[Qemu-devel] [Bug 1830872] [NEW] AARCH64 to ARMv7 mistranslation in TCG

* [Qemu-devel] [Bug 1830872] [NEW] AARCH64 to ARMv7 mistranslation in TCG
@ 2019-05-29  9:13 Laszlo Ersek (Red Hat)
  2019-05-29 12:08 ` [Qemu-devel] [Bug 1830872] " Philippe Mathieu-Daudé
                   ` (9 more replies)
  0 siblings, 10 replies; 27+ messages in thread
From: Laszlo Ersek (Red Hat) @ 2019-05-29  9:13 UTC (permalink / raw)
  To: qemu-devel

Public bug reported:

The following guest code:

https://github.com/tianocore/edk2/blob/3604174718e2afc950c3cc64c64ba5165c8692bd/MdePkg/Library/BaseMemoryLibOptDxe/AArch64/CopyMem.S

implements, in hand-optimized aarch64 assembly, the CopyMem() edk2 (EFI
Development Kit II) library function. (CopyMem() basically has memmove()
semantics, to provide a standard C analog here.) The relevant functions
are InternalMemCopyMem() and __memcpy().

When TCG translates this aarch64 code to x86_64, everything works fine.

When TCG translates this aarch64 code to ARMv7, the destination area of
the translated CopyMem() function becomes corrupted -- it differs from
the intended source contents. Namely, in every 4096 byte block, the
8-byte word at offset 4032 (0xFC0) is zeroed out in the destination,
instead of receiving the intended source value.

I'm attaching two hexdumps of the same destination area:

- "good.txt" is a hexdump of the destination area when CopyMem() was
  translated to x86_64,

- "bad.txt" is a hexdump of the destination area when CopyMem() was
  translated to ARMv7.

In order to assist with the analysis of this issue, I disassembled the
aarch64 binary with "objdump". Please find the listing in
"DxeCore.objdump", attached. The InternalMemCopyMem() function starts at
hex offset 2b2ec. The __memcpy() function starts at hex offset 2b180.

And, I ran the guest on the ARMv7 host with "-d
in_asm,op,op_opt,op_ind,out_asm". Please find the log in
"tcg.in_asm.op.op_opt.op_ind.out_asm.log", attached.

The TBs that correspond to (parts of) the InternalMemCopyMem() and
__memcpy() functions are scattered over the TCG log file, but the offset
between the "nice" disassembly from "DxeCore.objdump", and the in-RAM
TBs in the TCG log, can be determined from the fact that there is a
single prfm instruction in the entire binary. The instruction's offset
is 0x2b180 in "DxeCore.objdump" -- at the beginning of the __memcpy()
function --, and its RAM address is 0x472d2180 in the TCG log. Thus the
difference (= the load address of DxeCore.efi) is 0x472a7000.

QEMU was built at commit a4f667b67149 ("Merge remote-tracking branch
'remotes/cohuck/tags/s390x-20190521-3' into staging", 2019-05-21).

The reproducer command line is (on an ARMv7 host):

  qemu-system-aarch64 \
    -display none \
    -machine virt,accel=tcg \
    -nodefaults \
    -nographic \
    -drive if=pflash,format=raw,file=$prefix/share/qemu/edk2-aarch64-code.fd,readonly \
    -drive if=pflash,format=raw,file=$prefix/share/qemu/edk2-arm-vars.fd,snapshot=on \
    -cpu cortex-a57 \
    -chardev stdio,signal=off,mux=on,id=char0 \
    -mon chardev=char0,mode=readline \
    -serial chardev:char0

The apparent symptom is an assertion failure *in the guest*, such as

> ASSERT [DxeCore]
> /home/lacos/src/upstream/qemu/roms/edk2/MdePkg/Library/BaseLib/String.c(1090):
> Length < _gPcd_FixedAtBuild_PcdMaximumAsciiStringLength

but that is only a (distant) consequence of the CopyMem()
mistranslation, and resultant destination area corruption.

Originally reported in the following two mailing list messages:
- http://mid.mail-archive.com/9d2e260c-c491-03d2-9b8b-b57b72083f77@redhat.com
- http://mid.mail-archive.com/f1cec8c0-1a9b-f5bb-f951-ea0ba9d276ee@redhat.com

** Affects: qemu
     Importance: Undecided
         Status: New

** Attachment added: "aarch64-to-arm-mistranslation.tar.xz"
   https://bugs.launchpad.net/bugs/1830872/+attachment/5267323/+files/aarch64-to-arm-mistranslation.tar.xz

-- 
You received this bug notification because you are a member of qemu-
devel-ml, which is subscribed to QEMU.
https://bugs.launchpad.net/bugs/1830872

Title:
  AARCH64 to ARMv7 mistranslation in TCG

Status in QEMU:
  New

Bug description:
  The following guest code:

  https://github.com/tianocore/edk2/blob/3604174718e2afc950c3cc64c64ba5165c8692bd/MdePkg/Library/BaseMemoryLibOptDxe/AArch64/CopyMem.S

  implements, in hand-optimized aarch64 assembly, the CopyMem() edk2 (EFI
  Development Kit II) library function. (CopyMem() basically has memmove()
  semantics, to provide a standard C analog here.) The relevant functions
  are InternalMemCopyMem() and __memcpy().

  When TCG translates this aarch64 code to x86_64, everything works
  fine.

  When TCG translates this aarch64 code to ARMv7, the destination area of
  the translated CopyMem() function becomes corrupted -- it differs from
  the intended source contents. Namely, in every 4096 byte block, the
  8-byte word at offset 4032 (0xFC0) is zeroed out in the destination,
  instead of receiving the intended source value.

  I'm attaching two hexdumps of the same destination area:

  - "good.txt" is a hexdump of the destination area when CopyMem() was
    translated to x86_64,

  - "bad.txt" is a hexdump of the destination area when CopyMem() was
    translated to ARMv7.

  In order to assist with the analysis of this issue, I disassembled the
  aarch64 binary with "objdump". Please find the listing in
  "DxeCore.objdump", attached. The InternalMemCopyMem() function starts at
  hex offset 2b2ec. The __memcpy() function starts at hex offset 2b180.

  And, I ran the guest on the ARMv7 host with "-d
  in_asm,op,op_opt,op_ind,out_asm". Please find the log in
  "tcg.in_asm.op.op_opt.op_ind.out_asm.log", attached.

  The TBs that correspond to (parts of) the InternalMemCopyMem() and
  __memcpy() functions are scattered over the TCG log file, but the offset
  between the "nice" disassembly from "DxeCore.objdump", and the in-RAM
  TBs in the TCG log, can be determined from the fact that there is a
  single prfm instruction in the entire binary. The instruction's offset
  is 0x2b180 in "DxeCore.objdump" -- at the beginning of the __memcpy()
  function --, and its RAM address is 0x472d2180 in the TCG log. Thus the
  difference (= the load address of DxeCore.efi) is 0x472a7000.

  QEMU was built at commit a4f667b67149 ("Merge remote-tracking branch
  'remotes/cohuck/tags/s390x-20190521-3' into staging", 2019-05-21).

  The reproducer command line is (on an ARMv7 host):

    qemu-system-aarch64 \
      -display none \
      -machine virt,accel=tcg \
      -nodefaults \
      -nographic \
      -drive if=pflash,format=raw,file=$prefix/share/qemu/edk2-aarch64-code.fd,readonly \
      -drive if=pflash,format=raw,file=$prefix/share/qemu/edk2-arm-vars.fd,snapshot=on \
      -cpu cortex-a57 \
      -chardev stdio,signal=off,mux=on,id=char0 \
      -mon chardev=char0,mode=readline \
      -serial chardev:char0

  The apparent symptom is an assertion failure *in the guest*, such as

  > ASSERT [DxeCore]
  > /home/lacos/src/upstream/qemu/roms/edk2/MdePkg/Library/BaseLib/String.c(1090):
  > Length < _gPcd_FixedAtBuild_PcdMaximumAsciiStringLength

  but that is only a (distant) consequence of the CopyMem()
  mistranslation, and resultant destination area corruption.

  Originally reported in the following two mailing list messages:
  - http://mid.mail-archive.com/9d2e260c-c491-03d2-9b8b-b57b72083f77@redhat.com
  - http://mid.mail-archive.com/f1cec8c0-1a9b-f5bb-f951-ea0ba9d276ee@redhat.com

To manage notifications about this bug go to:
https://bugs.launchpad.net/qemu/+bug/1830872/+subscriptions

^ permalink raw reply	[flat|nested] 27+ messages in thread