bpf: raw tracepoint args v6

in assembler it looks like:
(gdb) disassemble __bpf_trace_xdp_exception
Dump of assembler code for function __bpf_trace_xdp_exception:
   0xffffffff81132080 <+0>:     mov    %ecx,%ecx
   0xffffffff81132082 <+2>:     jmpq   0xffffffff811231f0 <bpf_trace_run3>

where

TRACE_EVENT(xdp_exception,
        TP_PROTO(const struct net_device *dev,
                 const struct bpf_prog *xdp, u32 act),

The above assembler snippet is casting 32-bit 'act' field into 'u64'
to pass into bpf_trace_run3() and all other registers are passed as-is.
So all of ~500 of __bpf_trace_*() functions are only 5-10 _byte_ long
and in total this approach adds 7k bytes to .text and 8k bytes
to .rodata since I want them to appear in kallsyms, so I don't
have to add another 8-byte fields to 'struct trace_event_class'
and 'struct trace_event_call'

Such approach, I think, is the lowest overhead we can possibly have
while calling trace_xdp_exception() from kernel C code and
transitioning into bpf land.
Since many folks are starting to use tracepoint+bpf at speeds of
1M+ events per second this is very valuable optimization.

Since I'm not touching anything on ftrace or perf side,
I'm thinking to extend BPF_PROG_ATTACH command to specify
which tracepoint to attach to.
The user space will look like:

prog_fd = bpf_prog_load(...);
bpf_prog_attach(prog_fd, "xdp_exception");

On the kernel side I will walk kallsyms to
find __tracepoint_xdp_exception record, check that
tp->name == "xdp_exception",
then find __bpf_trace_xdp_exception() address (also in kallsyms),
and finally use tracepoint_probe_register() to connect the two.

Signed-off-by: Alexei Starovoitov <ast@kernel.org>