Message-ID: <20160413033649.7r3msnmo3trtq47z@treble>
Date:	Tue, 12 Apr 2016 22:36:49 -0500
From:	Josh Poimboeuf <jpoimboe@...hat.com>
To:	Denys Vlasenko <dvlasenk@...hat.com>
Cc:	Ingo Molnar <mingo@...nel.org>, Thomas Graf <tgraf@...g.ch>,
	Peter Zijlstra <peterz@...radead.org>,
	David Rientjes <rientjes@...gle.com>,
	Andrew Morton <akpm@...ux-foundation.org>,
	linux-kernel@...r.kernel.org, Arnd Bergmann <arnd@...db.de>
Subject: This patch triggers a bad gcc bug (was Re: [PATCH] force inlining of
 some byteswap operations)

On Thu, Feb 04, 2016 at 08:45:35PM +0100, Denys Vlasenko wrote:
> Sometimes gcc mysteriously doesn't inline
> very small functions we expect to be inlined. See
>     https://gcc.gnu.org/bugzilla/show_bug.cgi?id=66122
> 
> With this .config:
> http://busybox.net/~vda/kernel_config_OPTIMIZE_INLINING_and_Os,
> the following functions get deinlined many times.
> Examples of disassembly:
> 
> <get_unaligned_be16> (12 copies, 51 calls):
>        66 8b 07                mov    (%rdi),%ax
>        55                      push   %rbp
>        48 89 e5                mov    %rsp,%rbp
>        86 e0                   xchg   %ah,%al
>        5d                      pop    %rbp
>        c3                      retq
> 
> <get_unaligned_be32> (12 copies, 135 calls):
>        8b 07                   mov    (%rdi),%eax
>        55                      push   %rbp
>        48 89 e5                mov    %rsp,%rbp
>        0f c8                   bswap  %eax
>        5d                      pop    %rbp
>        c3                      retq
> 
> <get_unaligned_be64> (2 copies, 20 calls):
>        48 8b 07                mov    (%rdi),%rax
>        55                      push   %rbp
>        48 89 e5                mov    %rsp,%rbp
>        48 0f c8                bswap  %rax
>        5d                      pop    %rbp
>        c3                      retq
> 
> <__swab16p> (16 copies, 146 calls):
>        55                      push   %rbp
>        89 f8                   mov    %edi,%eax
>        86 e0                   xchg   %ah,%al
>        48 89 e5                mov    %rsp,%rbp
>        5d                      pop    %rbp
>        c3                      retq
> 
> <__swab32p> (43 copies, ~560 calls):
>        55                      push   %rbp
>        89 f8                   mov    %edi,%eax
>        0f c8                   bswap  %eax
>        48 89 e5                mov    %rsp,%rbp
>        5d                      pop    %rbp
>        c3                      retq
> 
> <__swab64p> (21 copies, 119 calls):
>        55                      push   %rbp
>        48 89 f8                mov    %rdi,%rax
>        48 0f c8                bswap  %rax
>        48 89 e5                mov    %rsp,%rbp
>        5d                      pop    %rbp
>        c3                      retq
> 
> <__swab32s> (6 copies, 47 calls):
>        8b 07                   mov    (%rdi),%eax
>        55                      push   %rbp
>        48 89 e5                mov    %rsp,%rbp
>        0f c8                   bswap  %eax
>        89 07                   mov    %eax,(%rdi)
>        5d                      pop    %rbp
>        c3                      retq
> 
> This patch fixes this via s/inline/__always_inline/.
> Code size decrease after the patch is ~4.5k:
> 
>     text     data      bss       dec     hex filename
> 92202377 20826112 36417536 149446025 8e85d89 vmlinux
> 92197848 20826112 36417536 149441496 8e84bd8 vmlinux5_swap_after
> 
> Signed-off-by: Denys Vlasenko <dvlasenk@...hat.com>
> Cc: Ingo Molnar <mingo@...nel.org>
> Cc: Thomas Graf <tgraf@...g.ch>
> Cc: Peter Zijlstra <peterz@...radead.org>
> Cc: David Rientjes <rientjes@...gle.com>
> Cc: Andrew Morton <akpm@...ux-foundation.org>
> Cc: linux-kernel@...r.kernel.org
> ---
>  include/uapi/linux/byteorder/big_endian.h    | 24 ++++++++++++------------
>  include/uapi/linux/byteorder/little_endian.h | 24 ++++++++++++------------
>  include/uapi/linux/swab.h                    | 10 +++++-----
>  3 files changed, 29 insertions(+), 29 deletions(-)

Hi,

This patch seems to trigger a gcc bug which can produce corrupt code.  I
discovered it when investigating an objtool warning reported by kbuild
bot:

  http://www.spinics.net/lists/linux-scsi/msg95481.html

>From the disassembly of drivers/scsi/qla2xxx/qla_attr.o:

0000000000002f53 <qla2x00_get_host_fabric_name>:
    2f53:       55                      push   %rbp
    2f54:       48 89 e5                mov    %rsp,%rbp

0000000000002f57 <qla2x00_get_fc_host_stats>:
    2f57:       55                      push   %rbp
    2f58:       b9 e8 00 00 00          mov    $0xe8,%ecx
    2f5d:       48 89 e5                mov    %rsp,%rbp
...

Note that qla2x00_get_host_fabric_name() is inexplicably truncated after
setting up the frame pointer.  It falls through to the next function, which is
very wrong.

I can recreate it with either gcc 5.3.1 or gcc 6.0 on linus/master with
the .config from the above link.

The call chain which appears to trigger the problem is:

qla2x00_get_host_fabric_name()
  wwn_to_u64()
    get_unaligned_be64()
      be64_to_cpup()
        __be64_to_cpup() <- changed to __always_inline by this patch

It occurs with the combination of the following two recent commits:

- bc27fb68aaad ("include/uapi/linux/byteorder, swab: force inlining of some byteswap operations")
- ef3fb2422ffe ("scsi: fc: use get/put_unaligned64 for wwn access")

I can confirm that reverting either patch makes the problem go away.
I'm planning on opening a gcc bug tomorrow.

> -static inline __u64 __be64_to_cpup(const __be64 *p)
> +static __always_inline __u64 __be64_to_cpup(const __be64 *p)
>  {
>  	return __swab64p((__u64 *)p);
>  }

-- 
Josh

Powered by blists - more mailing lists