ACPI: NUMA: Only parse CFMWS at boot when CXL_ACPI is on
PATCH v2 1/1 ACPI: NUMA: Only parse CFMWS at boot when CXL_ACPI is on
On CXL platforms, the Static Resource Affinity Table (SRAT) may not cover memory affinity information for all the CXL memory regions. Since each CXL memory region is enumerated via a CXL Fixed Memory Window Structure (CFMWS), during early boot the kernel parses the CFMWS tables to find all CXL memory regions and sets a NUMA node for each of them. This memory affinity information of CXL memory regions is later used by the CXL ACPI driver.
The CFMWS table doesn't provide the memory affinity information either. Currently the kernel assigns a 'faked' NUMA node for each CXL memory region, starting from the next node of the highest node that is enumerated via the SRAT. This can potentially increase the maximum NUMA node ID of the platform ('nr_node_ids') a lot. E.g., on a GNR platform with 4 NUMA nodes and 18 CFMWS tables, this bumps the 'nr_node_ids' to 22.
Increasing the 'nr_node_ids' has side effects. For instance, it is widely used by the kernel for "highest possible NUMA node" based memory allocations. It also impacts userspace ABIs, e.g., some NUMA memory related system calls such as 'get_mempolicy' which requires 'maxnode' not being smaller than the 'nr_node_ids'.
Currently parsing CFMWS tables and assigning faked NUMA node at boot is done unconditionally. However, if the CXL ACPI driver is not enabled, there will be no user of such memory affinity information of CXL memory regions.
Change to only parsing the CFMWS tables at boot when CXL_ACPI is enabled in Kconfig to avoid the unnecessary cost of bumping up 'nr_node_ids'.
E.g., on the aforementioned GNR platform, the "Slab" in /proc/meminfo is reduced with this change (when CXL_ACPI is off):
w/ this change w/o
Slab 900488 kB 923660 kB
---
v1 -> v2:
- Use Dan's suggestion to simplify the diff:
https://lore.kernel.org/linux-cxl/[email protected]/
Hi Alison, Gregory,
I didn't add your RB since the code now is different from that you reviewed. Appreciate if you can take a look again and provide the tag if the patch looks good to you.
--- drivers/acpi/numa/srat.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-)
drivers/acpi/numa/srat.c +5 -2
@@ -654,8 +654,11 @@ int __init acpi_numa_init(void) } last_real_pxm = fake_pxm; fake_pxm++; -acpi_table_parse_cedt (ACPI_CEDT_TYPE_CFMWS , acpi_parse_cfmws, - &fake_pxm); + +/* No need to expand numa nodes if CXL is disabled */ +if (IS_ENABLED (CONFIG_CXL_ACPI )) +acpi_table_parse_cedt (ACPI_CEDT_TYPE_CFMWS , acpi_parse_cfmws, + &fake_pxm);if (cnt <0 )return cnt;
base-commit: 084f843093bee5563b179fd4b630122ba820e0c7
Signature · 1 line
2.53.0
Quoted 8 lines · Kai Huang
On CXL platforms, the Static Resource Affinity Table (SRAT) may not cover memory affinity information for all the CXL memory regions. Since each CXL memory region is enumerated via a CXL Fixed Memory Window Structure (CFMWS), during early boot the kernel parses the CFMWS tables to find all CXL memory regions and sets a NUMA node for each of them. This memory affinity information of CXL memory regions is later used by the CXL ACPI driver.
... snip ...
Quoted 2 lines
Signed-off-by Kai Huang <[email protected]>
lgtm
Quoted 30 lines · On Mon, 9 Mar 2026 11:23:13 +1300 Kai Huang <[email protected]> wrote:
On CXL platforms, the Static Resource Affinity Table (SRAT) may not cover memory affinity information for all the CXL memory regions. Since each CXL memory region is enumerated via a CXL Fixed Memory Window Structure (CFMWS), during early boot the kernel parses the CFMWS tables to find all CXL memory regions and sets a NUMA node for each of them. This memory affinity information of CXL memory regions is later used by the CXL ACPI driver.
The CFMWS table doesn't provide the memory affinity information either. Currently the kernel assigns a 'faked' NUMA node for each CXL memory region, starting from the next node of the highest node that is enumerated via the SRAT. This can potentially increase the maximum NUMA node ID of the platform ('
nr_node_ids') a lot. E.g., on a GNR platform with 4 NUMA nodes and 18 CFMWS tables, this bumps the 'nr_node_ids' to 22.Increasing the '
nr_node_ids' has side effects. For instance, it is widely used by the kernel for "highest possible NUMA node" based memory allocations. It also impacts userspace ABIs, e.g., some NUMA memory related system calls such as 'get_mempolicy' which requires 'maxnode' not being smaller than the 'nr_node_ids'.Currently parsing CFMWS tables and assigning faked NUMA node at boot is done unconditionally. However, if the CXL ACPI driver is not enabled, there will be no user of such memory affinity information of CXL memory regions.
Change to only parsing the CFMWS tables at boot when
CXL_ACPIis enabled inKconfigto avoid the unnecessary cost of bumping up 'nr_node_ids'.
Quoted 8 lines
E.g., on the aforementioned GNR platform, the "Slab" in
/proc/meminfois reduced with this change (whenCXL_ACPIis off):w/ this change w/o
Slab 900488 kB 923660 kB
Signed-off-by Kai Huang <[email protected]>
Even without all the reasoning above, I'm keen to remove state that we know is pointless!
Quoted 44 lines · Kai Huang
On CXL platforms, the Static Resource Affinity Table (SRAT) may not cover memory affinity information for all the CXL memory regions. Since each CXL memory region is enumerated via a CXL Fixed Memory Window Structure (CFMWS), during early boot the kernel parses the CFMWS tables to find all CXL memory regions and sets a NUMA node for each of them. This memory affinity information of CXL memory regions is later used by the CXL ACPI driver.
The CFMWS table doesn't provide the memory affinity information either. Currently the kernel assigns a 'faked' NUMA node for each CXL memory region, starting from the next node of the highest node that is enumerated via the SRAT. This can potentially increase the maximum NUMA node ID of the platform ('
nr_node_ids') a lot. E.g., on a GNR platform with 4 NUMA nodes and 18 CFMWS tables, this bumps the 'nr_node_ids' to 22.Increasing the '
nr_node_ids' has side effects. For instance, it is widely used by the kernel for "highest possible NUMA node" based memory allocations. It also impacts userspace ABIs, e.g., some NUMA memory related system calls such as 'get_mempolicy' which requires 'maxnode' not being smaller than the 'nr_node_ids'.Currently parsing CFMWS tables and assigning faked NUMA node at boot is done unconditionally. However, if the CXL ACPI driver is not enabled, there will be no user of such memory affinity information of CXL memory regions.
Change to only parsing the CFMWS tables at boot when
CXL_ACPIis enabled inKconfigto avoid the unnecessary cost of bumping up 'nr_node_ids'.E.g., on the aforementioned GNR platform, the "Slab" in
/proc/meminfois reduced with this change (whenCXL_ACPIis off):w/ this change w/o
Slab 900488 kB 923660 kB
Signed-off-by Kai Huang <[email protected]>---
v1 -> v2:
- Use Dan's suggestion to simplify the diff:
Quoted 7 lines
https://lore.kernel.org/linux-cxl/[email protected]/
Hi Alison, Gregory,
I didn't add your RB since the code now is different from that you reviewed. Appreciate if you can take a look again and provide the tag if the patch looks good to you.
Dave Jiang
AppliedQuoted 30 lines · Kai Huang
On CXL platforms, the Static Resource Affinity Table (SRAT) may not cover memory affinity information for all the CXL memory regions. Since each CXL memory region is enumerated via a CXL Fixed Memory Window Structure (CFMWS), during early boot the kernel parses the CFMWS tables to find all CXL memory regions and sets a NUMA node for each of them. This memory affinity information of CXL memory regions is later used by the CXL ACPI driver.
The CFMWS table doesn't provide the memory affinity information either. Currently the kernel assigns a 'faked' NUMA node for each CXL memory region, starting from the next node of the highest node that is enumerated via the SRAT. This can potentially increase the maximum NUMA node ID of the platform ('
nr_node_ids') a lot. E.g., on a GNR platform with 4 NUMA nodes and 18 CFMWS tables, this bumps the 'nr_node_ids' to 22.Increasing the '
nr_node_ids' has side effects. For instance, it is widely used by the kernel for "highest possible NUMA node" based memory allocations. It also impacts userspace ABIs, e.g., some NUMA memory related system calls such as 'get_mempolicy' which requires 'maxnode' not being smaller than the 'nr_node_ids'.Currently parsing CFMWS tables and assigning faked NUMA node at boot is done unconditionally. However, if the CXL ACPI driver is not enabled, there will be no user of such memory affinity information of CXL memory regions.
Change to only parsing the CFMWS tables at boot when
CXL_ACPIis enabled inKconfigto avoid the unnecessary cost of bumping up 'nr_node_ids'.
Quoted 8 lines
E.g., on the aforementioned GNR platform, the "Slab" in
/proc/meminfois reduced with this change (whenCXL_ACPIis off):w/ this change w/o
Slab 900488 kB 923660 kB
Signed-off-by Kai Huang <[email protected]>
Applied to cxl/next1e1cd49ded59 ("ACPI: NUMA: Only parse CFMWS at boot when CXL_ACPI is on")
Quoted 38 lines
---
v1 -> v2:
- Use Dan's suggestion to simplify the diff:
https://lore.kernel.org/linux-cxl/[email protected]/
Hi Alison, Gregory,
I didn't add your RB since the code now is different from that you reviewed. Appreciate if you can take a look again and provide the tag if the patch looks good to you.
--- drivers/acpi/numa/srat.c | 7 +++++-- 1 file changed, 5 insertions(+), 2 deletions(-)drivers/acpi/numa/srat.c +5
-2@@ -654,8 +654,11 @@ int __init acpi_numa_init(void) } last_real_pxm = fake_pxm; fake_pxm++; -acpi_table_parse_cedt (ACPI_CEDT_TYPE_CFMWS , acpi_parse_cfmws, - &fake_pxm); + +/* No need to expand numa nodes if CXL is disabled */ +if (IS_ENABLED (CONFIG_CXL_ACPI )) +acpi_table_parse_cedt (ACPI_CEDT_TYPE_CFMWS , acpi_parse_cfmws, + &fake_pxm);if (cnt <0 )return cnt;base-commit:
084f843093bee5563b179fd4b630122ba820e0c7