Re: kexec_file overwrites reserved EFI ESRT memory

[Dave Young]

On 12/03/19 at 12:45am, Michael Weiser wrote:
> Hi Dave,
> 
> On Mon, Dec 02, 2019 at 05:05:20PM +0800, Dave Young wrote:
> 
> > > It seems a serious problem, the EFI modified memmap does not get an
> > > /proc/iomem resource update, but kexec_file relies on /proc/iomem in
> > > X86.
> > > 
> > > There is an question from Sai about why add_efi_memmap is not enabled by
> > > default:
> > > https://www.spinics.net/lists/linux-mm/msg185166.html
> 
> Incidentally, a data point I did not think to mention: I do boot the
> kernel as EFI application directly from the firmware as a boot entry
> with compiled in initrd and command line:
> 
> $ grep EFI nobak/kernel/linux/.config
> CONFIG_EFI=y
> CONFIG_EFI_STUB=y
> # CONFIG_EFI_MIXED is not set
> CONFIG_DMI_SCAN_MACHINE_NON_EFI_FALLBACK=y
> # EFI (Extensible Firmware Interface) Support
> CONFIG_EFI_VARS=m
> CONFIG_EFI_ESRT=y
> CONFIG_EFI_VARS_PSTORE=m
> # CONFIG_EFI_VARS_PSTORE_DEFAULT_DISABLE is not set
> CONFIG_EFI_RUNTIME_MAP=y
> # CONFIG_EFI_FAKE_MEMMAP is not set
> CONFIG_EFI_RUNTIME_WRAPPERS=y
> # CONFIG_EFI_BOOTLOADER_CONTROL is not set
> # CONFIG_EFI_CAPSULE_LOADER is not set
> # CONFIG_EFI_TEST is not set
> # CONFIG_EFI_RCI2_TABLE is not set
> # end of EFI (Extensible Firmware Interface) Support
> CONFIG_UEFI_CPER=y
> CONFIG_UEFI_CPER_X86=y
> CONFIG_EFI_EARLYCON=y
> CONFIG_EFI_PARTITION=y
> CONFIG_FB_EFI=y
> CONFIG_EFIVAR_FS=y
> # CONFIG_EFI_PGT_DUMP is not set
> 
> $ grep CMDLINE nobak/kernel/linux/.config
> CONFIG_CMDLINE_BOOL=y
> CONFIG_CMDLINE="root=UUID=97[...]e4 rd.luks.uuid=8a[...]c3 rd.luks.allow-discards=8a[...]c3 mem_sleep_default=deep resume=UUID=97[...]e4 resume_offset=96256 efi=debug memblock=debug"
> CONFIG_CMDLINE_OVERRIDE=y
> # CONFIG_BLK_CMDLINE_PARSER is not set
> # CONFIG_CMDLINE_PARTITION is not set
> CONFIG_FB_CMDLINE=y
> 
> $ efibootmgr -v
> BootCurrent: 000A
> Timeout: 2 seconds
> BootOrder: 000A,0009,0008,0005,0007,0006,0004,0002,0001,0000,0003
> [...]
> Boot0005* gentoo-5.4.0-next-20191127+-clear
> HD(1,GPT,e7[...]f2,0x800,0x64000)/File(\kernel-5.4.0-next-20191127+-clear)
> [...]
> Boot000A* gentoo-5.4.1-gentoo
> HD(1,GPT,e7[...]f2,0x800,0x64000)/File(\kernel-5.4.1-gentoo)
> 
> So there's no boot loader that could construct an e820 table for the
> kernel to consume. I understand it's then up to the EFI stub to come up
> with a e820 table from the EFI memory map.
> 
> > > Long time ago the add_efi_memmap is only enabled in case we explict
> > > enable it on cmdline, I'm not sure if we can do it by default, maybe we
> > > should.   Need opinion from X86 maintainers..
> > > Can you try below diff see if it works for you? (not tested, and need
> > > explicitly 'add_efi_memmap' in kernel cmdline param)
> 
> Neither adding add_efi_memmap nor adding your patch and setting that option
> does make the ESRT memory region appear in /proc/iomem. kexec_file still
> loads the kernel across the ESRT region.
> 

Hmm, sorry, my bad, actuall add_efi_memmap does not consider the
EFI_MEMORY_RUNTIME attribute, it only reads the memory descriptor types.

Will read your replied information later, did not get time today, but
probably below chunk can help?

diff --git a/arch/x86/platform/efi/quirks.c b/arch/x86/platform/efi/quirks.c
index 3b9fd679cea9..516307617621 100644
--- a/arch/x86/platform/efi/quirks.c
+++ b/arch/x86/platform/efi/quirks.c
@@ -293,6 +293,8 @@ void __init efi_arch_mem_reserve(phys_addr_t addr, u64 size)
 	early_memunmap(new, new_size);
 
 	efi_memmap_install(new_phys, num_entries);
+	e820__range_update(addr, size, E820_TYPE_RAM, E820_TYPE_RESERVED);
+	e820__update_table(e820_table);
 }
 
 /*

Thanks
Dave