diff mbox series

objtool: Fix UNWIND_HINT_{SAVE,RESTORE} across basic blocks

Message ID 20240227073527.avcm5naavbv3cj5s@treble
State New
Headers
Series objtool: Fix UNWIND_HINT_{SAVE,RESTORE} across basic blocks |

Commit Message

Josh Poimboeuf Feb. 27, 2024, 7:35 a.m. UTC 
  On Fri, Feb 23, 2024 at 04:45:22PM -0800, Kees Cook wrote:
> Josh, Peter, can you give a clue about what's happened here? For
> added context, the referenced commit is changing a noreturn function
> prototype[1], which I know gets some special objtool handling, but it's
> converting a pointer arg to a u8... I can't imagine what has gone weird
> here. :P

I think this convinced GCC to reshuffle some basic blocks, which
uncovered an objtool bug.  Or, objtool just wasn't smart enough, as the
warning says ;-)

This should fix it, absent any more warnings.  Kees, feel free to take
it in your branch if that helps.

---8<---

From: Josh Poimboeuf <jpoimboe@kernel.org>
Subject: [PATCH] objtool: Fix UNWIND_HINT_{SAVE,RESTORE} across basic blocks

If SAVE and RESTORE unwind hints are in different basic blocks, and
objtool sees the RESTORE before the SAVE, it errors out with:

  vmlinux.o: warning: objtool: vmw_port_hb_in+0x242: objtool isn't smart enough to handle this CFI save/restore combo

In such a case, defer following the RESTORE block until the
straight-line path gets followed later.

Fixes: 8faea26e6111 ("objtool: Re-add UNWIND_HINT_{SAVE_RESTORE}")
Reported-by: kernel test robot <lkp@intel.com>
Closes: https://lore.kernel.org/oe-kbuild-all/202402240702.zJFNmahW-lkp@intel.com/
Signed-off-by: Josh Poimboeuf <jpoimboe@kernel.org>
---
 tools/objtool/check.c | 12 ++++++++++++
 1 file changed, 12 insertions(+)

Comments

Kees Cook March 1, 2024, 6:33 a.m. UTC | #1 
On Mon, Feb 26, 2024 at 11:35:27PM -0800, Josh Poimboeuf wrote:
> On Fri, Feb 23, 2024 at 04:45:22PM -0800, Kees Cook wrote:
> > Josh, Peter, can you give a clue about what's happened here? For
> > added context, the referenced commit is changing a noreturn function
> > prototype[1], which I know gets some special objtool handling, but it's
> > converting a pointer arg to a u8... I can't imagine what has gone weird
> > here. :P
> 
> I think this convinced GCC to reshuffle some basic blocks, which
> uncovered an objtool bug.  Or, objtool just wasn't smart enough, as the
> warning says ;-)
> 
> This should fix it, absent any more warnings.  Kees, feel free to take
> it in your branch if that helps.

Thanks very much! I'll carry it in my tree.

-Kees
diff mbox series

Patch

diff --git a/tools/objtool/check.c b/tools/objtool/check.c
index 548ec3cd7c00..c4c2f75eadfd 100644
--- a/tools/objtool/check.c
+++ b/tools/objtool/check.c
@@ -3620,6 +3620,18 @@  static int validate_branch(struct objtool_file *file, struct symbol *func,
 				}
 
 				if (!save_insn->visited) {
+					/*
+					 * If the restore hint insn is at the
+					 * beginning of a basic block and was
+					 * branched to from elsewhere, and the
+					 * save insn hasn't been visited yet,
+					 * defer following this branch for now.
+					 * It will be seen later via the
+					 * straight-line path.
+					 */
+					if (!prev_insn)
+						return 0;
+
 					WARN_INSN(insn, "objtool isn't smart enough to handle this CFI save/restore combo");
 					return 1;
 				}