Merge bug24702

Merge bug24500

Defaults all centralised and isolated.

Merge bug22961 in (a dependency)

Add description

refactor (Matthias's suggestions)

Merge bug24606

drop messages in dl cycles

This is the fate of messages in cycles:
=WARNING REPORT==== 3-Feb-2012::12:54:34 ===
Message dropped. Dead-letter queues cycle detected: [<<"foo">>,<<"foo">>]

Umm, and those.

Build-Depend on erlang-nox and erlang-src.

merge default into bug24624 and use new rabbit_misc:format/2

refactor: extract string formatting function

Move cookie_hash/0 too.

Move rabbit_misc:nodeparts/1 and rabbit_misc:makenode/1 into rabbit_nodes

"to contact" isn't correct in the duplicate check case.

Unify the more resilient hostname check, refactor out this code into a new module.

Update copyright 2012

Clean up diagnostics code, clean up diagnostics messages.

Remove nodes file immediately after starting Mnesia.

Oops

Merge in default

Merge default in.

include some mirror master state info in status()
After careful inspection, I have determined that known_senders and
seen_status are the only items worth reporting on.

Quick hack at rabbitmqctl force_cluster_forget. Appears to work.

Committed some changes by mistake...

Cosmetic, restored reject/4 arguments to the previous order.

Merge heads

Merged bug24713 into default

merge bug24703 into default

Merge default.

Typo

simplify

Revert comment change

OTOH if we're not trying to cluster with anything we should probably be happy.

But Matthias points out that we can avoid the three-way-ness by changing the meaning of non-forced clustering so that you must only have to contact *one* disc node. This seems to work.

Reintroduce the three levels of forcedness. We do still need it.

Inline.

This doesn't start any more.

Base discnodeness on what we want to become, not what we are.

Change the semantics of Force to fit this particular bug. *meep*.

Don't claim to be adding mirror when we're not.

We don't actually need this two-stage startup, we can just fail to init/1.

Let's pay attention to whether we actually started or not.

Merge default in

Rebase 038bd3c185ae

Resurrect and remove redundant ensure_mnesia_running()

Junk bug24703

Two-stage-startup-with-mnesia-record, like rabbit_amqqueue_process.

Merged bug24671 into default

Better comments regarding gen/0, explicit range for phash/2.

Merge default.

Better comment for gen/0.

Update comment for gen/0.

Merge master in.

Change fash guid generation function to use phash + rotating blocks.

Merge bug22470

merge bug24715 into default

cosmetic: ensure_gm_heartbeat is not a gm callback

use same variable naming convention as in channel

don't leak queue monitors

Merge bug24718

silence epmd start

let erlang start epmd on windows

Backed out changeset 6bdb3b57d0c5

Renamede reject_tx/4 to reject/4.

try to start epmd as a background job

it helps to not throw away the state
it really does

Modify generic-unix packaging step to override rabbitmq-sys defaults.

Merge default in

Move defaults to rabbitmq-sys

Don't use timer:apply_interval/4.

It's a resource, it'll print 'queue' anyway.

merge default into bug24494

produce only one log entry for abnormal connection closure
instead of two

check that dlx is defined is dl-rk is set

Merged bug24664 into default

Merge bug24601 into default

Remove redundant template from html-to-website-xml.xsl

merge bug23706 into default

cosmetic

consistency

validate log levels on startup

merge default into bug22470

get rid of debug and message logging
since they have limited use
This does make the rabbit_log server somewhat superfluous, but it does
get a new lease of life in bug 22470

introduce log categories
thus opening pandora's box
this turned out to be simplest way of controlling the connection logging

Merge in default (and make error more readable)

prettier and more consistent connection log messages

move error_logger 'flow control' into application code
since tcp_acceptor no longer does any per-connection logging itself

do not log accepted connections in tcp_acceptor
Logging in such generic code should be very sparse. Let's leave that
to application code (rabbit_reader et al in case of rabbit) instead.

drop ssl upgrade message
it's not that important and just clutters the log

This is never used.

Finally: test that a ram node will not start by itself. You know - the point of the bug.

Test the reset-when-not-all-disc-nodes-are-up thing.

Propagate further.

Change CONFIG_FILE, LOG_BASE, MNESIA_BASE and ENABLED_PLUGINS_FILE defaults
- CONFIG_FILE to ${RABBIT_HOME}/rabbitmq
- LOG_BASE to ${RABBIT_HOME}/log
- MNESIA_BASE to ${RABBIT_HOME}/mnesia
- ENABLED_PLUGINS_FILE to ${RABBIT_HOME}/enabled_plugins

Merge bug24706 in

Move non-empty defaults to rabbitmq-env
- rationalise server-settings
- correctly override PLUGINS_DIR
- allow use of PID_FILE override
- simplify NODE_IP_ADDRESS and NODE_PORT setting

Delete extra whitespace.

Merge bug24435

Alignment

I think this should be a soft error. Although the spec does say to use command-invalid when the client attempts to declare an exchange of an unknown type, that's much more obviously the client's fault. So let's go with an old favourite.

Tweak language

Cons instead of append, reject does not change the channel state.

'case' statement style

Better error messages and style fixes.

reject doesn't need to know about the channel state.

Fix to reject_tx, list comprehension instead of map.

Flatten the list of lists.

check dlx routing key arg

overwrite malformed x-death headers

don't discard accumulated changes

use POSIX timestamps instead of human readable ones

merge default into bug20337

Remove ?CREDIT_CPU_BOUND and have /1 versions of ack and send instead.

Huh

Fix various tests

Executing basick.{nack,reject} transactionally.

Fail to start if no disc nodes are up (bug 24703), and fail to reset if any disc nodes are down (see bug 24376).

...however, we should still allow rabbitmqctl reset to work when some nodes are down. (This is what we did before, the difference between reset and force_reset is that force_reset works if mnesia is completely hosed).

Don't force on init() (which fixes bug 24703), plus a little refactor.

Forgot to add/remove files.

Change dummy exchange name from "inert" to "invalid"

Duplicate "exchange" in error message

Using rabbit_misc:rs/1 to format the exchange name.

Protocol error instead of warning when routing through inert exchange.

I originally raised protocol errors in bind/unbind as well but that defies
the purpose of this fixes since if a queue is bound to an exchange with invalid
type the boot sequence fails since it tries to re-bind the queue.

Dropping messages silently instead of emitting warnings.

Matthias noted that emitting a warning for every message that can easily clog
the logger.

Add stub for "inert" exchange to be used when exchange types are missing.

Right now it just emits warnings on startup and when a message is routed
through the exchange.

I need to handle the "add_binding" and "remove_bindings" better, probably
emitting a protocol error.

Also, I'm not sure the log warning are visible enough.

Show both log files.

Further clarity.

Improve display of boot errors.

Merge bug24653

refactor: use types exported by 'supervisor'
instead of defining our own

helper function for constructing tcp listener sup child specs
useful in e.g. stomp

use the correct inet types

use correct type for {M,F,A}
the built-in mfa() type stands for module-function-*arity*

refactor: check_tcp_listener_address/2 -> tcp_listener_addresses/1
...which is a more sensible name and signature

refactor: turn a condition into a guard

cosmetic: move helper functions where they belong

handle the remaining enabled_plugins_file error case more gracefully

don't complain about empty enabled_plugins file

Merged bug23432 into default

display posix error description

Improvement?

Don't fsync after inner nested transactions

More doc.

clean junk

Junk default (bug 24500 broke erlang client)

Resurrect default

5df9b13693d6 broke rabbitmqctl reset. Revert.

More credit for the queue -> msg_store case.

Conserve precious boot steps.

Remove underscores from the boot screen. How horrible!

Merge bug24500

Be more assertive

Remove doc:div and doc:toc, add show-in-this-page and reindent. Really this should have been in bug 23052. Ah well.

Merge bug23938

correct typo

Language tweak for mnesia_sync justification

Brief explanation justifying the existence of mnesia_sync

Merged bug24641 into default.

HiPE credit_flow - this gets MCM -x 1 -y 1 -i 5 from 29.5kHz to 30.8kHz.

For direct client we *do* want the reader pid (amqp_channel) rather than the conn_pid (amqp_direct_connection) here.

Correct comment and use spawn_link

Ignore bump_credit messages in rabbit_tests.

merge heads

two flavours of 'write'
Just like in the other applications of credit_flow:send/ack, we need
two flavours of the surrounding API function, i.e. 'write' and
'write_flow' in this case. That way existing call sites, such as in
rabbit_test can remain undisturbed.

more sensible arg order
and consistent var naming

Same fix as c412171663cc for rabbitmq-plugins

Flow control: queue slave -> msg_store.

Don't pass the client pid in for write.

Flow control: vq -> msg_store.

Change the guid related entries names in the process dict.

Change naming in rabbit_guid, explicit guid passing in string/0 and binary/0.

Specifically:
fast_guid() -> gen()
guid() -> secure_gen()
string_guid(Prefix) -> string(Guid, Prefix)
binstring_guid(Prefix) -> binary(Guid, Prefix)

Split guid/0 in guid/0 and the faster fast_guid/0.

The first function works like the old one (hashes the {serial, counter} each
time.

The second function (fast_guid/0) hashes the serial once, and the XORs it with
the counter each time a new guid is needed. Thus, it is more predictable but
should not produce less unique guids.

If the fast function is used when generating ids for messages, it leads to a
considerable speedup when testing with BroadcastMain (from 67k msg/s to
75k msg/s).

merge bug 24265 into default (Somehow support DNs in rabbitmq_auth_mechanism_ssl)

clearly this was tested thoroughly

make flow control explicit in the channel API
just as we did for the queue
As a result the existing call sites of rabbit_channel:do, e.g. in
rabbit_test and the erlang client, can be left alone.

Merge bug 24638

Merge bug19336 to junk

simplify publisher monitoring code

only monitor publishers when they publish with flow control

Merge bug24676

Merge bug24674

Merge bug21453 into junk

fix supervisor2 bug reported by dialyzer
The erroneous code is invoked when a child terminates normally at the
same time as it is being terminated by the supervisor. So
unsurprisingly we haven't seen this, and it's hard to test.

provide two flavours of rabbit_amqqueue:deliver
- one with flow control and one without
This allows us to leave the various existing call sites of
rabbit_amqqueue:deliver unchanged, and only subject the channel to
flow control. It also reduces the distance between the two places that
need to know which messages should be subject to flow control by
moving the rabbit_flow invocation logic from the channel into
rabbit_amqqueue.

merge default into bug24664

Merge bug24681

Remove unnecessary process

merge rabbit_router:deliver into rabbit_amqqueue:deliver

it's really not worthwhile exporting process_channel_frame

do not hard-wire the channel interaction into process_channel_frame
This makes that function more versatile, e.g. permitting the Erlang
client to skip the credit_flow:send step.

Reinstate a bit of code that accidentally got nuked in 9a87718c74df.

sender_down / receiver_down -> peer_down.

Merge heads of bug 24664 again.

rabbit_flow -> credit_flow

cosmetic(ish)

Revert b03fe604af24. Not enough sleep.

Merge heads of bug 24664

Move rabbit_flow:send invocation from router into channel; monitor all queues in the channel.