Re: This patch triggers a bad gcc bug (was Re: [PATCH] force inlining of some byteswap operations)

From: Denys Vlasenko
Date: Thu Apr 14 2016 - 13:09:20 EST


On 04/14/2016 05:57 PM, Josh Poimboeuf wrote:
> On Thu, Apr 14, 2016 at 05:29:06PM +0200, Denys Vlasenko wrote:
>> On 04/13/2016 07:10 PM, Josh Poimboeuf wrote:
>>>>>>>> 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.
>>>>>>>
>>>>>>> Wow, that's ... interesting.
>>>>>>>
>>>>>>>
>>>>>>>> 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.
>>>>>>>
>>>>>>>
>>>>>>> Note that if CONFIG_OPTIMIZE_INLINING is not set, _all_ "inline"
>>>>>>> keywords are in fact __always_inline, so the bug must be
>>>>>>> triggering
>>>>>>> even without the patch.
>>>>>>
>>>>>> Makes sense in theory, but the bug doesn't actually trigger when I
>>>>>> revert the patch and set CONFIG_OPTIMIZE_INLINING=n.
>>>>>>
>>>>>> Perhaps even more surprising, it doesn't trigger *with* the patch
>>>>>> and
>>>>>> CONFIG_OPTIMIZE_INLINING=n.
>>>>>
>>>>> [ Adding James to CC since this bug affects scsi. ]
>>>>>
>>>>> Here's the gcc bug:
>>>>>
>>>>> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70646
>>>>>
>>>>
>>>>
>>>> Actually, adding me doesn't help, I've added linux-scsi. The summary
>>>> is that there's a but in gcc-5.3.1 which is miscompiling qla_attr.c ...
>>>> this means we're going to have to ask the compiler version of reported
>>>> crashes.
>>>
>>> The bug isn't specific to a compiler version. I've seen it with gcc
>>> 5.3.1 and gcc 6.0. I haven't tried any older versions. And the gcc bug
>>> hasn't been resolved (or even investigated) yet.
>>>
>>> The bug is triggered by a combination of the above two commits from the
>>> 4.6 merge window, so presumably we'd need to revert one of them to avoid
>>> crashes in 4.6.
>>
>> The bug is indeed in the compiler. 4.9 and all later versions are affected.
>> gcc bugzilla now has a reproducer. In abridged form:
>>
>>
>> static inline __attribute__((always_inline)) u64 __swab64p(const u64 *p)
>> {
>> return (__builtin_constant_p((u64)(*p)) ? ((u64)( (((u64)(*p) & (u64)0x00000000000000ffULL) << 56) | (((u64)(*p) & (u64)0x000000000000ff00ULL) << 40) | (((u64)(*p) & (u64)0x0000000000ff0000ULL) << 24) | (((u64)(*p) & (u64)0x00000000ff000000ULL) << 8) | (((u64)(*p) & (u64)0x000000ff00000000ULL) >> 8) | (((u64)(*p) & (u64)0x0000ff0000000000ULL) >> 24) | (((u64)(*p) & (u64)0x00ff000000000000ULL) >> 40) | (((u64)(*p) & (u64)0xff00000000000000ULL) >> 56))) : __builtin_bswap64(*p));
>> }
>> static inline u64 wwn_to_u64(void *wwn)
>> {
>> return __swab64p(wwn);
>> }
>> static void qla2x00_get_host_fabric_name(struct Scsi_Host *shost)
>> {
>> scsi_qla_host_t *vha = shost_priv(shost);
>> u8 node_name[8] = { 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF, 0xFF};
>> u64 fabric_name = wwn_to_u64(node_name);
>> if (vha->device_flags & 0x1)
>> fabric_name = wwn_to_u64(vha->fabric_node_name);
>> (((struct fc_host_attrs *)(shost)->shost_data)->fabric_name) = fabric_name;
>> }
>
> Nice work with the reproducer!
>
>> Two (or more, there were more before simplification) levels of inlining
>> are necessary for bug to trigger in this example (folding to one level
>> makes it go away). "__attribute__((always_inline))" is necessary too.
>>
>>
>> Since we have lots of __always_inline anyway, this bug has a potential
>> to miscompile kernels regardless of CONFIG_OPTIMIZE_INLINING setting,
>> and with or without the patches mentioned above (they just happen
>> to create a reliable reproducer).
>
> Well, but setting !CONFIG_OPTIMIZE_INLINING makes the problem go away
> for some reason. It seems like the bug requires wwn_to_u64() being
> out-of-line and __swab64p() being in-line.

By my reading of what gcc gurus are talking there,
gcc has some new-ish machinery for discarding unreachable code.
Akin to not continuing code generation after a call to never-returning
function like exit(), but smarter (it can detect much less obvious
cases when some code path is not possible).

And it has a subtle bug. In this case, it decided that both branches
of ternary op ?: in __swab64p() are impossible, and therefore
__swab64p() is impossible.
(Which is, of course, bogus, that's why it's a bug).

Then this bogus decision was propagated through inlining
and gcc decided that entire qla2x00_get_host_fabric_name()
function is an impossible (unreachable) code path, and...
eliminated it all. Good boy :D

> In fact, the following patch seems to fix it:
>
> diff --git a/include/scsi/scsi_transport_fc.h b/include/scsi/scsi_transport_fc.h
> index bf66ea6..56b9e81 100644
> --- a/include/scsi/scsi_transport_fc.h
> +++ b/include/scsi/scsi_transport_fc.h
> @@ -796,7 +796,7 @@ fc_remote_port_chkready(struct fc_rport *rport)
> return result;
> }
>
> -static inline u64 wwn_to_u64(u8 *wwn)
> +static __always_inline u64 wwn_to_u64(u8 *wwn)
> {
> return get_unaligned_be64(wwn);
> }

It is not a guarantee.