Commit 4c3b2b24 authored by Michal 'vorner' Vaner's avatar Michal 'vorner' Vaner
Browse files

Merge #2671

Clean up docs/design/cc-protocol.txt

Conflicts:
	doc/design/cc-protocol.txt
parents bf86e0c9 5d3ce61a
protocol version 0x536b616e
The CC protocol
===============
DATA 0x01
HASH 0x02
LIST 0x03
NULL 0x04
TYPE_MASK 0x0f
We use our home-grown protocol for IPC between modules. There's a
central daemon routing the messages.
LENGTH_32 0x00
LENGTH_16 0x10
LENGTH_8 0x20
LENGTH_MASK 0xf0
MESSAGE ENCODING
----------------
When decoding, the entire message length must be known. If this is
transmitted over a raw stream such as TCP, this is usually encoded
with a 4-byte length followed by the message itself. If some other
wrapping is used (say as part of a different message structure) the
length of the message must be preserved and included for decoding.
The first 4 bytes of the message is the protocol version encoded
directly as a 4-byte value. Immediately following this is a HASH
element. The length of the hash element is the remainder of the
message after subtracting 4 bytes for the protocol version.
This initial HASH is intended to be used by the message routing system
if one is in use.
ITEM TYPES
Addressing
----------
There are four basic types encoded in this protocol. A simple data
blob (DATA), a tag-value series (HASH), an ordered list (LIST), and
a NULL type (which is used internally to encode DATA types which are
empty and can be used to indicate existance without data in a hash.)
Each item can be of any type, so a hash of hashes and hashes of lists
are typical.
All multi-byte integers which are encoded in binary are in network
byte order.
ITEM ENCODING
-------------
Each item is preceeded by a single byte which describes that item.
This byte contains the item type and item length encoding:
Thing Length Description
---------------- -------- ------------------------------------
TyLen 1 byte Item type and length encoding
Length variable Item data blob length
Item Data variable Item data blob
The TyLen field includes both the item data type and the item's
length. The length bytes are encoded depending on the length of data
portion, and the smallest data encoding type supported should be
used. Note that this length compression is used just for data
compactness. It is wasteful to encode the most common length (8-bit
length) as 4 bytes, so this method allows one byte to be used rather
than 4, three of which are nearly always zero.
HASH
----
This is a tag/value pair where each tag is an opaque unique blob and
the data elements are of any type. Hashes are not encoded in any
specific tag or item order.
The length of the HASH's data area is processed for tag/value pairs
until the entire area is consumed. Running out of data prematurely
indicates an incorrectly encoded message.
The data area consists of repeated items:
Thing Length Description
---------------- -------- ------------------------------------
Tag Length 1 byte The length of the tag.
Tag Variable The tag name
Item Variable Encoded item
The Tag Length field is always one byte, which limits the tag name to
255 bytes maximum. A tag length of zero is invalid.
LIST
----
A LIST is a list of items encoded and decoded in a specific order.
The order is chosen entirely by the source curing encoding.
The length of the LIST's data is consumed by the ITEMs it contains.
Running out of room prematurely indicates an incorrectly encoded
message.
The data area consists of repeated items:
Each connected client gets an unique address, called ``l-name''. A
message can be sent directly to such l-name, if it is known to the
sender.
Thing Length Description
-------------- ------ ----------------------------------------
Item Variable Encoded item
A client may subscribe to a group of communication. A message can be
broadcasted to a whole group instead of a single client. There's also
an instance parameter to addressing, but we didn't find any actual use
for it and it is not used for anything. It is left in the default `*`
for most of our code and should be done so in any new code. It wasn't
priority to remove it yet.
Wire format
-----------
DATA
----
Each message on the wire looks like this:
A DATA item is a simple blob of data. No further processing of this
data is performed by this protocol on these elements.
<message length><header length><header><body>
The data blob is the entire data area. The data area can be 0 or more
bytes long.
The message length is 4-byte unsigned integer in network byte order,
specifying the number of bytes of the rest of the message (eg. header
length, header and body put together).
It is typical to encode integers as strings rather than binary
integers. However, so long as both sender and recipient agree on the
format of the data blob itself, any blob encoding may be used.
The header length is 2-byte unsigned integer in network byte order,
specifying the length of the header.
The header is a string representation of single JSON object. It
specifies the type of message and routing information.
NULL
----
The body is the payload of the message. It takes the whole rest of
size of the message (so its length is message length - 2 - header
length). The content is not examined by the routing daemon, but the
clients expect it to be valid JSON object.
This data element indicates no data is actually present. This can be
used to indicate that a tag is present in a HASH but no data is
actually at that location, or in a LIST to indicate empty item
positions.
The body may be empty in case the message is not to be routed to
client, but it is instruction for the routing daemon. See message
types below.
There is no data portion of this type, and the encoded length is
ignored and is always zero.
The message is sent in this format to the routing daemon, the daemon
optionally modifies the headers and delivers it in the same format to
the recipient(s).
Note that this is different than a DATA element with a zero length.
The headers
-----------
The header object can contain following information:
EXAMPLE
-------
This is Ruby syntax, but should be clear enough for anyone to read.
Example data encoding:
{
"from" => "sender@host",
"to" => "recipient@host",
"seq" => 1234,
"data" => {
"list" => [ 1, 2, nil, "this" ],
"description" => "Fun for all",
},
}
Wire-format:
In this format, strings are not shown in hex, but are included "like
this." Descriptions are written (like this.)
Message Length: 0x64 (100 bytes)
Protocol Version: 0x53 0x6b 0x61 0x6e
(remaining length: 96 bytes)
0x04 "from" 0x21 0x0b "sender@host"
0x02 "to" 0x21 0x0e "recipient@host"
0x03 "seq" 0x21 0x04 "1234"
0x04 "data" 0x22
0x04 "list" 0x23
0x21 0x01 "1"
0x21 0x01 "2"
0x04
0x21 0x04 "this"
0x0b "description" 0x0b "Fun for all"
MESSAGE ROUTING
---------------
The message routing daemon uses the top-level hash to contain routing
instructions and additional control data. Not all of these are
required for various control message types; see the individual
descriptions for more information.
Tag Description
------- ----------------------------------------
msg Sender-supplied data
from sender's identity
group Group name this message is being sent to
instance Instance in this group
repl if present, this message is a reply.
seq sequence number, used in replies
to recipient or "*" for no specific receiver
type "send" for a channel message
"type" is a DATA element, which indicates to the message routing
system what the purpose of this message is.
|====================================================================================================
|Name |type |Description
|====================================================================================================
|from |string|Sender's l-name
|type |string|Type of the message. The routed message is "send".
|group |string|The group to deliver to.
|instance |string|Instance in the group. Purpose lost in history. Defaults to "*".
|to |string|Override recipient (group/instance ignored).
|seq |int |Tracking number of the message.
|reply |int |If present, contains a seq number of message this is a reply to.
|want_answer|bool |If present and true, the daemon generates error if there's no matching recipient.
|====================================================================================================
Types of messages
-----------------
Get Local Name (type "getlname")
--------------------------------
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Upon connection, this is the first message to be sent to the control
daemon. It will return the local name of this client. Each
connection gets its own unique local name, and local names are never
repeated. They should be considered opaque strings, in a format
useful only to the message routing system. They are used in replies
or to send to a specific destination.
Upon connection, this is the first message to be sent to the daemon.
It will return the local name of this client. Each connection gets
its own unique local name, and local names are never repeated. They
should be considered opaque strings, in a format useful only to the
message routing system. They are used in replies or to send to a
specific destination.
To request the local name, the only element included is the
"type" => "getlname"
{"type": "getlname"}
tuple. The response is also a simple, single tuple:
"lname" => "UTF-8 encoded local name blob"
{"lname" => "Opaque utf-8 string"}
Until this message is sent, no other types of messages may be sent on
this connection.
Regular Group Messages (type "send")
------------------------------------
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
When sending a message:
Message routed to other client. This one expects the body to be
non-empty.
"msg" is the sender supplied data. It is encoded as per its type.
It is a required field, but may be the NULL type if not needed.
In OpenReg, this was another wire format message, stored as an
ITEM_DATA. This was done to make it easy to decode the routing
information without having to decode arbitrary application-supplied
data, but rather treat this application data as an opaque blob.
Expected headers are:
"from" is a DATA element, and its value is a UTF-8 encoded sender
identity. It MUST be the "local name" supplied by the message
routing system upon connection. The message routing system will
enforce this, but will not add it. It is a required field.
* from
* group
* instance (set to "*" if no specific instance desired)
* seq (should be unique for the sender)
* to (set to "*" if not directed to specific client)
* reply (optional, only if it is reply)
* want_answer (optional, only when not a reply)
"group" is a DATA element, and its value is the UTF-8 encoded group
name this message is being transmitted to. It is a required field for
all messages of type "send".
A client does not see its own transmissions.
"instance" is a DATA element, and its value is the UTF-8 encoded
instance name, with "*" meaning all instances.
Group Subscriptions (type "subscribe")
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
"repl" is the sequence number being replied to, if this is a reply.
Indicates the sender wants to be included in the given group.
"seq" is a unique identity per client. That is, the <lname, seq>
tuple must be unique over the lifetime of the connection, or at least
over the lifetime of the expected reply duration.
Expected headers are:
"to" is a DATA element, and its value is a UTF-8 encoded recipient
identity. This must be a specific recipient name or "*" to indicate
"all listeners on this channel." It is a required field.
* group
* instance (leave at "*" for default)
When a message of type "send" is received by the client, all the data
is used as above. This indicates a message of the given type was
received.
There is no response to this message and the client is subscribed to
the given group and instance.
A client does not see its own transmissions. (XXXMLG Need to check this)
The group can be any utf-8 string and the group doesn't have to exist
before (it is created when at least one client is in it). A client may
be subscribed in multiple groups.
Group Unsubscribe (type "unsubscribe")
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Group Subscriptions (type "subscribe")
--------------------------------------
The headers to be included are "group" and "instance" and have the same
meaning as a "subscribe" message. Only, the client is removed from the
group.
A subscription requires the "group", "instance", and a flag to
indicate the subscription type ("subtype"). If instance is "*" the
instance name will be ignored when deciding to forward a message to
this client or not.
Transmitted messages
--------------------
"subtype" is a DATA element, and contains "normal" for normal channel
subscriptions, "meonly" for only those messages on a channel with the
recipient specified exactly as the local name, or "promisc" to receive
all channel messages regardless of other filters. As its name
implies, "normal" is for typical subscriptions, and "promisc" is
intended for channel message debugging.
These are the messages generally transmitted in the body of the
message.
There is no response to this message.
Command
~~~~~~~
It is a command from one process to another, to do something or send
some information. It is identified by a name and can optionally have
parameters. It'd look like this:
Group Unsubscribe (type "unsubscribe")
-------------------------------
{"command": ["name", <parameters>]}
The parameters may be omitted (then the array is 1 element long). If
present, it may be any JSON element. However, the most usual is an
object with named parameter values.
It is usually transmitted with the `want_answer` header turned on to
cope with the situation the remote end doesn't exist, and sent to a
group (eg. `to` with value of `*`).
Success reply
~~~~~~~~~~~~~
When the command is successful, the other side answers by a reply of
the following format:
{"result": [0, <result>]}
The result is the return value of the command. It may be any JSON
element and it may be omitted (for the case of ``void'' function).
The fields to be included are "group" and "instance" and have the same
meaning as a "subscribe" message.
This is transmitted with the `reply` header set to the `seq` number of
the original command. It is sent with the `to` header set.
There is no response to this message.
Error reply
~~~~~~~~~~~
In case something goes wrong, an error reply is sent. This is similar
as throwing an exception from local function. The format is similar:
Statistics (type "stats")
-------------------------
{"result": [ecode, "Error description"]}
Request statistics from the message router. No other fields are
inclued in the request.
The `ecode` is non-zero error code. Most of the current code uses `1`
for all errors. The string after that is mandatory and must contain a
human-readable description of the error.
The response contains a single element "stats" which is an opaque
element. This is used mostly for debugging, and its format is
specific to the message router. In general, some method to simply
dump raw messages would produce something useful during debugging.
The negative error codes are reserved for errors from the daemon.
Currently, only `-1` is used and it is generated when a message with
`reply` not included is sent, it has the `want_answer` header set to
`true` and there's no recipient to deliver the message to. This
usually means a command was sent to a non-existent recipient.
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment