| 1 |
|
| 2 |
.. _howitworks: |
| 3 |
|
| 4 |
How Mitogen Works |
| 5 |
================= |
| 6 |
|
| 7 |
Some effort is required to accomplish the seemingly magical feat of |
| 8 |
bootstrapping a remote Python process without any software installed on the |
| 9 |
remote machine. The steps involved are unlikely to be immediately obvious to |
| 10 |
the casual reader, and they required several iterations to discover, so we |
| 11 |
document them thoroughly below. |
| 12 |
|
| 13 |
|
| 14 |
The UNIX First Stage |
| 15 |
-------------------- |
| 16 |
|
| 17 |
To allow delivery of the bootstrap compressed using :py:mod:`zlib`, it is |
| 18 |
necessary for something on the remote to be prepared to decompress the payload |
| 19 |
and feed it to a Python interpreter [#f1]_. Since we would like to avoid |
| 20 |
writing an error-prone shell fragment to implement this, and since we must |
| 21 |
avoid writing to the remote machine's disk in case it is read-only, the Python |
| 22 |
process started on the remote machine by Mitogen immediately forks in order to |
| 23 |
implement the decompression. |
| 24 |
|
| 25 |
|
| 26 |
Python Command Line |
| 27 |
################### |
| 28 |
|
| 29 |
The Python command line sent to the host is a :mod:`zlib`-compressed [#f2]_ and |
| 30 |
base64-encoded copy of :py:meth:`mitogen.parent.Connection._first_stage`, |
| 31 |
which is carefully written to maximize it compatibility and minimize its size. |
| 32 |
A simplified illustration of the bootstrap command is |
| 33 |
|
| 34 |
.. code:: |
| 35 |
|
| 36 |
python -c 'exec(sys.argv[1].decode("base64").decode("zlib"))' <base64> ... |
| 37 |
|
| 38 |
The command-line arranges for the Python interpreter to decode the base64'd |
| 39 |
component, decompress it and execute it as Python code. Base64 is used since |
| 40 |
to protect against any special characters that may be interpreted by the system |
| 41 |
shell in use. |
| 42 |
|
| 43 |
|
| 44 |
Forking The First Stage |
| 45 |
####################### |
| 46 |
|
| 47 |
The first stage creates a UNIX pipe and saves a copy of the process's real |
| 48 |
``stdin`` file descriptor (used for communication with the master) so that it |
| 49 |
can be recovered by the bootstrapped process later. It then forks into a new |
| 50 |
process. |
| 51 |
|
| 52 |
After fork, the parent half overwrites its ``stdin`` with the read end of the |
| 53 |
pipe, and the child half writes the string ``MITOGEN0\n``, then begins reading the |
| 54 |
:py:mod:`zlib`-compressed payload supplied on ``stdin`` by the master, and |
| 55 |
writing the decompressed result to the write-end of the UNIX pipe. |
| 56 |
|
| 57 |
To allow recovery of ``stdin`` for reuse by the bootstrapped process for |
| 58 |
parent<->child communication, it is necessary for the first stage to avoid |
| 59 |
closing ``stdin`` or reading from it until EOF. Therefore, the master |
| 60 |
sends the :py:mod:`zlib`-compressed payload prefixed with an integer size, |
| 61 |
allowing reading by the first stage of exactly the required bytes. |
| 62 |
|
| 63 |
|
| 64 |
Configuring argv[0] |
| 65 |
################### |
| 66 |
|
| 67 |
Forking provides an excellent opportunity to tidy up the eventual Python |
| 68 |
interpreter, in particular, restarting it using a fresh command-line to get rid |
| 69 |
of the large base64-encoded first stage parameter, and to replace **argv[0]** |
| 70 |
with something descriptive. |
| 71 |
|
| 72 |
After configuring its ``stdin`` to point to the read end of the pipe, the |
| 73 |
fork parent re-executes Python with **argv[0]** composed of the Python |
| 74 |
interpreter path and a remote name supplied by the Mitogen parent. As no |
| 75 |
arguments are provided to this new execution of Python, and since ``stdin`` is |
| 76 |
connected to a pipe (whose write end is connected to the first stage), the |
| 77 |
Python interpreter begins reading source code to execute from the pipe |
| 78 |
connected to ``stdin``. |
| 79 |
|
| 80 |
|
| 81 |
Bootstrap Preparation |
| 82 |
##################### |
| 83 |
|
| 84 |
Now we have the mechanism in place to send a :py:mod:`zlib`-compressed script |
| 85 |
to the remote Python interpreter, it is time to choose what to send. |
| 86 |
|
| 87 |
The script sent is simply the source code for :py:mod:`mitogen.core`, with a |
| 88 |
single line suffixed to trigger execution of the |
| 89 |
:py:meth:`mitogen.core.ExternalContext.main` function. The encoded arguments |
| 90 |
to the main function include some additional details, such as the logging package |
| 91 |
level that was active in the parent process, and whether debugging or profiling |
| 92 |
are enabled. |
| 93 |
|
| 94 |
After the script source code is prepared, it is passed through |
| 95 |
:py:func:`mitogen.master.minimize_source` to strip it of docstrings and |
| 96 |
comments, while preserving line numbers. This reduces the compressed payload |
| 97 |
by around 20%. |
| 98 |
|
| 99 |
|
| 100 |
Preserving The `mitogen.core` Source |
| 101 |
#################################### |
| 102 |
|
| 103 |
One final trick is implemented in the first stage: after bootstrapping the new |
| 104 |
child, it writes a duplicate copy of the :py:mod:`mitogen.core` source it just |
| 105 |
used to bootstrap it back into another pipe connected to the child. The child's |
| 106 |
module importer cache is initialized with a copy of the source, so that |
| 107 |
subsequent bootstraps of children-of-children do not require the source to be |
| 108 |
fetched from the master a second time. |
| 109 |
|
| 110 |
|
| 111 |
Signalling Success |
| 112 |
################## |
| 113 |
|
| 114 |
Once the first stage has signalled ``MITO000\n``, the master knows it is ready |
| 115 |
to receive the compressed bootstrap. After decompressing and writing the |
| 116 |
bootstrap source to its parent Python interpreter, the first stage writes the |
| 117 |
string ``MITO001\n`` to ``stdout`` before exiting. The master process waits for |
| 118 |
this string before considering bootstrap successful and the child's ``stdio`` |
| 119 |
ready to receive messages. |
| 120 |
|
| 121 |
The signal value is 8 bytes to match the minimum chunk size required to |
| 122 |
disambiguate between lines containing an interesting token during SSH password |
| 123 |
authentication, a debug message from the SSH client itself, or a message from |
| 124 |
the first stage. |
| 125 |
|
| 126 |
|
| 127 |
ExternalContext.main() |
| 128 |
---------------------- |
| 129 |
|
| 130 |
.. automethod:: mitogen.core.ExternalContext.main |
| 131 |
|
| 132 |
|
| 133 |
Generating A Synthetic `mitogen` Package |
| 134 |
######################################## |
| 135 |
|
| 136 |
Since the bootstrap consists of the :py:mod:`mitogen.core` source code, and |
| 137 |
this code is loaded by Python by way of its main script (:mod:`__main__` |
| 138 |
module), initially the module layout in the child will be incorrect. |
| 139 |
|
| 140 |
The first step taken after bootstrap is to rearrange :py:data:`sys.modules` slightly |
| 141 |
so that :py:mod:`mitogen.core` appears in the correct location, and all |
| 142 |
classes defined in that module have their ``__module__`` attribute fixed up |
| 143 |
such that :py:mod:`cPickle` correctly serializes instance module names. |
| 144 |
|
| 145 |
Once a synthetic :py:mod:`mitogen` package and :py:mod:`mitogen.core` module |
| 146 |
have been generated, the bootstrap **deletes** `sys.modules['__main__']`, so |
| 147 |
that any attempt to import it (by :py:mod:`cPickle`) will cause the import to |
| 148 |
be satisfied by fetching the master's actual :mod:`__main__` module. This is |
| 149 |
necessary to allow master programs to be written as a self-contained Python |
| 150 |
script. |
| 151 |
|
| 152 |
|
| 153 |
Reaping The First Stage |
| 154 |
####################### |
| 155 |
|
| 156 |
After the bootstrap has called :py:func:`os.dup` on the copy of the ``stdin`` |
| 157 |
file descriptor saved by the first stage, it is closed. |
| 158 |
|
| 159 |
Additionally, since the first stage was forked prior to re-executing the Python |
| 160 |
interpreter, it will exist as a zombie process until the parent process reaps |
| 161 |
it. Therefore the bootstrap must call :py:func:`os.wait` soon after startup. |
| 162 |
|
| 163 |
|
| 164 |
Setup Logging |
| 165 |
############# |
| 166 |
|
| 167 |
The child's :py:mod:`logging` package root logger is configured to have the |
| 168 |
same log level as the root logger in the master, and |
| 169 |
:py:class:`mitogen.core.LogHandler` is installed to forward logs to the master |
| 170 |
context's :py:data:`FORWARD_LOG <mitogen.core.FORWARD_LOG>` handle. |
| 171 |
|
| 172 |
The log level is copied into the child to avoid generating a potentially large |
| 173 |
amount of network IO forwarding logs that will simply be filtered away once |
| 174 |
they reach the master. |
| 175 |
|
| 176 |
|
| 177 |
The Module Importer |
| 178 |
################### |
| 179 |
|
| 180 |
An instance of :py:class:`mitogen.core.Importer` is installed in |
| 181 |
:py:data:`sys.meta_path`, where Python's :keyword:`import` statement will |
| 182 |
execute it before attempting to find a module locally. |
| 183 |
|
| 184 |
|
| 185 |
Standard IO Redirection |
| 186 |
####################### |
| 187 |
|
| 188 |
Two instances of :py:class:`mitogen.core.IoLogger` are created, one for |
| 189 |
``stdout`` and one for ``stderr``. This class creates a UNIX pipe whose read |
| 190 |
end is added to the IO multiplexer, and whose write end is used to overwrite |
| 191 |
the handles inherited during process creation. |
| 192 |
|
| 193 |
Even without IO redirection, something must replace ``stdin`` and ``stdout``, |
| 194 |
otherwise it is possible for the stream used for communication between parent |
| 195 |
and child to be accidentally corrupted by subprocesses run by user code. |
| 196 |
|
| 197 |
The inherited ``stdin`` is replaced by a file descriptor pointing to |
| 198 |
``/dev/null``. |
| 199 |
|
| 200 |
Finally Python's :py:data:`sys.stdout` is reopened to ensure line buffering is |
| 201 |
active, so that ``print`` statements and suchlike promptly appear in the logs. |
| 202 |
|
| 203 |
|
| 204 |
Function Call Dispatch |
| 205 |
###################### |
| 206 |
|
| 207 |
.. currentmodule:: mitogen.core |
| 208 |
|
| 209 |
After all initialization is complete, the child's main thread sits in a loop |
| 210 |
reading from a :py:class:`Channel <mitogen.core.Channel>` connected to the |
| 211 |
:py:data:`CALL_FUNCTION <mitogen.core.CALL_FUNCTION>` handle. This handle is |
| 212 |
written to by |
| 213 |
:py:meth:`call() <mitogen.parent.Context.call>` |
| 214 |
and :py:meth:`call_async() <mitogen.parent.Context.call_async>`. |
| 215 |
|
| 216 |
:py:data:`CALL_FUNCTION <mitogen.core.CALL_FUNCTION>` only accepts requests |
| 217 |
from the context IDs listed in :py:data:`mitogen.parent_ids`, forming a chain |
| 218 |
of trust between the master and any intermediate context leading to the |
| 219 |
recipient of the message. In combination with :ref:`source-verification`, this |
| 220 |
is a major contributor to ensuring contexts running on compromised |
| 221 |
infrastructure cannot trigger code execution in siblings or any parent. |
| 222 |
|
| 223 |
|
| 224 |
Shutdown |
| 225 |
######## |
| 226 |
|
| 227 |
.. currentmodule:: mitogen.core |
| 228 |
|
| 229 |
When a context receives :py:data:`SHUTDOWN <mitogen.core.SHUTDOWN>` from its |
| 230 |
immediate parent, it closes its own :py:data:`CALL_FUNCTION |
| 231 |
<mitogen.core.CALL_FUNCTION>` :py:class:`Channel <mitogen.core.Channel>` before |
| 232 |
sending :py:data:`SHUTDOWN <mitogen.core.SHUTDOWN>` to any directly connected |
| 233 |
children. Closing the channel has the effect of causing |
| 234 |
:py:meth:`ExternalContext._dispatch_calls` to exit and begin joining on the |
| 235 |
broker thread. |
| 236 |
|
| 237 |
During shutdown, the master waits up to 5 seconds for children to disconnect |
| 238 |
gracefully before force disconnecting them, while children will use that time |
| 239 |
to call :py:meth:`socket.shutdown(SHUT_WR) <socket.socket.shutdown>` on their |
| 240 |
:py:class:`IoLogger <mitogen.core.IoLogger>` socket's write ends before |
| 241 |
draining any remaining data buffered on the read ends, and ensuring any |
| 242 |
deferred broker function callbacks have had a chance to complete, necessary to |
| 243 |
capture for example forwarding any remaining :py:mod:`logging` records. |
| 244 |
|
| 245 |
An alternative approach is to wait until the IoLogger socket is completely |
| 246 |
closed, with some hard timeout, but this necessitates greater discipline than |
| 247 |
is common in infrastructure code (how often have you forgotten to redirect |
| 248 |
stderr to ``/dev/null`` when starting a daemon process?), so needless |
| 249 |
irritating delays would often be experienced during program termination. |
| 250 |
|
| 251 |
If the main thread (responsible for function call dispatch) fails to shut down |
| 252 |
gracefully, because some user function is hanging, it will still be cleaned up |
| 253 |
since as the final step in broker shutdown, the broker sends |
| 254 |
:py:mod:`signal.SIGTERM <signal>` to its own process. |
| 255 |
|
| 256 |
|
| 257 |
.. _stream-protocol: |
| 258 |
|
| 259 |
Stream Protocol |
| 260 |
--------------- |
| 261 |
|
| 262 |
.. currentmodule:: mitogen.core |
| 263 |
|
| 264 |
Once connected, a basic framing protocol is used to communicate between |
| 265 |
parent and child. Integers use big endian in their encoded form. |
| 266 |
|
| 267 |
.. list-table:: |
| 268 |
:header-rows: 1 |
| 269 |
:widths: auto |
| 270 |
|
| 271 |
* - Field |
| 272 |
- Size |
| 273 |
- Description |
| 274 |
|
| 275 |
* - `magic` |
| 276 |
- 2 |
| 277 |
- Integer 0x4d49 (``MI``), used to detect stream corruption. |
| 278 |
|
| 279 |
* - `dst_id` |
| 280 |
- 4 |
| 281 |
- Integer target context ID. :py:class:`Router` delivers messages |
| 282 |
locally when their `dst_id` matches :py:data:`mitogen.context_id`, |
| 283 |
otherwise they are routed up or downstream. |
| 284 |
|
| 285 |
* - `src_id` |
| 286 |
- 4 |
| 287 |
- Integer source context ID. Used as the target of replies if any are |
| 288 |
generated. |
| 289 |
|
| 290 |
* - `auth_id` |
| 291 |
- 4 |
| 292 |
- The context ID under whose authority the message is acting. See |
| 293 |
:py:ref:`source-verification`. |
| 294 |
|
| 295 |
* - `handle` |
| 296 |
- 4 |
| 297 |
- Integer target handle in the destination context. This is one of the |
| 298 |
:py:ref:`standard-handles`, or a dynamically generated handle used to |
| 299 |
receive a one-time reply, such as the return value of a function call. |
| 300 |
|
| 301 |
* - `reply_to` |
| 302 |
- 4 |
| 303 |
- Integer target handle to direct any reply to this message. Used to |
| 304 |
receive a one-time reply, such as the return value of a function call, |
| 305 |
or to signal a special condition for the message. :ref:`See below |
| 306 |
<reply_to_values>` for special values for this field. |
| 307 |
|
| 308 |
* - `length` |
| 309 |
- 4 |
| 310 |
- Length of the data part of the message. |
| 311 |
|
| 312 |
* - `data` |
| 313 |
- n/a |
| 314 |
- Message data, which may be raw or pickled. |
| 315 |
|
| 316 |
|
| 317 |
|
| 318 |
.. _standard-handles: |
| 319 |
|
| 320 |
Standard Handles |
| 321 |
################ |
| 322 |
|
| 323 |
Masters listen on the following handles: |
| 324 |
|
| 325 |
.. _FORWARD_LOG: |
| 326 |
.. currentmodule:: mitogen.core |
| 327 |
.. data:: FORWARD_LOG |
| 328 |
|
| 329 |
Receives `(logger_name, level, msg)` 3-tuples and writes them to the |
| 330 |
master's ``mitogen.ctx.<context_name>`` logger. |
| 331 |
|
| 332 |
.. _GET_MODULE: |
| 333 |
.. currentmodule:: mitogen.core |
| 334 |
.. data:: GET_MODULE |
| 335 |
|
| 336 |
Receives the name of a module to load `fullname`, locates the source code |
| 337 |
for `fullname`, and routes one or more :py:data:`LOAD_MODULE` messages back |
| 338 |
towards the sender of the :py:data:`GET_MODULE` request. If lookup fails, |
| 339 |
:data:`None` is sent instead. |
| 340 |
|
| 341 |
See :ref:`import-preloading` for a deeper discussion of |
| 342 |
:py:data:`GET_MODULE`/:py:data:`LOAD_MODULE`. |
| 343 |
|
| 344 |
.. _ALLOCATE_ID: |
| 345 |
.. currentmodule:: mitogen.core |
| 346 |
.. data:: ALLOCATE_ID |
| 347 |
|
| 348 |
Replies to any message sent to it with a newly allocated range of context |
| 349 |
IDs, to allow children to safely start their own contexts. Presently IDs are |
| 350 |
allocated in batches of 1000 from a 32 bit range, allowing up to 4.2 million |
| 351 |
parent contexts to be created and destroyed before the associated Router |
| 352 |
must be recreated. |
| 353 |
|
| 354 |
This is handled by :class:`mitogen.master.IdAllocator` in the master |
| 355 |
process, and messages are sent to it from |
| 356 |
:class:`mitogen.parent.ChildIdAllocator` in children. |
| 357 |
|
| 358 |
Children listen on the following handles: |
| 359 |
|
| 360 |
.. _LOAD_MODULE: |
| 361 |
.. currentmodule:: mitogen.core |
| 362 |
.. data:: LOAD_MODULE |
| 363 |
|
| 364 |
Receives `(pkg_present, path, compressed, related)` tuples, composed of: |
| 365 |
|
| 366 |
* **pkg_present**: Either :data:`None` for a plain ``.py`` module, or a |
| 367 |
list of canonical names of submodules existing witin this package. For |
| 368 |
example, a :py:data:`LOAD_MODULE` for the :py:mod:`mitogen` package would |
| 369 |
return a list like: `["mitogen.core", "mitogen.fakessh", |
| 370 |
"mitogen.master", ..]`. This list is used by children to avoid generating |
| 371 |
useless round-trips due to Python 2.x's :keyword:`import` statement |
| 372 |
behavior. |
| 373 |
* **path**: Original filesystem where the module was found on the master. |
| 374 |
* **compressed**: :py:mod:`zlib`-compressed module source code. |
| 375 |
* **related**: list of canonical module names on which this module appears |
| 376 |
to depend. Used by children that have ever started any children of their |
| 377 |
own to preload those children with :py:data:`LOAD_MODULE` messages in |
| 378 |
response to a :py:data:`GET_MODULE` request. |
| 379 |
|
| 380 |
.. _CALL_FUNCTION: |
| 381 |
.. currentmodule:: mitogen.core |
| 382 |
.. data:: CALL_FUNCTION |
| 383 |
|
| 384 |
Receives `(chain_id, mod_name, class_name, func_name, args, kwargs)` |
| 385 |
6-tuples from :class:`mitogen.parent.CallChain`, imports ``mod_name``, then |
| 386 |
attempts to execute `class_name.func_name(\*args, \**kwargs)`. |
| 387 |
|
| 388 |
* `chain_id`: if not :data:`None`, an identifier unique to the originating |
| 389 |
:class:`mitogen.parent.CallChain`. When set, if an exception occurs |
| 390 |
during a call, future calls with the same ID automatically fail with the |
| 391 |
same exception without ever executing, and failed calls with no |
| 392 |
`reply_to` set are not dumped to the logging framework as they otherwise |
| 393 |
would. This is used to implement pipelining. |
| 394 |
|
| 395 |
When this channel is closed (by way of receiving a dead message), the |
| 396 |
child's main thread begins graceful shutdown of its own :py:class:`Broker` |
| 397 |
and :py:class:`Router`. |
| 398 |
|
| 399 |
.. _SHUTDOWN: |
| 400 |
.. currentmodule:: mitogen.core |
| 401 |
.. data:: SHUTDOWN |
| 402 |
|
| 403 |
When received from a child's immediate parent, causes the broker thread to |
| 404 |
enter graceful shutdown, including sending a dead message to the child's |
| 405 |
main thread, causing it to join on the exit of the broker thread. |
| 406 |
|
| 407 |
The final step of a child's broker shutdown process sends |
| 408 |
:py:mod:`signal.SIGTERM <signal>` to itself, ensuring the process dies even |
| 409 |
if the main thread was hung executing user code. |
| 410 |
|
| 411 |
Each context is responsible for sending :py:data:`SHUTDOWN` to each of its |
| 412 |
directly connected children in response to the master sending |
| 413 |
:py:data:`SHUTDOWN` to it, and arranging for the connection to its parent |
| 414 |
to be closed shortly thereafter. |
| 415 |
|
| 416 |
|
| 417 |
Masters, and children that have ever been used to create a descendent child |
| 418 |
also listen on the following handles: |
| 419 |
|
| 420 |
.. _ADD_ROUTE: |
| 421 |
.. currentmodule:: mitogen.core |
| 422 |
.. data:: ADD_ROUTE |
| 423 |
|
| 424 |
Receives `target_id` integer from downstream, describing an ID allocated to |
| 425 |
a recently constructed child. The receiver verifies no existing route |
| 426 |
exists to `target_id` before updating its local table to route messages for |
| 427 |
`target_id` via the stream from which the :py:data:`ADD_ROUTE` message was |
| 428 |
received. |
| 429 |
|
| 430 |
.. _DEL_ROUTE: |
| 431 |
.. currentmodule:: mitogen.core |
| 432 |
.. data:: DEL_ROUTE |
| 433 |
|
| 434 |
Receives `target_id` integer from downstream, verifies a route exists to |
| 435 |
`target_id` via the stream on which the message was received, removes that |
| 436 |
route from its local table, triggers the ``disconnect`` signal on any |
| 437 |
:class:`mitogen.core.Context` instance in the local process, then |
| 438 |
propagates the message upward towards its own parent. |
| 439 |
|
| 440 |
.. currentmodule:: mitogen.core |
| 441 |
.. data:: DETACHING |
| 442 |
|
| 443 |
Sent to inform a parent that user code has invoked |
| 444 |
:meth:`ExternalContext.detach` to decouple the lifecycle of a directly |
| 445 |
connected context and its subtree from the running program. |
| 446 |
|
| 447 |
A child usually shuts down immediately if it loses its parent connection, |
| 448 |
and parents usually terminate any related Python/SSH subprocess on |
| 449 |
disconnection. Receiving :data:`DETACHING` informs the parent the |
| 450 |
connection will soon drop, but the process intends to continue life |
| 451 |
independently, and to avoid terminating the related subprocess if that |
| 452 |
subprocess is the child itself. |
| 453 |
|
| 454 |
Non-master parents also listen on the following handles: |
| 455 |
|
| 456 |
.. currentmodule:: mitogen.core |
| 457 |
.. data:: GET_MODULE |
| 458 |
|
| 459 |
As with master's ``GET_MODULE``, except this implementation |
| 460 |
(:py:class:`mitogen.master.ModuleForwarder`) serves responses using |
| 461 |
:py:class:`mitogen.core.Importer`'s cache before forwarding the request to |
| 462 |
its parent context. The response is cached by each context in turn before |
| 463 |
being forwarded on to the child context that originally made the request. |
| 464 |
In this way, the master need never re-send a module it has already sent to |
| 465 |
a direct descendant. |
| 466 |
|
| 467 |
.. currentmodule:: mitogen.core |
| 468 |
.. data:: FORWARD_MODULE |
| 469 |
|
| 470 |
Receives `(context, fullname)` tuples from its parent and arranges for a |
| 471 |
:data:`LOAD_MODULE` to be sent towards `context` for the module `fullname` |
| 472 |
and any related modules. The module must already have been delivered to the |
| 473 |
current context by its parent in a prior :data:`LOAD_MODULE` message. |
| 474 |
|
| 475 |
If the receiver is the immediate parent of `context`, then only |
| 476 |
:data:`LOAD_MODULE` is sent to the child. Otherwise :data:`LOAD_MODULE` is |
| 477 |
sent to the next closest parent if the module has not previously been sent |
| 478 |
on that stream, followed by a copy of the :data:`FORWARD_MODULE` message. |
| 479 |
|
| 480 |
This message is used to recursively preload indirect children with modules, |
| 481 |
ensuring they are cached and deduplicated at each hop in the chain leading |
| 482 |
to the target context. |
| 483 |
|
| 484 |
.. _reply_to_values: |
| 485 |
|
| 486 |
Special values for the `reply_to` field: |
| 487 |
|
| 488 |
.. _IS_DEAD: |
| 489 |
.. currentmodule:: mitogen.core |
| 490 |
.. autodata:: IS_DEAD |
| 491 |
|
| 492 |
|
| 493 |
|
| 494 |
Additional handles are created to receive the result of every function call |
| 495 |
triggered by :py:meth:`call_async() <mitogen.parent.Context.call_async>`. |
| 496 |
|
| 497 |
|
| 498 |
Use of Pickle |
| 499 |
############# |
| 500 |
|
| 501 |
The current implementation uses the Python :py:mod:`cPickle` module, with a |
| 502 |
restrictive class whitelist to prevent triggering undesirable code execution. |
| 503 |
The primary reason for using :py:mod:`cPickle` is that it is computationally |
| 504 |
efficient, and avoids including a potentially large body of serialization code |
| 505 |
in the bootstrap. |
| 506 |
|
| 507 |
The pickler will instantiate only built-in types and one of 3 constructor |
| 508 |
functions, to support unpickling :py:class:`CallError |
| 509 |
<mitogen.core.CallError>`, :py:class:`mitogen.core.Sender`,and |
| 510 |
:py:class:`Context <mitogen.core.Context>`. |
| 511 |
|
| 512 |
The choice of Pickle is one area to be revisited later. All accounts suggest it |
| 513 |
cannot be used securely, however few of those accounts appear to be expert, and |
| 514 |
none mention any additional attacks that would not be prevented by using a |
| 515 |
restrictive class whitelist. |
| 516 |
|
| 517 |
|
| 518 |
The IO Multiplexer |
| 519 |
------------------ |
| 520 |
|
| 521 |
Since we must include our IO multiplexer as part of the bootstrap, |
| 522 |
off-the-shelf implementations are for the most part entirely inappropriate. For |
| 523 |
example, a minimal copy of Twisted weighs in at around 440KiB and is composed |
| 524 |
of approximately 115 files. Even if we could arrange for an entire Python |
| 525 |
package to be transferred during bootstrap, this minimal configuration is |
| 526 |
massive in comparison to Mitogen's solution, multiplies quickly in the |
| 527 |
presence of many machines, and would require manually splitting up the parts of |
| 528 |
Twisted that we would like to use. |
| 529 |
|
| 530 |
|
| 531 |
|
| 532 |
.. _routing: |
| 533 |
|
| 534 |
Message Routing |
| 535 |
--------------- |
| 536 |
|
| 537 |
Routing assumes it is impossible to construct a tree such that one of a |
| 538 |
context's parents will not know the ID of a target the context is attempting to |
| 539 |
communicate with. |
| 540 |
|
| 541 |
When :py:class:`mitogen.core.Router` receives a message, it checks the IDs |
| 542 |
associated with its directly connected streams for a potential route. If any |
| 543 |
stream matches, either because it directly connects to the target ID, or |
| 544 |
because the master sent an :py:data:`ADD_ROUTE <mitogen.core.ADD_ROUTE>` |
| 545 |
message associating it, then the message will be forwarded down the tree using |
| 546 |
that stream. |
| 547 |
|
| 548 |
If the message does not match any :py:data:`ADD_ROUTE <mitogen.core.ADD_ROUTE>` |
| 549 |
message or stream, instead it is forwarded upwards to the immediate parent, and |
| 550 |
recursively by each parent in turn until one is reached that knows how to |
| 551 |
forward the message down the tree. |
| 552 |
|
| 553 |
When a parent establishes a new child, it sends a corresponding |
| 554 |
:py:data:`ADD_ROUTE <mitogen.core.ADD_ROUTE>` message towards its parent, which |
| 555 |
recursively forwards it up towards the root. |
| 556 |
|
| 557 |
Parents keep note of all routes associated with each stream they connect with, |
| 558 |
and trigger ``DEL_ROUTE`` messages propagated upstream for each route |
| 559 |
associated with that stream if the stream is disconnected for any reason. |
| 560 |
|
| 561 |
|
| 562 |
Example |
| 563 |
####### |
| 564 |
|
| 565 |
.. image:: images/context-tree.svg |
| 566 |
:class: mitogen-full-width |
| 567 |
|
| 568 |
In the diagram, when ``node12b`` is creating the ``sudo:node12b:webapp`` |
| 569 |
context, it must send ``ADD_ROUTE`` messages to ``rack12``, which will |
| 570 |
propagate it to ``dc1``, and recursively to ``bastion``, and ``master``; |
| 571 |
``node12b`` does not require an ``ADD_ROUTE`` message since it has a stream |
| 572 |
directly connected to the new context. |
| 573 |
|
| 574 |
Since Mitogen streams are strictly ordered, it is never possible for a parent |
| 575 |
to receive a message from a newly constructed child before receiving a |
| 576 |
corresponding ``ADD_ROUTE`` sent by the child's parent, describing how to reply |
| 577 |
to it. |
| 578 |
|
| 579 |
When ``sudo:node12b:webapp`` wants to send a message to |
| 580 |
``sudo:node22a:webapp``, the message will be routed as follows: |
| 581 |
|
| 582 |
``sudo:node12b:webapp -> node12b -> rack12 -> dc1 -> bastion -> dc2 -> rack22 -> node22a -> sudo:node22a:webapp`` |
| 583 |
|
| 584 |
.. image:: images/route.svg |
| 585 |
:class: mitogen-full-width |
| 586 |
|
| 587 |
|
| 588 |
Disconnect Propagation |
| 589 |
###################### |
| 590 |
|
| 591 |
To ensure timely shutdown when a failure occurs, where some context is awaiting |
| 592 |
a response from another context that has become disconnected, |
| 593 |
:class:`mitogen.core.Router` additionally records the destination context ID of |
| 594 |
every message received on a particular stream. |
| 595 |
|
| 596 |
When ``DEL_ROUTE`` is generated locally or received on some other stream, |
| 597 |
:class:`mitogen.parent.RouteMonitor` uses this to find every stream that ever |
| 598 |
communicated with the route that is about to go away, and forwards the message |
| 599 |
to each found. |
| 600 |
|
| 601 |
The recipient ``DEL_ROUTE`` handler in turn uses the message to find any |
| 602 |
:class:`mitogen.core.Context` in the local process corresponding to the |
| 603 |
disappearing route, and if found, fires a ``disconnected`` event on it. |
| 604 |
|
| 605 |
Any interested party, such as :class:`mitogen.core.Receiver`, may subscribe to |
| 606 |
the event and use it to abort any threads that were asleep waiting for a reply |
| 607 |
that will never arrive. |
| 608 |
|
| 609 |
|
| 610 |
.. _source-verification: |
| 611 |
|
| 612 |
Source Verification |
| 613 |
################### |
| 614 |
|
| 615 |
Before forwarding or dispatching a message it has received, |
| 616 |
:py:class:`mitogen.core.Router` first looks up the corresponding |
| 617 |
:py:class:`mitogen.core.Stream` it would use to send responses towards the |
| 618 |
context ID listed in the `auth_id` field, and if the looked up stream does not |
| 619 |
match the stream on which the message was received, the message is discarded |
| 620 |
and a warning is logged. |
| 621 |
|
| 622 |
This creates a trust chain leading up to the root of the tree, preventing |
| 623 |
downstream contexts from injecting messages appearing to be from the master or |
| 624 |
any more trustworthy parent. In this way, privileged functionality such as |
| 625 |
:py:data:`CALL_FUNCTION <mitogen.core.CALL_FUNCTION>` can base trust decisions |
| 626 |
on the accuracy of :py:ref:`auth_id <stream-protocol>`. |
| 627 |
|
| 628 |
The `auth_id` field is separate from `src_id` in order to support granting |
| 629 |
privilege to contexts that do not follow the tree's natural trust chain. This |
| 630 |
supports cases where siblings are permitted to execute code on one another, or |
| 631 |
where isolated processes can connect to a listener and communicate with an |
| 632 |
already established established tree, such as where a :mod:`mitogen.unix` |
| 633 |
client receives the same privilege as the process it connects to. |
| 634 |
|
| 635 |
|
| 636 |
Differences Between Master And Child Brokers |
| 637 |
############################################ |
| 638 |
|
| 639 |
The main difference between :py:class:`mitogen.core.Broker` and |
| 640 |
:py:class:`mitogen.master.Broker` is that when the stream connection to the |
| 641 |
parent is lost in a child, the broker will trigger its own shutdown. |
| 642 |
|
| 643 |
|
| 644 |
The Module Importer |
| 645 |
------------------- |
| 646 |
|
| 647 |
:py:class:`mitogen.core.Importer` is still a work in progress, as there |
| 648 |
are a variety of approaches to implementing it, and the present implementation |
| 649 |
is not pefectly efficient in every case. |
| 650 |
|
| 651 |
It operates by intercepting :keyword:`import` statements via |
| 652 |
:py:data:`sys.meta_path`, asking Python if it can satisfy the import by itself, |
| 653 |
and if not, indicating to Python that it is capable of loading the module. |
| 654 |
|
| 655 |
In :py:meth:`load_module() <mitogen.core.Importer.load_module>` an RPC is |
| 656 |
started to the parent context, requesting the module source code by way of a |
| 657 |
:py:data:`GET_MODULE <mitogen.core.GET_MODULE>`. If the parent context does not |
| 658 |
have the module available, it recursively forwards the request upstream, while |
| 659 |
avoiding duplicate requests for the same module from its own threads and any |
| 660 |
child contexts. |
| 661 |
|
| 662 |
|
| 663 |
Neutralizing :py:mod:`__main__` |
| 664 |
############################### |
| 665 |
|
| 666 |
To avoid accidental execution of the :py:mod:`__main__` module's code in a |
| 667 |
slave context, when serving the source of the main module, Mitogen removes any |
| 668 |
code occurring after the first conditional that looks like a standard |
| 669 |
:py:mod:`__main__` execution guard: |
| 670 |
|
| 671 |
.. code-block:: python |
| 672 |
|
| 673 |
# Code that looks like this is stripped from __main__. |
| 674 |
if __name__ == '__main__': |
| 675 |
run_some_code() |
| 676 |
|
| 677 |
To further avoid accidental execution, Mitogen will refuse to serve |
| 678 |
:mod:`__main__` to children if no execution guard is found, as it is common |
| 679 |
that no guard is present during early script prototyping. |
| 680 |
|
| 681 |
These are hacks, but they are the safest and least annoying found to solve the |
| 682 |
problem. |
| 683 |
|
| 684 |
|
| 685 |
Avoiding Negative Imports |
| 686 |
######################### |
| 687 |
|
| 688 |
In Python 2.x where relative imports are the default, a large number of import |
| 689 |
requests will be made for modules that do not exist. For example: |
| 690 |
|
| 691 |
.. code-block:: python |
| 692 |
|
| 693 |
# mypkg/__init__.py |
| 694 |
|
| 695 |
import sys |
| 696 |
import os |
| 697 |
|
| 698 |
In Python 2.x, Python will first try to load :py:mod:`mypkg.sys` and |
| 699 |
:py:mod:`mypkg.os`, which do not exist, before falling back on :py:mod:`sys` |
| 700 |
and :py:mod:`os`. |
| 701 |
|
| 702 |
These negative imports present a challenge, as they introduce a large number of |
| 703 |
pointless network round-trips. Therefore in addition to the |
| 704 |
:py:mod:`zlib`-compressed source, for packages the master sends along a list of |
| 705 |
child modules known to exist. |
| 706 |
|
| 707 |
Before indicating it can satisfy an import request, |
| 708 |
:py:class:`mitogen.core.Importer` first checks to see if the module belongs to |
| 709 |
a package it has previously imported, and if so, ignores the request if the |
| 710 |
module does not appear in the enumeration of child modules belonging to the |
| 711 |
package that was provided by the master. |
| 712 |
|
| 713 |
|
| 714 |
.. _import-preloading: |
| 715 |
|
| 716 |
Import Preloading |
| 717 |
################# |
| 718 |
|
| 719 |
.. currentmodule:: mitogen.core |
| 720 |
|
| 721 |
To further avoid round-trips, when a module or package is requested by a child, |
| 722 |
its bytecode is scanned in the master to find all the module's |
| 723 |
:keyword:`import` statements, and of those, which associated modules appear to |
| 724 |
have been loaded in the master's :py:data:`sys.modules`. |
| 725 |
|
| 726 |
The :py:data:`sys.modules` check is necessary to handle various kinds of |
| 727 |
conditional execution, for example, when a module's code guards an |
| 728 |
:keyword:`import` statement based on the active Python runtime version, |
| 729 |
operating system, or optional third party dependencies. |
| 730 |
|
| 731 |
Before replying to a child's request for a module with dependencies: |
| 732 |
|
| 733 |
* If the request is for a package, any dependent modules used by the package |
| 734 |
that appear within the package itself are known to be missing from the child, |
| 735 |
since the child requested the top-level package module, therefore they are |
| 736 |
pre-loaded into the child using :py:data:`LOAD_MODULE` messages before |
| 737 |
sending the :py:data:`LOAD_MODULE` message for the requested package module |
| 738 |
itself. In this way, the child will already have dependent modules cached by |
| 739 |
the time it receives the requested module, avoiding one round-trip for each |
| 740 |
dependency. |
| 741 |
|
| 742 |
For example, when a child requests the :py:mod:`django` package, and the master |
| 743 |
determines the :py:mod:`django` module code in the master has :keyword:`import` |
| 744 |
statements for :py:mod:`django.utils`, :py:mod:`django.utils.lru_cache`, and |
| 745 |
:py:mod:`django.utils.version`, |
| 746 |
and that execution of the module code on the master caused those modules to |
| 747 |
appear in the master's :py:data:`sys.modules`, there is high probability |
| 748 |
execution of the :py:mod:`django` module code in the child will cause the |
| 749 |
same modules to be loaded. Since all those modules exist within the |
| 750 |
:py:mod:`django` package, and we already know the child lacks that package, |
| 751 |
it is safe to assume the child will make follow-up requests for those modules |
| 752 |
too. |
| 753 |
|
| 754 |
In the example, 4 round-trips are replaced by 1 round-trip. |
| 755 |
|
| 756 |
For any package module ever requested by a child, the parent keeps a note of |
| 757 |
the name of the package for one final optimization: |
| 758 |
|
| 759 |
* If the request is for a sub-module of a package, and it is known the child |
| 760 |
loaded the package's implementation from the parent, then any dependent |
| 761 |
modules of the requested module at any nesting level within the package that |
| 762 |
is known to be missing are sent using :py:data:`LOAD_MODULE` messages before |
| 763 |
sending the :py:data:`LOAD_MODULE` message for the requested module, avoiding |
| 764 |
1 round-trip for each dependency within the same top-level package. |
| 765 |
|
| 766 |
For example, when a child has previously requested the :py:mod:`django` |
| 767 |
package module, the parent knows the package was completely absent on the |
| 768 |
child. Therefore when the child subsequently requests the |
| 769 |
:py:mod:`django.db` package module, it is safe to assume the child will |
| 770 |
generate subsequent :py:data:`GET_MODULE` requests for the 2 |
| 771 |
:py:mod:`django.conf`, 3 :py:mod:`django.core`, 2 :py:mod:`django.db`, 3 |
| 772 |
:py:mod:`django.dispatch`, and 7 :py:mod:`django.utils` indirect dependencies |
| 773 |
for :py:mod:`django.db`. |
| 774 |
|
| 775 |
In the example, 17 round-trips are replaced by 1 round-trip. |
| 776 |
|
| 777 |
The method used to detect import statements is similar to the standard library |
| 778 |
:py:mod:`modulefinder` module: rather than analyze module source code, |
| 779 |
:ref:`IMPORT_NAME <python:bytecodes>` opcodes are extracted from the module's |
| 780 |
bytecode. This is since clean source analysis methods (:py:mod:`ast` and |
| 781 |
:py:mod:`compiler`) are an order of magnitude slower, and incompatible across |
| 782 |
major Python versions. |
| 783 |
|
| 784 |
|
| 785 |
Concurrency |
| 786 |
########### |
| 787 |
|
| 788 |
Duplicate requests must never be issued to the parent, either due to a local |
| 789 |
import or any :py:data:`GET_MODULE` originating from a child. This lets parents |
| 790 |
assume a module requested once by a downstream connection need never be |
| 791 |
re-sent, for example, if it appears as a preloading dependency in a subsequent |
| 792 |
:py:data:`GET_MODULE`, or had been requested immediately after being sent as a |
| 793 |
preloading dependency for an unrelated request by a descendent. |
| 794 |
|
| 795 |
Therefore each tree layer must deduplicate :py:data:`GET_MODULE` requests, and |
| 796 |
synchronize their descendents and local threads on corresponding |
| 797 |
:py:data:`LOAD_MODULE` responses from the parent. |
| 798 |
|
| 799 |
In each context, pending requests are serialized by a |
| 800 |
:py:class:`threading.Lock` within :py:class:`mitogen.core.Importer`, which may |
| 801 |
only be held for operations that cannot block, since :py:class:`ModuleForwarder |
| 802 |
<mitogen.master.ModuleForwarder>` must acquire it while synchronizing |
| 803 |
:py:data:`GET_MODULE` requests from children on the IO multiplexer thread. |
| 804 |
|
| 805 |
|
| 806 |
Requests From Local Threads |
| 807 |
~~~~~~~~~~~~~~~~~~~~~~~~~~~ |
| 808 |
|
| 809 |
When Mitogen begins satisfying an import, it is known the module has never been |
| 810 |
imported in the local process. :py:class:`Importer <mitogen.core.Importer>` |
| 811 |
executes under the runtime importer lock, ensuring :py:keyword:`import` |
| 812 |
statements executing in local threads are serialized. |
| 813 |
|
| 814 |
.. note:: |
| 815 |
|
| 816 |
In Python 2, :py:exc:`ImportError` is raised when :py:keyword:`import` is |
| 817 |
attempted while the runtime import lock is held by another thread, |
| 818 |
therefore imports must be serialized by only attempting them from the main |
| 819 |
(:py:data:`CALL_FUNCTION`) thread. |
| 820 |
|
| 821 |
The problem is most likely to manifest in third party libraries that lazily |
| 822 |
import optional dependencies at runtime from a non-main thread. The |
| 823 |
workaround is to explicitly import those dependencies from the main thread |
| 824 |
before initializing the third party library. |
| 825 |
|
| 826 |
This was fixed in Python 3.5, but Python 3.x is not yet supported. See |
| 827 |
`Python Issue #9260`_. |
| 828 |
|
| 829 |
.. _Python Issue #9260: https://bugs.python.org/issue9260 |
| 830 |
|
| 831 |
While holding its own lock, :py:class:`Importer <mitogen.core.Importer>` |
| 832 |
checks if the source is not yet cached, determines if an in-flight |
| 833 |
:py:data:`GET_MODULE` exists for it, starting one if none exists, adds itself |
| 834 |
to a list of callbacks fired when a corresponding :py:data:`LOAD_MODULE` |
| 835 |
arrives from the parent, then sleeps waiting for the callback. |
| 836 |
|
| 837 |
When the source becomes available, the module is constructed on the calling |
| 838 |
thread using the best practice documented in `PEP 302`_. |
| 839 |
|
| 840 |
.. _PEP 302: https://www.python.org/dev/peps/pep-0302/ |
| 841 |
|
| 842 |
|
| 843 |
Requests From Children |
| 844 |
~~~~~~~~~~~~~~~~~~~~~~ |
| 845 |
|
| 846 |
As with local imports, when :py:data:`GET_MODULE` is received from a child, |
| 847 |
while holding the :py:class:`Importer <mitogen.core.Importer>` lock, |
| 848 |
:py:class:`ModuleForwarder <mitogen.master.ModuleForwarder>` checks if the |
| 849 |
source is not yet cached, determines if an in-flight :py:data:`GET_MODULE` |
| 850 |
toward the parent exists for it, starting one if none exists, then adds a |
| 851 |
completion handler to the list of callbacks fired when a corresponding |
| 852 |
:py:data:`LOAD_MODULE` arrives from the parent. |
| 853 |
|
| 854 |
When the source becomes available, the completion handler issues corresponding |
| 855 |
:py:data:`LOAD_MODULE` messages toward the child for the requested module after |
| 856 |
any required for dependencies known to be absent from the child. |
| 857 |
|
| 858 |
Since intermediaries do not know a module's dependencies until the module's |
| 859 |
source arrives, it is not possible to preemptively issue :py:data:`LOAD_MODULE` |
| 860 |
for those dependencies toward a requesting child as they become available from |
| 861 |
the parent at the intermediary. This creates needless network serialization and |
| 862 |
latency that should be addressed in a future design. |
| 863 |
|
| 864 |
|
| 865 |
Child Module Enumeration |
| 866 |
######################## |
| 867 |
|
| 868 |
Package children are enumerated using :py:func:`pkgutil.iter_modules`. |
| 869 |
|
| 870 |
|
| 871 |
Use Of Threads |
| 872 |
-------------- |
| 873 |
|
| 874 |
The package always runs the IO multiplexer in a thread. This is so the |
| 875 |
multiplexer retains control flow in order to shut down gracefully, say, if the |
| 876 |
user's code has hung and the master context has disconnected. |
| 877 |
|
| 878 |
While it is possible for the IO multiplexer to recover control of a hung |
| 879 |
function call on UNIX using for example :py:mod:`signal.SIGALRM <signal>`, this |
| 880 |
mechanism is not portable to non-UNIX operating systems, and does not work in |
| 881 |
every case, for example when Python blocks signals during a variety of |
| 882 |
:py:mod:`threading` package operations. |
| 883 |
|
| 884 |
At some point it is likely Mitogen will be extended to support children running |
| 885 |
on Windows. When that happens, it would be nice if the process model on Windows |
| 886 |
and UNIX did not differ, and in fact the code used on both were identical. |
| 887 |
|
| 888 |
|
| 889 |
.. _waking-sleeping-threads: |
| 890 |
|
| 891 |
Waking Sleeping Threads |
| 892 |
####################### |
| 893 |
|
| 894 |
Due to fundamental deficiencies in Python 2's threading implementation, it is |
| 895 |
not possible to block waiting on synchronization objects sanely. Two major |
| 896 |
problems exist: |
| 897 |
|
| 898 |
* Sleeping with no timeout set causes signals to be blocked, preventing the |
| 899 |
user from terminating the process using CTRL+C. |
| 900 |
|
| 901 |
* Sleeping with a timeout set internally makes use of polling, with an |
| 902 |
exponential backoff that eventually results in the thread sleeping |
| 903 |
unconditionally in 50ms increments. . This is a huge source of latency that |
| 904 |
quickly multiplies. |
| 905 |
|
| 906 |
As the UNIX self-pipe trick must already be employed to wake the broker thread |
| 907 |
from its select loop, Mitogen reuses this technique to wake any thread |
| 908 |
synchronization primitive exposed by the library, embodied in a queue-like |
| 909 |
abstraction called a :py:class:`mitogen.core.Latch`. |
| 910 |
|
| 911 |
Unfortunately it is commonplace for hosts to enforce severe per-process file |
| 912 |
descriptors limits, so aside from being inefficient, it is impossible in the |
| 913 |
usual case to create a pair of descriptors for every waitable object, which for |
| 914 |
example includes the result of every single asynchronous function call. |
| 915 |
|
| 916 |
For this reason self-pipes are created on a per-thread basis, with their |
| 917 |
associated :py:func:`socketpairs <socket.socketpair>` kept in thread-local |
| 918 |
storage. When a latch wishes to sleep its thread, this pair is created |
| 919 |
on-demand and temporarily associated with it only for the duration of the |
| 920 |
sleep. |
| 921 |
|
| 922 |
Python's garbage collector is relied on to clean up by calling the pair's |
| 923 |
destructor on thread exit. There does not otherwise seem to be a robust method |
| 924 |
to trigger cleanup code on arbitrary threads. |
| 925 |
|
| 926 |
To summarize, file descriptor usage is bounded by the number of threads rather |
| 927 |
than the number of waitables, which is a much smaller number, however it also |
| 928 |
means that Mitogen requires twice as many file descriptors as there are user |
| 929 |
threads, with a minimum of 4 required in any configuration. |
| 930 |
|
| 931 |
|
| 932 |
Latch Internals |
| 933 |
~~~~~~~~~~~~~~~ |
| 934 |
|
| 935 |
.. currentmodule:: mitogen.core |
| 936 |
|
| 937 |
Attributes: |
| 938 |
|
| 939 |
* `lock` – :py:class:`threading.Lock`. |
| 940 |
* `queue` – items waiting to be dequeued. |
| 941 |
* `sleeping` – write sides of the socketpairs for each sleeping thread, and |
| 942 |
threads in the process of waking from sleep. |
| 943 |
* `waking` – integer number of `sleeping` threads in the process of waking up. |
| 944 |
* `closed` – boolean defaulting to :py:data:`False`. Every time `lock` |
| 945 |
is acquired, `closed` must be tested, and if it is :py:data:`True`, |
| 946 |
:py:class:`LatchError` must be thrown. |
| 947 |
|
| 948 |
|
| 949 |
Latch.put() |
| 950 |
~~~~~~~~~~~ |
| 951 |
|
| 952 |
:py:meth:`Latch.put` operates by: |
| 953 |
|
| 954 |
1. Acquiring `lock`. |
| 955 |
2. Appending the item on to `queue`. |
| 956 |
3. If `waking` is less than the length of `sleeping`, write a byte to the |
| 957 |
socket at `sleeping[waking]` and increment `waking`. |
| 958 |
|
| 959 |
In this way each thread is woken only once, and receives each element according |
| 960 |
to when its socket was placed on `sleeping`. |
| 961 |
|
| 962 |
|
| 963 |
Latch.close() |
| 964 |
~~~~~~~~~~~~~ |
| 965 |
|
| 966 |
:py:meth:`Latch.close` acquires `lock`, sets `closed` to :py:data:`True`, then |
| 967 |
writes a byte to every `sleeping[waking]` socket, while incrementing `waking`, |
| 968 |
until no more unwoken sockets exist. Per above, on waking from sleep, after |
| 969 |
removing itself from `sleeping`, each sleeping thread tests if `closed` is |
| 970 |
:py:data:`True`, and if so throws :py:class:`LatchError`. |
| 971 |
|
| 972 |
It is necessary to ensure at most one byte is delivered on each socket, even if |
| 973 |
the latch is being torn down, as the sockets outlive the scope of a single |
| 974 |
latch, and must never have extraneous data buffered on them, as this will cause |
| 975 |
unexpected wakeups if future latches sleep on the same thread. |
| 976 |
|
| 977 |
|
| 978 |
Latch.get() |
| 979 |
~~~~~~~~~~~ |
| 980 |
|
| 981 |
:py:meth:`Latch.get` is far more intricate, as there are many outcomes to |
| 982 |
handle. Queue ordering is strictly first-in first-out, and threads always |
| 983 |
receive items in the order they are requested, as they become available. |
| 984 |
|
| 985 |
**1. Non-empty, No Waiters, No sleep** |
| 986 |
On entry `lock` is taken, and if `queue` is non-empty, and `sleeping` is |
| 987 |
empty, it is safe to return `queue`'s first item without blocking. |
| 988 |
|
| 989 |
**2. Non-empty, Waiters Present, Queue > Waiters, No sleep** |
| 990 |
When `sleeping` is non-empty but there are more items than sleeping |
| 991 |
threads, it is safe to pop `queue[len(sleeping)]` without blocking. |
| 992 |
|
| 993 |
**3. Non-empty, Waiters Present, Queue <= Waiters** |
| 994 |
In this case `sleeping` is non-empty and there are no surplus items. It is |
| 995 |
not safe to pop any item even though we are holding `lock`, as it would |
| 996 |
starve waking threads of their position in favour of the calling thread, |
| 997 |
since scheduling uncertainty exists between a thread waking from |
| 998 |
:py:func:`select.select` and re-acquiring `lock`. |
| 999 |
|
| 1000 |
This avoids the need for a retry loop for waking threads, and a thread |
| 1001 |
being continually re-woken to discover `queue` drained by a thread that |
| 1002 |
never slept. |
| 1003 |
|
| 1004 |
**4. Sleep** |
| 1005 |
Since no surplus items existed, the thread adds its socket to `sleeping` |
| 1006 |
before releasing `lock`, and sleeping in :py:func:`select.select` waiting |
| 1007 |
for timeout, or a write from :py:meth:`Latch.put` or |
| 1008 |
:py:meth:`Latch.close`. |
| 1009 |
|
| 1010 |
If :py:func:`select.select` throws an exception, the exception must be |
| 1011 |
caught and re-raised only after some of the wake steps below have |
| 1012 |
completed. |
| 1013 |
|
| 1014 |
**5. Wake, Non-empty** |
| 1015 |
On wake `lock` is re-acquired, the socket is removed from `sleeping` after |
| 1016 |
noting its index, and :py:class:`TimeoutError` is thrown if `waking` |
| 1017 |
indicates :py:meth:`Latch.put()` nor :py:meth:`Latch.close` have yet to |
| 1018 |
send a wake byte to that index. The byte is then read off, |
| 1019 |
:py:class:`LatchError` is thrown if `closed` is :py:data:`True`, otherwise |
| 1020 |
the queue item corresponding to the thread's index is popped and returned. |
| 1021 |
|
| 1022 |
It is paramount that in every case, if a byte was written to the socket, |
| 1023 |
that the byte is read away. The socket is reused by subsequent latches |
| 1024 |
sleeping on the same thread, and unexpected wakeups are triggered if |
| 1025 |
extraneous data remains buffered on the socket. |
| 1026 |
|
| 1027 |
It is also necessary to favour the synchronized `waking` variable over the |
| 1028 |
return value of :py:func:`select.select`, as scheduling uncertainty |
| 1029 |
introduces a race between the select timing out, and :py:meth:`Latch.put()` |
| 1030 |
or :py:meth:`Latch.close` writing a wake byte before :py:meth:`Latch.get` |
| 1031 |
has re-acquired `lock`. |
| 1032 |
|
| 1033 |
|
| 1034 |
.. rubric:: Footnotes |
| 1035 |
|
| 1036 |
.. [#f1] Although some connection methods such as SSH support compression, and |
| 1037 |
Mitogen enables SSH compression by default, there are circumstances where |
| 1038 |
disabling SSH compression is desirable, and many scenarios for future |
| 1039 |
connection methods where transport-layer compression is not supported at |
| 1040 |
all. |
| 1041 |
|
| 1042 |
.. [#f2] Compression may seem redundant, however it is basically free and reducing IO |
| 1043 |
is always a good idea. The 33% / 200 byte saving may mean the presence or |
| 1044 |
absence of an additional frame on the network, or in real world terms after |
| 1045 |
accounting for SSH overhead, around a 2% reduced chance of a stall during |
| 1046 |
connection setup due to a dropped frame. |