More hygienic linking [LWN.net]

More hygienic linking

Posted Mar 31, 2024 3:53 UTC (Sun) by danobi (subscriber, #102249)
In reply to: More hygienic linking by epa
Parent article: A backdoor in xz

I was wondering that as well and tried to get ifunc to behave like that. Couldn't get it to work. But then I read https://gynvael.coldwind.pl/?lang=en&id=782 and saw this bit:

((head -c +$N > /dev/null 2>&1) && head -c +$W) > liblzma_la-crc64-fast.o

Looks like the binary payload just replaces one of the ifunc implementations at build time. At least that's my guess.

to post comments

More hygienic linking

Posted Apr 1, 2024 11:24 UTC (Mon) by excors (subscriber, #95769) [Link] (1 responses)

I think there's a bit more than that. The injected.txt script in Andres Freund's email includes:

V='[...]extern int _get_cpuid(int, void*, void*, void*, void*, void*);\nstatic inline bool _is_arch_extension_supported(void) { int success = 1; uint32_t r[4]; success = _get_cpuid(1, &r[0], &r[1], &r[2], &r[3], [...]'
[...]
sed "/return is_arch_extension_supported()/ c\return _is_arch_extension_supported()" $top_srcdir/src/liblzma/check/crc64_fast.c

In the upstream code, is_arch_extension_supported calls __get_cpuid (provided by GCC). The backdoored build script modifies crc64_fast.c so it calls _get_cpuid (one underscore) instead. _get_cpuid is defined in liblzma_la-crc64-fast.o, which is a binary hidden in the test files.

If I understand correctly, the loader executes the ifunc resolver function (crc64_resolve) when liblzma is loaded. crc64_resolve calls [_]is_arch_extension_supported to determine if it can use the optimised version with the CLMUL instruction. But now that actually calls _get_cpuid which initialises the runtime backdoor.

So, there's nothing funny happening with the linker and there's no symbol overriding. The functions are "replaced" using sed before compiling.

(I guess they could have simply had the build script insert an __attribute__((constructor)) function, but anyone debugging or profiling a liblzma-using application might have been surprised to see liblzma functions getting executed at startup. Using ifuncs makes it less suspicious - liblzma has a good reason to run crc64_resolve and ...get_cpuid, clearly visible in the upstream code (modulo some underscores), so they probably hoped anyone investigating would stop at that point.)