diff mbox series

perf test: Switch basic bpf filtering test to use syscall tracepoint

Message ID 20230123083224.276404-1-naveen.n.rao@linux.vnet.ibm.com (mailing list archive)
State Not Applicable
Delegated to: BPF
Headers show
Series perf test: Switch basic bpf filtering test to use syscall tracepoint | expand

Checks

Context Check Description
netdev/tree_selection success Not a local patch

Commit Message

Naveen N. Rao Jan. 23, 2023, 8:32 a.m. UTC
BPF filtering tests can sometime fail. Running the test in verbose mode
shows the following:
  $ sudo perf test 42
  42: BPF filter                                                      :
  42.1: Basic BPF filtering                                           : FAILED!
  42.2: BPF pinning                                                   : Skip
  42.3: BPF prologue generation                                       : Skip
  $ perf --version
  perf version 4.18.0-425.3.1.el8.ppc64le
  $ sudo perf test -v 42
  42: BPF filter                                                      :
  42.1: Basic BPF filtering                                           :
  --- start ---
  test child forked, pid 711060
  ...
  bpf: config 'func=do_epoll_wait' is ok
  Looking at the vmlinux_path (8 entries long)
  Using /usr/lib/debug/lib/modules/4.18.0-425.3.1.el8.ppc64le/vmlinux for symbols
  Open Debuginfo file: /usr/lib/debug/.build-id/81/56f5a07f92ccb62c5600ba0e4aacfb5f3a7534.debug
  Try to find probe point from debuginfo.
  Matched function: do_epoll_wait [4ef8cb0]
  found inline addr: 0xc00000000061dbe4
  Probe point found: __se_compat_sys_epoll_pwait+196
  found inline addr: 0xc00000000061d9f4
  Probe point found: __se_sys_epoll_pwait+196
  found inline addr: 0xc00000000061d824
  Probe point found: __se_sys_epoll_wait+36
  Found 3 probe_trace_events.
  Opening /sys/kernel/tracing//kprobe_events write=1
  ...
  BPF filter result incorrect, expected 56, got 56 samples
  test child finished with -1
  ---- end ----
  BPF filter subtest 1: FAILED!

The statement above about the result being incorrect looks weird, and it
is due to that particular perf build missing commit 3e11300cdfd5f1
("perf test: Fix bpf test sample mismatch reporting"). In reality, due
to commit 4b04e0decd2518 ("perf test: Fix basic bpf filtering test"),
perf expects there to be 56*3 samples.

However, the number of samples we receive is going to be dependent on
where the probes are installed, which is dependent on where
do_epoll_wait gets inlined. On s390x, it looks like probes at all the
inlined locations are hit. But, that is not the case on ppc64le.

Fix this by switching the test to instead use the syscall tracepoint.
This ensures that we will only ever install a single event enabling us
to reliably determine the sample count.

Reported-by: Disha Goel <disgoel@linux.vnet.ibm.com>
Signed-off-by: Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
---
 tools/perf/tests/bpf-script-example.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)


base-commit: 5670ebf54bd26482f57a094c53bdc562c106e0a9

Comments

Arnaldo Carvalho de Melo Jan. 23, 2023, 12:58 p.m. UTC | #1
Em Mon, Jan 23, 2023 at 02:02:24PM +0530, Naveen N. Rao escreveu:
> BPF filtering tests can sometime fail. Running the test in verbose mode
> shows the following:

Thanks, applied.

- Arnaldo

>   $ sudo perf test 42
>   42: BPF filter                                                      :
>   42.1: Basic BPF filtering                                           : FAILED!
>   42.2: BPF pinning                                                   : Skip
>   42.3: BPF prologue generation                                       : Skip
>   $ perf --version
>   perf version 4.18.0-425.3.1.el8.ppc64le
>   $ sudo perf test -v 42
>   42: BPF filter                                                      :
>   42.1: Basic BPF filtering                                           :
>   --- start ---
>   test child forked, pid 711060
>   ...
>   bpf: config 'func=do_epoll_wait' is ok
>   Looking at the vmlinux_path (8 entries long)
>   Using /usr/lib/debug/lib/modules/4.18.0-425.3.1.el8.ppc64le/vmlinux for symbols
>   Open Debuginfo file: /usr/lib/debug/.build-id/81/56f5a07f92ccb62c5600ba0e4aacfb5f3a7534.debug
>   Try to find probe point from debuginfo.
>   Matched function: do_epoll_wait [4ef8cb0]
>   found inline addr: 0xc00000000061dbe4
>   Probe point found: __se_compat_sys_epoll_pwait+196
>   found inline addr: 0xc00000000061d9f4
>   Probe point found: __se_sys_epoll_pwait+196
>   found inline addr: 0xc00000000061d824
>   Probe point found: __se_sys_epoll_wait+36
>   Found 3 probe_trace_events.
>   Opening /sys/kernel/tracing//kprobe_events write=1
>   ...
>   BPF filter result incorrect, expected 56, got 56 samples
>   test child finished with -1
>   ---- end ----
>   BPF filter subtest 1: FAILED!
> 
> The statement above about the result being incorrect looks weird, and it
> is due to that particular perf build missing commit 3e11300cdfd5f1
> ("perf test: Fix bpf test sample mismatch reporting"). In reality, due
> to commit 4b04e0decd2518 ("perf test: Fix basic bpf filtering test"),
> perf expects there to be 56*3 samples.
> 
> However, the number of samples we receive is going to be dependent on
> where the probes are installed, which is dependent on where
> do_epoll_wait gets inlined. On s390x, it looks like probes at all the
> inlined locations are hit. But, that is not the case on ppc64le.
> 
> Fix this by switching the test to instead use the syscall tracepoint.
> This ensures that we will only ever install a single event enabling us
> to reliably determine the sample count.
> 
> Reported-by: Disha Goel <disgoel@linux.vnet.ibm.com>

> Signed-off-by: Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
> ---
>  tools/perf/tests/bpf-script-example.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/tools/perf/tests/bpf-script-example.c b/tools/perf/tests/bpf-script-example.c
> index 7981c69ed1b456..b638cc99d5ae56 100644
> --- a/tools/perf/tests/bpf-script-example.c
> +++ b/tools/perf/tests/bpf-script-example.c
> @@ -43,7 +43,7 @@ struct {
>  	__type(value, int);
>  } flip_table SEC(".maps");
>  
> -SEC("func=do_epoll_wait")
> +SEC("syscalls:sys_enter_epoll_pwait")
>  int bpf_func__SyS_epoll_pwait(void *ctx)
>  {
>  	int ind =0;
> 
> base-commit: 5670ebf54bd26482f57a094c53bdc562c106e0a9
> -- 
> 2.39.1
>
kajoljain Jan. 30, 2023, 8:58 a.m. UTC | #2
On 1/23/23 14:02, Naveen N. Rao wrote:
> BPF filtering tests can sometime fail. Running the test in verbose mode
> shows the following:
>   $ sudo perf test 42
>   42: BPF filter                                                      :
>   42.1: Basic BPF filtering                                           : FAILED!
>   42.2: BPF pinning                                                   : Skip
>   42.3: BPF prologue generation                                       : Skip
>   $ perf --version
>   perf version 4.18.0-425.3.1.el8.ppc64le
>   $ sudo perf test -v 42
>   42: BPF filter                                                      :
>   42.1: Basic BPF filtering                                           :
>   --- start ---
>   test child forked, pid 711060
>   ...
>   bpf: config 'func=do_epoll_wait' is ok
>   Looking at the vmlinux_path (8 entries long)
>   Using /usr/lib/debug/lib/modules/4.18.0-425.3.1.el8.ppc64le/vmlinux for symbols
>   Open Debuginfo file: /usr/lib/debug/.build-id/81/56f5a07f92ccb62c5600ba0e4aacfb5f3a7534.debug
>   Try to find probe point from debuginfo.
>   Matched function: do_epoll_wait [4ef8cb0]
>   found inline addr: 0xc00000000061dbe4
>   Probe point found: __se_compat_sys_epoll_pwait+196
>   found inline addr: 0xc00000000061d9f4
>   Probe point found: __se_sys_epoll_pwait+196
>   found inline addr: 0xc00000000061d824
>   Probe point found: __se_sys_epoll_wait+36
>   Found 3 probe_trace_events.
>   Opening /sys/kernel/tracing//kprobe_events write=1
>   ...
>   BPF filter result incorrect, expected 56, got 56 samples
>   test child finished with -1
>   ---- end ----
>   BPF filter subtest 1: FAILED!

Patch looks good to me.

Reviewed-by: Kajol Jain<kjain@linux.ibm.com>

Thanks,
Kajol Jain
> 
> The statement above about the result being incorrect looks weird, and it
> is due to that particular perf build missing commit 3e11300cdfd5f1
> ("perf test: Fix bpf test sample mismatch reporting"). In reality, due
> to commit 4b04e0decd2518 ("perf test: Fix basic bpf filtering test"),
> perf expects there to be 56*3 samples.
> 
> However, the number of samples we receive is going to be dependent on
> where the probes are installed, which is dependent on where
> do_epoll_wait gets inlined. On s390x, it looks like probes at all the
> inlined locations are hit. But, that is not the case on ppc64le.
> 
> Fix this by switching the test to instead use the syscall tracepoint.
> This ensures that we will only ever install a single event enabling us
> to reliably determine the sample count.
> 
> Reported-by: Disha Goel <disgoel@linux.vnet.ibm.com>
> Signed-off-by: Naveen N. Rao <naveen.n.rao@linux.vnet.ibm.com>
> ---
>  tools/perf/tests/bpf-script-example.c | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> diff --git a/tools/perf/tests/bpf-script-example.c b/tools/perf/tests/bpf-script-example.c
> index 7981c69ed1b456..b638cc99d5ae56 100644
> --- a/tools/perf/tests/bpf-script-example.c
> +++ b/tools/perf/tests/bpf-script-example.c
> @@ -43,7 +43,7 @@ struct {
>  	__type(value, int);
>  } flip_table SEC(".maps");
>  
> -SEC("func=do_epoll_wait")
> +SEC("syscalls:sys_enter_epoll_pwait")
>  int bpf_func__SyS_epoll_pwait(void *ctx)
>  {
>  	int ind =0;
> 
> base-commit: 5670ebf54bd26482f57a094c53bdc562c106e0a9
Arnaldo Carvalho de Melo Jan. 30, 2023, 2:09 p.m. UTC | #3
Em Mon, Jan 30, 2023 at 02:28:49PM +0530, kajoljain escreveu:
> 
> 
> On 1/23/23 14:02, Naveen N. Rao wrote:
> > BPF filtering tests can sometime fail. Running the test in verbose mode
> > shows the following:
> >   $ sudo perf test 42
> >   42: BPF filter                                                      :
> >   42.1: Basic BPF filtering                                           : FAILED!
> >   42.2: BPF pinning                                                   : Skip
> >   42.3: BPF prologue generation                                       : Skip
> >   $ perf --version
> >   perf version 4.18.0-425.3.1.el8.ppc64le
> >   $ sudo perf test -v 42
> >   42: BPF filter                                                      :
> >   42.1: Basic BPF filtering                                           :
> >   --- start ---
> >   test child forked, pid 711060
> >   ...
> >   bpf: config 'func=do_epoll_wait' is ok
> >   Looking at the vmlinux_path (8 entries long)
> >   Using /usr/lib/debug/lib/modules/4.18.0-425.3.1.el8.ppc64le/vmlinux for symbols
> >   Open Debuginfo file: /usr/lib/debug/.build-id/81/56f5a07f92ccb62c5600ba0e4aacfb5f3a7534.debug
> >   Try to find probe point from debuginfo.
> >   Matched function: do_epoll_wait [4ef8cb0]
> >   found inline addr: 0xc00000000061dbe4
> >   Probe point found: __se_compat_sys_epoll_pwait+196
> >   found inline addr: 0xc00000000061d9f4
> >   Probe point found: __se_sys_epoll_pwait+196
> >   found inline addr: 0xc00000000061d824
> >   Probe point found: __se_sys_epoll_wait+36
> >   Found 3 probe_trace_events.
> >   Opening /sys/kernel/tracing//kprobe_events write=1
> >   ...
> >   BPF filter result incorrect, expected 56, got 56 samples
> >   test child finished with -1
> >   ---- end ----
> >   BPF filter subtest 1: FAILED!
> 
> Patch looks good to me.
> 
> Reviewed-by: Kajol Jain<kjain@linux.ibm.com>

Thanks, added to that cset, as it is still just on tmp.perf/core.

- Arnaldo
diff mbox series

Patch

diff --git a/tools/perf/tests/bpf-script-example.c b/tools/perf/tests/bpf-script-example.c
index 7981c69ed1b456..b638cc99d5ae56 100644
--- a/tools/perf/tests/bpf-script-example.c
+++ b/tools/perf/tests/bpf-script-example.c
@@ -43,7 +43,7 @@  struct {
 	__type(value, int);
 } flip_table SEC(".maps");
 
-SEC("func=do_epoll_wait")
+SEC("syscalls:sys_enter_epoll_pwait")
 int bpf_func__SyS_epoll_pwait(void *ctx)
 {
 	int ind =0;