On Thu, Apr 23, 2020 at 12:32:32PM -0700, Kees Cook wrote: > On Thu, Apr 23, 2020 at 01:28:45PM -0500, Josh Poimboeuf wrote: > > On Thu, Apr 23, 2020 at 11:06:14AM -0700, Kees Cook wrote: > > > Regardless, the type_mismatch it triggered for misalignment and > > > object-size checks, and the alignment check is likely going to always > > > misfire on x86. The randconfig includes that config: > > > > > > CONFIG_UBSAN_ALIGNMENT=y > > > > > > So perhaps the config should be strengthened to disallow it under > > > COMPILE_TEST? > > > > > > config UBSAN_ALIGNMENT > > > def_bool !UBSAN_NO_ALIGNMENT > > > depends on !COMPILE_TEST > > > > But this issue basically makes UBSAN_TRAP unusable, right? Should > > UBSAN_ALIGNMENT and UBSAN_TRAP be made mutually exclusive? > > Well, no, not UBSAN_TRAP alone. It seems that -fsanitize=alignment (and > possibly -fsanitize=object-size) is likely to produce false positives. And > that was already well understood. With UBSAN_TRAP, though, those functions > become unusable. Part of the goal of adding UBSAN_TRAP was to be able to > extract the non-false-positive sanitizer options out of UBSAN_MISC so > they could be used in production without the large debug text overhead. > (And the "alignment" checker had long since already been separate and > set to false by default on x86, since it's going to trip over everything > since it shouldn't get run on such hardware for any sane reason that > I know of.) So, while UBSAN_TRAP with UBSAN_ALIGNMENT is going to be > especially bad, UBSAN_ALIGNMENT at all is likely bad enough on x86, so > I think it'd be best to make it a negative depends on COMPILE_TEST (but > leave it available if someone is doing some really unusual UBSAN work). I think I'm still missing why UBSAN_ALIGNMENT is even possible on x86. But at least it sounds like your fix will keep the objtool reports from coming in :-) -- Josh
