G'Day Masami,
On Tue, Jun 30, 2015 at 12:37 AM, Masami Hiramatsu
<masami.hiramatsu...@hitachi.com> wrote:
> On 2015/06/30 0:00, Arnaldo Carvalho de Melo wrote:
>> Em Sun, Jun 28, 2015 at 05:46:51PM +0200, Benjamin King escreveu:
>>> Hi Brendan
>>>
>>>> Is there a trick to getting perf to probe a user-level address without
>>>> debuginfo? Eg (on Linux 4.0):
>>>> [...]
>>>> I can do this using ftrace ok, eg, "p:tick_0x583 /root/tick:0x583"
>>>> works. Thanks,
>>>
>>> Not quite what you have asked for, but you can add the probe via ftrace and
>>> then use it from perf. Probes from /sys/kernel/debug/tracing/uprobe_events
>>> will show up in 'perf list' as well.
>>
>> Masami,
>>
>> Is this already possible?
>
> Yes, it should be possible. However, I don't recommend you to
> probe the address inside a function without debuginfo. You must
> check the instruction boundary in that case.
Ah, because otherwise I could accidentally trace at a non-zero offsite
inside an instruction? That probably explains the only failure I saw
on uprobes on linux 4.0.
Here's the original use case:
# readelf -n /root/node/node
[...]
stapsdt 0x00000082 NT_STAPSDT (SystemTap probe descriptors)
Provider: node
Name: http__client__request
Location: 0x0000000000bf48ff, Base: 0x0000000000f22464, Semaphore:
0x0000000001243024
Arguments: 8@%rax 8@%rdx 8@-136(%rbp) -4@-140(%rbp) 8@-72(%rbp)
8@-80(%rbp) -4@-144(%rbp)
[...]
These user SDT probes (or "USDT probes") can't currently be traced
using perf (although Hemant Kumar was working on a patch).
I can trace them via ftrace, and have already written a helper script.
But that's another story. I'm interested in doing this from perf as
well! (Without needing to create the probes via ftrace.)
I have their instruction offset above from readelf (and need to
subtract the base load address). Here's what doesn't work:
A) Tracing via the absolute address. Eg, "perf probe -x node '@+0x7f48ff'
B) Tracing via the offset from main. Eg, "perf probe -x node
'main+0x...", since perf complains I'm tracing at an offset bigger
than main.
C) Tracing via the offset from the function that contains the probe.
In this case, because the function name is too big! (Due to C++) Eg:
# ./perf probe -x /root/node/node --filter="*"
'_ZN4node26DTRACE_HTTP_CLIENT_REQUESTERKN2v820FunctionCallbackInfoINS0_5ValueEEE+0x42f'
Added new event:
Error: Failed to add events.
I've attached a trivial patch for (C), it just increases the buffer
size in __add_probe_trace_events() from 64 to 128. It's still a bit of
a nuisance to resolve the containing function (addr2line) when I have
the absolute address...
Brendan
# diff -u linux-4.0.0+/tools/perf/util/probe-event.c
linux-4.0.0+brendan/tools/perf/util/probe-event.c
--- linux-4.0.0+/tools/perf/util/probe-event.c 2015-03-31 18:13:30.001009754
+0000
+++ linux-4.0.0+brendan/tools/perf/util/probe-event.c 2015-06-30
18:20:57.914308026 +0000
@@ -2383,7 +2383,7 @@
{
int i, fd, ret;
struct probe_trace_event *tev = NULL;
- char buf[64];
+ char buf[128];
const char *event, *group;
struct strlist *namelist;
LIST_HEAD(blacklist);
@@ -2437,7 +2437,7 @@
group = PERFPROBE_GROUP;
/* Get an unused new event name */
- ret = get_new_event_name(buf, 64, event,
+ ret = get_new_event_name(buf, 128, event,
namelist, allow_suffix);
if (ret < 0)
break;
