work around gcc bugs with 'asm goto' with outputs
authorLinus Torvalds <torvalds@linux-foundation.org>
Fri, 9 Feb 2024 20:39:31 +0000 (12:39 -0800)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 9 Feb 2024 23:57:48 +0000 (15:57 -0800)
commit4356e9f841f7fbb945521cef3577ba394c65f3fc
tree87bb7c1de6e813f9d7e9e57bb54c9116d3f83137
parent9ed18b0b7765ec3aa5f10843d1c3daaa1a225337
work around gcc bugs with 'asm goto' with outputs

We've had issues with gcc and 'asm goto' before, and we created a
'asm_volatile_goto()' macro for that in the past: see commits
3f0116c3238a ("compiler/gcc4: Add quirk for 'asm goto' miscompilation
bug") and a9f180345f53 ("compiler/gcc4: Make quirk for
asm_volatile_goto() unconditional").

Then, much later, we ended up removing the workaround in commit
43c249ea0b1e ("compiler-gcc.h: remove ancient workaround for gcc PR
58670") because we no longer supported building the kernel with the
affected gcc versions, but we left the macro uses around.

Now, Sean Christopherson reports a new version of a very similar
problem, which is fixed by re-applying that ancient workaround.  But the
problem in question is limited to only the 'asm goto with outputs'
cases, so instead of re-introducing the old workaround as-is, let's
rename and limit the workaround to just that much less common case.

It looks like there are at least two separate issues that all hit in
this area:

 (a) some versions of gcc don't mark the asm goto as 'volatile' when it
     has outputs:

        https://gcc.gnu.org/bugzilla/show_bug.cgi?id=98619
        https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110420

     which is easy to work around by just adding the 'volatile' by hand.

 (b) Internal compiler errors:

        https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110422

     which are worked around by adding the extra empty 'asm' as a
     barrier, as in the original workaround.

but the problem Sean sees may be a third thing since it involves bad
code generation (not an ICE) even with the manually added 'volatile'.

but the same old workaround works for this case, even if this feels a
bit like voodoo programming and may only be hiding the issue.

Reported-and-tested-by: Sean Christopherson <seanjc@google.com>
Link: https://lore.kernel.org/all/20240208220604.140859-1-seanjc@google.com/
Cc: Nick Desaulniers <ndesaulniers@google.com>
Cc: Uros Bizjak <ubizjak@gmail.com>
Cc: Jakub Jelinek <jakub@redhat.com>
Cc: Andrew Pinski <quic_apinski@quicinc.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
35 files changed:
arch/arc/include/asm/jump_label.h
arch/arm/include/asm/jump_label.h
arch/arm64/include/asm/alternative-macros.h
arch/arm64/include/asm/jump_label.h
arch/csky/include/asm/jump_label.h
arch/loongarch/include/asm/jump_label.h
arch/mips/include/asm/jump_label.h
arch/parisc/include/asm/jump_label.h
arch/powerpc/include/asm/jump_label.h
arch/powerpc/include/asm/uaccess.h
arch/powerpc/kernel/irq_64.c
arch/riscv/include/asm/arch_hweight.h
arch/riscv/include/asm/bitops.h
arch/riscv/include/asm/checksum.h
arch/riscv/include/asm/cpufeature.h
arch/riscv/include/asm/jump_label.h
arch/riscv/lib/csum.c
arch/s390/include/asm/jump_label.h
arch/sparc/include/asm/jump_label.h
arch/um/include/asm/cpufeature.h
arch/x86/include/asm/cpufeature.h
arch/x86/include/asm/jump_label.h
arch/x86/include/asm/rmwcc.h
arch/x86/include/asm/special_insns.h
arch/x86/include/asm/uaccess.h
arch/x86/kvm/svm/svm_ops.h
arch/x86/kvm/vmx/vmx.c
arch/x86/kvm/vmx/vmx_ops.h
arch/xtensa/include/asm/jump_label.h
include/linux/compiler-gcc.h
include/linux/compiler_types.h
net/netfilter/nft_set_pipapo_avx2.c
samples/bpf/asm_goto_workaround.h
tools/arch/x86/include/asm/rmwcc.h
tools/include/linux/compiler_types.h