| lists.openwall.net | lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC | |
|
Open Source and information security mailing list archives
| ||
|
Date: Wed, 22 Mar 2017 21:00:23 +0900
From: Taeung Song <treeze.taeung@...il.com>
To: Arnaldo Carvalho de Melo <arnaldo.melo@...il.com>
Cc: linux-kernel@...r.kernel.org, Jiri Olsa <jolsa@...nel.org>,
Namhyung Kim <namhyung@...nel.org>,
Ingo Molnar <mingo@...nel.org>,
Peter Zijlstra <peterz@...radead.org>,
Wang Nan <wangnan0@...wei.com>,
Masami Hiramatsu <mhiramat@...nel.org>,
Jiri Olsa <jolsa@...hat.com>
Subject: Re: [PATCH 2/4] perf annotate: Avoid division by zero when
calculating percent
On 03/21/2017 11:21 PM, Arnaldo Carvalho de Melo wrote:
> Em Tue, Mar 21, 2017 at 11:14:07AM -0300, Arnaldo Carvalho de Melo escreveu:
>> Em Tue, Mar 21, 2017 at 07:20:20AM +0900, Taeung Song escreveu:
>>> And,
>>> I tested by perf-stat on the same situation as below.
>>>
>>> $ perf stat -e "{cycles,page-faults,branch-misses}" ./old <input.txt
>>> 6623856
>>
>> Please always try to spell out all the steps needed to get to some
>> result, for instance, in this case the info above, that you are asking
>> for three counters to be recorded at once probably has the key to
>> reproduce this, as I think that you may run your workload and sometimes
>> not get one page fault, leading tho that division by zero, but I have to
>> try to reproduce it now that I have this clue.
>
> Bingo!
>
> [root@...et ~]# perf record -e "{cycles,page-faults,branch-misses}" usleep 1
> [ perf record: Woken up 1 times to write data ]
> [ perf record: Captured and wrote 0.020 MB perf.data (20 samples) ]
> [root@...et ~]# perf annotate -l --stdio 2>&1 | grep -w nan | head
> 100.00 -nan 100.00 msr.h:104
> 0.00 -nan 0.00 : ffffffff81060bf0: mov %edi,%ecx
> 0.00 -nan 0.00 : ffffffff81060bf2: mov %esi,%eax
> 0.00 -nan 0.00 : ffffffff81060bf4: wrmsr
> msr.h:104 100.00 -nan 100.00 : ffffffff81060bf6: nopl 0x0(%rax,%rax,1)
> 0.00 -nan 0.00 : ffffffff81060bfb: retq
> 0.00 -nan 0.00 : ffffffff81060bfc: push %rbp
> 0.00 -nan 0.00 : ffffffff81060bfd: mov %esi,%eax
> 0.00 -nan 0.00 : ffffffff81060bff: mov %rdx,%rsi
> 0.00 -nan 0.00 : ffffffff81060c02: shl $0x20,%rsi
> [root@...et ~]#
>
> So, if we don't use -l/--print-line we get:
>
> [root@...et ~]# perf annotate --stdio --no-source native_write_msr
> Percent | Source code & Disassembly of vmlinux for cycles (3 samples)
> -------------------------------------------------------------------------------------------
> : ffffffff81060bf0 <native_write_msr>:
> : __wrmsr():
> 0.00 0.00 0.00 : ffffffff81060bf0: mov %edi,%ecx
> 0.00 0.00 0.00 : ffffffff81060bf2: mov %esi,%eax
> 0.00 0.00 0.00 : ffffffff81060bf4: wrmsr
> : arch_static_branch():
> 100.00 0.00 100.00 : ffffffff81060bf6: nopl 0x0(%rax,%rax,1)
> : native_write_msr():
> 0.00 0.00 0.00 : ffffffff81060bfb: retq
> 0.00 0.00 0.00 : ffffffff81060bfc: push %rbp
> 0.00 0.00 0.00 : ffffffff81060bfd: mov %esi,%eax
> 0.00 0.00 0.00 : ffffffff81060bff: mov %rdx,%rsi
> 0.00 0.00 0.00 : ffffffff81060c02: shl $0x20,%rsi
> 0.00 0.00 0.00 : ffffffff81060c06: mov %eax,%edx
> 0.00 0.00 0.00 : ffffffff81060c08: mov %rsp,%rbp
> 0.00 0.00 0.00 : ffffffff81060c0b: or %rdx,%rsi
> 0.00 0.00 0.00 : ffffffff81060c0e: xor %edx,%edx
> 0.00 0.00 0.00 : ffffffff81060c10: callq ffffffff81457df0 <do_trace_write_msr>
> 0.00 0.00 0.00 : ffffffff81060c15: pop %rbp
> 0.00 0.00 0.00 : ffffffff81060c16: retq
> [root@...et ~]#
>
> I.e. one column per event in that group, and no -nan, all is well.
The reason is the below if statement at util/annotate.c:910.
881 double disasm__calc_percent(struct annotation *notes, int evidx, s64
offset,
...
904 struct sym_hist *h = annotation__histogram(notes,
evidx);
905 unsigned int hits = 0;
906
907 while (offset < end)
908 hits += h->addr[offset++];
909
910 if (h->sum) {
911 *nr_samples = hits;
912 percent = 100.0 * hits / h->sum;
913 }
914 }
The if statement 'if (h->sum)' prevent division by zero, so all is well.
So I sent this patch because of same reason
adding the if statement 'if (h->sum)' in symbol__get_source_line().
But will resend v2 with more clearly commit log message!
I think previous log message is a bit insufficient..
Thanks,
Taeung
>
> But if we ask for --lines it gets b0rked:
>
> [root@...et ~]# perf annotate --stdio --no-source native_write_msr --print-line
>
> Sorted summary for file /lib/modules/4.11.0-rc2+/build/vmlinux
> ----------------------------------------------
>
> 100.00 -nan 100.00 msr.h:104
> Percent | Source code & Disassembly of vmlinux for cycles (3 samples)
> -------------------------------------------------------------------------------------------
> : ffffffff81060bf0 <native_write_msr>:
> : __wrmsr():
> 0.00 -nan 0.00 : ffffffff81060bf0: mov %edi,%ecx
> 0.00 -nan 0.00 : ffffffff81060bf2: mov %esi,%eax
> 0.00 -nan 0.00 : ffffffff81060bf4: wrmsr
> : arch_static_branch():
> msr.h:104 100.00 -nan 100.00 : ffffffff81060bf6: nopl 0x0(%rax,%rax,1)
> : native_write_msr():
> 0.00 -nan 0.00 : ffffffff81060bfb: retq
> 0.00 -nan 0.00 : ffffffff81060bfc: push %rbp
> 0.00 -nan 0.00 : ffffffff81060bfd: mov %esi,%eax
> 0.00 -nan 0.00 : ffffffff81060bff: mov %rdx,%rsi
> 0.00 -nan 0.00 : ffffffff81060c02: shl $0x20,%rsi
> 0.00 -nan 0.00 : ffffffff81060c06: mov %eax,%edx
> 0.00 -nan 0.00 : ffffffff81060c08: mov %rsp,%rbp
> 0.00 -nan 0.00 : ffffffff81060c0b: or %rdx,%rsi
> 0.00 -nan 0.00 : ffffffff81060c0e: xor %edx,%edx
> 0.00 -nan 0.00 : ffffffff81060c10: callq ffffffff81457df0 <do_trace_write_msr>
> 0.00 -nan 0.00 : ffffffff81060c15: pop %rbp
> 0.00 -nan 0.00 : ffffffff81060c16: retq
> [root@...et ~]#
>
> I don't have that much time to continue pursuing this, try to continue
> investigating, I'll try to help and to fix this later.
>
> - Arnaldo
>
Powered by blists - more mailing lists