1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
|
AARCH64 to ARMv7 mistranslation in TCG
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://<email address hidden>
- http://<email address hidden>
|