Re: [PATCH v3] mm, slub: change run-time assertion in kmalloc_index() to compile-time

[Nathan Chancellor]

On Sat, May 15, 2021 at 11:24:25PM +0200, Vlastimil Babka wrote:
> On 5/15/21 11:09 PM, Hyeonggon Yoo wrote:
> > Hello Vlastimil, recently kbuild-all test bot reported compile error on
> > clang 10.0.1, with defconfig.
> 
> Hm yes, catching some compiler bug was something that was noted to be
> possible to happen.
> 
> > Nathan Chancellor wrote:
> >> I think this happens because arch_prepare_optimized_kprobe() calls kzalloc()
> >> with a size of MAX_OPTINSN_SIZE, which is
> >>
> >> #define MAX_OPTINSN_SIZE                                \
> >>       (((unsigned long)optprobe_template_end -        \
> >>          (unsigned long)optprobe_template_entry) +     \
> >>         MAX_OPTIMIZED_LENGTH + JMP32_INSN_SIZE)
> > 
> >> and the optprobe_template_{end,entry} are not evaluated as constants.
> >>
> >> I am not sure what the solution is. There seem to be a growing list of issues
> >> with LLVM 10 that were fixed in LLVM 11, which might necessitate requiring
> >> LLVM 11 and newer to build the kernel, given this affects a defconfig.
> >> Cheers,
> >> Nathan
> > 
> > 
> > I think it's because kmalloc compiles successfully when size is constant,
> > and kmalloc_index isn't. so I think compiler seems to be confused.
> > 
> > currently if size is non-constant, kmalloc calls dummy function __kmalloc,
> > which always returns NULL.
> 
> That's a misunderstanding. __kmalloc() is not a dummy function, you
> probably found only the header declaration.
> 
> > so what about changing kmalloc to do compile-time assertion too, and track
> > all callers that are calling kmalloc with non-constant argument.
> 
> kmalloc() is expected to be called with both constant and non-constant
> size. __builtin_constant_p() is used to determine which implementation
> to use. One based on kmalloc_index(), other on __kmalloc().
> 
> It appears clang 10.0.1 is mistakenly evaluating __builtin_constant_p()
> as true. Probably something to do with LTO, because MAX_OPTINSN_SIZE
> seems it could be a "link-time constant".

This happens with x86_64 defconfig so LTO is not involved.

However, the explanation makes sense, given that the LLVM change I
landed on changes the sparse conditional constant propagation pass,
which I believe can influence how LLVM handles __builtin_constant_p().

> Maybe we could extend Marco Elver's followup patch that uses
> BUILD_BUG_ON vs BUG() depending on size_is_constant parameter. It could
> use BUG() also if the compiler is LLVM < 11 or something. What would be
> the proper code for this condition?

This should work I think:

diff --git a/include/linux/slab.h b/include/linux/slab.h
index 9d316aac0aba..1b653266f2aa 100644
--- a/include/linux/slab.h
+++ b/include/linux/slab.h
@@ -413,7 +413,7 @@ static __always_inline unsigned int __kmalloc_index(size_t size,
 	if (size <=  16 * 1024 * 1024) return 24;
 	if (size <=  32 * 1024 * 1024) return 25;
 
-	if (size_is_constant)
+	if ((IS_ENABLED(CONFIG_CC_IS_GCC) || CONFIG_CLANG_VERSION > 110000) && size_is_constant)
 		BUILD_BUG_ON_MSG(1, "unexpected size in kmalloc_index()");
 	else
 		BUG();