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
|
performance: 0.604
device: 0.574
other: 0.545
socket: 0.479
semantic: 0.462
network: 0.446
vnc: 0.436
PID: 0.398
graphic: 0.376
permissions: 0.359
boot: 0.342
files: 0.342
debug: 0.312
KVM: 0.248
No check for unaligned data access in ARM32 instructions
hi
According to the ARM documentation, there are alignment requirements of load/store instructions. Alignment fault should be raised if the alignment check is failed. However, it seems that QEMU doesn't implement this, which is against the documentation of ARM. For example, the instruction LDRD/STRD/LDREX/STREX must check the address is word alignment no matter what value the SCTLR.A is.
I attached a testcase, which contains a instruction at VA 0x10240: ldrd r0,[pc.#1] in the main function. QEMU can successfully load the data in the unaligned address. The test is done in QEMU 5.1.0. I can provide more testcases for the other instructions if you need. Many thanks.
To patch this, we need a check while we translate the instruction to tcg. If the address is unaligned, a signal number (i.e., SIGBUS) should be raised.
Regards
Muhui
We don't implement SCTLR.A, but you're right that we should be
checking the mandatory alignments.
Note! Any fix will only apply to system mode (qemu-system-arm)
and not user-only mode (qemu-arm).
Thanks for confirmation.
Btw: I was wondering why the fix will only apply to system mode rather than user-only mode. Unaligned data access is not permitted in user level programs, either.
Because for user-only, we cheat and use host load/store
operations directly. This makes for much faster emulation
but imposes a number of limitations -- including ignoring
of the alignment bits on hosts that have native unaligned
accesses.
As a corollary, when running user-only on a host that
enforces alignment, you cannot emulate a guest that
*allows* unaligned accesses.
Proposed patches:
https://<email address hidden>/
Richard's patches have been merged (see https://git.qemu.org/?p=qemu.git;a=commitdiff;h=4d753eb5fb03ee7bc71ec and the following ones), so I'm setting the state to "Fix committed" now.
|