[PATCH 3/4] efi_loader: preserve installer images in pmem

[Heinrich Schuchardt]

On 10/25/24 13:14, Sughosh Ganu wrote:
> From: Ilias Apalodimas <ilias.apalodimas at linaro.org>
>
> One of the problems OS installers face, when running in EFI, is that
> the mounted ISO after calling ExitBootServices goes away. For some
> distros this is a problem since they rely on finding some core packages
> before continuing the installation. Distros have works around this --
> e.g Fedora has a special kernel command line parameter called
> inst.stage2 [0].
>
> ACPI has NFIT and NVDIMM support to provide ramdisks to the OS, but we
> don't have anything in place for DTs. Linux and device trees have support
> for persistent memory devices.
>
> It's worth noting that for linux to instantiate the /dev/pmemX device,
> the memory described in the pmem node has to be omitted from the EFI
> memory map we hand over to the OS if ZONE_DEVICES and SPARSEMEM is
> enabled. With those enabled the pmem driver ends up calling
> devm_memremap_pages() instead of devm_memremap(). The latter works
> whether the memory is omitted or marked as reserved, but mapping pages
> only works if the memory is omitted.
>
> On top of that, depending on how the kernel is configured, that memory
> area must be page aligned or 2MB aligned. PowerPC is an exception here
> and requires 16MB alignment, but since we don't have EFI support for
> it, limit the alignment to 2MB.
>
> Ensure that the ISO image is 2MB aligned and remove the region
> occupied by the image from the EFI memory map.
>
> Signed-off-by: Ilias Apalodimas <ilias.apalodimas at linaro.org>
> Signed-off-by: Sughosh Ganu <sughosh.ganu at linaro.org>
> ---
>   lib/efi_loader/efi_bootmgr.c | 22 +++++++++++++++++-----
>   1 file changed, 17 insertions(+), 5 deletions(-)
>
> diff --git a/lib/efi_loader/efi_bootmgr.c b/lib/efi_loader/efi_bootmgr.c
> index a3aa2b8d1b..16f75555f6 100644
> --- a/lib/efi_loader/efi_bootmgr.c
> +++ b/lib/efi_loader/efi_bootmgr.c
> @@ -18,6 +18,8 @@
>   #include <efi_loader.h>
>   #include <efi_variable.h>
>   #include <asm/unaligned.h>
> +#include <linux/kernel.h>
> +#include <linux/sizes.h>
>
>   static const struct efi_boot_services *bs;
>   static const struct efi_runtime_services *rs;
> @@ -358,13 +360,16 @@ static efi_status_t prepare_loaded_image(u16 *label, ulong addr, ulong size,
>   	}
>
>   	/*
> -	 * TODO: expose the ramdisk to OS.
> -	 * Need to pass the ramdisk information by the architecture-specific
> -	 * methods such as 'pmem' device-tree node.
> +	 * Linux supports 'pmem' which allows OS installers to find, reclaim
> +	 * the mounted images and continue the installation since the contents
> +	 * of the pmem region are treated as local media.
> +	 *
> +	 * The memory regions used for it needs to be carved out of the EFI
> +	 * memory map.
>   	 */
> -	ret = efi_add_memory_map(addr, size, EFI_RESERVED_MEMORY_TYPE);
> +	ret = efi_remove_memory_map(addr, size, EFI_CONVENTIONAL_MEMORY);
>   	if (ret != EFI_SUCCESS) {
> -		log_err("Memory reservation failed\n");
> +		log_err("Failed to reserve memory\n");
>   		goto err;
>   	}
>
> @@ -486,6 +491,13 @@ static efi_status_t try_load_from_uri_path(struct efi_device_path_uri *uridp,
>   		ret = EFI_INVALID_PARAMETER;
>   		goto err;
>   	}
> +	/*
> +	 * Depending on the kernel configuration, pmem memory area must be page
> +	 * aligned or 2MB aligned. PowerPC is an exception here and requires
> +	 * 16MB alignment, but since we don't have EFI support for it, limit
> +	 * the alignment to 2MB.
> +	 */
> +	image_size = ALIGN(image_size, SZ_2M);

The code regarding .iso and .img handling seems to be misplaced. Why
should we treat a file loaded from a block device differently to a file
loaded from the network?

Best regards

Heinrich

>
>   	/*
>   	 * If the file extension is ".iso" or ".img", mount it and try to load