Re: [PATCH] kbuild: clang: Support building UM with SUBARCH=i386

[Kees Cook <kees@kernel.org>]

On Wed, Mar 05, 2025 at 03:45:54PM +0100, Nathan Chancellor wrote:
> On Tue, Mar 04, 2025 at 09:07:57AM -0800, Kees Cook wrote:
> > On Tue, Mar 04, 2025 at 03:51:19PM +0100, Thomas Weißschuh wrote:
> > > No, it doesn't.
> > > 
> > > Running tests with:
> > > $ .kunit/linux kunit.filter_glob=overflow.DEFINE_FLEX_test kunit.enable=1 mem=1G console=tty kunit_shutdown=halt
> > > [15:48:30] =================== overflow (1 subtest) ===================
> > > [15:48:30] # DEFINE_FLEX_test: EXPECTATION FAILED at lib/overflow_kunit.c:1200
> > > [15:48:30] Expected __builtin_dynamic_object_size(two_but_zero, 0) == expected_raw_size, but
> > > [15:48:30]     __builtin_dynamic_object_size(two_but_zero, 0) == 12 (0xc)
> > > [15:48:30]     expected_raw_size == 8 (0x8)
> > > [15:48:30] [FAILED] DEFINE_FLEX_test
> > > [15:48:30] # module: overflow_kunit
> > > [15:48:30] ==================== [FAILED] overflow =====================
> > > [15:48:30] ============================================================
> > > [15:48:30] Testing complete. Ran 1 tests: failed: 1
> > > [15:48:31] Elapsed time: 43.985s total, 0.001s configuring, 43.818s building, 0.133s running
> > > 
> > > If I force CONFIG_CC_HAS_COUNTED_BY=n then the test succeeds.
> > > Clang 19.1.7 from the Arch Linux repos.
> > 
> > I wasn't seeing with Clang 20 from git:
> > ClangBuiltLinux clang version 20.0.0git (git@github.com:llvm/llvm-project.git 72901fe19eb1e55d0ee1c380ab7a9f57d2f187c5)
> > 
> > But I do see the error with ToT Clang:
> > ClangBuiltLinux clang version 21.0.0git (git@github.com:llvm/llvm-project.git eee3db5421040cfc3eae6e92ed714650a6f741fa)
> > 
> > Clang 17.1: (does not support counted_by)
> > 
> >     # DEFINE_FLEX_test: missing counted_by
> >     # DEFINE_FLEX_test: sizeof(two_but_zero): 8
> >     # DEFINE_FLEX_test: __struct_size(two_but_zero): 12
> >     # DEFINE_FLEX_test: __member_size(two_but_zero): 12
> >     # DEFINE_FLEX_test: __member_size(two_but_zero->array): 4
> > 
> > Clang 19.1.1: (actually is _does_ support counted_by, but Linux disables it)
> > 
> >     # DEFINE_FLEX_test: missing counted_by
> >     # DEFINE_FLEX_test: sizeof(two_but_zero): 8
> >     # DEFINE_FLEX_test: __struct_size(two_but_zero): 12
> >     # DEFINE_FLEX_test: __member_size(two_but_zero): 12
> >     # DEFINE_FLEX_test: __member_size(two_but_zero->array): 4
> > 
> > GCC 13.3:
> > 
> >     # DEFINE_FLEX_test: missing counted_by
> >     # DEFINE_FLEX_test: sizeof(two_but_zero): 8
> >     # DEFINE_FLEX_test: __struct_size(two_but_zero): 12
> >     # DEFINE_FLEX_test: __member_size(two_but_zero): 12
> >     # DEFINE_FLEX_test: __member_size(two_but_zero->array): 4
> > 
> > Clang 21 (ToT):
> > 
> >     # DEFINE_FLEX_test: has counted_by
> >     # DEFINE_FLEX_test: sizeof(two_but_zero): 8
> >     # DEFINE_FLEX_test: __struct_size(two_but_zero): 12
> >     # DEFINE_FLEX_test: __member_size(two_but_zero): 12
> >     # DEFINE_FLEX_test: __member_size(two_but_zero->array): 0
> > 
> > GCC 15 (ToT):
> > 
> >     # DEFINE_FLEX_test: has counted_by
> >     # DEFINE_FLEX_test: sizeof(two_but_zero): 8
> >     # DEFINE_FLEX_test: __struct_size(two_but_zero): 12
> >     # DEFINE_FLEX_test: __member_size(two_but_zero): 12
> >     # DEFINE_FLEX_test: __member_size(two_but_zero->array): 0
> > 
> > It seems like the on-stack sizes with __bdos all agree now, regardless
> > of the used compiler features. It is only the array size calculation
> > that now gets masked by counted_by. (i.e. the stack size is overridden
> > by the zero "count" for the array elements.)
> > 
> > I'll send a fix for the test...
> 
> Just for my own understanding, is this because of the adjustment that
> Bill did to the __bdos() calculation in [1]? I think that tracks because
> the version of LLVM 20 that you have is pretty old and does not have
> that change. I know for a fact I tested the original change to the
> overflow KUnit test to adjust the expected calculation result and it
> passed but it was before that change as well. If I use a current version
> of LLVM 20, I see the failure. If I allow LLVM 18 to use __counted_by(),
> the test passes with it. Not that it truly matters but it does explain
> how we got to this point.

Yes, totally! This is exactly how I got there too. Great; thank you for
summarizing! :)

-- 
Kees Cook