GitLab

Given 1k requests per second (assuming one stream per request), we can
easily loop around the stream ID after less than 2 months. 32bits is
plenty (usually) for connection-scoped stream IDs because individual
connections don't usually last that long, but isn't enough for a
_global_ stream ID.

Given that there's no reason for these to be 32bit IDs, let's just make
them 64bits.

* support for forced direct connections.

Co-authored-by: Aarsh Shah <aarshkshah1992@gmail.com>
Co-authored-by: Raúl Kripalani <raul@protocol.ai>

* Rank dial addresss.

Otherwise, we can modify the context after/while the process is shutting down.

fixes #189

only cancel dials when an outbound connection succeeds. this may result in duplicate connections, but it's better to have two connections than dropping both of them and ending up with zero connections.

1. Always return the caller's context error if relevant.
2. Don't return "context canceled" when we're just shutting down.
3. Don't claim that the context deadline has been exceeded when the dial timeout
is canceled.

Avoid logging about closed listeners, etc., when shutting down.

License: MIT
Signed-off-by: Jakub Sztandera <kubuxu@protonmail.ch>

This can be quite an overhead in cases of high connection rates.
The main overhead is thread blocking syscall causing a lot of context
switching.

License: MIT
Signed-off-by: Jakub Sztandera <kubuxu@protonmail.ch>

This reverts commit 3dded3dc.

The global per-peer dial timeout had a significant drawback: When dialing many
peers, this timeout could cause libp2p to cancel dials while they were still
stuck in the limiter.

A better but more complicated approach is a time budget system but we can
implement that later.

This change simply applies the limit to each `DialPeer`/`NewStream` call
independently and makes it easy to override. While old timeout tried to
account for how much we're willing to spend dialing a single peer, this new
timeout tries to account for the amount of time a single "client" is willing to
wait for a dial to complete before they no longer care.

Removes:

* go-libp2p-conn logic (moved to transports)
* go-libp2p-peerstream (moved here)

Changes:

* New transport interface.
* Explicit over implicit (the implicit automagic now all lives in go-libp2p):
  * No more default transports, muxers, etc.
  * No more fallback dialer. Transports are now *required*.
* Stream opening:
  * Connection picking logic (instead of just picking the first).
  * Tries harder to open a stream of some connections happen to be closed.
* Stream closing:
  * No longer treats half-closed streams as fully closed. Users *must* read the
    an EOF or reset the stream for it to be garbage collected.
* No more polling for dead connections.
* Multiplexers are now *non-optional*. Really, they haven't been optional for a
  while but we still pretended that they were.
* No more Network type alias. It added a bunch of code and didn't really provide
  anything but an alternative set of methods that do the same thing.
* Notifications:
  * New guarantee: connection open notifications will complete before connection
    close notifications begin.
  * Given that, notifications are now delivered in parallel. No more
    notification backlogs blocking connection closing/opening.

We call this method in `FindProvidersAsync`. Really, we shouldn't even be doing
that (but this is "good enough" for now).

It doesn't exist anymore. This is now handled entirely though contexts.

* Add in the reset method and use it instead of Close. Close now only closes one
  side of the connection.