Re: [PATCH -tip 3/3] perf-probe: Use the actual address as a hintfor uprobes

From: Masami Hiramatsu
Date: Sun Dec 22 2013 - 16:54:51 EST


(2013/12/21 3:03), Arnaldo Carvalho de Melo wrote:
> Em Fri, Dec 20, 2013 at 10:03:02AM +0000, Masami Hiramatsu escreveu:
>> Use the actual address of tracepoint as a hint to find
>> different local symbols. Since sometimes there are local
>> symbols which have same name, it is impossible to determine
>> which symbol should be used. This saves the actual address
>> from debuginfo and use it as a hint later.
>>
>> The reason why we don't use the address directly is that
>> the address stored in the debuginfo has some section offset.
>> Thus this just uses the last 12 bits of the address as a
>> hint when searching the symbol in the map.
>
> Again, can you provide an example of cases where before this patch the
> problem happens, with actual command output?

Ok, here is what the problem I've met,

Without this patch;
# perf probe -vfn -x perf scnprintf 2>&1
[...]
Writing event: p:probe_perf/scnprintf /kbuild/ksrc/linux-2.6/tools/perf/perf:0x8fd10
Writing event: p:probe_perf/scnprintf_1 /kbuild/ksrc/linux-2.6/tools/perf/perf:0xc02d0
Writing event: p:probe_perf/scnprintf_2 /kbuild/ksrc/linux-2.6/tools/perf/perf:0xc02d0
Writing event: p:probe_perf/scnprintf_3 /kbuild/ksrc/linux-2.6/tools/perf/perf:0x9ea00
Writing event: p:probe_perf/scnprintf_4 /kbuild/ksrc/linux-2.6/tools/perf/perf:0x25f36
Writing event: p:probe_perf/scnprintf_5 /kbuild/ksrc/linux-2.6/tools/perf/perf:0x29f90
Writing event: p:probe_perf/scnprintf_6 /kbuild/ksrc/linux-2.6/tools/perf/perf:0x29f90
Writing event: p:probe_perf/scnprintf_7 /kbuild/ksrc/linux-2.6/tools/perf/perf:0xc02d0
Writing event: p:probe_perf/scnprintf_8 /kbuild/ksrc/linux-2.6/tools/perf/perf:0xc02d0
Writing event: p:probe_perf/scnprintf_9 /kbuild/ksrc/linux-2.6/tools/perf/perf:0xc02d0

You can see there are many 0xc02d0, which is the last symbol of scnprintf in perf.
# nm perf | grep scnprintf
[...]
00000000004a3a30 t scnprintf
000000000041dad0 t scnprintf
0000000000425480 t scnprintf
0000000000442b20 t scnprintf
0000000000448b20 t scnprintf
00000000004525e0 t scnprintf
00000000004555f0 t scnprintf
00000000004678c0 t scnprintf
00000000004c02d0 t scnprintf

With this patch,
# perf probe -vfn -x perf scnprintf 2>&1
[...]
Writing event: p:probe_perf/scnprintf /kbuild/ksrc/linux-2.6/tools/perf/perf:0x1a820
Writing event: p:probe_perf/scnprintf_1 /kbuild/ksrc/linux-2.6/tools/perf/perf:0x1dad0
Writing event: p:probe_perf/scnprintf_2 /kbuild/ksrc/linux-2.6/tools/perf/perf:0x25480
Writing event: p:probe_perf/scnprintf_3 /kbuild/ksrc/linux-2.6/tools/perf/perf:0x25560
Writing event: p:probe_perf/scnprintf_4 /kbuild/ksrc/linux-2.6/tools/perf/perf:0x25f36
Writing event: p:probe_perf/scnprintf_5 /kbuild/ksrc/linux-2.6/tools/perf/perf:0x29f90
Writing event: p:probe_perf/scnprintf_6 /kbuild/ksrc/linux-2.6/tools/perf/perf:0x3e050
Writing event: p:probe_perf/scnprintf_7 /kbuild/ksrc/linux-2.6/tools/perf/perf:0x48b20
Writing event: p:probe_perf/scnprintf_8 /kbuild/ksrc/linux-2.6/tools/perf/perf:0x48b20
Writing event: p:probe_perf/scnprintf_9 /kbuild/ksrc/linux-2.6/tools/perf/perf:0x525e0

Oops, I thought 12bits are enough, but it seems not...

BTW, I'm not sure why debuginfo and nm shows symbol address + 0x400000,
and why the perf's map/symbol can remove this offset. Could you tell me
how it works?
If I can get the offset (0x400000) from binary, I don't need this kind
of ugly hacks...

Thank you,

>
> Then output from the same command with the patch applied, showing how it
> works after the fix.
>
> This way users can more quickly use your work and, among others, I can
> help you by validating it in my test machines.
>
> Thanks,
>
> - Arnaldo
>
>> Signed-off-by: Masami Hiramatsu <masami.hiramatsu.pt@xxxxxxxxxxx>
>> ---
>> tools/perf/util/probe-event.c | 16 +++++++++++++---
>> tools/perf/util/probe-event.h | 1 +
>> tools/perf/util/probe-finder.c | 1 +
>> 3 files changed, 15 insertions(+), 3 deletions(-)
>>
>> diff --git a/tools/perf/util/probe-event.c b/tools/perf/util/probe-event.c
>> index e27cecb..e5fbda3 100644
>> --- a/tools/perf/util/probe-event.c
>> +++ b/tools/perf/util/probe-event.c
>> @@ -187,11 +187,18 @@ static int init_user_exec(void)
>> }
>>
>> static const char *__target_symbol;
>> +static unsigned long __target_hint;
>> static struct symbol *__result_sym;
>> +#define HINT_MASK 0xfff
>>
>> static int filter_target_symbol(struct map *map __maybe_unused,
>> struct symbol *sym)
>> {
>> + /* Check the last bits is same */
>> + if (__target_hint)
>> + if ((sym->start & HINT_MASK) != (__target_hint & HINT_MASK))
>> + return 0;
>> +
>> if (strcmp(__target_symbol, sym->name) == 0) {
>> __result_sym = sym;
>> return 0;
>> @@ -201,7 +208,7 @@ static int filter_target_symbol(struct map *map __maybe_unused,
>>
>> /* Find the offset of the symbol in the executable binary */
>> static int find_symbol_offset(const char *exec, const char *function,
>> - unsigned long *offs)
>> + unsigned long hint, unsigned long *offs)
>> {
>> struct symbol *sym;
>> struct map *map = NULL;
>> @@ -218,6 +225,7 @@ static int find_symbol_offset(const char *exec, const char *function,
>> pr_debug("Search %s in %s\n", function, exec);
>> __target_symbol = function;
>> __result_sym = NULL;
>> + __target_hint = hint;
>> if (map__load(map, filter_target_symbol)) {
>> pr_err("Failed to find %s in %s.\n", function, exec);
>> goto out;
>> @@ -357,8 +365,10 @@ static int add_exec_to_probe_trace_events(struct probe_trace_event *tevs,
>> return 0;
>>
>> for (i = 0; i < ntevs && ret >= 0; i++) {
>> + offset = tevs[i].point.address - tevs[i].point.offset;
>> /* Get proper offset */
>> - ret = find_symbol_offset(exec, tevs[i].point.symbol, &offset);
>> + ret = find_symbol_offset(exec, tevs[i].point.symbol,
>> + offset, &offset);
>> if (ret < 0)
>> break;
>> offset += tevs[i].point.offset;
>> @@ -2418,7 +2428,7 @@ static int convert_name_to_addr(struct perf_probe_event *pev, const char *exec)
>> goto out;
>> }
>>
>> - ret = find_symbol_offset(exec, pp->function, &vaddr);
>> + ret = find_symbol_offset(exec, pp->function, 0, &vaddr);
>> if (ret < 0)
>> goto out;
>>
>> diff --git a/tools/perf/util/probe-event.h b/tools/perf/util/probe-event.h
>> index f9f3de8..d481c46 100644
>> --- a/tools/perf/util/probe-event.h
>> +++ b/tools/perf/util/probe-event.h
>> @@ -12,6 +12,7 @@ struct probe_trace_point {
>> char *symbol; /* Base symbol */
>> char *module; /* Module name */
>> unsigned long offset; /* Offset from symbol */
>> + unsigned long address; /* Actual address of the trace point */
>> bool retprobe; /* Return probe flag */
>> };
>>
>> diff --git a/tools/perf/util/probe-finder.c b/tools/perf/util/probe-finder.c
>> index ffb657f..7db7e05 100644
>> --- a/tools/perf/util/probe-finder.c
>> +++ b/tools/perf/util/probe-finder.c
>> @@ -729,6 +729,7 @@ static int convert_to_trace_point(Dwarf_Die *sp_die, Dwfl_Module *mod,
>> return -ENOENT;
>> }
>> tp->offset = (unsigned long)(paddr - sym.st_value);
>> + tp->address = (unsigned long)paddr;
>> tp->symbol = strdup(symbol);
>> if (!tp->symbol)
>> return -ENOMEM;
>>
>


--
Masami HIRAMATSU
IT Management Research Dept. Linux Technology Center
Hitachi, Ltd., Yokohama Research Laboratory
E-mail: masami.hiramatsu.pt@xxxxxxxxxxx


--
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/