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/c...
...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.