Re: objtool crashes on clang output (drivers/hwmon/pmbus/adm1275.o)

From: Josh Poimboeuf <jpoimboe@redhat.com>
To: Arnd Bergmann <arnd@arndb.de>
Cc: Jann Horn <jannh@google.com>,
	Peter Zijlstra <peterz@infradead.org>,
	Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
	clang-built-linux <clang-built-linux@googlegroups.com>,
	Nick Desaulniers <ndesaulniers@google.com>
Subject: Re: objtool crashes on clang output (drivers/hwmon/pmbus/adm1275.o)
Date: Fri, 12 Jul 2019 08:57:55 -0500	[thread overview]
Message-ID: <20190712135755.7qa4wxw3bfmwn5rp@treble> (raw)
In-Reply-To: <CAK8P3a1_8kjzamn6_joBbZTO8NeGn0E3O+MZ+bcOQ0HkkRHXRQ@mail.gmail.com>

On Fri, Jul 12, 2019 at 09:51:35AM +0200, Arnd Bergmann wrote:
> I no longer see any of the "can't find switch jump table" in last
> nights randconfig
> builds. I do see one other rare warning, see attached object file:
> 
> fs/reiserfs/do_balan.o: warning: objtool: replace_key()+0x158: stack
> state mismatch: cfa1=7+40 cfa2=7+56
> fs/reiserfs/do_balan.o: warning: objtool: balance_leaf()+0x2791: stack
> state mismatch: cfa1=7+176 cfa2=7+192
> fs/reiserfs/ibalance.o: warning: objtool: balance_internal()+0xe8f:
> stack state mismatch: cfa1=7+240 cfa2=7+248
> fs/reiserfs/ibalance.o: warning: objtool:
> internal_move_pointers_items()+0x36f: stack state mismatch: cfa1=7+152
> cfa2=7+144
> fs/reiserfs/lbalance.o: warning: objtool:
> leaf_cut_from_buffer()+0x58b: stack state mismatch: cfa1=7+128
> cfa2=7+112
> fs/reiserfs/lbalance.o: warning: objtool:
> leaf_copy_boundary_item()+0x7a9: stack state mismatch: cfa1=7+104
> cfa2=7+96
> fs/reiserfs/lbalance.o: warning: objtool:
> leaf_copy_items_entirely()+0x3d2: stack state mismatch: cfa1=7+120
> cfa2=7+128
> 
> I suspect this comes from the calls to the __reiserfs_panic() noreturn function,
> but have not actually looked at the object file.

Looking at one of the examples:

    2346:       0f 85 6a 01 00 00       jne    24b6 <leaf_copy_items_entirely+0x3a8>
    ...
    23b1:       e9 2a 01 00 00          jmpq   24e0 <leaf_copy_items_entirely+0x3d2>
    ...
    24b6:       31 ff                   xor    %edi,%edi
    24b8:       48 c7 c6 00 00 00 00    mov    $0x0,%rsi
                        24bb: R_X86_64_32S      .rodata.str1.1
    24bf:       48 c7 c2 00 00 00 00    mov    $0x0,%rdx
                        24c2: R_X86_64_32S      .rodata.str1.1+0x127b
    24c6:       48 c7 c1 00 00 00 00    mov    $0x0,%rcx
                        24c9: R_X86_64_32S      .rodata.str1.1+0x1679
    24cd:       41 b8 90 01 00 00       mov    $0x190,%r8d
    24d3:       49 c7 c1 00 00 00 00    mov    $0x0,%r9
                        24d6: R_X86_64_32S      .rodata.str1.1+0x127b
    24da:       b8 00 00 00 00          mov    $0x0,%eax
    24df:       55                      push   %rbp
    24e0:       41 52                   push   %r10
    24e2:       e8 00 00 00 00          callq  24e7 <leaf_item_bottle>
                        24e3: R_X86_64_PC32     __reiserfs_panic-0x4

Objtool is correct this time: There *is* a stack state mismatch at
0x24e0.  The stack size is different at 0x24e0, depending on whether it
came from 0x2346 or from 0x23b1.

In this case it's not a problem for code flow, because the basic block
is a dead end.

But it *is* a problem for unwinding.  The location of the previous stack
frame is nondeterministic.

And that's extra important for calls to noreturn functions, because they
often dump the stack before exiting.

So it looks like a compiler bug to me.

-- 
Josh