ebpf connection tracker: perf map fixes

[alban]

This adds fixes in the perf map that receives the tcp events from the ebpf probes:

• gobpf: use the correct timestamp function instead of assuming the timestamp is at offset zero in the struct
• gobpf: fix loop on events that could drop events
• gobpf: add a go channel to notify lost events
• gobpf: make the size of the ring buffers configurable instead of using 8 page (32KB)
• tcptracer-bpf: change the size of the ring buffer "maps/tcp_event_ipv4" to 256 pages (1MB)
• tcptracer-bpf: use the go channel to notify lost events
• scope: add a callback on lost events: log the error and stop the ebpf tracker correctly. So if it ever happens, we will know about it.

This is to address the address the following failure that happened a few weeks ago:

<probe> ERRO: 2017/04/13 07:34:17.299127 tcp tracer received event with timestamp 677884430530033 even though the last timestamp was 677884430530738. Stopping the eBPF tracker.
<probe> WARN: 2017/04/13 07:34:17.337241 ebpf tracker died, gently falling back to proc scanning

---

This is a "work in progress" because this is vendoring unmerged branches in gobpf and tcptracer-bpf.

• elf: perf ordering fixes iovisor/gobpf#43
• perf maps: add configurable parameters iovisor/gobpf#44
• perf map fixes tcptracer-bpf#37

---

I tried to reduce the size of the perf ring buffers to one page (4KB) in order to test it. I made lots of connections in parallel in 4 terminals (but on the same cpu to stress one perf ring buffer):

for i in $(seq 1 10000) ; do echo -n "$i " ; taskset --cpu-list 2 wget -O /dev/null http://172.17.0.2 2>/dev/null; done

Then, I checked that the fallback worked correctly:

<probe> ERRO: 2017/05/09 09:17:40.798823 tcp tracer lost 7 events. Stopping the eBPF tracker
<probe> WARN: 2017/05/09 09:17:41.701317 ebpf tracker died, gently falling back to proc scanning

---

/cc @iaguis @2opremio

[alban]

I rebased and revendored gobpf + tcptracer-bpf since PRs have been merged.

I'm keeping the "WIP" title because I'd like to test again to make sure nothing was wrong in the rebase/revendor/git-subtrees changes.

[alban]

I tried to quickly create plenty of connections to nginx on several cpus with this script (in several terminals):

cpu=0
for i in $(seq 1 10000) ; do
  echo -n "$i "
  taskset --cpu-list $cpu wget -O /dev/null http://172.17.0.2 2>/dev/null
done
echo

And I didn't notice any problems in the logs.

However, chrome quickly took all the cpu and crashed:

[2opremio]

Mind creating a separate ticket for this?

[2opremio]

(It's not a real scenario, but it's worth keeping track of it)

[alban]

Mind creating a separate ticket for this?

#2517

[alban]

@2opremio PTAL