* The issue with `perf report -s comm`
@ 2017-09-21 14:51 Alexander Pozdneev
2017-10-04 6:13 ` [RFC] perf callchain: Compare dsos (as well) for CCKEY_FUNCTION Ravi Bangoria
0 siblings, 1 reply; 7+ messages in thread
From: Alexander Pozdneev @ 2017-09-21 14:51 UTC (permalink / raw)
To: linux-perf-users
Hello,
I have the following issue with perf `report -s comm` that seems to be
a bug. This is how the code looks like:
int main() {
double a, b, c;
a = do_things_main(WAIT_TIME * 2); // 40% of time, a.out
b = do_things1(WAIT_TIME); // 20% of time, libdo1.so
c = do_things2(WAIT_TIME * 2); // 20% of time, libdo2.so
return (int)(a + b + c);
}
This is how I run it:
$ LD_LIBRARY_PATH="." perf record -g ./a.out
[ perf record: Woken up 3 times to write data ]
[ perf record: Captured and wrote 0.743 MB perf.data (11911 samples) ]
This is the regular output of `perf report`:
$ perf report | grep -v '#' | head -n 12
39.99% 39.99% a.out a.out [.] main
|
---main
39.94% 39.94% a.out libdo2.so [.] do_things2
|
---do_things2
20.05% 20.05% a.out libdo1.so [.] do_things1
|
---do_things1
It looks exactly as expected. If I sort by `dso`, the output is also OK:
$ perf report -s dso | grep -v '#'
39.99% 39.99% a.out
|
---main
39.94% 39.94% libdo2.so
|
---do_things2
20.05% 20.05% libdo1.so
|
---do_things1
However, If I try to sort by `comm`, the output looks strange:
$ perf report -s comm | grep -v '#'
100.00% 100.00% a.out
|
|--59.99%--do_things1
|
--39.99%--main
Specifically, `do_things2()` is missing in the report.
Here is the full source code of this example:
https://github.com/pozdneev/perf-report-s-comm-bug
Originally, Louis Stuber (IBM Client Center Montpellier) discovered
this behaviour when we have been playing with Flame Graphs
(http://www.brendangregg.com/blog/2016-04-30/linux-perf-folded.html).
This is the list of Linux/perf versions that I've tried:
* Ubuntu 14.04.5 (3.19.0-80-generic), perf version 3.19.8-ckt22
* RHEL 7.3 ppc64le (3.10.0-514.el7.ppc64le), perf version
3.10.0-514.el7.ppc64le.debug
* My colleague confirms that he observes the same behavior with kernel
version 4.13.
Could you please check if the issue remains in the current development
version of Linux perf?
Thanks!
Alexander
^ permalink raw reply [flat|nested] 7+ messages in thread
* [RFC] perf callchain: Compare dsos (as well) for CCKEY_FUNCTION
2017-09-21 14:51 The issue with `perf report -s comm` Alexander Pozdneev
@ 2017-10-04 6:13 ` Ravi Bangoria
2017-10-04 13:08 ` Jiri Olsa
0 siblings, 1 reply; 7+ messages in thread
From: Ravi Bangoria @ 2017-10-04 6:13 UTC (permalink / raw)
To: pozdneyev, acme, linux-kernel
Cc: linux-perf-users, peterz, mingo, alexander.shishkin, yao.jin, ak,
jolsa, kjlx, milian.wolff, zhangmengting, Ravi Bangoria
Two functions from different binaries can have same start
address. Thus, comparing only start address in match_chain()
leads to inconsistent callchains. Fix this by adding a check
for dsos as well.
Ex, https://www.spinics.net/lists/linux-perf-users/msg04067.html
Reported-by: Alexander Pozdneev <pozdneyev@gmail.com>
Signed-off-by: Ravi Bangoria <ravi.bangoria@linux.vnet.ibm.com>
---
tools/perf/util/callchain.c | 9 ++++++++-
1 file changed, 8 insertions(+), 1 deletion(-)
diff --git a/tools/perf/util/callchain.c b/tools/perf/util/callchain.c
index 510b513..6d5a483 100644
--- a/tools/perf/util/callchain.c
+++ b/tools/perf/util/callchain.c
@@ -678,6 +678,9 @@ static enum match_result match_chain(struct callchain_cursor_node *node,
{
struct symbol *sym = node->sym;
u64 left, right;
+ struct dso *left_dso = NULL;
+ struct dso *right_dso = NULL;
+
if (callchain_param.key == CCKEY_SRCLINE) {
enum match_result match = match_chain_srcline(node, cnode);
@@ -689,12 +692,16 @@ static enum match_result match_chain(struct callchain_cursor_node *node,
if (cnode->ms.sym && sym && callchain_param.key == CCKEY_FUNCTION) {
left = cnode->ms.sym->start;
right = sym->start;
+ if (cnode->ms.map && node->map) {
+ left_dso = cnode->ms.map->dso;
+ right_dso = node->map->dso;
+ }
} else {
left = cnode->ip;
right = node->ip;
}
- if (left == right) {
+ if (left == right && left_dso == right_dso) {
if (node->branch) {
cnode->branch_count++;
--
1.8.3.1
^ permalink raw reply related [flat|nested] 7+ messages in thread
* Re: [RFC] perf callchain: Compare dsos (as well) for CCKEY_FUNCTION
2017-10-04 6:13 ` [RFC] perf callchain: Compare dsos (as well) for CCKEY_FUNCTION Ravi Bangoria
@ 2017-10-04 13:08 ` Jiri Olsa
2017-10-05 3:50 ` [PATCH] " Ravi Bangoria
0 siblings, 1 reply; 7+ messages in thread
From: Jiri Olsa @ 2017-10-04 13:08 UTC (permalink / raw)
To: Ravi Bangoria
Cc: pozdneyev, acme, linux-kernel, linux-perf-users, peterz, mingo,
alexander.shishkin, yao.jin, ak, jolsa, kjlx, milian.wolff,
zhangmengting
On Wed, Oct 04, 2017 at 11:43:08AM +0530, Ravi Bangoria wrote:
> Two functions from different binaries can have same start
> address. Thus, comparing only start address in match_chain()
> leads to inconsistent callchains. Fix this by adding a check
> for dsos as well.
>
> Ex, https://www.spinics.net/lists/linux-perf-users/msg04067.html
>
> Reported-by: Alexander Pozdneev <pozdneyev@gmail.com>
> Signed-off-by: Ravi Bangoria <ravi.bangoria@linux.vnet.ibm.com>
> ---
> tools/perf/util/callchain.c | 9 ++++++++-
> 1 file changed, 8 insertions(+), 1 deletion(-)
>
> diff --git a/tools/perf/util/callchain.c b/tools/perf/util/callchain.c
> index 510b513..6d5a483 100644
> --- a/tools/perf/util/callchain.c
> +++ b/tools/perf/util/callchain.c
> @@ -678,6 +678,9 @@ static enum match_result match_chain(struct callchain_cursor_node *node,
> {
> struct symbol *sym = node->sym;
> u64 left, right;
> + struct dso *left_dso = NULL;
> + struct dso *right_dso = NULL;
> +
>
> if (callchain_param.key == CCKEY_SRCLINE) {
> enum match_result match = match_chain_srcline(node, cnode);
> @@ -689,12 +692,16 @@ static enum match_result match_chain(struct callchain_cursor_node *node,
> if (cnode->ms.sym && sym && callchain_param.key == CCKEY_FUNCTION) {
> left = cnode->ms.sym->start;
> right = sym->start;
> + if (cnode->ms.map && node->map) {
> + left_dso = cnode->ms.map->dso;
> + right_dso = node->map->dso;
makes sense.. but why not to get those maps separately?
if (cnode->ms.map)
left_dso = cnode->ms.map->dso;
if (node->map) {
right_dso = node->map->dso;
I'd think that if one is missing, it's most likely different
map/dso and you want to fail the == check
jirka
> + }
> } else {
> left = cnode->ip;
> right = node->ip;
> }
>
> - if (left == right) {
> + if (left == right && left_dso == right_dso) {
> if (node->branch) {
> cnode->branch_count++;
>
> --
> 1.8.3.1
>
^ permalink raw reply [flat|nested] 7+ messages in thread
* [PATCH] perf callchain: Compare dsos (as well) for CCKEY_FUNCTION
2017-10-04 13:08 ` Jiri Olsa
@ 2017-10-05 3:50 ` Ravi Bangoria
2017-10-05 4:13 ` Namhyung Kim
0 siblings, 1 reply; 7+ messages in thread
From: Ravi Bangoria @ 2017-10-05 3:50 UTC (permalink / raw)
To: pozdneyev, acme, linux-kernel, jolsa
Cc: linux-perf-users, peterz, mingo, alexander.shishkin, yao.jin, ak,
kjlx, milian.wolff, zhangmengting, Ravi Bangoria
Two functions from different binaries can have same start
address. Thus, comparing only start address in match_chain()
leads to inconsistent callchains. Fix this by adding a check
for dsos as well.
Ex, https://www.spinics.net/lists/linux-perf-users/msg04067.html
Reported-by: Alexander Pozdneev <pozdneyev@gmail.com>
Signed-off-by: Ravi Bangoria <ravi.bangoria@linux.vnet.ibm.com>
---
tools/perf/util/callchain.c | 9 ++++++++-
1 file changed, 8 insertions(+), 1 deletion(-)
diff --git a/tools/perf/util/callchain.c b/tools/perf/util/callchain.c
index be09d77..6d7f645 100644
--- a/tools/perf/util/callchain.c
+++ b/tools/perf/util/callchain.c
@@ -685,6 +685,9 @@ static enum match_result match_chain(struct callchain_cursor_node *node,
{
struct symbol *sym = node->sym;
u64 left, right;
+ struct dso *left_dso = NULL;
+ struct dso *right_dso = NULL;
+
if (callchain_param.key == CCKEY_SRCLINE) {
enum match_result match = match_chain_srcline(node, cnode);
@@ -696,12 +699,16 @@ static enum match_result match_chain(struct callchain_cursor_node *node,
if (cnode->ms.sym && sym && callchain_param.key == CCKEY_FUNCTION) {
left = cnode->ms.sym->start;
right = sym->start;
+ if (cnode->ms.map)
+ left_dso = cnode->ms.map->dso;
+ if (node->map)
+ right_dso = node->map->dso;
} else {
left = cnode->ip;
right = node->ip;
}
- if (left == right) {
+ if (left == right && left_dso == right_dso) {
if (node->branch) {
cnode->branch_count++;
--
1.8.3.1
^ permalink raw reply related [flat|nested] 7+ messages in thread
* Re: [PATCH] perf callchain: Compare dsos (as well) for CCKEY_FUNCTION
2017-10-05 3:50 ` [PATCH] " Ravi Bangoria
@ 2017-10-05 4:13 ` Namhyung Kim
2017-10-05 9:12 ` [PATCH v2] " Ravi Bangoria
0 siblings, 1 reply; 7+ messages in thread
From: Namhyung Kim @ 2017-10-05 4:13 UTC (permalink / raw)
To: Ravi Bangoria
Cc: pozdneyev, acme, linux-kernel, jolsa, linux-perf-users, peterz,
mingo, alexander.shishkin, yao.jin, ak, kjlx, milian.wolff,
zhangmengting, kernel-team
Hi,
On Thu, Oct 05, 2017 at 09:20:21AM +0530, Ravi Bangoria wrote:
> Two functions from different binaries can have same start
> address. Thus, comparing only start address in match_chain()
> leads to inconsistent callchains. Fix this by adding a check
> for dsos as well.
>
> Ex, https://www.spinics.net/lists/linux-perf-users/msg04067.html
>
> Reported-by: Alexander Pozdneev <pozdneyev@gmail.com>
> Signed-off-by: Ravi Bangoria <ravi.bangoria@linux.vnet.ibm.com>
> ---
> tools/perf/util/callchain.c | 9 ++++++++-
> 1 file changed, 8 insertions(+), 1 deletion(-)
>
> diff --git a/tools/perf/util/callchain.c b/tools/perf/util/callchain.c
> index be09d77..6d7f645 100644
> --- a/tools/perf/util/callchain.c
> +++ b/tools/perf/util/callchain.c
> @@ -685,6 +685,9 @@ static enum match_result match_chain(struct callchain_cursor_node *node,
> {
> struct symbol *sym = node->sym;
> u64 left, right;
> + struct dso *left_dso = NULL;
> + struct dso *right_dso = NULL;
> +
>
> if (callchain_param.key == CCKEY_SRCLINE) {
> enum match_result match = match_chain_srcline(node, cnode);
> @@ -696,12 +699,16 @@ static enum match_result match_chain(struct callchain_cursor_node *node,
> if (cnode->ms.sym && sym && callchain_param.key == CCKEY_FUNCTION) {
> left = cnode->ms.sym->start;
> right = sym->start;
> + if (cnode->ms.map)
> + left_dso = cnode->ms.map->dso;
> + if (node->map)
> + right_dso = node->map->dso;
AFAIK having a symbol guarantees having a map too. So it could simply
be:
left_dso = cnode->ms.map->dso;
right_dso = node->map->dso;
Thanks,
Namhyung
> } else {
> left = cnode->ip;
> right = node->ip;
> }
>
> - if (left == right) {
> + if (left == right && left_dso == right_dso) {
> if (node->branch) {
> cnode->branch_count++;
>
> --
> 1.8.3.1
>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-perf-users" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at http://vger.kernel.org/majordomo-info.html
^ permalink raw reply [flat|nested] 7+ messages in thread
* [PATCH v2] perf callchain: Compare dsos (as well) for CCKEY_FUNCTION
2017-10-05 4:13 ` Namhyung Kim
@ 2017-10-05 9:12 ` Ravi Bangoria
2017-10-05 9:26 ` Jiri Olsa
0 siblings, 1 reply; 7+ messages in thread
From: Ravi Bangoria @ 2017-10-05 9:12 UTC (permalink / raw)
To: pozdneyev, acme, linux-kernel, jolsa, namhyung
Cc: linux-perf-users, peterz, mingo, alexander.shishkin, yao.jin, ak,
kjlx, milian.wolff, zhangmengting, Ravi Bangoria
Two functions from different binaries can have same start
address. Thus, comparing only start address in match_chain()
leads to inconsistent callchains. Fix this by adding a check
for dsos as well.
Ex, https://www.spinics.net/lists/linux-perf-users/msg04067.html
Reported-by: Alexander Pozdneev <pozdneyev@gmail.com>
Signed-off-by: Ravi Bangoria <ravi.bangoria@linux.vnet.ibm.com>
---
Changes in v2:
- Remove unnecessary checks for 'map'
tools/perf/util/callchain.c | 6 +++++-
1 file changed, 5 insertions(+), 1 deletion(-)
diff --git a/tools/perf/util/callchain.c b/tools/perf/util/callchain.c
index be09d77..a971caf 100644
--- a/tools/perf/util/callchain.c
+++ b/tools/perf/util/callchain.c
@@ -685,6 +685,8 @@ static enum match_result match_chain(struct callchain_cursor_node *node,
{
struct symbol *sym = node->sym;
u64 left, right;
+ struct dso *left_dso = NULL;
+ struct dso *right_dso = NULL;
if (callchain_param.key == CCKEY_SRCLINE) {
enum match_result match = match_chain_srcline(node, cnode);
@@ -696,12 +698,14 @@ static enum match_result match_chain(struct callchain_cursor_node *node,
if (cnode->ms.sym && sym && callchain_param.key == CCKEY_FUNCTION) {
left = cnode->ms.sym->start;
right = sym->start;
+ left_dso = cnode->ms.map->dso;
+ right_dso = node->map->dso;
} else {
left = cnode->ip;
right = node->ip;
}
- if (left == right) {
+ if (left == right && left_dso == right_dso) {
if (node->branch) {
cnode->branch_count++;
--
1.8.3.1
^ permalink raw reply related [flat|nested] 7+ messages in thread
* Re: [PATCH v2] perf callchain: Compare dsos (as well) for CCKEY_FUNCTION
2017-10-05 9:12 ` [PATCH v2] " Ravi Bangoria
@ 2017-10-05 9:26 ` Jiri Olsa
0 siblings, 0 replies; 7+ messages in thread
From: Jiri Olsa @ 2017-10-05 9:26 UTC (permalink / raw)
To: Ravi Bangoria
Cc: pozdneyev, acme, linux-kernel, jolsa, namhyung, linux-perf-users,
peterz, mingo, alexander.shishkin, yao.jin, ak, kjlx,
milian.wolff, zhangmengting
On Thu, Oct 05, 2017 at 02:42:34PM +0530, Ravi Bangoria wrote:
> Two functions from different binaries can have same start
> address. Thus, comparing only start address in match_chain()
> leads to inconsistent callchains. Fix this by adding a check
> for dsos as well.
>
> Ex, https://www.spinics.net/lists/linux-perf-users/msg04067.html
>
> Reported-by: Alexander Pozdneev <pozdneyev@gmail.com>
> Signed-off-by: Ravi Bangoria <ravi.bangoria@linux.vnet.ibm.com>
> ---
> Changes in v2:
> - Remove unnecessary checks for 'map'
Acked-by: Jiri Olsa <jolsa@kernel.org>
thanks,
jirka
>
> tools/perf/util/callchain.c | 6 +++++-
> 1 file changed, 5 insertions(+), 1 deletion(-)
>
> diff --git a/tools/perf/util/callchain.c b/tools/perf/util/callchain.c
> index be09d77..a971caf 100644
> --- a/tools/perf/util/callchain.c
> +++ b/tools/perf/util/callchain.c
> @@ -685,6 +685,8 @@ static enum match_result match_chain(struct callchain_cursor_node *node,
> {
> struct symbol *sym = node->sym;
> u64 left, right;
> + struct dso *left_dso = NULL;
> + struct dso *right_dso = NULL;
>
> if (callchain_param.key == CCKEY_SRCLINE) {
> enum match_result match = match_chain_srcline(node, cnode);
> @@ -696,12 +698,14 @@ static enum match_result match_chain(struct callchain_cursor_node *node,
> if (cnode->ms.sym && sym && callchain_param.key == CCKEY_FUNCTION) {
> left = cnode->ms.sym->start;
> right = sym->start;
> + left_dso = cnode->ms.map->dso;
> + right_dso = node->map->dso;
> } else {
> left = cnode->ip;
> right = node->ip;
> }
>
> - if (left == right) {
> + if (left == right && left_dso == right_dso) {
> if (node->branch) {
> cnode->branch_count++;
>
> --
> 1.8.3.1
>
^ permalink raw reply [flat|nested] 7+ messages in thread
end of thread, other threads:[~2017-10-05 9:26 UTC | newest]
Thread overview: 7+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2017-09-21 14:51 The issue with `perf report -s comm` Alexander Pozdneev
2017-10-04 6:13 ` [RFC] perf callchain: Compare dsos (as well) for CCKEY_FUNCTION Ravi Bangoria
2017-10-04 13:08 ` Jiri Olsa
2017-10-05 3:50 ` [PATCH] " Ravi Bangoria
2017-10-05 4:13 ` Namhyung Kim
2017-10-05 9:12 ` [PATCH v2] " Ravi Bangoria
2017-10-05 9:26 ` Jiri Olsa
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).