2010-08-18 07:45:54

by Lin Ming

Subject: libpfm4 faults on P4 machine

On P4 machine, intel_x86_arch_support.get_event_info(...) faults because
intel_x86_arch_support.pe is NULL.

intel_x86_arch_support.pe is set at run time, but P4 does not have arch
event support, so intel_x86_arch_support.pe is always NULL.

#gdb libpfm4/examples/check_events

Program received signal SIGSEGV, Segmentation fault.
[Switching to Thread 0xb785caf0 (LWP 3950)]
0x0804d74b in pfm_intel_x86_get_event_info (this=0x81c5f80, idx=0, info=0xbff45f14)
at pfmlib_intel_x86.c:667
667 info->desc = pe[idx].desc;
(gdb) bt
#0 0x0804d74b in pfm_intel_x86_get_event_info (this=0x81c5f80, idx=0, info=0xbff45f14)
at pfmlib_intel_x86.c:667
#1 0x08049cd3 in pfmlib_parse_event (event=0x804fb7b "PERF_COUNT_HW_CPU_CYCLES", d=0xbff45f7c)
at pfmlib_common.c:676
#2 0x08049f3f in pfm_get_event_encoding (str=0x804fb7b "PERF_COUNT_HW_CPU_CYCLES", dfl_plm=8,
fstr=0xbff464d4, idx=0xbff464c8, codes=0xbff464d0, count=0xbff464cc) at pfmlib_common.c:877
#3 0x08048dc1 in main (argc=<value optimized out>, argv=0xbff46584) at check_events.c:118
(gdb) list
662
663 /*
664 * pmu and idx filled out by caller
665 */
666 info->name = pe[idx].name;
667 info->desc = pe[idx].desc;
668 info->code = pe[idx].code;
669 info->equiv = pe[idx].equiv;
670
671 /* unit masks + modifiers */
(gdb)

Any fix?

Lin Ming

2010-08-18 09:21:17

by Stephane Eranian

Subject: Re: libpfm4 faults on P4 machine

Lin,

It should not have chosen intel_x86_arch to begin with.
That's where the bug is. Let me fix this.


On Wed, Aug 18, 2010 at 9:46 AM, Lin Ming <[email protected]> wrote:
> On P4 machine, intel_x86_arch_support.get_event_info(...) faults because
> intel_x86_arch_support.pe is NULL.
>
> intel_x86_arch_support.pe is set at run time, but P4 does not have arch
> event support, so intel_x86_arch_support.pe is always NULL.
>
> #gdb libpfm4/examples/check_events
>
> Program received signal SIGSEGV, Segmentation fault.
> [Switching to Thread 0xb785caf0 (LWP 3950)]
> 0x0804d74b in pfm_intel_x86_get_event_info (this=0x81c5f80, idx=0, info=0xbff45f14)
>    at pfmlib_intel_x86.c:667
> 667             info->desc  = pe[idx].desc;
> (gdb) bt
> #0  0x0804d74b in pfm_intel_x86_get_event_info (this=0x81c5f80, idx=0, info=0xbff45f14)
>    at pfmlib_intel_x86.c:667
> #1  0x08049cd3 in pfmlib_parse_event (event=0x804fb7b "PERF_COUNT_HW_CPU_CYCLES", d=0xbff45f7c)
>    at pfmlib_common.c:676
> #2  0x08049f3f in pfm_get_event_encoding (str=0x804fb7b "PERF_COUNT_HW_CPU_CYCLES", dfl_plm=8,
>    fstr=0xbff464d4, idx=0xbff464c8, codes=0xbff464d0, count=0xbff464cc) at pfmlib_common.c:877
> #3  0x08048dc1 in main (argc=<value optimized out>, argv=0xbff46584) at check_events.c:118
> (gdb) list
> 662
> 663             /*
> 664              * pmu and idx filled out by caller
> 665              */
> 666             info->name  = pe[idx].name;
> 667             info->desc  = pe[idx].desc;
> 668             info->code  = pe[idx].code;
> 669             info->equiv = pe[idx].equiv;
> 670
> 671             /* unit masks + modifiers */
> (gdb)
>
> Any fix?
>
> Lin Ming
>
>

2010-08-18 09:42:21

by Stephane Eranian

Subject: Re: libpfm4 faults on P4 machine

On Wed, Aug 18, 2010 at 11:21 AM, stephane eranian
<[email protected]> wrote:
> Lin,
>
> It should not have chosen  intel_x86_arch to begin with.
> That's where the bug is. Let me fix this.
>
>
Fixed in the git tree, just pull again.
Thanks for reporting this issue.

> On Wed, Aug 18, 2010 at 9:46 AM, Lin Ming <[email protected]> wrote:
>> On P4 machine, intel_x86_arch_support.get_event_info(...) faults because
>> intel_x86_arch_support.pe is NULL.
>>
>> intel_x86_arch_support.pe is set at run time, but P4 does not have arch
>> event support, so intel_x86_arch_support.pe is always NULL.
>>
>> #gdb libpfm4/examples/check_events
>>
>> Program received signal SIGSEGV, Segmentation fault.
>> [Switching to Thread 0xb785caf0 (LWP 3950)]
>> 0x0804d74b in pfm_intel_x86_get_event_info (this=0x81c5f80, idx=0, info=0xbff45f14)
>>    at pfmlib_intel_x86.c:667
>> 667             info->desc  = pe[idx].desc;
>> (gdb) bt
>> #0  0x0804d74b in pfm_intel_x86_get_event_info (this=0x81c5f80, idx=0, info=0xbff45f14)
>>    at pfmlib_intel_x86.c:667
>> #1  0x08049cd3 in pfmlib_parse_event (event=0x804fb7b "PERF_COUNT_HW_CPU_CYCLES", d=0xbff45f7c)
>>    at pfmlib_common.c:676
>> #2  0x08049f3f in pfm_get_event_encoding (str=0x804fb7b "PERF_COUNT_HW_CPU_CYCLES", dfl_plm=8,
>>    fstr=0xbff464d4, idx=0xbff464c8, codes=0xbff464d0, count=0xbff464cc) at pfmlib_common.c:877
>> #3  0x08048dc1 in main (argc=<value optimized out>, argv=0xbff46584) at check_events.c:118
>> (gdb) list
>> 662
>> 663             /*
>> 664              * pmu and idx filled out by caller
>> 665              */
>> 666             info->name  = pe[idx].name;
>> 667             info->desc  = pe[idx].desc;
>> 668             info->code  = pe[idx].code;
>> 669             info->equiv = pe[idx].equiv;
>> 670
>> 671             /* unit masks + modifiers */
>> (gdb)
>>
>> Any fix?
>>
>> Lin Ming
>>
>>
>