Re: tracepoint filter problems

[Peter Zijlstra <peterz@infradead.org>]

On Thu, Oct 29, 2015 at 12:27:42PM -0300, Arnaldo Carvalho de Melo wrote:
> Hi,
> 
> 	In 'perf trace' we have an equivalent to strace's -e option, i.e. a way
> to state which syscalls one wants to see, so, when we do:
> 
> [root@felicio linux]# trace -e open
>      0.348 ( 0.005 ms): ls/6887 open(filename: /etc/ld.so.cache, flags: CLOEXEC                       ) = 3
>      0.363 ( 0.005 ms): ls/6887 open(filename: /lib64/libselinux.so.1, flags: CLOEXEC                 ) = 3
>      0.400 ( 0.006 ms): ls/6887 open(filename: /lib64/libcap.so.2, flags: CLOEXEC                     ) = 3
>      0.427 ( 0.006 ms): ls/6887 open(filename: /lib64/libacl.so.1, flags: CLOEXEC                     ) = 3
>      0.454 ( 0.006 ms): ls/6887 open(filename: /lib64/libc.so.6, flags: CLOEXEC                       ) = 3
>      0.497 ( 0.006 ms): ls/6887 open(filename: /lib64/libpcre.so.1, flags: CLOEXEC                    ) = 3
>      0.533 ( 0.006 ms): ls/6887 open(filename: /lib64/liblzma.so.5, flags: CLOEXEC                    ) = 3
>      0.561 ( 0.006 ms): ls/6887 open(filename: /lib64/libdl.so.2, flags: CLOEXEC                      ) = 3
>      0.586 ( 0.006 ms): ls/6887 open(filename: /lib64/libattr.so.1, flags: CLOEXEC                    ) = 3
>      0.611 ( 0.006 ms): ls/6887 open(filename: /lib64/libpthread.so.0, flags: CLOEXEC                 ) = 3
>      0.859 ( 0.007 ms): ls/6887 open(filename: /usr/lib/locale/locale-archive, flags: CLOEXEC         ) = 3
> ^C[root@felicio linux]#
> 
> I.e. tracing system wide, looking for any program that calls the open syscall, it works, as it
> used the PERF_EVENT_IOC_SET_FILTER to set the relevant 'id == 486' to filter it:
> 
> [root@felicio linux]# head -2 /sys/kernel/debug/tracing/events/syscalls/sys_enter_open/format 
> name: sys_enter_open
> ID: 486
> 
> Ok, this is how it should be, all works, but if I try instead calling a program
> that in turn calls another, the filter will apply for the first in line, but
> not for the calls made by another program called from it.
> 
> In the example below 'perf trace' will ask to see just the 'open' syscall, and
> it works for the started workload, namely 'perf record', but then it'll call
> 'sleep 100000000000' and there the filter doesn't apply, bummer :-\
> 
> The inherit bit is set, sure, as we can see when adding -vv to the 'perf trace'
> call, ideas?

--- a/kernel/events/core.c
+++ b/kernel/events/core.c
@@ -6939,6 +6939,10 @@ static int perf_tp_filter_match(struct p
 {
 	void *record = data->raw->data;
 
+	/* only top level events have filters set */
+	if (event->parent)
+		event = event->parent;
+
 	if (likely(!event->filter) || filter_match_preds(event->filter, record))
 		return 1;
 	return 0;