Re: [PATCH -tip] kasan: Emit different calls for instrumentable memintrinsics

[Marco Elver]

On Mon, 13 Feb 2023 at 13:36, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
> On Mon, Feb 13, 2023 at 12:01:40PM +0100, Jakub Jelinek wrote:
>
> > The current gcc behavior is that operations like aggregate copies, or
> > clearing which might or might not need memcpy/memset/memmove under the hood
> > later are asan instrumented before the operation (in order not to limit the
> > choices on how it will be expanded), uses of builtins (__builtin_ prefixed
> > or not) are also instrumented before the calls unless they are one of the
> > calls that is recognized as always instrumented.  None for hwasan,
> > for asan:
> > index, memchr, memcmp, memcpy, memmove, memset, strcasecmp, strcat, strchr,
> > strcmp, strcpy, strdup, strlen, strncasecmp, strncat, strncmp, strcspn,
> > strpbrk, strspn, strstr, strncpy
> > and for those builtins gcc disables inline expansion and enforces a library
> > call (but until the expansion they are treated in optimizations like normal
> > builtins and so could be say DCEd, or their aliasing behavior is considered
> > etc.).  kasan behaves the same I think.
> >
> > Now, I think libasan only has __asan_ prefixed
> > __asan_memmove, __asan_memset and __asan_memcpy, nothing else, so most of
> > the calls from the above list even can't be prefixed.

Correct, right now libasan only does memmove, memset, and memcpy. I
don't think it'll ever do more, at least not in the near future.

> > So, do you want for --param asan-kernel-mem-intrinsic-prefix=1 to __asan_
> > prefix just memcpy/memmove/memset and nothing else?

Yes.

> > Is it ok to emit
> > memcpy/memset/memmove from aggregate operations which are instrumented
> > already at the caller (and similarly is it ok to handle those operations
> > inline)?

Yes, I think that's fair.

> I'm thinking it is trivial to add more __asan prefixed functions as
> needed, while trying to untangle the trainwreck created by assuming the
> normal functions are instrumented is much more work.

For the kernel param, I'd only do memcpy/memmove/memset, as those are
the most brittle ones. The string functions are instrumented on most
architectures through lib/string.c being instrumented.

Thanks,
-- Marco