svn commit: r262466 - head/sys/cddl/dev/systrace

[Justin Hibbits]

On Wed, Feb 26, 2014 at 9:52 AM, Justin Hibbits <jhibbits at freebsd.org> wrote:
> On Tue, Feb 25, 2014 at 5:35 PM, Mark Johnston <markj at freebsd.org> wrote:
>> On Tue, Feb 25, 2014 at 03:17:56PM -0800, Justin Hibbits wrote:
>>> I think this broke powerpc building.  I see the following build failure:
>>>
>>> cc1: warnings being treated as errors
>>> /home/chmeee/freebsd/head/sys/modules/dtrace/systrace/../../../cddl/dev/systrace/systrace.c:
>>> In function 'systrace_probe':
>>> /home/chmeee/freebsd/head/sys/modules/dtrace/systrace/../../../cddl/dev/systrace/systrace.c:218:
>>> warning: function called through a non-compatible type
>>> /home/chmeee/freebsd/head/sys/modules/dtrace/systrace/../../../cddl/dev/systrace/systrace.c:218:
>>> note: if this code is reached, the program will abort
>>>
>>
>> Hi Justin,
>>
>> Sorry about this. I've reverted the commit.
>>
>> I realize that the change introduced undefined behaviour, but a similar
>> trick is used elsewhere in the DTrace code to pass extra arguments at a
>> probe site. Calling dtrace_probe() through a function pointer (patch
>> below) makes the warning go away, but I don't really understand why.
>> clang doesn't emit warnings in either case.
>>
>> Thanks,
>> -Mark
>>
>> diff --git a/sys/cddl/dev/systrace/systrace.c b/sys/cddl/dev/systrace/systrace.c
>> index 83f0793..5f4b82f 100644
>> --- a/sys/cddl/dev/systrace/systrace.c
>> +++ b/sys/cddl/dev/systrace/systrace.c
>> @@ -168,8 +168,8 @@ static dtrace_pops_t systrace_pops = {
>>  static struct cdev             *systrace_cdev;
>>  static dtrace_provider_id_t    systrace_id;
>>
>> -typedef void (*systrace_dtrace_probe)(dtrace_id_t, uintptr_t, uintptr_t,
>> -    uintptr_t, uintptr_t, uintptr_t, uintptr_t, uintptr_t, uintptr_t);
>> +typedef void (*systrace_probe_t)(dtrace_id_t, uintptr_t, uintptr_t, uintptr_t,
>> +    uintptr_t, uintptr_t, uintptr_t, uintptr_t, uintptr_t);
>>
>>  #if !defined(LINUX_SYSTRACE)
>>  /*
>> @@ -214,8 +214,9 @@ systrace_probe(u_int32_t id, int sysnum, struct sysent *sysent, void *params,
>>         }
>>
>>         /* Process the probe using the converted argments. */
>> -       ((systrace_dtrace_probe)(dtrace_probe))(id, uargs[0], uargs[1],
>> -           uargs[2], uargs[3], uargs[4], uargs[5], uargs[6], uargs[7]);
>> +       systrace_probe_t probe = (systrace_probe_t)dtrace_probe;
>> +       probe(id, uargs[0], uargs[1], uargs[2], uargs[3], uargs[4],
>> +           uargs[5], uargs[6], uargs[7]);
>>  }
>>
>>  #endif
>
>
> Hi Mark,
>
> I think this patch works because it circumvents gcc's variable
> tracking.  With the first patch, gcc knew the function that was being
> called, and knew it was undefined behavior.  With the second patch, it
> only knows at that time that you're calling through a function
> pointer.  It's completely forgotten that the function pointer is
> pointing to that function.  Just a guess.  I'll give it a go and let
> you know if it still complains.
>
> Thanks!
>
> - Justin

Just tested and confirmed it builds successfully.

- Justin