Re: [PATCH v5 00/11] perf: Support multiple system call tables in the build

[Arnaldo Carvalho de Melo <acme@kernel.org>]

On Fri, Mar 14, 2025 at 02:26:41PM -0300, Arnaldo Carvalho de Melo wrote:
> it finds the pair, but then its sc->args has a bogus pointer... I'll see
> where this isn't being initialized...

Breakpoint 4, trace__find_usable_bpf_prog_entry (trace=0x7fffffffa510, sc=0x1046f10) at builtin-trace.c:3874
3874			bool is_candidate = false;
(gdb) n
3876			if (pair == NULL || pair == sc ||
(gdb) p pair
$7 = (struct syscall *) 0x1083c50
(gdb) p pair->name
$8 = 0x81478e "accept4"
(gdb) n
3877			    pair->bpf_prog.sys_enter == trace->skel->progs.syscall_unaugmented)
(gdb) p i
$9 = 1
(gdb) n
3876			if (pair == NULL || pair == sc ||
(gdb) n
3880			printf("sc=%p\n", sc); fflush(stdout);
(gdb) n
sc=0x1046f10
3881			printf("sc->name=%p\n", sc->name); fflush(stdout);
(gdb) n
sc->name=0x6c66202c786c3830
3882			printf("sc->nr_args=%d, sc->args=%p\n", sc->nr_args, sc->args); fflush(stdout);
(gdb) p sc->nr_args
$10 = 1935635045
(gdb) p sc->args
$11 = (struct tep_format_field *) 0x257830203a6e656c
(gdb) p *sc
$12 = {e_machine = 540697702, id = 807761968, tp_format = 0x657075202c786c38, nr_args = 1935635045, args_size = 1634427759, bpf_prog = {sys_enter = 0x257830203a726464, 
    sys_exit = 0x7075202c786c3830}, is_exit = 101, is_open = 101, nonexistent = 114, use_btf = 95, args = 0x257830203a6e656c, 
  name = 0x6c66202c786c3830 <error: Cannot access memory at address 0x6c66202c786c3830>, fmt = 0x257830203a736761, arg_fmt = 0x786c3830}
(gdb) 

Ok, ran out of time, but if I simple avoid the second loop in:

static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace, int e_machine)


I.e. the one that starts with:

        /*
         * Now lets do a second pass looking for enabled syscalls without
         * an augmenter that have a signature that is a superset of another
         * syscall with an augmenter so that we can auto-reuse it.

This:

diff --git a/tools/perf/builtin-trace.c b/tools/perf/builtin-trace.c
index e0434f7dc67cb988..3664bb512c70cabf 100644
--- a/tools/perf/builtin-trace.c
+++ b/tools/perf/builtin-trace.c
@@ -3989,6 +3989,8 @@ static int trace__init_syscalls_bpf_prog_array_maps(struct trace *trace, int e_m
                        goto out;
        }
 
+       return 0;
+
        /*
         * Now lets do a second pass looking for enabled syscalls without
         * an augmenter that have a signature that is a superset of another
⬢ [acme@toolbox perf-tools-next]$ 


Then all works, we don't reuse any BPF program, but then that is an
heuristic anyway, that is tried becuase landlock_add_rule has a pointer
argument:

root@number:~# perf trace -e landlock_add_rule perf test -w landlock
     0.000 ( 0.003 ms): perf/71034 landlock_add_rule(ruleset_fd: 11, rule_type: LANDLOCK_RULE_PATH_BENEATH, rule_attr: 0x7fff6f2bb550, flags: 45) = -1 EINVAL (Invalid argument)
     0.004 ( 0.001 ms): perf/71034 landlock_add_rule(ruleset_fd: 11, rule_type: LANDLOCK_RULE_NET_PORT, rule_attr: 0x7fff6f2bb540, flags: 45) = -1 EINVAL (Invalid argument)
root@number:~# perf test enum
105: perf trace enum augmentation tests                              : Ok
root@number:~#

So its some sort of syncronization on the various new tables, sorted by
name, etc that then when iterating over the syscalls ends up using a sc
that is not initialized.

- Arnaldo