On Fri, Jul 20, 2018 at 03:36:41PM -0700, Nick Desaulniers wrote:
native_save_fl() is marked static inline, but by using it as a function pointer in arch/x86/kernel/paravirt.c, it MUST be outlined.
paravirt's use of native_save_fl() also requires that no GPRs other than %rax are clobbered.
Compilers have different heuristics which they use to emit stack guard code, the emittance of which can break paravirt's callee saved assumption by clobbering %rcx.
Marking a function definition extern inline means that if this version cannot be inlined, then the out-of-line version will be preferred. By having the out-of-line version be implemented in assembly, it cannot be instrumented with a stack protector, which might violate custom calling conventions that code like paravirt rely on.
The semantics of extern inline has changed since gnu89. This means that folks using GCC versions >= 5.1 may see symbol redefinition errors at link time for subdirs that override KBUILD_CFLAGS (making the C standard used implicit) regardless of this patch. This has been cleaned up earlier in the patch set, but is left as a note in the commit message for future travelers.
Reports: https://lkml.org/lkml/2018/5/7/534 https://github.com/ClangBuiltLinux/linux/issues/16
Discussion: https://bugs.llvm.org/show_bug.cgi?id=37512 https://lkml.org/lkml/2018/5/24/1371
Thanks to the many folks that participated in the discussion.
Acked-by: Juergen Gross jgross@suse.com Debugged-by: Alistair Strachan astrachan@google.com Debugged-by: Matthias Kaehlcke mka@chromium.org Reported-by: Sedat Dilek sedat.dilek@gmail.com Signed-off-by: Nick Desaulniers ndesaulniers@google.com Suggested-by: Arnd Bergmann arnd@arndb.de Suggested-by: H. Peter Anvin hpa@zytor.com Suggested-by: Tom Stellar tstellar@redhat.com Tested-by: Sedat Dilek sedat.dilek@gmail.com
Backport for 4.4. 4.4 is missing commit 784d5699eddc "x86: move exports to actual definitions" which doesn't apply cleanly, and not really worth backporting IMO. It's simpler to change this patch from upstream:
- #include <asm-generic/export.h>
rather than
- #include <asm/export.h>
Yeah, that makes sense, thanks for the backport. I'll queue it up after the next round of stable kernels comes out in a few days.
greg k-h