This should have been done a *long* time ago, back when I initially
added Websocket support. This is the first part of two in improving
loop handler support with regards to socket closure.
Reason may include: {normal, shutdown} for the most normal shutdown,
{normal, timeout} for a loop handler timeout shutdown, or {error, _}
if an error occured.
We now always send a failure reason (bad protocol, bad encoding, etc.)
unless the closure was initiated by the client and it didn't send a
close code.
We now check that the close frames have a payload that is valid UTF-8,
unless they don't have a payload at all.
We now do not crash the process anymore when bad opcodes are sent, or
when the opcode 0 is sent before fragmentation was initiated.
Overall this makes us closer to full compliance with the RFC.
Good in theory, but implementations may vary. If something stops
working after this commit we might need some tweaks to support
existing clients.
Please try it and give feedback.
It was only used by Safari 5.0.1 and possibly 5.1. Their market share
is dropping as we speak. It was also insecure (disabled in Firefox
for that reason).
This will allow us to make much more efficient and cleaner code for
the rest of the Websocket versions we support (drafts 7 to 17 + RFC),
which are pretty much all versions seen in the wild excluding the
one we're removing here.
This behavior can be enabled with the `compress` protocol option.
See the `compress_response` example for more details.
All tests are now ran with and without compression for both HTTP
and HTTPS.
This changes the behavior of the `timeout` protocol option to
mean "Time in which the full request line and headers must be
received". The default of 5s should be fine for all normal uses.
This change has no noticeable impact on performance and is thus
enabled by default for everyone. It can be disabled by setting
`timeout` to `infinity` although that is definitely not encouraged.
Inspired by the contribution from @naryl on github.
It was added to help with response body streaming functions.
But it was a clumsy solution that we discarded in favor of
passing socket and transport to said function. It was also
very odd compared to the rest of the cowboy_req interface.
If you used this function before, worry not, here's its
proper equivalent.
[Socket, Transport] = cowboy_req:get([socket, transport], Req)
First and foremost: yes, you can still use Cowboy as a rebar dependency.
This commit only removes the use of rebar when *developing* Cowboy, not
when *using* Cowboy.
Over the past two years I went from very happy with rebar to unsatisfied
and most recently found it counter productive in many ways, from having
insane default configuration to various unefficient operations. The earlier
reversal from 'rebar ct' to 'ct_run' made my workflow much more natural,
as I always needed to look at 'logs/raw.log' to find out what was wrong,
anyway. Why not let 'ct_run' output it directly instead? Removing rebar
made my life easier.
If you wonder why I don't patch rebar, there's two reasons. First is that
the direction taken by rebar isn't compatible with my views, and this
would be a huge fight to steer it in another direction. I got other,
more important fights to make. Second is that I'd rather patch OTP so
that everyone benefits from it, not just users of rebar.
Anyway this isn't my personal blog so I will stop babbling here. There's
a few important things to note relative to this commit:
* You don't need rebar to work on Cowboy anymore
* The eunit tests are now ran through common_test
Ping me if it doesn't work out for you.
