Differences between revisions 57 and 58
Revision 57 as of 2013-08-19 18:15:07
Size: 11402
Editor: mpm
Comment:
Revision 58 as of 2014-03-07 11:11:48
Size: 11437
Editor: IdanKamara
Comment: remove known issue
Deletions are marked like this. Additions are marked like this.
Line 201: Line 201:
 * server needs a repository to start  * --(server needs a repository to start)-- (fixed in Cset:e811b93f2cb1)

Command Server

A server that allows communication with Mercurial's API over a pipe.

1. Problem statement

Mercurial presents several barriers for third-party applications wishing to automate interaction:

  • It is licensed under the GPL, so third-party tools using its internal APIs directly must also be GPL
  • It is written in Python, which makes it difficult to interface via other languages
  • It does not make any stability guarantees for its internal API
  • Documentation is fairly minimal

The usual answer to these problems is to use its command line API which is:

  • Language-neutral
  • Guaranteed stable
  • Well-documented
  • No licensing issues

The two primary downsides of interfacing with the command line directly are:

  • Significant performance overhead for launching Mercurial repeatedly
  • Text parsing required

2. Licensing

The Mercurial developers specifically designed this command server with the intention that users of it could write clients to the command server and release clients that use the existing command server under licenses of their choosing. However, if you modify Mercurial to export new functionality via the command server, that adds obligations for you under GPL. If you have questions about this, please contact us.

3. Command server approach

The goal of the command server is to facilitate the creation of wrapper libraries that are:

  • Friendly to a variety of languages (see libraries below)

  • Guaranteed stable
  • Low-overhead
  • Conveniently licensed

This is done by allowing third-party applications and libraries to communicate with Mercurial over a pipe that eliminates the per-command start-up overhead. Libraries can then encapsulate the command generation and parsing to present a language-appropriate API to these commands. This strategy is similar to how applications typically communicate with SQL servers.

The command server is available with Mercurial versions 1.9 or higher.

4. Protocol

All communication with the server is done on stdin/stdout. The byte order used by the server is big-endian.

Data sent from the server is channel based, meaning a (channel, length) pair is sent before the actual data. The channel is a single character, while the length is an unsigned int (4 bytes). In the examples below, the length field is in plain text.

o
1234
<data: 1234 bytes>

that is 1234 bytes sent on channel 'o', with 1234 bytes of data following.

When starting the server, it will send a hello message on the 'o' channel. The message is sent as one chunk. It is composed of a \n separated list where each item is of the format:

<field name>: <field data>

<field name> is limited to [a-z0-9]+, and <field data> is field specific (cannot contain new lines).

Known fields are:

  • capabilities: a space separated list of strings representing what commands the server supports.
  • encoding: a string, the servers encoding.

capabilities: capability1 capability2 ... capabilityN\n
encoding: UTF-8

At the most basic level, the server will support the 'runcommand' capability.

Clients should ignore unknown fields in the hello message, in case a new version of the server decides to update it with some important information.

More on channels below.

4.1. Encoding

Strings are encoded by default in Mercurial's local encoding. At the moment the encoding cannot be changed after server startup. To set it at startup, use HGENCODING. To query the server's encoding, see the 'getencoding' command.

Clients wanting to use Unicode should specify a UTF-8 encoding, but be aware that some responses will mix UTF-8 metadata and raw file contents. See EncodingStrategy for more information.

4.2. Channels

Channels are divided into two, required and optional. Required channels identifiers are uppercase. They cannot be ignored. If a client encounters an unexpected required channel, it should abort.

Optional channels identifiers are lowercase, and their data can be ignored.

Optional:

  • 'o'utput channel: most of the communication happens on this channel. When running commands, output Mercurial writes to stdout is written to this channel.
  • 'e'rror channel: when running commands, this correlates to stderr.
  • 'r'esult channel: the server uses this channel to tell the client that a command finished by writing its return value (command specific).
  • 'd'ebug channel: used when the server is started with logging to '-'.

Required:

  • 'I'nput channel: the length field here tells the client how many bytes to send.
  • 'L'ine based input channel: the client should send a single line of input (no more than length bytes). This channel is used when Mercurial interacts with the user or when iterating over stdin. Data sent should include the line separator (\n or \r\n).

Input should be sent on stdin in the following format:

length
data

length = 0 sent by the client is interpreted as EOF by the server. The server will not ask for more than 4kb per request as to not fill up the pipe.

Note: Mercurial checks if stdin points to a terminal device to determine if it can communicate with the user (unless the config value ui.interactive is set). Most of the time when the command server is being run as a child process, stdin is not a terminal device. In that case it is needed to explicitly tell Mercurial to be interactive by setting ui.interactive=True.

4.3. Commands

The server is running on an endless loop (until stdin is closed) waiting for commands. A command request looks like this:

commandname\n
<command specific request>

The server aborts upon unknown commands. Clients are expected to check what commands are supported by the server by consulting the capabilities.

4.3.1. runcommand

Run the command specified by a list of \0-terminated strings. An unsigned int indicating the length of the arguments should be sent before the list. Example:

runcommand\n
8
log\0
-l\0
5

Which corresponds to running 'hg log -l 5'.

The server responds with input/output generated by Mercurial on the matching channels. When the command returns, the server writes the return code (signed integer) of the command to the 'r'esult channel.

4.3.2. getencoding

Returns the servers encoding on the result channel.

client:

getencoding\n

server responds with:

r
5
ascii

4.4. Examples

4.4.1. runcommand

Complete example of a client running 'hg summary', right after starting the server:

(text in the server column is <channel>: <length>, where length is really 4 byte unsigned ints, not plain text like below)

server

client

notes

connected, waiting for hello message

o: 52
capabilities: runcommand getencoding\n
encoding: UTF-8

server is waiting for a command

runcommand\n
7
summary

client talks to server on stdin

starts running command

o: 27
parent: 14571:17c0cb1045e5

o: 3
tip

o: 1
\n

o: 53
paper, coal: display diffstat on the changeset page\n

o: 16
branch: default\n

o: 16
commit: (clean)\n

o: 18
update: (current)\n

r: 4
0

server finished running command, writes ret on the 'r' channel to the client

closes server stdin

client disconnects

server exits

client waits for server to exit

And another one with activity on the input channels too by running 'import -':

(starting after client read the hello message)

server

client

notes

server is waiting for a command

getencoding\n

r: 5
UTF-8

server responds with the encoding, then waits for the next command

runcommand\n
8
import\0
-

starts running command

o: 26
applying patch from stdin\n

l: 4096

server tells client to send it a line

21
# HG changeset patch\n

client responds with <length><line>

l: 4096

server processes line, asks for another one

...this goes on until the client has no more input

l: 4096

0

it responds with length=0

r: 4
0

server finished running command, writes ret on the 'r' channel to the client

closes server stdin

client disconnects

server exits

client waits for server to exit

5. Known issues

  • server needs a repository to start (fixed in e811b93f2cb1)

  • loading/unloading of extensions using --config does not work
  • aliases set using --config are permanent
  • server doesn't notice changes to hgrc files
  • output generated by an extension during ext/repo/uisetup does not conform to the command protocol
  • --time doesn't work

6. Example client

This is a minimal Python example to illustrate how to establish a connection and execute a command.

   1 import sys, struct, subprocess
   2 
   3 # connect to the server
   4 server = subprocess.Popen(['hg', '--config', 'ui.interactive=True', 'serve', '--cmdserver', 'pipe'],
   5                           stdin=subprocess.PIPE, stdout=subprocess.PIPE)
   6 
   7 def readchannel(server):
   8     channel, length = struct.unpack('>cI', server.stdout.read(5))
   9     if channel in 'IL': # input
  10         return channel, length
  11     return channel, server.stdout.read(length)
  12 
  13 def writeblock(data):
  14     server.stdin.write(struct.pack('>I', len(data)))
  15     server.stdin.write(data)
  16     server.stdin.flush()
  17 
  18 # read the hello block
  19 hello = readchannel(server)
  20 print "hello block:", repr(hello)
  21 
  22 # write the command
  23 server.stdin.write('runcommand\n')
  24 writeblock('\0'.join(sys.argv[1:]))
  25 
  26 # receive the response
  27 while True:
  28     channel, val = readchannel(server)
  29     if channel == 'o':
  30         print "output:", repr(val)
  31     elif channel == 'e':
  32         print "error:", repr(val)
  33     elif channel == 'r':
  34         print "exit code:", struct.unpack(">l", val)[0]
  35         break
  36     elif channel == 'L':
  37         print "(line read request)"
  38         writeblock(sys.stdin.readline(val))
  39     elif channel == 'I':
  40         print "(block read request)"
  41         writeblock(sys.stdin.read(val))
  42     else:
  43         print "unexpected channel:", channel, val
  44         if channel.isupper(): # required?
  45             break
  46 
  47 # shut down the server
  48 server.stdin.close()

7. Libraries

A list of client libraries using the command server (feel free to add yours here):


CategoryDeveloper

CommandServer (last edited 2022-12-23 22:42:51 by gavenkoa)