pmem
remove no longer supported efi_fake_mem= kernel cmdline option
Kernel support for efi_fake_mem was removed in 6.11. 37aee82c213d ("x86/efi: Drop support for fake EFI memory maps")
The ndctl tests stopped depending on it since ndctl v80: dfa937e2eb26 ("test/daxctl-create.sh: use CXL DAX regions instead of efi_fake_mem")
Remove it from run_qemu.sh
git grep efi_mem shows that this feature is unfortunately quite "invasive" :-(
I'm afraid removing it requires some minimal understanding of it (which I don't have now).
The good news though: --efi-mems 0 has precedence over --preset and drops efi_fake_mem most of the time! Except when using --nfit-test, this last one has even more precedence.
Questions:
- How does it hurt? I mean, isn't
efi_fake_memsimply ignored by v6.11 and above? - Should it be just dropped or replaced by something newer and better? (Which means not replacing it could "hurt" somehow?)
Some code reviews / context :
- https://lkml.org/lkml/2015/10/12/348 (2015)
- https://lore.kernel.org/all/157309100651.1579826.4018037842830501297.stgit@dwillia2-desk3.amr.corp.intel.com/ (2019)
- https://lore.kernel.org/linux-efi/[email protected]/t/#u (2024 drop)
If it is harmless, might also be worth keeping it around a little bit longer in case anyone wants to run (or bisect) with slightly older kernels.
If it is harmless, might also be worth keeping it around a little bit longer in case anyone wants to run (or bisect) with slightly older kernels.
That was my typical reaction but then I thought: how come passing the test suite after v6.11 needs nothing instead? What was the point of the efi_fake_mem values before v6.11 if they can just be ignored after?
For some manual testing that is not in the test suite(s) maybe?
https://lore.kernel.org/linux-efi/[email protected]/t/#u seems to recommend memmap as a replacement?
That was my typical reaction but then I thought: how come passing the test suite after v6.11 needs nothing instead? What was the point of the
efi_fake_memvalues before v6.11 if they can just be ignored after?
The test suite passes post 6.11 because of this change Alison referred to: https://lore.kernel.org/all/519161e23a43e530dbcffac203ecbbb897aa5342.1724813664.git.alison.schofield@intel.com/
..which uses a cxl_test device instead of an efi_fake_mem device for the one test that was using it (daxctl-create.sh).
Maybe it makes sense to remove it from a default run_qemu run, either via the default preset, or via --nfit-test.
Also thinking about the above patch a bit more, this introduces a dependency on cxl_test for nfit_test usage - I'm not sure if this is the first time that's happened, but with the above patch, --nfit-test might need to imply --cxl-test?
another option, if the nfit-test dependency on cxl is problematic, is to perhaps use memmap= to replace what efi_fake_mem was doing:
memmap=<size>%<offset>-<oldtype>+<newtype>
Just need to be careful that unlike efi_fake_mem which is safer in that it only adjust attributes without touching the memory type, memmap= is not so careful. However, I have not tried this "type change" syntax, only the the pmem syntax which is a large foot-gun if you get it wrong:
`memmap=nn[KMG]!ss[KMG,EARLY]`
