Fix trace.py USDT argument filtering

[palmtenor]

@goldshtn fixed general USDT support in trace.py in #698 . However, filtering is still not able to function since we are not reading the arguments correctly.

This Diffs replaces argX variables in filtering conditions to argX_filter, and use the bpf_readarg_N macros to read argument values into argX_filter variables.

In order to define the variables correctly, this Diff also exposes the ability to get USDT argument type (i.e., size) in Python USDT class. I think this would be useful elsewhere as well.

Tested by defining an USDT probe with two parameters in my test binary and run

trace.py 'u:pathToBinary:theProbeName (arg1!=0) "Arg1 %lld Arg2 %s", arg1, arg2'

Generated BPF program looks like:

int probe_contextSwitch_1(struct pt_regs *ctx)
{

        u32 __pid = bpf_get_current_pid_tgid();
        if (__pid == 3787616) { return 0; }

        uint64_t arg1_filter;
        bpf_usdt_readarg(1, ctx, &arg1_filter);

        if (!((arg1_filter!=0))) return 0;

        struct probe_contextSwitch_1_data_t __data = {0};
        __data.timestamp_ns = bpf_ktime_get_ns();
        __data.pid = bpf_get_current_pid_tgid();
        bpf_get_current_comm(&__data.comm, sizeof(__data.comm));
        u64 arg1;
        bpf_usdt_readarg(1, ctx, &arg1);
        __data.v0 = (long long)arg1;
        u64 arg2;
        bpf_usdt_readarg(2, ctx, &arg2);

        if (arg2 != 0) {
                bpf_probe_read(&__data.v1, sizeof(__data.v1), (void *)arg2);
        }

        probe_contextSwitch_1_events.perf_submit(ctx, &__data, sizeof(__data));
        return 0;
}

[4ast]

Teng, Thanks for the fix!
@goldshtn please review.

[goldshtn]

Thanks, this looks good. I totally forgot about the filters, and it's really great that you were able to fix them cleanly. Getting the argument type for each argument will be useful for @4ast 's request to print this information in -v -v mode.

[goldshtn]

Why not just check arg[0:3] == "arg" as we do elsewhere? It's just that depending on Probe.aliases is a bit fragile, because the arg aliases are not really used for USDT probes and might change or move at some point.

[palmtenor]

Ideally we should do this. The issue here is for filters, unlike the return (printf) values, we don't actually parse them. We just take what's in between brackets in the CLI argument and put it in the if statement directly, with some direct find-and-replace on variable names (as you can see here). We would have to parse the filter expression in order to get type and expr as we do for printf values:(

[goldshtn]

Right, forgot about that. The filter is indeed quite unstructured. Good to go then.

[goldshtn]

Maybe "Unable to determine type of {} " to be a bit clearer. (Add "of".)

[palmtenor]

Done:)