[v2,0/6] Manual definition of Soft Reserved memory devices

Message

Dan Williams March 22, 2020, 4:12 p.m. UTC

Changes since v1 [1]:
- Kill the ifdef'ery in arch/x86/mm/numa.c (Rafael)

- Add a dummy phys_to_target_node() for ARM64 (0day-robot)

- Initialize ->child and ->sibling to NULL in the resource returned by
  find_next_iomem_res() (Inspired by Tom's feedback even though it does
  not set them like he suggested)

- Collect Ard's Ack

[1]: http://lore.kernel.org/r/[e-mail]

---

My primary motivation is making the dax_kmem facility useful to
shipping platforms that have performance differentiated memory, but
may not have EFI-defined soft-reservations / HMAT (or
non-EFI-ACPI-platform equivalent). I'm anticipating HMAT enabled
platforms where the platform firmware policy for what is
soft-reserved, or not, is not the policy the system owner would pick.
I'd also highlight Joao's work [2] (see the TODO section) as an
indication of the demand for custom carving memory resources and
applying the device-dax memory management interface.

Given the current dearth of systems that supply an ACPI HMAT table, and
the utility of being able to manually define device-dax "hmem" instances
via the efi_fake_mem= option, relax the requirements for creating these
devices. Specifically, add an option (numa=nohmat) to optionally disable
consideration of the HMAT and update efi_fake_mem= to behave like
memmap=nn!ss in terms of delimiting device boundaries.

[2]: https://lore.kernel.org/lkml/[e-mail]/

With Ard's and Rafael's ack I'd feel ok taking this through the nvdimm
tree, please holler if anything still needs some fixups.

Dependencies:

b2ca916ce392 ACPI: NUMA: Up-level "map to online node" functionality
4fcbe96e4d0b mm/numa: Skip NUMA_NO_NODE and online nodes in numa_map_to_online_node()
575e23b6e13c powerpc/papr_scm: Switch to numa_map_to_online_node()
1e5d8e1e47af x86/mm: Introduce CONFIG_NUMA_KEEP_MEMINFO
5d30f92e7631 x86/NUMA: Provide a range-to-target_node lookup facility
7b27a8622f80 libnvdimm/e820: Retrieve and populate correct 'target_node' info

Tested with:

        numa=nohmat efi_fake_mem=4G@9G:0x40000,4G@13G:0x40000

...to create to device-dax instances:

	# daxctl list -RDu
	[
	  {
	    "path":"\/platform\/hmem.1",
	    "id":1,
	    "size":"4.00 GiB (4.29 GB)",
	    "align":2097152,
	    "devices":[
	      {
	        "chardev":"dax1.0",
	        "size":"4.00 GiB (4.29 GB)",
	        "target_node":3,
	        "mode":"devdax"
	      }
	    ]
	  },
	  {
	    "path":"\/platform\/hmem.0",
	    "id":0,
	    "size":"4.00 GiB (4.29 GB)",
	    "align":2097152,
	    "devices":[
	      {
	        "chardev":"dax0.0",
	        "size":"4.00 GiB (4.29 GB)",
	        "target_node":2,
	        "mode":"devdax"
	      }
	    ]
	  }
	]


---

Dan Williams (6):
      x86/numa: Cleanup configuration dependent command-line options
      x86/numa: Add 'nohmat' option
      efi/fake_mem: Arrange for a resource entry per efi_fake_mem instance
      ACPI: HMAT: Refactor hmat_register_target_device to hmem_register_device
      resource: Report parent to walk_iomem_res_desc() callback
      ACPI: HMAT: Attach a device for each soft-reserved range

 arch/arm64/mm/numa.c                |   13 +++++
 arch/x86/include/asm/numa.h         |    8 +++
 arch/x86/kernel/e820.c              |   16 +++++-
 arch/x86/mm/numa.c                  |   10 +---
 arch/x86/mm/numa_emulation.c        |    3 +
 arch/x86/xen/enlighten_pv.c         |    2 -
 drivers/acpi/numa/hmat.c            |   76 +++++----------------------
 drivers/acpi/numa/srat.c            |    9 +++
 drivers/dax/Kconfig                 |    5 ++
 drivers/dax/Makefile                |    3 -
 drivers/dax/hmem/Makefile           |    6 ++
 drivers/dax/hmem/device.c           |   97 +++++++++++++++++++++++++++++++++++
 drivers/dax/hmem/hmem.c             |    2 -
 drivers/firmware/efi/x86_fake_mem.c |   12 +++-
 include/acpi/acpi_numa.h            |   14 +++++
 include/linux/dax.h                 |    8 +++
 kernel/resource.c                   |   11 +++-
 17 files changed, 209 insertions(+), 86 deletions(-)
 create mode 100644 drivers/dax/hmem/Makefile
 create mode 100644 drivers/dax/hmem/device.c
 rename drivers/dax/{hmem.c => hmem/hmem.c} (98%)

base-commit: 7b27a8622f802761d5c6abd6c37b22312a35343c

Comments

Rafael J. Wysocki March 25, 2020, 10:02 a.m. UTC | #1

On Sun, Mar 22, 2020 at 5:28 PM Dan Williams <[e-mail]> wrote:
>
> Changes since v1 [1]:
> - Kill the ifdef'ery in arch/x86/mm/numa.c (Rafael)
>
> - Add a dummy phys_to_target_node() for ARM64 (0day-robot)
>
> - Initialize ->child and ->sibling to NULL in the resource returned by
>   find_next_iomem_res() (Inspired by Tom's feedback even though it does
>   not set them like he suggested)
>
> - Collect Ard's Ack
>
> [1]: http://lore.kernel.org/r/[e-mail]
>
> ---
>
> My primary motivation is making the dax_kmem facility useful to
> shipping platforms that have performance differentiated memory, but
> may not have EFI-defined soft-reservations / HMAT (or
> non-EFI-ACPI-platform equivalent). I'm anticipating HMAT enabled
> platforms where the platform firmware policy for what is
> soft-reserved, or not, is not the policy the system owner would pick.
> I'd also highlight Joao's work [2] (see the TODO section) as an
> indication of the demand for custom carving memory resources and
> applying the device-dax memory management interface.
>
> Given the current dearth of systems that supply an ACPI HMAT table, and
> the utility of being able to manually define device-dax "hmem" instances
> via the efi_fake_mem= option, relax the requirements for creating these
> devices. Specifically, add an option (numa=nohmat) to optionally disable
> consideration of the HMAT and update efi_fake_mem= to behave like
> memmap=nn!ss in terms of delimiting device boundaries.
>
> [2]: https://lore.kernel.org/lkml/[e-mail]/
>
> With Ard's and Rafael's ack I'd feel ok taking this through the nvdimm
> tree, please holler if anything still needs some fixups.

Acked-by: Rafael J. Wysocki <[e-mail]>

for the whole series.

Thanks!