perf trace: Handle raw_syscalls:sys_enter just like the BPF_OUTPUT augmented event
tip-bot for Arnaldo Carvalho de Melo
merge botCommit-ID: b119970aa541091e405373399690c24ead9d2920 Gitweb: https://git.kernel.org/tip/b119970aa541091e405373399690c24ead9d2920 Author: Arnaldo Carvalho de Melo <[email protected]> AuthorDate: Tue, 16 Jul 2019 11:53:03 -0300 Committer: Arnaldo Carvalho de Melo <[email protected]> CommitDate: Mon, 29 Jul 2019 18:34:41 -0300
perf trace: Handle raw_syscalls:sys_enter just like the BPF_OUTPUT augmented event
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 payload.
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()) = 0x563b88677000I.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 one.
Fix it by setting trace__sys_enter() as the handler for raw_syscalls:sys_enter and setup the tp_field tracepoint field accessors.
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.
--- tools/perf/builtin-trace.c | 15 +++++++++++++++ 1 file changed, 15 insertions(+)
tools/perf/builtin-trace.c +15 -0
@@ -4128,7 +4128,22 @@ int cmd_trace(int argc, const char **argv)if (perf_evsel__init_augmented_syscall_tp (augmented, evsel) ||perf_evsel__init_augmented_syscall_tp_args (augmented)) 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_" )) {