[010/107] perf trace: Handle raw_syscalls:sys_enter just like the BPF_OUTPUT augmented event
diff mbox series

Message ID 20190730025610.22603-11-acme@kernel.org
State Accepted
Commit b119970aa541091e405373399690c24ead9d2920
Headers show
  • [001/107] perf include bpf: Add bpf_tail_call() prototype
Related show

Commit Message

Arnaldo Carvalho de Melo July 30, 2019, 2:54 a.m. UTC
From: Arnaldo Carvalho de Melo <acme@redhat.com>

So, we use a PERF_COUNT_SW_BPF_OUTPUT to output the augmented sys_enter
payload, i.e. to output more than just the raw syscall args, and if
something goes wrong when handling an unfiltered syscall, we bail out
and just return 1 in the bpf program associated with
raw_syscalls:sys_enter, meaning, don't filter that tracepoint, in which
case what will appear in the perf ring buffer isn't the BPF_OUTPUT
event, but the original raw_syscalls:sys_enter event with its normal

Now that we're switching to using a bpf_tail_call +
BPF_MAP_TYPE_PROG_ARRAY we're going to use this in the common case, so a
bug where raw_syscalls:sys_enter wasn't being handled by
trace__sys_enter() surfaced and for  that case, instead of using the
strace-like augmenter (trace__sys_enter()), we continued to use the
normal generic tracepoint handler:

  (gdb) p evsel
  $2 = (struct perf_evsel *) 0xc03e40
  (gdb) p evsel->name
  $3 = 0xbc56c0 "raw_syscalls:sys_enter"
  (gdb) p ((struct perf_evsel *) 0xc03e40)->name
  $4 = 0xbc56c0 "raw_syscalls:sys_enter"
  (gdb) p ((struct perf_evsel *) 0xc03e40)->handler
  $5 = (void *) 0x495eb3 <trace__event_handler>

This resulted in this:

     0.027 raw_syscalls:sys_enter:NR 12 (0, 7fcfcac64c9b, 4d, 7fcfcac64c9b, 7fcfcac6ce00, 19)
     ... [continued]: brk())                = 0x563b88677000

I.e. only the sys_exit tracepoint was being properly handled, but since
the sys_enter went to the generic trace__event_handler() we printed it
using libtraceevent's formatter instead of 'perf trace's strace-like

Fix it by setting trace__sys_enter() as the handler for
raw_syscalls:sys_enter and setup the tp_field tracepoint field

Now, to test it we just make raw_syscalls:sys_enter return 1 right after
checking if the pid is filtered, making it not use
bpf_perf_output_event() but rather ask for the tracepoint not to be
filtered and the result is the expected one:

  brk(NULL)                               = 0x556f42d6e000

I.e. raw_syscalls:sys_enter returns 1, gets handled by
trace__sys_enter() and gets it combined with the raw_syscalls:sys_exit
in a strace-like way.

Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Jiri Olsa <jolsa@kernel.org>
Cc: Luis Cláudio Gonçalves <lclaudio@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Link: https://lkml.kernel.org/n/tip-0mkocgk31nmy0odknegcby4z@git.kernel.org
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
 tools/perf/builtin-trace.c | 15 +++++++++++++++
 1 file changed, 15 insertions(+)

diff mbox series

diff --git a/tools/perf/builtin-trace.c b/tools/perf/builtin-trace.c
index fb8b8e78d7b5..872c9cc982a5 100644
--- a/tools/perf/builtin-trace.c
+++ b/tools/perf/builtin-trace.c
@@ -4128,7 +4128,22 @@  int cmd_trace(int argc, const char **argv)
 				if (perf_evsel__init_augmented_syscall_tp(augmented, evsel) ||
 					goto out;
+				/*
+				 * Augmented is __augmented_syscalls__ BPF_OUTPUT event
+				 * Above we made sure we can get from the payload the tp fields
+				 * that we get from syscalls:sys_enter tracefs format file.
+				 */
 				augmented->handler = trace__sys_enter;
+				/*
+				 * Now we do the same for the *syscalls:sys_enter event so that
+				 * if we handle it directly, i.e. if the BPF prog returns 0 so
+				 * as not to filter it, then we'll handle it just like we would
+				 * for the BPF_OUTPUT one:
+				 */
+				if (perf_evsel__init_augmented_syscall_tp(evsel, evsel) ||
+				    perf_evsel__init_augmented_syscall_tp_args(evsel))
+					goto out;
+				evsel->handler = trace__sys_enter;
 			if (strstarts(perf_evsel__name(evsel), "syscalls:sys_exit_")) {