tree e3220cd334c147a9eb9d4ee2512d45f7d9863587
parent 6e6867093de35141f0a76b66ac13f9f2e2c8e77a
author Matt Fleming <matt@codeblueprint.co.uk> 1457695163 +0000
committer Ingo Molnar <mingo@kernel.org> 1457798265 +0100

x86/efi: Fix boot crash by always mapping boot service regions into new EFI page tables

Some machines have EFI regions in page zero (physical address
0x00000000) and historically that region has been added to the e820
map via trim_bios_range(), and ultimately mapped into the kernel page
tables. It was not mapped via efi_map_regions() as one would expect.

Alexis reports that with the new separate EFI page tables some boot
services regions, such as page zero, are not mapped. This triggers an
oops during the SetVirtualAddressMap() runtime call.

For the EFI boot services quirk on x86 we need to memblock_reserve()
boot services regions until after SetVirtualAddressMap(). Doing that
while respecting the ownership of regions that may have already been
reserved by the kernel was the motivation behind this commit:

  7d68dc3f1003 ("x86, efi: Do not reserve boot services regions within reserved areas")

That patch was merged at a time when the EFI runtime virtual mappings
were inserted into the kernel page tables as described above, and the
trick of setting ->numpages (and hence the region size) to zero to
track regions that should not be freed in efi_free_boot_services()
meant that we never mapped those regions in efi_map_regions(). Instead
we were relying solely on the existing kernel mappings.

Now that we have separate page tables we need to make sure the EFI
boot services regions are mapped correctly, even if someone else has
already called memblock_reserve(). Instead of stashing a tag in
->numpages, set the EFI_MEMORY_RUNTIME bit of ->attribute. Since it
generally makes no sense to mark a boot services region as required at
runtime, it's pretty much guaranteed the firmware will not have
already set this bit.

For the record, the specific circumstances under which Alexis
triggered this bug was that an EFI runtime driver on his machine was
responding to the EVT_SIGNAL_VIRTUAL_ADDRESS_CHANGE event during
SetVirtualAddressMap().

The event handler for this driver looks like this,

  sub rsp,0x28
  lea rdx,[rip+0x2445] # 0xaa948720
  mov ecx,0x4
  call func_aa9447c0  ; call to ConvertPointer(4, & 0xaa948720)
  mov r11,QWORD PTR [rip+0x2434] # 0xaa948720
  xor eax,eax
  mov BYTE PTR [r11+0x1],0x1
  add rsp,0x28
  ret

Which is pretty typical code for an EVT_SIGNAL_VIRTUAL_ADDRESS_CHANGE
handler. The "mov r11, QWORD PTR [rip+0x2424]" was the faulting
instruction because ConvertPointer() was being called to convert the
address 0x0000000000000000, which when converted is left unchanged and
remains 0x0000000000000000.

The output of the oops trace gave the impression of a standard NULL
pointer dereference bug, but because we're accessing physical
addresses during ConvertPointer(), it wasn't. EFI boot services code
is stored at that address on Alexis' machine.

Reported-by: Alexis Murzeau <amurzeau@gmail.com>
Signed-off-by: Matt Fleming <matt@codeblueprint.co.uk>
Cc: Andy Lutomirski <luto@amacapital.net>
Cc: Ard Biesheuvel <ard.biesheuvel@linaro.org>
Cc: Ben Hutchings <ben@decadent.org.uk>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Brian Gerst <brgerst@gmail.com>
Cc: Denys Vlasenko <dvlasenk@redhat.com>
Cc: H. Peter Anvin <hpa@zytor.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Maarten Lankhorst <maarten.lankhorst@canonical.com>
Cc: Matthew Garrett <mjg59@srcf.ucam.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Raphael Hertzog <hertzog@debian.org>
Cc: Roger Shimizu <rogershimizu@gmail.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: linux-efi@vger.kernel.org
Link: http://lkml.kernel.org/r/1457695163-29632-2-git-send-email-matt@codeblueprint.co.uk
Link: https://bugs.debian.org/cgi-bin/bugreport.cgi?bug=815125
Signed-off-by: Ingo Molnar <mingo@kernel.org>