A problem with __memset() is that, at least for me, it always ends
up being a call. There is a use case where we need to write only 1
byte, so I thought that introducing a call there (when compiling
without KMSAN) would be unacceptable.

Wonder what happens with that use case if we e.g. build with fortify-source.
Calling memset() for a single byte might be indicating the code is not hot.

quoted

...

quoted

+__no_sanitize_memory
+static inline void *memset_no_sanitize_memory(void *s, int c,
size_t n)
+{
+       return memset(s, c, n);
+}

I think depending on the compiler optimizations this might end up
being a call to normal memset, that would still change the shadow
bytes.

Interesting, do you have some specific scenario in mind? I vaguely
remember that in the past there were cases when sanitizer annotations
were lost after inlining, but I thought they were sorted out?

Sanitizer annotations are indeed lost after inlining, and we cannot do
much about that.
They are implemented using function attributes, and if a function
dissolves after inlining, we cannot possibly know which instructions
belonged to it.

Consider the following example (also available at
https://godbolt.org/z/5r7817G8e):

==================================
void *kmalloc(int size);

__attribute__((no_sanitize("kernel-memory")))
__attribute__((always_inline))
static void *memset_nosanitize(void *s, int c, int n) {
  return __builtin_memset(s, c, n);
}

void *do_something_nosanitize(int size) {
  void *ptr = kmalloc(size);
  memset_nosanitize(ptr, 0, size);
  return ptr;
}

void *do_something_sanitize(int size) {
  void *ptr = kmalloc(size);
  __builtin_memset(ptr, 0, size);
  return ptr;
}
==================================

If memset_nosanitize() has __attribute__((always_inline)), the
compiler generates the same LLVM IR calling __msan_memset() for both
do_something_nosanitize() and do_something_sanitize().
If we comment out this attribute, do_something_nosanitize() calls
memset_nosanitize(), which doesn't have the sanitize_memory attribute.

But even now __builtin_memset() is still calling __msan_memset(),
because __attribute__((no_sanitize("kernel-memory"))) somewhat
counterintuitively still preserves some instrumentation (see
include/linux/compiler-clang.h for details).
Replacing __attribute__((no_sanitize("kernel-memory"))) with
__attribute__((disable_sanitizer_instrumentation)) fixes this
situation:

define internal fastcc noundef ptr @memset_nosanitize(void*, int,
int)(ptr noundef returned writeonly %s, i32 noundef %n) unnamed_addr
#2 {
entry:
%conv = sext i32 %n to i64
tail call void @llvm.memset.p0.i64(ptr align 1 %s, i8 0, i64 %conv, i1 false)
ret ptr %s
}

And, in any case, if this were to happen, would not it be considered a
compiler bug that needs fixing there, and not in the kernel?

As stated above, I don't think this is more or less working as intended.
If we really want the ability to inline __memset(), we could transform
it into memset() in non-sanitizer builds, but perhaps having a call is
also acceptable?