mbox series

[v2,0/6] ACPI: Only create NUMA nodes from entries in SRAT or SRAT emulation.

Message ID 20200717175959.899775-1-Jonathan.Cameron@huawei.com
Headers show
Series ACPI: Only create NUMA nodes from entries in SRAT or SRAT emulation. | expand

Message

Jonathan Cameron July 17, 2020, 5:59 p.m. UTC
Here, I will use the term Proximity Domains for the ACPI description and
NUMA Nodes for the in kernel representation.

ACPI 6.3 included a clarification that only Static Resource Allocation
Structures in SRAT may define the existence of proximity domains
(sec 5.2.16). This clarification closed a possible interpretation that
other parts of ACPI (e.g. DSDT _PXM, NFIT etc) could define new proximity
domains that were not also mentioned in SRAT structures.

In practice the kernel has never allowed this alternative interpretation as
such nodes are only partially initialized. This is architecture specific
but to take an example, on x86 alloc_node_data has not been called.
Any use of them for node specific allocation, will result in a crash as the
infrastructure to fallback to a node with memory is not setup.

We ran into a problem when enabling _PXM handling for PCI devices and found
there were boards out there advertising devices in proximity domains that
didn't exist [2].

The fix suggested in this series is to replace instances that should not
'create' new nodes with pxm_to_node.  This function needs a some additional
hardening against invalid inputs to make sure it is safe for use in these
new callers.

Patch 1 Hardens pxm_to_node() against numa_off, and pxm entry being too large.

Patch 2-4 change the various callers not related to SRAT entries so that they
set this parameter to false, so do not attempt to initialize a new NUMA node
if the relevant one does not already exist.

Patch 5 is a function rename to reflect change in functionality of
acpi_map_pxm_to_online_node() as it no longer creates a new map, but just does a
lookup of existing maps.

Patch 6 covers the one place we do not allow the full flexibility defined
in the ACPI spec.  For SRAT GIC Interrupt Translation Service (ITS) Affinity
Structures, on ARM64, the driver currently makes an additional pass of SRAT
later in the boot than the one used to identify NUMA domains.
Note, this currently means that an ITS placed in a proximity domain that is
not defined by another SRAT structure will result in the a crash.

To avoid this crash with minimal changes we do not create new NUMA nodes based
on this particular entry type.  Any current platform trying to do this will not
boot, so this is an improvement, if perhaps not a perfect solution.

[1] Note in ACPI Specification 6.3 5.2.16 System Resource Affinity Table (SRAT)
[2] https://patchwork.kernel.org/patch/10597777/

Thanks to Bjorn Helgaas for review of v1 and Barry Song for internal reviews that
lead to a slightly different approach for this v2.

Changes since v1.
* Use pxm_to_node for what was previously the path using acpi_map_pxm_to_node
  with create==false. (Barry)
* Broke patch up into an initial noop stage followed by patches (Bjorn)
  to update each type of case in which partial creation of NUMA nodes is prevented.
* Added patch 5 to rename function to reflect change of functionality.
* Updated descriptions (now mostly in individual patches) inline with Bjorn's comments.

Jonathan Cameron (6):
  ACPI: Add out of bounds and numa_off protections to pxm_to_node
  ACPI: Do not create new NUMA domains from ACPI static tables that are
    not SRAT
  ACPI: Remove side effect of partly creating a node in
    acpi_map_pxm_to_online_node
  ACPI: rename acpi_map_pxm_to_online_node to pxm_to_online_node
  ACPI: Remove side effect of partly creating a node in acpi_get_node
  irq-chip/gic-v3-its: Fix crash if ITS is in a proximity domain without
    processor or memory

 drivers/acpi/arm64/iort.c        |  2 +-
 drivers/acpi/nfit/core.c         |  6 ++----
 drivers/acpi/numa/hmat.c         |  4 ++--
 drivers/acpi/numa/srat.c         |  4 ++--
 drivers/iommu/intel/dmar.c       |  2 +-
 drivers/irqchip/irq-gic-v3-its.c |  7 ++++++-
 include/linux/acpi.h             | 15 +++++++--------
 7 files changed, 21 insertions(+), 19 deletions(-)

-- 
2.19.1

Comments

Hanjun Guo July 20, 2020, 2:02 a.m. UTC | #1
On 2020/7/18 1:59, Jonathan Cameron wrote:
> Here, I will use the term Proximity Domains for the ACPI description and

> NUMA Nodes for the in kernel representation.

> 

> ACPI 6.3 included a clarification that only Static Resource Allocation

> Structures in SRAT may define the existence of proximity domains

> (sec 5.2.16). This clarification closed a possible interpretation that

> other parts of ACPI (e.g. DSDT _PXM, NFIT etc) could define new proximity

> domains that were not also mentioned in SRAT structures.

> 

> In practice the kernel has never allowed this alternative interpretation as

> such nodes are only partially initialized. This is architecture specific

> but to take an example, on x86 alloc_node_data has not been called.

> Any use of them for node specific allocation, will result in a crash as the

> infrastructure to fallback to a node with memory is not setup.

> 

> We ran into a problem when enabling _PXM handling for PCI devices and found

> there were boards out there advertising devices in proximity domains that

> didn't exist [2].

> 

> The fix suggested in this series is to replace instances that should not

> 'create' new nodes with pxm_to_node.  This function needs a some additional

> hardening against invalid inputs to make sure it is safe for use in these

> new callers.

> 

> Patch 1 Hardens pxm_to_node() against numa_off, and pxm entry being too large.

> 

> Patch 2-4 change the various callers not related to SRAT entries so that they

> set this parameter to false, so do not attempt to initialize a new NUMA node

> if the relevant one does not already exist.

> 

> Patch 5 is a function rename to reflect change in functionality of

> acpi_map_pxm_to_online_node() as it no longer creates a new map, but just does a

> lookup of existing maps.

> 

> Patch 6 covers the one place we do not allow the full flexibility defined

> in the ACPI spec.  For SRAT GIC Interrupt Translation Service (ITS) Affinity

> Structures, on ARM64, the driver currently makes an additional pass of SRAT

> later in the boot than the one used to identify NUMA domains.

> Note, this currently means that an ITS placed in a proximity domain that is

> not defined by another SRAT structure will result in the a crash.

> 

> To avoid this crash with minimal changes we do not create new NUMA nodes based

> on this particular entry type.  Any current platform trying to do this will not

> boot, so this is an improvement, if perhaps not a perfect solution.


Make sense to me,

Reviewed-by: Hanjun Guo <guohanjun@huawei.com>
Jonathan Cameron July 28, 2020, 4:20 p.m. UTC | #2
On Sat, 18 Jul 2020 01:59:53 +0800
Jonathan Cameron <Jonathan.Cameron@huawei.com> wrote:

> Here, I will use the term Proximity Domains for the ACPI description and

> NUMA Nodes for the in kernel representation.

> 

> ACPI 6.3 included a clarification that only Static Resource Allocation

> Structures in SRAT may define the existence of proximity domains

> (sec 5.2.16). This clarification closed a possible interpretation that

> other parts of ACPI (e.g. DSDT _PXM, NFIT etc) could define new proximity

> domains that were not also mentioned in SRAT structures.

> 

> In practice the kernel has never allowed this alternative interpretation as

> such nodes are only partially initialized. This is architecture specific

> but to take an example, on x86 alloc_node_data has not been called.

> Any use of them for node specific allocation, will result in a crash as the

> infrastructure to fallback to a node with memory is not setup.

> 

> We ran into a problem when enabling _PXM handling for PCI devices and found

> there were boards out there advertising devices in proximity domains that

> didn't exist [2].

> 

> The fix suggested in this series is to replace instances that should not

> 'create' new nodes with pxm_to_node.  This function needs a some additional

> hardening against invalid inputs to make sure it is safe for use in these

> new callers.

> 

> Patch 1 Hardens pxm_to_node() against numa_off, and pxm entry being too large.

> 

> Patch 2-4 change the various callers not related to SRAT entries so that they

> set this parameter to false, so do not attempt to initialize a new NUMA node

> if the relevant one does not already exist.

> 

> Patch 5 is a function rename to reflect change in functionality of

> acpi_map_pxm_to_online_node() as it no longer creates a new map, but just does a

> lookup of existing maps.

> 

> Patch 6 covers the one place we do not allow the full flexibility defined

> in the ACPI spec.  For SRAT GIC Interrupt Translation Service (ITS) Affinity

> Structures, on ARM64, the driver currently makes an additional pass of SRAT

> later in the boot than the one used to identify NUMA domains.

> Note, this currently means that an ITS placed in a proximity domain that is

> not defined by another SRAT structure will result in the a crash.

> 

> To avoid this crash with minimal changes we do not create new NUMA nodes based

> on this particular entry type.  Any current platform trying to do this will not

> boot, so this is an improvement, if perhaps not a perfect solution.

> 

> [1] Note in ACPI Specification 6.3 5.2.16 System Resource Affinity Table (SRAT)

> [2] https://patchwork.kernel.org/patch/10597777/

> 

> Thanks to Bjorn Helgaas for review of v1 and Barry Song for internal reviews that

> lead to a slightly different approach for this v2.


Thanks Barry / Hanjun,

Anyone else have time to take a look?

I'm happy to bring it back after the merge window closes, but would like to know
if people are happy with the general approach.  I'm keen to finally be able to
resolve that issue with _PXM and PCI. 

Thanks,

Jonathan

> 

> Changes since v1.

> * Use pxm_to_node for what was previously the path using acpi_map_pxm_to_node

>   with create==false. (Barry)

> * Broke patch up into an initial noop stage followed by patches (Bjorn)

>   to update each type of case in which partial creation of NUMA nodes is prevented.

> * Added patch 5 to rename function to reflect change of functionality.

> * Updated descriptions (now mostly in individual patches) inline with Bjorn's comments.

> 

> Jonathan Cameron (6):

>   ACPI: Add out of bounds and numa_off protections to pxm_to_node

>   ACPI: Do not create new NUMA domains from ACPI static tables that are

>     not SRAT

>   ACPI: Remove side effect of partly creating a node in

>     acpi_map_pxm_to_online_node

>   ACPI: rename acpi_map_pxm_to_online_node to pxm_to_online_node

>   ACPI: Remove side effect of partly creating a node in acpi_get_node

>   irq-chip/gic-v3-its: Fix crash if ITS is in a proximity domain without

>     processor or memory

> 

>  drivers/acpi/arm64/iort.c        |  2 +-

>  drivers/acpi/nfit/core.c         |  6 ++----

>  drivers/acpi/numa/hmat.c         |  4 ++--

>  drivers/acpi/numa/srat.c         |  4 ++--

>  drivers/iommu/intel/dmar.c       |  2 +-

>  drivers/irqchip/irq-gic-v3-its.c |  7 ++++++-

>  include/linux/acpi.h             | 15 +++++++--------

>  7 files changed, 21 insertions(+), 19 deletions(-)

>