zonelist:
> 
>   Oops: general protection fault, probably for non-canonical address 0xdffffc00000006e3
>   KASAN: probably user-memory-access in range [0x0000000000003718-0x000000000000371f]
>   Workqueue: async async_run_entry_fn
>   RIP: __alloc_frozen_pages_noprof+0x288
>    new_slab
>    ___slab_alloc
>    __kmalloc_node_track_caller_noprof
>    devm_kmalloc
>    cxl_mock_mem_probe
> 
> Map the requested node to an online one, as dax_hmem already does
> for its target node.
> 
> Fixes: cf1f6877b088 ("cxl/memdev: Add numa_node attribute")
> Fixes: e41c8452b9b2 ("tools/testing/cxl: Add a single-port host-bridge regression config")
> Fixes: c9435dbee119 ("tools/testing/cxl: Add an RCH topology")
> Signed-off-by: Davidlohr Bueso <dave@stgolabs.net>
> ---
> Found while regression testing unrelated work.

Hi Davidlohr,

Thanks for fixing up cxl-test!
My short response is:
	Reviewed-by: Alison Schofield <alison.schofield@intel.com>
	Tested-by: Alison Schofield <alison.schofield@intel.com>

And the long version is, I tested this and it failed somewhere
else. I have CONFIG_PINCTRL=y so the first node aware allocation 
lands in the device core, not in cxl_mock_mem_probe(). Looks like:

BUG: unable to handle page fault for address: 0000000000002d18
    RIP: 0010:__alloc_frozen_pages_noprof+0x120/0x310
     allocate_slab
     new_slab
     __slab_alloc_node
     __kmalloc_node_track_caller_noprof
     devm_kmalloc
     pinctrl_bind_pins
     really_probe

We (me) had not seen this because our cxl_test guests always have two or more
CPU-and-memory nodes, plus memory-only and nvdimm nodes on top. Node 1 is
online, so numa_map_to_online_node() is a no-op.

It reproduces with a single -numa node holding the CPUs and all memory.
That is one proximity domain, node_possible_map == {0}, no pgdat for node 1.

These settings appear as different presets for folks using run_qemu [1].
The 'tiny' option must be selected to hit this.

I'm thinking about how to work 'tiny' into the regular regression runs.

Good find Davidlohr!


-- Alison

[1] https://github.com/pmem/run_qemu


> 
>  tools/testing/cxl/test/cxl.c | 6 +++---
>  1 file changed, 3 insertions(+), 3 deletions(-)
> 
> diff --git a/tools/testing/cxl/test/cxl.c b/tools/testing/cxl/test/cxl.c
> index ef92dd35e030..5156516a82d2 100644
> --- a/tools/testing/cxl/test/cxl.c
> +++ b/tools/testing/cxl/test/cxl.c
> @@ -1748,7 +1748,7 @@ static int cxl_mem_init(void)
>  		if (!pdev)
>  			goto err_mem;
>  		pdev->dev.parent = &dport->dev;
> -		set_dev_node(&pdev->dev, i % 2);
> +		set_dev_node(&pdev->dev, numa_map_to_online_node(i % 2));
>  
>  		rc = cxl_mock_platform_device_add(pdev, &cxl_mem[i]);
>  		if (rc)
> @@ -1763,7 +1763,7 @@ static int cxl_mem_init(void)
>  		if (!pdev)
>  			goto err_single;
>  		pdev->dev.parent = &dport->dev;
> -		set_dev_node(&pdev->dev, i % 2);
> +		set_dev_node(&pdev->dev, numa_map_to_online_node(i % 2));
>  
>  		rc = cxl_mock_platform_device_add(pdev, &cxl_mem_single[i]);
>  		if (rc)
> @@ -1779,7 +1779,7 @@ static int cxl_mem_init(void)
>  		if (!pdev)
>  			goto err_rcd;
>  		pdev->dev.parent = &rch->dev;
> -		set_dev_node(&pdev->dev, i % 2);
> +		set_dev_node(&pdev->dev, numa_map_to_online_node(i % 2));
>  
>  		rc = cxl_mock_platform_device_add(pdev, &cxl_rcd[i]);
>  		if (rc)
> -- 
> 2.39.5
>