Re: [PATCH 1/9] perf, tools: Support handling complete branch stacks as histograms v6
From: Namhyung Kim
Date: Sun May 25 2014 - 22:45:36 EST
Hi Andi,
On Fri, 23 May 2014 14:35:03 -0700, Andi Kleen wrote:
> On Mon, May 19, 2014 at 05:21:15PM +0900, Namhyung Kim wrote:
>> This is gone with 540476de74c9 ("perf tools: Remove
>> symbol_conf.use_callchain check").
>
> The patchkit applies to tip/perf/core.
The commit 540476de74c9 is also in the tip/perf/core. Please check
machine_resolve_callchain_sample().
>
>> > + * Check for overlap into the callchain.
>> > + * The return address is one off compared to
>> > + * the branch entry. To adjust for this
>> > + * assume the calling instruction is not longer
>> > + * than 8 bytes.
>> > + */
>> > + if (be[i].from < chain->ips[first_call] &&
>> > + be[i].from >= chain->ips[first_call] - 8)
>> > + first_call++;
>>
>> It seems that you need to check chain->ips[first_call] is greater than
>> PERF_CONTEXT_MAX and use such value as the cpumode...
>
> I don't understand the comment. The only IP that gets resolved is the from/to.
> And add_callchain_ip does it own resolution.
>
> Wouldn't make any sense to get it from first_call
Okay, let me explain it this way..
You're checking the branch stack with normal callchain to find overlap
by comparing the 'from' address and the address in chain->ips[]. But
chain->ips[0] doesn't contain a valid address but a PERF_CONTEXT_XXX for
cpumode of subsequent callchains. So the first_call of 0 won't do
anything meaningful for you and it'd still contain overlapped
callchains.
$ perf --version
perf version 3.15.rc4.g816bf8
$ perf record -b -g ./tcall
$ perf report -D | grep -A35 SAMPLE
4748858059190923 0x3608 [0x240]: PERF_RECORD_SAMPLE(IP, 0x1): 31914/31914: 0xffffffff81043ffa period: 1 addr: 0
... chain: nr:17
..... 0: ffffffffffffff80
..... 1: ffffffff81043ffa
..... 2: ffffffff81029d40
..... 3: ffffffff81025554
..... 4: ffffffff811246c7
..... 5: ffffffff81125d69
..... 6: ffffffff811280dc
..... 7: ffffffff811a4266
..... 8: ffffffff811a4bb1
..... 9: ffffffff811f1a4f
..... 10: ffffffff811a344c
..... 11: ffffffff811a49bb
..... 12: ffffffff811a4ac8
..... 13: ffffffff811a4d3d
..... 14: ffffffff81664689
..... 15: fffffffffffffe00
..... 16: 0000003153ebca47
... branch stack: nr:16
..... 0: ffffffff81029d3b -> ffffffff81043ff0
..... 1: ffffffff810280c9 -> ffffffff81029d18
..... 2: ffffffff81043ffd -> ffffffff810280be
..... 3: 0000000000000000 -> 0000000000000000
..... 4: 0000000000000000 -> 0000000000000000
..... 5: 0000000000000000 -> 0000000000000000
..... 6: 0000000000000000 -> 0000000000000000
..... 7: 0000000000000000 -> 0000000000000000
..... 8: 0000000000000000 -> 0000000000000000
..... 9: 0000000000000000 -> 0000000000000000
..... 10: 0000000000000000 -> 0000000000000000
..... 11: 0000000000000000 -> 0000000000000000
..... 12: 0000000000000000 -> 0000000000000000
..... 13: 0000000000000000 -> 0000000000000000
..... 14: 0000000000000000 -> 0000000000000000
..... 15: 0000000000000000 -> 0000000000000000
As you can see, chain->ips[0] is ffffffffffffff80 (= -128) that is
defined as PERF_CONTEXT_KERNEL. And in this case nr of branch stack is
16 but it's actually 3. I guess you need to ignore 0 entries..
Also perf report seems to fail to resolve symbols/srclines in branch
stack (possibly due to missing cpumode) and find loops.
$ perf report --branch-history --stdio
...
0.00% native_writ [k] native_write_msr_safe [kernel.kallsyms]
|
---0xffffffff81043ff0
0xffffffff81029d3b
0xffffffff81029d18
0xffffffff810280c9
0xffffffff810280be
0xffffffff81043ffd
|
|--99.77%-- 0xffffffff81043ff0
| 0xffffffff81029d3b
| 0xffffffff81029d18
| 0xffffffff810280c9
| 0xffffffff810280be
| 0xffffffff81043ffd
| 0
| 0
| |
| |--91.43%-- native_write_msr_safe +10
| | intel_pmu_enable_all +80
| | x86_pmu_enable +628
| | perf_pmu_enable +39
| | perf_event_context_sched_in +121
| | perf_event_comm +364
| | set_task_comm +102
| | setup_new_exec +129
| | load_elf_binary +1007
| |
| --8.57%-- 0
| 0
| |
| |--91.78%-- native_write_msr_safe +10
| | intel_pmu_enable_all +80
| | x86_pmu_enable +628
| | perf_pmu_enable +39
| | perf_event_context_sched_in +121
| | perf_event_comm +364
| | set_task_comm +102
| | setup_new_exec +129
| |
| --8.22%-- 0
| 0
| native_write_msr_safe +10
| intel_pmu_enable_all +80
| x86_pmu_enable +628
| perf_pmu_enable +39
| perf_event_context_sched_in +121
| perf_event_comm +364
| set_task_comm +102
--0.23%-- [...]
>
>>
>>
>> > + } else
>> > + be[i] = branch->entries[branch->nr - i - 1];
>> > + }
>> > +
>> > + nr = remove_loops(be, nr);
>> > +
>> > + for (i = 0; i < nr; i++) {
>> > + err = add_callchain_ip(machine, thread, parent,
>> > + root_al,
>> > + -1, be[i].to);
>> > + if (!err)
>> > + err = add_callchain_ip(machine, thread,
>> > + parent, root_al,
>> > + -1, be[i].from);
>>
>> ... for here.
>>
>>
>> > + if (err == -EINVAL)
>> > + break;
>> > + if (err)
>> > + return err;
>> > + }
>> > + chain_nr -= nr;
>>
>> It seems it could make some callchain nodes being ignored. What if a
>> case like small callchains with matches to only 2 nodes in the LBR?
>>
>> nr = 16, chain_nr = 10 and first_call = 2
>
> The chain_nr variable is just to handle it when the user
> specified a max_stack value. nr is always capped to max_stack too.
> If lbr size is >= max_stack it will end up being 0 or negative and the
> following loop to add normal call stack entries will do nothing.
>
> I think that's the correct behavior.
Hmm.. I assumed user didn't specify a max_stack but failed to understand
how the code prevents from ignoring remaining 8 callchain nodes.
Thanks,
Namhyung
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/