On 06.02.23 02:02, Dan Williams wrote:
Summary:
--------
CXL RAM support allows for the dynamic provisioning of new CXL RAM
regions, and more routinely, assembling a region from an existing
configuration established by platform-firmware. The latter is motivated
by CXL memory RAS (Reliability, Availability and Serviceability)
support, that requires associating device events with System Physical
Address ranges and vice versa.
The 'Soft Reserved' policy rework arranges for performance
differentiated memory like CXL attached DRAM, or high-bandwidth memory,
to be designated for 'System RAM' by default, rather than the device-dax
dedicated access mode. That current device-dax default is confusing and
surprising for the Pareto of users that do not expect memory to be
quarantined for dedicated access by default. Most users expect all
'System RAM'-capable memory to show up in FREE(1).
Details:
--------
Recall that the Linux 'Soft Reserved' designation for memory is a
reaction to platform-firmware, like EFI EDK2, delineating memory with
the EFI Specific Purpose Memory attribute (EFI_MEMORY_SP). An
alternative way to think of that attribute is that it specifies the
*not* general-purpose memory pool. It is memory that may be too precious
for general usage or not performant enough for some hot data structures.
However, in the absence of explicit policy it should just be 'System
RAM' by default.
Rather than require every distribution to ship a udev policy to assign
dax devices to dax_kmem (the device-memory hotplug driver) just make
that the kernel default. This is similar to the rationale in:
commit 8604d9e534a3 ("memory_hotplug: introduce CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE")
With this change the relatively niche use case of accessing this memory
via mapping a device-dax instance can be achieved by building with
CONFIG_MEMORY_HOTPLUG_DEFAULT_ONLINE=n, or specifying
memhp_default_state=offline at boot, and then use:
daxctl reconfigure-device $device -m devdax --force
...to shift the corresponding address range to device-dax access.
The process of assembling a device-dax instance for a given CXL region
device configuration is similar to the process of assembling a
Device-Mapper or MDRAID storage-device array. Specifically, asynchronous
probing by the PCI and driver core enumerates all CXL endpoints and
their decoders. Then, once enough decoders have arrived to a describe a
given region, that region is passed to the device-dax subsystem where it
is subject to the above 'dax_kmem' policy. This assignment and policy
choice is only possible if memory is set aside by the 'Soft Reserved'
designation. Otherwise, CXL that is mapped as 'System RAM' becomes
immutable by CXL driver mechanisms, but is still enumerated for RAS
purposes.
This series is also available via:
https://git.kernel.org/pub/scm/linux/kernel/git/cxl/cxl.git/log/?h=for-6.3/cxl-ram-region
...and has gone through some preview testing in various forms.
My concern would be that in setups with a lot of CXL memory
(soft-reserved), having that much offline memory during boot might make
the kernel run out of memory. After all, offline memory consumes memory
for the memmap.
Is the assumption that something like that cannot happen because we'll
never ever have that much soft-reserved memory? :)
Note that this is a concern only applies when not using auto-onlining in
the kernel during boot, which (IMHO) is or will be the default in the
future.
--
Thanks,
David / dhildenb
