Annotation of wikisrc/users/maya.mdwn, revision 1.4
1.1 maya 1: # Tips for using analyzers on NetBSD #
3: ## Address Sanitizer (+UBsan) (preliminary) ##
1.3 maya 5: ASan reports memory violations, and detects many off-by-ones. It seems to produce very high quality reports.
1.1 maya 6:
1.3 maya 7: It only needs to be compiled on the resulting binary. It cannot compile static objects so requires some fiddling with makefiles to disable those.
1.1 maya 8:
9: I've been running it on netbsd tests in the following manner:
11: cd /usr/src/tests/lib/libc
12: env USETOOLS=never MK_SSP=no HAVE_SSP=no CFLAGS="-fno-omit-frame-pointer -O0 -g -ggdb -U_FORTIFY_SOURCE -fsanitize=address -fsanitize=undefined" LDFLAGS="-lasan -lubsan" make -j20
1.3 maya 14: env ASAN_OPTIONS=alloc_dealloc_mismatch=0 LD_PRELOAD="/usr/lib/libasan.so /usr/lib/libubsan.so" atf-run # 
1.1 maya 15:
1.3 maya 16: sysctl -w security.pax.aslr.enabled=0 # 
1.2 maya 17:
1.1 maya 18: 1. Seems like this is a cause of worse reports, as in-library functions
19: are not intercepted.
21: 2. Not even close to canonical commands, should probably be improved.
1.3 maya 23: 3. ASAN_OPTIONS=alloc_dealloc_mismatch=0 is because atf-run itself triggers a bug. Should have a look at it so this option doesn't need to be disabled.
1.2 maya 24:
1.3 maya 25: 4. To workaround "Shadow memory range interleaves with an existing memory mapping. ASan cannot proceed correctly". An alternative is to LD_PRELOAD and LD_LIBRARY_PATH an entirely separate libc, ld.elf_so, etc.
27: 5. ASan can be wrong about which function is faulting, if we're talking about functions internal to the library. Running it on a separate file or in GDB can be helpful. Worth trying if the failure doesn't make sense.
1.1 maya 28:
29: Important note: ASan should not be run on production systems. [It can pose a
30: security risk](http://www.openwall.com/lists/oss-security/2016/02/17/9).
32: ## Coverity ##
34: Coverity is a static analyzer.
36: You can see a part of its output in coverity-updates@, and a lot more
37: if you go to the website (sign up with your netbsd email or poke someone
38: for access).
40: A lot of the reports are about strncpy/strcpy or in code that belongs
41: to GCC (in the case of userland), you can tackle this by limiting
42: results to a particular directory (click the folder icon). You can
43: also sort by issue.
45: Some suggestions for things to focus on, as there are many defects
1.4 ! maya 47:
1.1 maya 48: - Setuid programs
49: - Anything kernel
50: - Stuff that runs as root
51: - Library or other code you know well already
52: - Drivers for hardware you actually own and can test
54: Future ideas:
1.2 maya 56: - GCC could be told to add ASan flags for all shared objects, making it easier to build world with those flags
1.1 maya 57: - We could run all of NetBSD with ASan for some real world tests.
1.2 maya 58: - ASan for kernel? (subr_kmem.c has some flags which do some of the work, could it do more?)
1.1 maya 59: - Fuzzers are cool.
CVSweb for NetBSD wikisrc <wikimaster@NetBSD.org> software: FreeBSD-CVSweb