Re: [PATCH 07/43] compiler_attributes.h: add __disable_sanitizer_instrumentation

From
Marco Elver <elver@google.com>
Date
2021-12-15 13:33:51
Also in
linux-mm, lkml
Reply to
Mark Rutland, 2021-12-15
A

On Wed, 15 Dec 2021 at 14:24, Mark Rutland [off-list ref] wrote:
On Tue, Dec 14, 2021 at 05:20:14PM +0100, Alexander Potapenko wrote:
quoted
The new attribute maps to
__attribute__((disable_sanitizer_instrumentation)), which will be
supported by Clang >= 14.0. Future support in GCC is also possible.

This attribute disables compiler instrumentation for kernel sanitizer
tools, making it easier to implement noinstr. It is different from the
existing __no_sanitize* attributes, which may still allow certain types
of instrumentation to prevent false positives.
When you say the __no_sanitize* attributes allow some instrumentation, does
that apply to any of the existing KASAN/KCSAN/KCOV support, or just for KMSAN?

The documentation just says the same as the commit message:

| This is not the same as __attribute__((no_sanitize(...))), which depending on
| the tool may still insert instrumentation to prevent false positive reports.

... which implies the other instrumentation might not be suprressed.

I ask because architectures which select ARCH_WANTS_NO_INSTR *need* to be able
to suppress all instrumentation. It's fine if that means they need a new
version of clang for KMSAN, but if there's latent instrumentation we have more
bugs to fix first...
Thus far, none of the existing K*SANs added other instrumentation.
Apart from KMSAN here, this will change with KCSAN's barrier
instrumentation, which is why this patch is also part of KCSAN's
upcoming changes -- recall I said I fixed barrier instrumentation for
arm64 as well, this is how :-)

See https://lore.kernel.org/all/20211130114433.2580590-26-elver@google.com/ (local)
how I resolved it for KCSAN on architectures that don't have objtool.

I expect this patch will be dropped from the KMSAN series once it
reaches mainline through the KCSAN changes.

Also note, this applies only for bug-detection tools, that may want to
avoid false positives. So by definition, it is irrelevant for KCOV
(which had its own attribute woes a while back though). Yeah, it's
been a long road to get the compilers to play along ... :-/

Thanks,
-- Marco
Thread (74 messages) 74 messages, 7 authors, last reply 2022-03-25
Keyboard shortcuts
hback out one level
jnext message in thread / next series-diff position
kprevious message in thread / previous series-diff position
ldrill in
Escclose help / fold thread tree
?toggle this help