GitLab

We had a very nasty problem: handshakes were serial so incoming
dials would wait for each other to finish handshaking. this was
particularly problematic when handshakes hung-- nodes would not
recover quickly. This led to gateways not bootstrapping peers
fast enough.

The approach taken here is to do what crypto/tls does:
defer the handshake until Read/Write[1]. There are a number of
reasons why this is _the right thing to do_:
- it delays handshaking until it is known to be necessary (doing io)
- it "accepts" before the handshake, getting the handshake out of the
  critical path entirely.
- it defers to the user's parallelization of conn handling. users
  must implement this in some way already so use that, instead of
  picking constants surely to be wrong (how many handshakes to run
  in parallel?)

[0] http://golang.org/src/crypto/tls/conn.go#L886

We now consider debugerrors harmful: we've run into cases where
debugerror.Wrap() hid valuable error information (err == io.EOF?).
I've removed them from the main code, but left them in some tests.
Go errors are lacking, but unfortunately, this isn't the solution.

It is possible that debugerros.New or debugerrors.Errorf should
remain still (i.e. only remove debugerrors.Wrap) but we don't use
these errors often enough to keep.

reuseport is a hack. It is necessary for us to do certain kinds of
tcp nat traversal. Ideally, reuseport would be available in go:

  https://github.com/golang/go/issues/9661

But until that issue is fixed, we're stuck with this. In some cases,
reuseport is strictly a detriment: nodes are not NATed. This commit
introduces an ENV var IPFS_REUSEPORT that can be set to false to
avoid using reuseport entirely:

  IPFS_REUSEPORT=false ipfs daemon

This approach addresses our current need. It could become a config
var if necessary. If reuseport continues to give problems, we should
look into improving it.

humanize bandwidth output

instrument conn.Conn for bandwidth metrics

add poll command for continuous bandwidth reporting

move bandwidth tracking onto multiaddr net connections

another mild refactor of recording locations

address concerns from PR

lower mock nodes in race test due to increased goroutines per connection

- Modified Godeps/Godeps.json by hand
- [TEST] Updated welcome docs hash to sharness
- [TEST] Updated contact doc
- [TEST] disabled breaking test (t0080-repo refs local)

- updated go-ctxgroup and goprocess
ctxgroup: AddChildGroup was changed to AddChild. Used in two files:
- p2p/net/mock/mock_net.go
- routing/dht/dht.go

- updated context from hg repo to git
prev. commit in hg was ad01a6fcc8a19d3a4478c836895ffe883bd2ceab. (context: make parentCancelCtx iterative)
represents commit 84f8955a887232b6308d79c68b8db44f64df455c in git repo

- updated context to master (b6fdb7d8a4ccefede406f8fe0f017fb58265054c)

Aaron Jacobs (2):
net/context: Don't accept a context in the DoSomethingSlow example.
context: Be clear that users must cancel the result of WithCancel.

Andrew Gerrand (1):
go.net: use golang.org/x/... import paths

Bryan C. Mills (1):
net/context: Don't leak goroutines in Done example.

Damien Neil (1):
context: fix removal of cancelled timer contexts from parent

David Symonds (2):
context: Fix WithValue example code.
net: add import comments.

Sameer Ajmani (1):
context: fix TestAllocs to account for ints in interfaces

We were half-way with this. there's no way for net.Dialers
to respect contexts, so we have to let the dial finish in
the background.

thanks @whyrusleeping for finding these.

Use of the ratelimiter should be conscious of the ratelimiter's
potential closing. any loops that add work to ratelimiter
should (a) only do so if the rate limiter is not closed,
or (b) prevent limiter while work is added
(i.e. use limiter.Go(addWorkHere))

Let's save log.Error for things the user can take action on.
Moved all our diagnostics to log.Debug. We can ideally reduce them
even further.

so that we dont get stuck dialing bad addrs first always.

This will mitigate the fd explosion, but slow down dials majorly
as any peer with more addresses than the rate limit will have
to wait a whole dial timeout (~15s)

This addr manager should seriously help with the addrsplosion
problem.

Network now signals when it successfully listens on some address
or when an address shuts down. This will be used to establish and
close nat port mappings. It could also be used to notify peers
of address changes.

network interface now allows setting Listeners after the fact.
This is useful to create the network and start listening as
separate steps. And to keep the network up to date on new
addresses the node might have to listen to.

network interface now allows setting Listeners after the fact.
This is useful to create the network and start listening as
separate steps. And to keep the network up to date on new
addresses the node might have to listen to.

dialing 4000 connections somehow keeps choking both travis and
jenkins. dialing this down to 500

- attempts set to 1 now.
- timeouts must account for that.

+ fixed race

This commit turns all dial logs into log.Events.

Everything's great except for one problem:
The LoggableMap I'm using does not print out things
correctly. I gave it peer.IDs, and Multiaddrs
and both got logged as nothing `{}` (didn't even call
their String() methods!) So, for now, this function
encodes it when called... This is wrong and should be
fixed before being merged in. Otherwise we  will be
constantly encoding peer.IDs and Multiaddrs without
needing to.

@briantigerchow how do you suggest doing this?
I don't know my way around your Loggable.

should be returning false.

increased the timeout for other CI too