[ Impact ]
* If you use UEFI http boot to boot an image (rather than an EFI
executable) and get all the way to a normal userspace, you can
access the boot image as /dev/pmem0. But this is not accessible in
the initrd; presumably some modules are missing.
* This is desirable because then you can just feed an installer ISO to
a machine via http boot and the installer just works as normal
* Add support for physical pmem devices, and simulation thereof with
the memmap kernel command line parameter
* The initrd is larger
[ Test Plan ]
* unpack an initrd on a Jammy system with the generic kernel
metapackage with unmkinitramfs
* observe that the directories kernel/drivers/{nvdimm,dax,acpi/nfit}
are not present
* install the updated initramfs-tools packages from proposed
* again unpack an initrd on a Jammy system with the generic kernel
metapackage with unmkinitramfs
* observe that the directories kernel/drivers/{nvdimm,dax,acpi/nfit}
are present now
* reboot to confirm that the system still boots
* modify /etc/default/grub GRUB_CMDLINE_LINUX_DEFAULT to contain a
memmap entry - memmap=1G!4G seems to work on many systems over 4G of
RAM, or do `dmesg | grep BIOS-e820` to observe the memory regions
and select a usable one.
* update-grub and reboot again
* a /dev/pmem device should now be present on the system
[ Where problems could occur ]
* The growth of the files in /boot will accelerate issues for users
who have a dedicated boot partition that is not large enough
[ Other Info ]
* Details on the memmap kernel command line parameter:
https://www.kernel.org/doc/Documentation/admin-guide/kernel-parameters.txt
* PMEM simulation with memmap:
https://docs.pmem.io/persistent-memory/getting-started-guide/creating-development-environments/linux-environments/linux-memmap
[ Original Bug Description ]
If you use UEFI http boot to boot an image (rather than an EFI executable) and get all the way to a normal userspace, you can access the boot image as /dev/pmem0. But this is not accessible in the initrd; presumably some modules are missing. Dimitri added some modules that are clearly going to be necessary (kernel/drivers/nvdimm) in 0.140ubuntu14 and I added kernel/drivers/dax too in local experiments but this appears not to be enough to get it to appear.
This is desirable because then you can just feed an installer ISO to a machine via http boot and the installer just works as normal (the speed and, uh, quality, of the implementation of HTTP in a given machine's firmware may mean this isn't always the best option but it would be nice if it worked in case someone's machine actually does this well).
You may need "nfit" module in the initrd too.