[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: qemu-ppc64-linux-user: signal handler endianness in elfv2

From: Vincent Fazio
Subject: Re: qemu-ppc64-linux-user: signal handler endianness in elfv2
Date: Mon, 10 Feb 2020 08:07:20 -0600
User-agent: Mozilla/5.0 (X11; Linux x86_64; rv:68.0) Gecko/20100101 Thunderbird/68.4.1


Thanks for the response.

On 2/10/20 1:08 AM, David Gibson wrote:
On Sat, Feb 08, 2020 at 12:57:40PM -0600, Vincent Fazio wrote:
I ran into a crash scenario in ppc64-linux-user when trying to run openssl
compiled with gcc & musl which emits an ELFv2 binary. The target CPU was a
big endian model (e6500)
Uh.. so I'm pretty sure ABIv2 is little-endian specific.  If you want
BE binaries, I think you need to configure your toolchain differently.

This is _generally_ the case. MUSL libc uses ELFv2 ABI for both big and little endian, so it's apparently assumed that gcc is set with --with-abi=elfv2. This is how it's currently configured with Buildroot.


OpenSSL recently had a similar assumption patched here: https://github.com/openssl/openssl/commit/098404128383ded87ba390dd74ecd9e2ffa6f530

I imagine the most frequent test case is running ppc64le on an x86_64 host where the swap never even occurs due to the endianness match.

But if the fix is to just not do the swap for ELFv2 binaries because the function pointer is always aligned with the host's memory endianness due to how functions addresses are calculated in ELFv2, i don't mind submitting a patch.

If QEMU decides this is an unsupported case, that's probably fine because it's likely not in use in large scale.


vfazio@vfazio1 ~/development/buildroot/output/target :( $ gdb
(gdb) run -d guest_errors -cpu e6500 -E 
lib/libc.so bin/bash
Starting program: /home/vfazio/development/qemu/ppc64-linux-user/qemu-ppc64
-d guest_errors -cpu e6500 -E 
lib/libc.so bin/bash
[Thread debugging using libthread_db enabled]
Using host libthread_db library "/lib/x86_64-linux-gnu/libthread_db.so.1".
[New Thread 0x7ffff7ff9700 (LWP 27186)]
invalid/unsupported opcode: 1f - 0c - 1a - 00 (7c005698) 0000000010028e58 0

Thread 1 "qemu-ppc64" received signal SIGSEGV, Segmentation fault.
0x00000000600a9257 in ldl_he_p (ptr=0x4c2c061000000000) at
351        __builtin_memcpy(&r, ptr, sizeof(r));

(gdb) !readelf -h bin/bash | grep abi
   Flags:                             0x2, abiv2
(gdb) !readelf -h lib/libc.so | grep abi
   Flags:                             0x2, abiv2

openssl does "probing" for PPC features (Altivec, Crypto, etc) by trying to
execute possibly illegal instructions and handling them... hence that
SIGILL. The problem is when it comes to jumping to the signal handler

#0  0x00000000600a9257 in ldl_he_p (ptr=0x4c2c061000000000) at
#1  0x00000000600a92fe in ldl_be_p (ptr=0x4c2c061000000000) at
#2  0x00000000600c0790 in translator_ldl_swap (env=0x62923150,
pc=5488768711647035392, do_swap=false) at
#3  0x000000006011c1ab in ppc_tr_translate_insn (dcbase=0x7fffffffd250,
cs=0x6291ae80) at /home/vfazio/development/qemu/target/ppc/translate.c:7856
#4  0x000000006005ae70 in translator_loop (ops=0x60805fc0 <ppc_tr_ops>,
db=0x7fffffffd250, cpu=0x6291ae80, tb=0x60a5f900
<static_code_gen_buffer+1681600>, max_insns=512)
     at /home/vfazio/development/qemu/accel/tcg/translator.c:102

The handler has what looks like an LE address but then swaps it since the
host is LE and the target is BE

setup_rt_frame (sig=4, ka=0x628c5f80 <sigact_table+96>, info=0x62931658,
set=0x7fffffffd2f8, env=0x62923150) at
575            qemu_log_mask(LOG_GUEST_ERROR, "sa_handler NIP to "
TARGET_FMT_lx "\n", ka->_sa_handler);
(gdb) p/x ka->_sa_handler
$11 = 0x10062c4c
579            env->nip = tswapl((target_ulong) ka->_sa_handler);
(gdb) p/x env->nip
$12 = 0x4c2c061000000000

the memcpy later faults because that's obviously not a valid address from
which to grab the PPC instructions that need to be translated.

I'm not familiar with the ELFv2 ABI or PPC assembly in general, so I'm not
sure what to expect here. Typically ELFv2 was reserved for ppc64le, but musl
uses ELFv2 for all ppc64 targets. This likely wouldn't be an issue in
ppc64LE since x86_64 is LE as well and no swap would take place.

Is the signal handler address tied to the endianess of the host? I noticed
there was no swapping in elfload.c so wasn't sure...removing the swap gets
me a little further in program execution but i do eventually crash (but that
may not be related to this specific issue).


Vincent Fazio
Embedded Software Engineer - Linux
Extreme Engineering Solutions, Inc

reply via email to

[Prev in Thread] Current Thread [Next in Thread]