controller.py 57.4 KB
Newer Older
1
2
3
4
#!/usr/bin/python3


import asyncio
5
import collections
6
from   concurrent.futures   import ThreadPoolExecutor
7
import contextlib
8
import datetime
9
import enum
BAIRE Anthony's avatar
BAIRE Anthony committed
10
import itertools
BAIRE Anthony's avatar
BAIRE Anthony committed
11
import json
12
13
14
import logging
import re
import os
15
import shlex
16
17
18
import signal
import socket
import sys
19
20
21
import time
import threading
import traceback
22
23
24

import docker
import MySQLdb
BAIRE Anthony's avatar
BAIRE Anthony committed
25
26
from   sqlalchemy import desc
import sqlalchemy.orm.scoping
27
import yaml
28

29
import config_reader
30
from database import *
31
from shared_swarm import SharedSwarmClient, ShuttingDown
32

33
34
HOST_PATH="/vol/host/"

35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
##################################################
# default number of concurrent tasks

# pushing images from the sandbox to the registry
NB_PUSH_SANDBOX = 2

# pulling images from the registry to the sandbox
NB_PULL_SANDBOX = 2

# pulling images from the registry to the swarm
NB_PULL_SWARM   = 2

# sandbox actions (start, stop, commit, ...)
NB_SANDBOX_TASKS = 4

# jobs
NB_JOB_TASKS     = 4


# default thread pool (used by most of the tasks)
default_executor = ThreadPoolExecutor(10)

##################################################

BAIRE Anthony's avatar
BAIRE Anthony committed
59
log = logging.getLogger("controller")
60

61
62
63
64
65
assert MySQLdb.threadsafety >= 1


class Error(Exception):
    pass
66

BAIRE Anthony's avatar
BAIRE Anthony committed
67
68
69
70
71
72
73
74
75
76
77
78
79
def docker_check_error(func, *k, **kw):
    """Wrapper for docker-py methods that produce a stream

    Methods producing a stram (eg: push, build) do not report all errors
    by raising exceptions. Some errors are reported later in the stream.

    This function parses the stream and raise Error() if needed.
    """
    for elem in func(*k, stream=True, **kw):
        js = json.loads(elem.decode())
        if "error" in js:
            raise Error("push error: " + js["error"])

BAIRE Anthony's avatar
BAIRE Anthony committed
80
81
82
83
84
85
86
87
88
@contextlib.contextmanager
def docker_warning(msg, *k, ignore=None):
    """Catch docker errors and issue a warning instead"""
    try:
        yield
    except docker.errors.APIError as e:
        if ignore is None or not isinstance(e, ignore):
            k += e,
            log.warning(msg + " (%s)", *k)
BAIRE Anthony's avatar
BAIRE Anthony committed
89

BAIRE Anthony's avatar
BAIRE Anthony committed
90
91
@contextlib.contextmanager
def report_error(fmt, *k):
BAIRE Anthony's avatar
doc    
BAIRE Anthony committed
92
93
94
95
96
97
98
    """Context manager for logging exceptions

    This function logs exceptions (when leaving the context) with log.error()
    (if the exception inherit from Error) or log.exception() otherwise.

    The log message is prepended with the string generated by: fmt % k
    """
BAIRE Anthony's avatar
BAIRE Anthony committed
99
100
101
102
103
104
105
106
    try:
        yield
    except Exception as e:
        msg = fmt % k
        log_func = log.error if isinstance(e, Error) else log.exception
        log_func("%s (%s)", msg,
                traceback.format_exception_only(type(e), e)[-1].strip())
        raise
107

BAIRE Anthony's avatar
BAIRE Anthony committed
108
109
110
111
112
113
114
115
116
117
118
119
120
def disable_future_warning(fut):
    """Add a dummy callback to a future to prevent asyncio warnings

    Return: the future

    The asynicio module log a warning message when the result of a future is
    not used. This function installs a dummy callback to the future so as to
    avoid this warning. This is useful for futures whose result *may* be
    ignored by the application.
    """
    fut.add_done_callback(lambda f: f.exception())
    return fut

121

122
123
124
125
126
127
128
129
130
def auto_create_task(func):
    """Decorator for forcing the creation of a task when a coroutine function is called

    Return a wrappers that calls the function, create the task on the fly and
    return it. Also it installs a no-op callback to avoid warnings in case the
    result is not used.
    """
    assert asyncio.iscoroutinefunction(func)

BAIRE Anthony's avatar
BAIRE Anthony committed
131
132
    return (lambda *k, **kw:
        disable_future_warning(asyncio.async(func(*k, **kw))))
BAIRE Anthony's avatar
BAIRE Anthony committed
133

134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
def cascade_future(src, dst):
    """propagate the result of a future to another future
    
    This function installs a callback to the future `src`, that propagates its
    result to the future `dst`.
    """
    def callback(fut):
        ex = fut.exception()
        if ex is None:
            dst.set_result(fut.result())
        else:
            dst.set_exception(ex)

    if src.done():
        callback(src)
    else:
        src.add_done_callback(callback)

BAIRE Anthony's avatar
todo    
BAIRE Anthony committed
152

153
class Manager:
BAIRE Anthony's avatar
BAIRE Anthony committed
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
    """A class for scheduling asynchronous jobs on a collection of keys

    (the one-line summary is not very helpful, sorry)
    

    Job submission

    The job is scheduled on key KEY by calling .process(KEY), the function
    returns an asyncio.Future object (which will receive the result)

    If .process(KEY) is called while a job is already running for KEY, then the
    manager arranges the job to be run a second time in a row for the same key


    Job implementation

    This is an abstract class. The job shall be implemented as the coroutine
    named ._process() in the inherited class.

    The manager guarantees that ._process() cannot be called multiple times
    concurrently for the same key

    In case a shutdown is requested (see .shutdown()):
     - all jobs that are not yet started are cancelled
     - running jobs continue until they return or until they try to acquire the
       internal semaphore (which raises ShuttingDown())


    Concurrency

    The job tasks are all started immediately (whatever the number of existing
    jobs)

    It is possible to limit the concurrency to a maximum number of parallel
    tasks. Manager provides an internal semaphore (the number of tokens is set
    in .__init__())

    The semaphore can be locked in two ways:
     - by locking the manager:

            with (yield from self): ... concurrency limited to `nb_tokens`
            tasks ...

     - by calling .run_in_executor (this is for running non-async code):

            yield from self.run_in_executor(func, args...)

    Shutdown

    When .shutdown() is called the Manager ensures that all jobs are properly
    terminated
     - it cancels all jobs that are not yet started
     - it prevents starting new jobs
     - it lets running pending jobs, but interrupts them if they when ther try
       to acquire the internal semaphore

    All cancel/interrupted tasks have their future raise ShuttingDown().

    .shutdown() returns after all pending jobs are terminated.


    Thread safety

    - Manager is *not* thread safe, all public methods must be called from the
      same thread
    """
220

221
    class _Handle:
BAIRE Anthony's avatar
BAIRE Anthony committed
222
        __slots__ = "key", "cur", "nxt"
223

224
    def __init__(self, nb_tokens=1, *, executor = default_executor, interruptible = False):
BAIRE Anthony's avatar
BAIRE Anthony committed
225
        # {key: _TaskHandler}
226
        self._handles = {}
227
        self._semaphore = asyncio.Semaphore(nb_tokens)
228
        self._shutdown = asyncio.Future()
229
        self._executor = executor
230
        self._interruptible = interruptible
231
232
233
234
235
236
237
238
239
240

    def _create_task(self, hnd):
        assert hnd.nxt is None
        def reset():
            assert not hnd.cur.done()
            nxt = hnd.nxt
            if nxt is not None:
                cascade_future(hnd.cur, nxt)
                hnd.nxt = None

BAIRE Anthony's avatar
BAIRE Anthony committed
241
        hnd.cur = asyncio.async(self._process(hnd.key, reset))
242
        hnd.cur.add_done_callback(lambda fut: self._done(hnd))
BAIRE Anthony's avatar
BAIRE Anthony committed
243
        log.debug("task scheduled %r %r", self, hnd.key)
244
        return hnd.cur
245

BAIRE Anthony's avatar
BAIRE Anthony committed
246
247
248
249
250
    def process(self, key):
        """Schedule the job to be run on key `key`

        returns an asyncio.Future (that will provide the result of ._process())
        """
251
        if self._shutdown.done():
252
253
            return self._shutdown

BAIRE Anthony's avatar
BAIRE Anthony committed
254
        hnd = self._handles.get(key)
255
256
        if hnd is None:
            # create handle
BAIRE Anthony's avatar
BAIRE Anthony committed
257
258
            self._handles[key] = hnd = self._Handle()
            hnd.key = key
259
260
261
262
263
264
265
266
267
268
269
270
271
272
            hnd.cur = None
            hnd.nxt = None

        if hnd.cur is None:
            # create task
            return self._create_task(hnd)
        else:
            # reschedule task
            if hnd.nxt is None:
                hnd.nxt = asyncio.Future()
            return hnd.nxt

    def _done(self, hnd):

BAIRE Anthony's avatar
BAIRE Anthony committed
273
        assert hnd is self._handles.get(hnd.key)
274
275
276
277
        assert hnd.cur.done()

        try:
            hnd.cur.result()
278
279
        except ShuttingDown:
            pass
280
        except Exception:
BAIRE Anthony's avatar
BAIRE Anthony committed
281
            log.exception("task %r %r unhandled exception", self, hnd.key)
282
283
284
285

        nxt     = hnd.nxt
        hnd.cur = hnd.nxt = None
        if nxt is None:
BAIRE Anthony's avatar
BAIRE Anthony committed
286
            del self._handles[hnd.key]
287
        else:
288
            cascade_future(self._create_task(hnd), nxt)
289

290

291
292
    @asyncio.coroutine
    def __iter__(self):
BAIRE Anthony's avatar
BAIRE Anthony committed
293
294
295
296
297
298
299
300
301
302
        """Coroutine for locking the internal semaphore
        
        Usage:
            with (yield from manager):
              ...

        Warning:
            after a shutdown is initiated, this function will always raise
            ShuttingDown()
        """
303
        ctx = yield from iter(self._semaphore)
304
        if self._shutdown.done():
305
306
307
308
309
310
            with ctx:
                raise ShuttingDown()
        return ctx


    @asyncio.coroutine
311
    def run_in_executor(self, *k, lock=True):
BAIRE Anthony's avatar
BAIRE Anthony committed
312
313
314
315
316
        """Run a function in a separate thread (with limited concurrency)

        This function locks the internal semaphore and runs the provided
        functionc call in a separate thread (using the executor)
        """
317
318
319
320
321
322
323
324

        def run():
            coro = asyncio.get_event_loop().run_in_executor(self._executor, *k)
            if self._interruptible:
                return next(asyncio.as_completed((coro, self._shutdown)))
            else:
                return coro
        
325
326
        if lock:
            with (yield from self):
327
                return (yield from run())
328
        else:
329
            return (yield from run())
330
331


BAIRE Anthony's avatar
BAIRE Anthony committed
332
    @asyncio.coroutine
BAIRE Anthony's avatar
BAIRE Anthony committed
333
    def _process(self, key, reset):
BAIRE Anthony's avatar
doc    
BAIRE Anthony committed
334
335
336
337
338
339
340
341
342
343
344
        """Actual implementation of the job (to be reimplemented in inherited classes)

        The Manager class guarantees that this function cannot be called
        multiple times concurrently on the same key (in case the same key is
        submitted multiple times, they Manager will call this function a second
        time after it has terminated).

        `reset` is a function that may be called to reset the 'dirty' state of
        this key (this is to avoid calling ._process() a second time if not
        necessary)
        """
345
346
        raise NotImplementedError()

347
    @asyncio.coroutine
348
    def shutdown(self):
BAIRE Anthony's avatar
BAIRE Anthony committed
349
350
351
352
        """Initiate a graceful shutdown

        This coroutine terminates once all tasks are properly terminated.
        """
353
354
        exc = ShuttingDown()
        self._shutdown.set_exception(exc)
355
        self._shutdown.exception()  # to avoid asyncio warnings
356
357
358
        # cancel all 'next' tasks
        for hnd in self._handles.values():
            if hnd.nxt is not None:
359
                hnd.nxt.set_exception(exc)
360
361
                hnd.nxt = None

362
363
364
365
        if not self._interruptible:
            yield from asyncio.gather(
                    *(h.cur for h in self._handles.values() if h.cur is not None),
                    return_exceptions=True)
BAIRE Anthony's avatar
BAIRE Anthony committed
366

367
368

class SandboxManager(Manager):
BAIRE Anthony's avatar
BAIRE Anthony committed
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
    """Manager for sandbox operation

    This manager handles all sandbox operations (start, stop, commit).

    Operations are requested asynchronously in the database:
     - Webapp.sandbox_state=starting (for starting a sandbox)
     - Webapp.sandbox_state=stopping (for stopping a sandbox)
     - WebappVersion with state=sandbox (for committing a new image)

    All state changes on Webapp.sandbox_state are atomic (e.g: if the user
    requests a stop while the sandbox is starting then the manager finishes
    with starting the sandbox but does not update the state, and it runs
    immediately again to stop the sandbox)

    Whatever is the value of Webapp.sandbox_state, the manager first examines
    commit requests and makes the commit if requested.

    If the container already exists before starting the webapp (not possible in
    normal operations), then a recovery image is committed first.

    When a commit is successful. The image manager is notified for pushing the
    image to the registry and (if it is not a recovery image) to pull it to the
    swarm (because this image will very likely be used soon).

    State changes:

     - sandbox start:
            starting->running     (normal case)
            starting->start_error (error case)

     - sandbox stop:
            stopping->idle        (normal case)
            stopping->stop_error  (error case)

     - image commit:
            sandbox->committed  (normal case)
            sandbox->error      (error case)
            (none)->committed   (recovery version) 
    """
408

409
    def __init__(self, ctrl, nb_threads = NB_SANDBOX_TASKS):
410
        super().__init__(nb_threads)
411
412
413
        self.ctrl = ctrl


BAIRE Anthony's avatar
BAIRE Anthony committed
414
415
    def inspect_sandbox(self, webapp):
        try:
BAIRE Anthony's avatar
BAIRE Anthony committed
416
417
            return self.ctrl.sandbox.inspect_container(
                    self.ctrl.gen_sandbox_name(webapp))
BAIRE Anthony's avatar
BAIRE Anthony committed
418
419
420
        except docker.errors.NotFound:
            return None

BAIRE Anthony's avatar
BAIRE Anthony committed
421
422
423
424
425
426

    @staticmethod
    def filter_commit_version(query, webapp_id):
        """Narrow a WebappVersion query to select the candidate versions to be committed"""
        return (query.
                filter_by(webapp_id=webapp_id,
BAIRE Anthony's avatar
BAIRE Anthony committed
427
                         state = int(VersionState.SANDBOX))
BAIRE Anthony's avatar
BAIRE Anthony committed
428
429
430
                    )

    def _start(self, webapp, version):
431
432
433
434
435
436
        """Start a webapp sandbox

        (to be executed in a thread pool)
        """

        ctrl = self.ctrl
BAIRE Anthony's avatar
BAIRE Anthony committed
437
        ses  = ctrl.session
438
439

        # prepare sandbox parameters
BAIRE Anthony's avatar
BAIRE Anthony committed
440
441
442

        # docker image
        if version is None:
443
            image = "%s:%s" % (ctrl.gen_factory_name(webapp.docker_os),
BAIRE Anthony's avatar
BAIRE Anthony committed
444
                    webapp.docker_os.version)
445
        else:
BAIRE Anthony's avatar
BAIRE Anthony committed
446
            image = "%s:%s" % (webapp.image_name, version.number)
BAIRE Anthony's avatar
BAIRE Anthony committed
447
448

        log.debug("sandbox %r: using image %r", webapp.docker_name, image)
449
450
451
452
453
454
455
456
457
458
459
460

        # safety checks
        # (because docker_name is used it the paths of the external volumes
        if ("/" in webapp.docker_name) or (webapp.docker_name in ("", ".", "..")):
            raise Error("malformatted docker_name")

        uid = webapp.id + 2000
        if uid < 2000:
            # just for safety
            raise Error("bad webapp id")

        # remove stale container (if any)
461
462
        if self.inspect_sandbox(webapp) is not None:
            self._stop(webapp)
463

BAIRE Anthony's avatar
BAIRE Anthony committed
464
        container = webapp.sandbox_name
465
466
        try:
            # prepare the sandbox
467
468
469
470
            # (create ssh keys)
            ctrl.check_host_path("isdir", ctrl.toolbox_path)
            ctrl.check_host_path("isdir", ctrl.sandbox_path)

BAIRE Anthony's avatar
BAIRE Anthony committed
471
            ctrl.sandbox.create_container("busybox:latest", name=container,
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
                    command = ["/bin/sh", "-c", """
set -ex

export PATH="$PATH:/.toolbox/bin"

# clean sandbox dir
rm -rf {sbx}

# create dirs
for dir in {sbx} {etc} {run}
do
    mkdir -p            ${{dir}}
    chown {uid}:65534   ${{dir}}
    chmod 0700          ${{dir}}
done

# xauth file
touch               {run}/XAuthority
chown {uid}:65534   {run}/XAuthority
chmod 0600          {run}/XAuthority

# generate ssh keys
(for type in ecdsa ed25519 rsa
do
    key={etc}/ssh_host_${{type}}_key
    [ -f $key ] || ssh-keygen -N '' -f $key -t $type >&2
    
    echo -n '{hostname}. ' | cat - ${{key}}.pub
done) > {etc}/ssh_known_hosts

# known_host file for allgo-shell
ssh-keygen -H -f {etc}/ssh_known_hosts
chmod 0644       {etc}/ssh_known_hosts
rm -f            {etc}/ssh_known_hosts.old

# authentication key for allgo-shell
rm -f               {etc}/identity
ssh-keygen -N '' -f {etc}/identity
chown {uid}:65534   {etc}/identity

# forced shell for the sshd config 
cat > {etc}/shell <<EOF
#!/bin/sh

export PATH="\$PATH:/.toolbox/bin"

uid=\`id -u\`
shell="\`getent passwd \$uid 2>/dev/null| cut -d : -f 7\`"
if [ -z "\$shell" ] ; then
    shell=/bin/sh
fi

if [ -n "\$SSH_ORIGINAL_COMMAND" ] ; then
    exec "\$shell" -c "\$SSH_ORIGINAL_COMMAND"
else
    exec "\$shell"
fi

EOF
chmod 755 {etc}/shell

# sshd config
cat > {etc}/sshd_config <<EOF
Port 22
Protocol 2

# turned off because it requires creating a 'sshd' user inside the sandbox
UsePrivilegeSeparation no

StrictModes no

ForceCommand /.sandbox/etc/ssh/shell

PermitRootLogin without-password

PubkeyAuthentication yes
AuthorizedKeysFile  /.sandbox/etc/ssh/identity.pub .ssh/authorized_keys .ssh/authorized_keys2

ChallengeResponseAuthentication no
PasswordAuthentication          no

X11Forwarding yes
X11DisplayOffset 10
PrintMotd no
PrintLastLog no
TCPKeepAlive yes

# Allow client to pass locale environment variables
AcceptEnv LANG LC_*

Subsystem sftp internal-sftp

UsePAM no
EOF
                    """.format(uid=uid,
                        hostname = "%s-sandbox-%s" % (ctrl.env, webapp.docker_name),
                        sbx = "/mnt/%s"         % webapp.docker_name,
                        etc = "/mnt/%s/etc/ssh" % webapp.docker_name,
                        run = "/mnt/%s/run"     % webapp.docker_name,
                        )],
572
                    host_config = ctrl.sandbox.create_host_config(
573
574
575
576
                        binds   = {
                            ctrl.sandbox_path: {"bind": "/mnt"},
                            ctrl.toolbox_path: {"bind": "/.toolbox", "mode": "ro"},
                    }))
577
578
            ctrl.sandbox.start(container)
            if ctrl.sandbox.wait(container):
579
580
                log.debug("sandbox %s output:\n%s", webapp.docker_name,
                        ctrl.sandbox.logs(container).decode(errors="replace"))
581
582
583
584
                raise Error("sandbox preparation failed")
            ctrl.sandbox.remove_container(container)

            # create and start the sandbox
585
586
587
588
589

            etc_dir = os.path.join(ctrl.sandbox_path, webapp.docker_name, "etc")
            run_dir = os.path.join(ctrl.sandbox_path, webapp.docker_name, "run")
            ctrl.check_host_path("isdir", etc_dir)
            ctrl.check_host_path("isdir", run_dir)
590
591
592
593

            if version is None and webapp.entrypoint:
                # prepend instructions to initialise a dummy entrypoint
                dn, bn = os.path.split(webapp.entrypoint)
BAIRE Anthony's avatar
todo    
BAIRE Anthony committed
594
                # FIXME: do nothing if entrypoint already exists
595
                prepare = """
596
                    {mkdir}
597
                    test -f {entrypoint} || cat > {entrypoint} <<EOF
598
599
600
601
602
603
604
#!/bin/sh
echo
echo "This is app '{name}' called with parameters '\$@'"
echo
echo "The workdir contains:"
ls -l
EOF
605
                    chmod 0755 -- {entrypoint}
606

607
                """.format( entrypoint  = shlex.quote(webapp.entrypoint),
608
                            name        = webapp.docker_name,
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
                            mkdir       = (("mkdir -p -- %s" % shlex.quote(dn)) if dn else ""))
            else:
                prepare = ""

            command = ["/bin/sh", "-c", """
set -x
export PATH="$PATH:/.toolbox/bin"

{prepare}

# xauth file (needed for X11 forwarding)
touch       /root/.Xauthority
chmod 600   /root/.Xauthority

exec /.toolbox/bin/sshd -D
            """.format(prepare=prepare)]
625

626
            ctrl.sandbox.create_container(image, name=container, hostname=container,
627
                    command = command,
628
629
                    host_config = ctrl.sandbox.create_host_config(
                        binds = {
630
631
632
                            etc_dir: {"bind": "/.sandbox/etc", "mode": "ro"},
                            run_dir: {"bind": "/.sandbox/run", "mode": "rw"},
                            ctrl.toolbox_path: {"bind": "/.toolbox", "mode": "ro"},
633
634
635
                            },
                        # TODO: maybe drop other caps
                        cap_drop = ["NET_RAW"],
636
                        restart_policy = {"Name": "unless-stopped"},
637
                        network_mode = ctrl.sandbox_network,
638
639
640
641
642
                        ))

            ctrl.sandbox.start(container)

        except:
BAIRE Anthony's avatar
BAIRE Anthony committed
643
644
            with docker_warning("cleanup error: unable to remove container %r",
                    container, ignore=docker.errors.NotFound):
645
646
647
                ctrl.sandbox.remove_container(container, force=True)
            raise

BAIRE Anthony's avatar
BAIRE Anthony committed
648
    def _commit(self, webapp, versions):
649
650
651
652
653
654
655
656
657
658
659
        """Commit a webapp sandbox

        (to be executed in a thread pool)

        The image version is looked up in webapp_versions (where state==sandbox).
        
        In case of any error, a recovery version is committed instead (to avoid
        loosing the work done inside the sandbox) and the candidates are put in
        error state.
        """
        ctrl = self.ctrl
BAIRE Anthony's avatar
BAIRE Anthony committed
660
        ses  = ctrl.session
661

BAIRE Anthony's avatar
BAIRE Anthony committed
662
663
664
665
666
667
        # pre-commit checks
        # - ensure that there is exactly one candidate webapp_version with
        #   state=sandbox
        # - otherwise:
        #   - put all candidates in error state
        #   - create a recovery version
668

BAIRE Anthony's avatar
BAIRE Anthony committed
669
670
671
672
673
        # version be committed
        version = None
        
        # error msg (if any)
        error = None
674

BAIRE Anthony's avatar
BAIRE Anthony committed
675
676
        # version ids to be recovered
        recover = ()
677

BAIRE Anthony's avatar
BAIRE Anthony committed
678
679
680
681
682
683
        if len(versions) == 1:
            # normal case (sandbox commit)
            version = versions[0]
            if not version.number:
                error   = "empty version number"
                recover = version.id,
684

BAIRE Anthony's avatar
BAIRE Anthony committed
685
686
687
        elif not versions:
            # sandbox rollback (when user drops a sandbox without committing a new image)
            error     = "dangling sandbox"
688

BAIRE Anthony's avatar
BAIRE Anthony committed
689
690
691
692
693
        else:
            # multiple candidates (should never happen)
            error    = "multiple candidate versions (%s)" % (
                ", ".join(map(repr, sorted(v.number for v in versions))))
            recover = tuple(v.id for v in versions)
694

BAIRE Anthony's avatar
BAIRE Anthony committed
695
696
        # TODO: make 'sandbox' a reserved name

BAIRE Anthony's avatar
BAIRE Anthony committed
697
698
699
        if error:
            changelog = "pre-commit error: " + error
            log.error("sandbox %r version id %r: %s", webapp.docker_name, recover, changelog)
700

BAIRE Anthony's avatar
BAIRE Anthony committed
701
            with ses.begin():
702
703
                # put all candidates in 'error state'
                if recover:
BAIRE Anthony's avatar
BAIRE Anthony committed
704
                    ses.execute('''UPDATE webapp_versions
BAIRE Anthony's avatar
fix sql    
BAIRE Anthony committed
705
706
                            SET changelog=CONCAT(changelog, " [", :changelog, "]"), state=:state
                            WHERE id IN :ids''', dict(changelog=changelog, ids=recover,
BAIRE Anthony's avatar
BAIRE Anthony committed
707
                                state=int(VersionState.ERROR)))
708
709

                # create a recovery version
BAIRE Anthony's avatar
BAIRE Anthony committed
710
711
712
713
714
                version = WebappVersion(
                        webapp_id = webapp.id,
                        number    = time.strftime("recovery-%Y%m%d-%H%M%S"),
                        changelog = changelog,
                        published = False,
BAIRE Anthony's avatar
BAIRE Anthony committed
715
                        state     = int(VersionState.SANDBOX))
BAIRE Anthony's avatar
BAIRE Anthony committed
716
717
718
                ses.add(version)
            ses.refresh(version)
            ses.expunge(version)
719

BAIRE Anthony's avatar
BAIRE Anthony committed
720
        assert version is not None
721

BAIRE Anthony's avatar
BAIRE Anthony committed
722
        # commit the docker image
723

BAIRE Anthony's avatar
BAIRE Anthony committed
724
725
        log.debug("dicts %r %r", webapp.__dict__, version.__dict__)
        log.info("commit sandbox %r version %r", webapp.docker_name, version.number)
726

BAIRE Anthony's avatar
BAIRE Anthony committed
727
728
729
730
        container = webapp.sandbox_name
        next_state = image_size = None
        try:
            # stop the container (if stopping or if creating a new sandbox)
BAIRE Anthony's avatar
BAIRE Anthony committed
731
            if webapp.sandbox_state in (SandboxState.STOPPING, SandboxState.STARTING):
BAIRE Anthony's avatar
BAIRE Anthony committed
732
733
               ctrl.sandbox.stop(container)
               ctrl.sandbox.wait(container)
734

BAIRE Anthony's avatar
BAIRE Anthony committed
735
736
            # commit
            cid = ctrl.sandbox.commit(container, webapp.image_name, version.number)
BAIRE Anthony's avatar
BAIRE Anthony committed
737
            next_state = VersionState.COMMITTED
BAIRE Anthony's avatar
BAIRE Anthony committed
738
739
740
741
742
743
744
            image_size = ctrl.sandbox.inspect_image(cid)["Size"]

            return version, error

        except docker.errors.NotFound:
            error = "commit error: container not found %r" % container
            log.error("%s", error)
BAIRE Anthony's avatar
BAIRE Anthony committed
745
            next_state = VersionState.ERROR
BAIRE Anthony's avatar
BAIRE Anthony committed
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
            image_size = 0
            ses.execute('''UPDATE webapp_versions
                    SET changelog=CONCAT(changelog, " [commit error: sandbox is down]")
                    WHERE id=%d''' % version.id)

            # here we do not propagate the error to allow starting/stopping the
            # sandbox immediately (without going through sandbox_state=:error)
            return None, error

        except Exception as e:
            log.exception("sandbox %r version %r: unexpected commit error (sandbox may still be recovered)",
                    webapp.docker_name, version.number)
            raise

        finally:
            # NOTE: if anything unexpected happens, the version is
            # left in state 'sandbox' and we propagate the exception to
            # ensure the work done inside the sandbox is not dropped
            # and the sandbox is pute in 'error' state
            #
            # The error will not be reported to the user. If this is an issue,
            # then the solution would be to create another error state (to be
            # used when the sandbox is still there).
            # 
            if next_state is not None:
                with ses.begin():
                    ses.execute("UPDATE webapp_versions SET state=%d, docker_image_size=%d WHERE id=%d"
                            % (next_state, image_size, version.id))
BAIRE Anthony's avatar
BAIRE Anthony committed
774

775

BAIRE Anthony's avatar
BAIRE Anthony committed
776
    @asyncio.coroutine
BAIRE Anthony's avatar
BAIRE Anthony committed
777
    def _manage_commit(self, webapp, versions, *, force=False):
BAIRE Anthony's avatar
BAIRE Anthony committed
778
779
780
781
782
783
784
785
        """Manage sandbox commit (if needed) and notify the image manager

        The commit is performed if one of these conditions is fulfilled:
        - a commit was requested (there is at least one WebappVersion entry
          with state=sandbox for this app)
        - a docker container exists (for this sandbox) and force is true
        """

BAIRE Anthony's avatar
BAIRE Anthony committed
786
        if not (versions or (force and self.inspect_sandbox(webapp) is not None)):
BAIRE Anthony's avatar
BAIRE Anthony committed
787
788
789
            return
        
        # do the commit
BAIRE Anthony's avatar
BAIRE Anthony committed
790
        version, error = yield from self.run_in_executor(self._commit, webapp, versions)
BAIRE Anthony's avatar
BAIRE Anthony committed
791
792

        # trigger push/pull operations (but do not wait)
BAIRE Anthony's avatar
BAIRE Anthony committed
793
        if version is not None:
794
795

            # push to the registry
BAIRE Anthony's avatar
BAIRE Anthony committed
796
            self.ctrl.image_manager.push(version.id)
797
798
799
800

            if not error:
                # preemptive pull to the swarm
                # (the image may be needed soon)
BAIRE Anthony's avatar
BAIRE Anthony committed
801
                self.ctrl.image_manager.pull(version.id, swarm=True)
802
803


BAIRE Anthony's avatar
BAIRE Anthony committed
804
805
    def _stop(self, webapp):
        """Stop a webapp sandbox
806

BAIRE Anthony's avatar
BAIRE Anthony committed
807
        (to be executed in a thread pool)
808
809
810
        """
        try:
            # FIXME: remove volumes (v=True) too ?
BAIRE Anthony's avatar
BAIRE Anthony committed
811
            self.ctrl.sandbox.remove_container(webapp.sandbox_name, force=True)
812
813
814
        except docker.errors.NotFound:
            pass

815
816

    @asyncio.coroutine
817
818
819
820
    def _process(self, webapp_id, reset):
        ctrl = self.ctrl

        log.debug("process sandbox %d", webapp_id)
BAIRE Anthony's avatar
BAIRE Anthony committed
821
822
823
824
825
826
827
828

        ses = ctrl.session
        with ses.begin():
            # current state of the sandbox + load docker os
            webapp = ses.query(Webapp).filter_by(id=webapp_id).one()
            webapp.docker_os

            # version to be started
829
            sandbox_version = webapp.sandbox_version
BAIRE Anthony's avatar
BAIRE Anthony committed
830
831
832
833
834
835
836
837
838
839

            # requested commits
            commit_versions = self.filter_commit_version(ses.query(WebappVersion), webapp_id).all()

            ses.expunge_all()

        # docker name of the sandbox & image
        webapp.sandbox_name = ctrl.gen_sandbox_name(webapp)
        webapp.image_name   = ctrl.gen_image_name(webapp)

BAIRE Anthony's avatar
BAIRE Anthony committed
840
        phase = "inspect"
841
        next_state = fail_state = None
842
        try:
BAIRE Anthony's avatar
BAIRE Anthony committed
843
            if webapp.sandbox_state == SandboxState.STARTING:
844
                # start the sandbox
BAIRE Anthony's avatar
BAIRE Anthony committed
845
                phase = "start"
BAIRE Anthony's avatar
BAIRE Anthony committed
846
847
                next_state = SandboxState.RUNNING
                fail_state = SandboxState.START_ERROR
848

BAIRE Anthony's avatar
BAIRE Anthony committed
849
                # commit (if a sandbox exists)
BAIRE Anthony's avatar
BAIRE Anthony committed
850
                yield from self._manage_commit(webapp, commit_versions, force=True)
BAIRE Anthony's avatar
BAIRE Anthony committed
851

BAIRE Anthony's avatar
BAIRE Anthony committed
852
                if sandbox_version is not None:
853
854
855
856
857
858
                    # ensure version belongs to this application
                    if sandbox_version.webapp_id != webapp.id:
                        raise Error("invalid version id %d (belongs to webapp %d)" % (
                            sandbox_version.id, sandbox_version.webapp_id))

                    # pull requested image
BAIRE Anthony's avatar
BAIRE Anthony committed
859
                    yield from ctrl.image_manager.pull(sandbox_version.id)
860
861
862
863
864
                else:
                    # pull image
                    yield from ctrl.image_manager.sandbox_pull_manager.process((
                        ctrl.gen_factory_name(webapp.docker_os),
                        webapp.docker_os.version))
BAIRE Anthony's avatar
BAIRE Anthony committed
865
866

                # start sandbox
BAIRE Anthony's avatar
BAIRE Anthony committed
867
                yield from self.run_in_executor(self._start, webapp, sandbox_version)
868

BAIRE Anthony's avatar
BAIRE Anthony committed
869
            elif webapp.sandbox_state == SandboxState.STOPPING:
870
                # stop the sandbox
BAIRE Anthony's avatar
BAIRE Anthony committed
871
                phase = "stop"
BAIRE Anthony's avatar
BAIRE Anthony committed
872
873
                next_state = SandboxState.IDLE
                fail_state = SandboxState.STOP_ERROR
874

BAIRE Anthony's avatar
BAIRE Anthony committed
875
                # commit (if requested)
BAIRE Anthony's avatar
BAIRE Anthony committed
876
                yield from self._manage_commit(webapp, commit_versions)
877

BAIRE Anthony's avatar
BAIRE Anthony committed
878
                yield from self.run_in_executor(self._stop, webapp)
879
880

            else:
BAIRE Anthony's avatar
BAIRE Anthony committed
881
882
                # commit (if requested)
                phase = "commit"
BAIRE Anthony's avatar
BAIRE Anthony committed
883
                yield from self._manage_commit(webapp, commit_versions)
BAIRE Anthony's avatar
BAIRE Anthony committed
884

885
886
887
888
        except ShuttingDown:
            next_state = None
            log.info("sandbox %r %s aborted (controller shutdown)", webapp.docker_name, phase)

889
890
891
892
        except BaseException as e:
            next_state = fail_state

            log_func = log.error if isinstance(e, (docker.errors.APIError, Error)) else log.exception
BAIRE Anthony's avatar
BAIRE Anthony committed
893
            log_func ("sandbox %r %s error (%s)", webapp.docker_name, phase,
894
895
896
897
898
899
900
901
                    traceback.format_exception_only(type(e), e)[-1].strip())

        finally:
            if next_state is not None:
                # atomically update the sandbox state in the db
                # (in case another action is requested during the process, eg: the user
                #  stops the sandbox while it is not fully started)
                log.info("sandbox %r is now in state %r", webapp.docker_name, next_state.name)
BAIRE Anthony's avatar
BAIRE Anthony committed
902
903
904
                with ses.begin():
                    ses.execute("UPDATE webapps SET sandbox_state=%d WHERE id=%d AND sandbox_state=%d" %
                            (next_state, webapp_id, webapp.sandbox_state))
905

906
907
908
            log.debug("done    sandbox %d", webapp_id)


BAIRE Anthony's avatar
BAIRE Anthony committed
909
class JobManager(Manager):
910
    class JobInfo:
911
        __slots__ = "job_id", "ver_id", "ctr_id", "version", "ctr_name", "client", "cpu", "mem", "node_id"
912

913
914
    def __init__(self, ctrl, bigmem_apps=()):
        super().__init__(0)
BAIRE Anthony's avatar
BAIRE Anthony committed
915
916
        self.ctrl = ctrl

917
918
919
920
921
922
923
924
        self.bigmem_apps = bigmem_apps


    @asyncio.coroutine
    def __iter__(self):
        raise NotImplementedError()


925
    def _create_job(self, info):
BAIRE Anthony's avatar
BAIRE Anthony committed
926
927
        ctrl = self.ctrl
        ses  = ctrl.session
928
929
930
        tmp_img = None

        assert info.ctr_id is None
BAIRE Anthony's avatar
BAIRE Anthony committed
931
932
933

        try:
            with ses.begin():
934
                job = ses.query(Job).filter_by(id=info.job_id).one()
BAIRE Anthony's avatar
BAIRE Anthony committed
935
936
937
                webapp = job.webapp

                log.info("start job %d (%s:%s)",
938
                        info.job_id, webapp.docker_name, info.version)
BAIRE Anthony's avatar
BAIRE Anthony committed
939

BAIRE Anthony's avatar
BAIRE Anthony committed
940
                job.state = int(JobState.RUNNING)       # pragma: nobranch (TODO: remove (coverage bug))
BAIRE Anthony's avatar
BAIRE Anthony committed
941
942
943

            
            repo = ctrl.gen_image_name(webapp)
944
            image = "%s:%s" % (repo, info.version)
BAIRE Anthony's avatar
BAIRE Anthony committed
945
946
947
948

            job_path = ctrl.gen_job_path(job)
            log.debug("job.path: %r", job_path)

949
950
951
            if info.ver_id is None:
                assert info.version == "sandbox"
                image = tmp_img = info.client.commit(ctrl.gen_sandbox_name(webapp), repo, info.version)["Id"]
BAIRE Anthony's avatar
BAIRE Anthony committed
952
953
954
955
956
            
            # TODO use another workdir
            # TODO use another uid

            ctrl.check_host_path("isdir", job_path)
957
            hc = ctrl.sandbox.create_host_config(
958
                        binds = {job_path: {"bind": "/tmp"}},
959
960
961
                        cap_drop = ["all"],
                        # FIXME: CAP_DAC_OVERRIDE needed because all nfs files have uid,gid=1000,1000
                        cap_add = ["dac_override"],
962
963
964
                        cpu_quota   = (None if info.cpu is None else (info.cpu * 1024)),
                        cpu_period  = (None if info.cpu is None else 1024),
#                        cpu_shares = info.cpu,
965
#                        mem_reservation = ctrl.mem_soft_limit,
966
                        mem_limit = info.mem,
967
968
969
970
                    )
            if ctrl.mem_soft_limit:
                # TODO: upgrade docker-py (and use create_host_config)
                hc["MemoryReservation"] = ctrl.mem_soft_limit
971
972
973
974
            # NOTE: cpu_shares has a different meaining in docker swarm and docker engine
            #  - swarm:  nb of cpus
            #  - engine: 1/1024 share of the total cpu resouces of the machine
            # engine requires  cpu_share>1
975
            if ctrl.cpu_shares:
BAIRE Anthony's avatar
BAIRE Anthony committed
976
                # TODO: upgrade docker-py (and use create_host_config)
977
                hc["CpuShares"] = info.cpu
978
            log.debug("host_config %r", hc)
979
            info.ctr_id = info.client.create_container(image, name=info.ctr_name,
980
                    working_dir = "/tmp",
981
                    # NOTE: the command line is a little complex, but this is
BAIRE Anthony's avatar
todos    
BAIRE Anthony committed
982
                    #   to ensure that (TODO write tests for this):
983
984
985
986
987
988
989
990
991
                    #   - no output is lost (we go though a pipe in case the
                    #     app has multiple processes writing to stdout/stderr
                    #     concurrently)
                    #   - we get the exit code of the app (not the exit code of
                    #     cat)
                    #   - we are failsafe (if fifo creation fails then the app
                    #     is run anyway, with the exit code of cat)
                    #   - we have no unusual dependencies (only sh, cat and
                    #     mkfifo)
992
                    command = ["/bin/sh", "-c", """
993
                                fifo=/.allgo.fifo.{job_id}
994
995
996
997
998
999
                                if mkfifo "$fifo"
                                then
                                    exec cat <"$fifo" >allgo.log &
                                    exec "$@" >"$fifo" 2>&1 &
                                    wait %1
                                    wait %2
1000
                                    rm "$fifo"
1001
1002
1003
                                else
                                    "$@" 2>&1 | cat >allgo.log
                                fi
1004
1005
1006
1007
                                failcnt="`cat /sys/fs/cgroup/memory/memory.failcnt`"
                                if [ "$failcnt" -ne 0 ] ; then
                                    echo "WARNING: out of memory (memory.failcnt=$failcnt)" >>allgo.log
                                fi
BAIRE Anthony's avatar
BAIRE Anthony committed
1008
                        """.format(job_id=job.id),
1009
                        "job%d" % job.id, webapp.entrypoint] + shlex.split(job.param),
BAIRE Anthony's avatar
BAIRE Anthony committed
1010

1011
                    labels = {"allgo.tmp_img": tmp_img or ""},
1012
                    environment=["constraint:node==" + info.node_id],
1013
                    host_config = hc)["Id"]
1014
            info.client.start(info.ctr_id)
1015
1016
1017
1018
1019

            with ses.begin():
                # save the container_id into the db
                job.container_id = info.ctr_id

1020
        except:
BAIRE Anthony's avatar
todo    
BAIRE Anthony committed
1021
            #TODO introduce a state JobState.ERROR
1022
1023
            self._remove_job(info, tmp_img=tmp_img)
            raise
BAIRE Anthony's avatar
BAIRE Anthony committed
1024
1025


1026
1027
    def _remove_job(self, info, *, tmp_img=None):
        ses = self.ctrl.session
BAIRE Anthony's avatar
BAIRE Anthony committed
1028

1029
1030
1031
        # TODO: report launch errors to the user
        # TODO: report exit code to the user
        # TODO: use another uid
BAIRE Anthony's avatar
BAIRE Anthony committed
1032

1033
1034
1035
1036
1037
1038
1039
        def parse_docker_timestamp(value):
            return datetime.datetime.strptime(
                    # limit the precision to the microsecond
                    # (othewise strptime fails)
                    re.sub(r"(\.\d{,6})\d*Z$", r"\1Z", value),
                    # iso8601 format
                    "%Y-%m-%dT%H:%M:%S.%fZ")
BAIRE Anthony's avatar
BAIRE Anthony committed
1040

1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
        with ses.begin():
            job = ses.query(Job).filter_by(id=info.job_id).one()

            exec_time = 0.0
            if info.ctr_id is not None:
                try:
                    js = info.client.inspect_container(info.ctr_id)
                except docker.errors.NotFound:
                    pass
                else:
                    started_at  = js["State"].get("StartedAt", "0001-")
                    finished_at = js["State"].get("FinishedAt")
BAIRE Anthony's avatar
BAIRE Anthony committed
1053

1054
1055
1056
1057
1058
1059
                    # default docker date is '0001-01-01T00:00:00Z'
                    if not started_at.startswith("0001-"):
                        try:
                            exec_time =( parse_docker_timestamp(finished_at)
                                        - parse_docker_timestamp(started_at)
                                        ).total_seconds()
BAIRE Anthony's avatar
BAIRE Anthony committed
1060
1061
                        except Exception: # pragma: nocover
                            log.exception("job %d: unable to compute exec time", info.job_id)
BAIRE Anthony's avatar
BAIRE Anthony committed
1062

1063
                    if tmp_img is None:
BAIRE Anthony's avatar
BAIRE Anthony committed
1064
                        tmp_img = js["Config"]["Labels"].get("allgo.tmp_img") or None
BAIRE Anthony's avatar
BAIRE Anthony committed
1065

1066
1067
                with docker_warning("job %d: cleanup error: unable to remove container", info.job_id):
                    info.client.remove_container(info.ctr_id)
BAIRE Anthony's avatar
BAIRE Anthony committed
1068
1069

            if tmp_img is not None:
1070
1071
1072
1073
1074
                with docker_warning("job %d: cleanup error: unable to remove tmp image", info.job_id):
                    info.client.remove_image(tmp_img)

            job.exec_time = exec_time
            job.state     = int(JobState.DONE)
1075
            job.container_id = None
1076
1077
1078
1079
1080
1081
1082
1083

        log.info("stop  job %d (duration %fs)", info.job_id, exec_time)


    @asyncio.coroutine
    def _finish_job(self, info):
        # wait for container to terminate
        if info.ctr_id is not None:
1084
            yield from info.client.wait_async(info.ctr_id)
BAIRE Anthony's avatar
BAIRE Anthony committed
1085

1086
1087
        # remove container
        yield from self.run_in_executor(self._remove_job, info, lock=False)
BAIRE Anthony's avatar
BAIRE Anthony committed
1088

1089
    @asyncio.coroutine
BAIRE Anthony's avatar
BAIRE Anthony committed
1090
1091
1092
1093
1094
1095
1096
1097
    def _process(self, job_id, reset):
        ctrl = self.ctrl
        ses  = ctrl.session
        log.debug("process job id %d", job_id)

        with ses.begin():
            # query db
            job = ses.query(Job).filter_by(id=job_id).first()
BAIRE Anthony's avatar
BAIRE Anthony committed
1098
            if job is None:     # pragma: nocover
1099
1100
                # unknown job
                log.warning("unknown job id %d", job_id)
BAIRE Anthony's avatar
BAIRE Anthony committed
1101
                return
BAIRE Anthony's avatar
BAIRE Anthony committed
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111

            state = JobState(job.state)

            if job.webapp is None:
                log.error("job %d: webapp id %r not found", job_id, job.webapp_id)
                if state == JobState.WAITING:       # pragma: nobranch
                    job.state = int(JobState.DONE)
                    job.exec_time = 0
                # TODO report error to the user ?
                return
1112
1113

            docker_name = job.webapp.docker_name
1114
1115
1116
1117
            
            info = self.JobInfo()
            info.job_id     = job_id
            info.ctr_id     = None
1118
            info.node_id    = None
1119
1120
            info.version    = job.version
            info.ctr_name   = ctrl.gen_job_name(job)
BAIRE Anthony's avatar
BAIRE Anthony committed
1121

1122
1123
1124
1125
1126
            # NOTE: .cpu .mem are the amount of cpu/mem requested when creating
            # a new job. They do not apply to already created jobs (by a
            # previous instance of the controller)
            info.cpu        = ctrl.cpu_shares
            info.mem        = ctrl.bigmem_hard_limit if docker_name in self.bigmem_apps else ctrl.mem_hard_limit
1127

BAIRE Anthony's avatar
BAIRE Anthony committed
1128
1129

            if job.version == "sandbox":
1130
                info.client  = ctrl.sandbox
BAIRE Anthony's avatar
BAIRE Anthony committed
1131
            else:
1132
1133
1134
1135
1136
                info.client  = ctrl.swarm


            if state == JobState.WAITING:
                # job is not yet started
1137
1138
1139
1140
                if job.container_id is not None:
                    log.warning("job %d is in state WAITING but already has a container id: %r (will be ignored)",
                            job.id, job.container_id)

1141
1142
1143
1144
1145
1146
                if job.version == "sandbox":
                    # to be run in the sandbox
                    info.ver_id = None
                else:
                    # to be run in the swarm

1147
                    # Find the wanted WebappVersion
1148
1149
1150
1151
                    #TODO: replace version_id with webapp_version_id
                    ver = ses.query(WebappVersion).filter_by(
                            webapp_id = job.webapp_id,
                            number    = job.version).filter(
1152
                                WebappVersion.state.in_((
BAIRE Anthony's avatar
BAIRE Anthony committed
1153
1154
                                    int(VersionState.COMMITTED),
                                    int(VersionState.READY)))
1155
1156
1157
                            ).order_by(
                                    WebappVersion.state.desc(),
                                    WebappVersion.id.desc()).first()
1158
1159
1160
1161
1162
1163
1164
1165
1166
                    if ver is None:
                        log.error("job %d: webapp %r version %r not found",
                                job_id, job.webapp.docker_name, job.version)

                        job.state = int(JobState.DONE)
                        # TODO report error to the user
                        return
                    info.ver_id = ver.id

BAIRE Anthony's avatar
BAIRE Anthony committed
1167
            elif state == JobState.RUNNING:     # pragma: nobranch
1168
1169
1170
1171
1172
1173
                # job is already started

                # we do not care about the actual version_id *but* we need to
                # know whether we are in the swarm or in the sandbox
                info.ver_id = None if job.version == "sandbox" else -1

1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
                if job.container_id is None:
                    # look up container id (if it is not yet in the db)
                    # FIXME: to be removed after migration (the container id is now stored in the db at creation time)
                    try:
                        info.ctr_id = job.container_id = info.client.inspect_container(
                                "%s-job-%d-%s" % (ctrl.env, job.id, job.webapp.docker_name))["Id"]
                    except docker.errors.NotFound:
                        pass
                else:
                    # check the presence of the container and validates its id against its name
                    info.ctr_id = ctrl.check_job_container(info.client, job)
1185
                    
1186
            else:
1187
                # unexpected state
1188
1189
                if state != JobState.DONE:
                    log.warning("job id %d is in unexpected state %s", job_id, state.name)
1190
1191
                return

BAIRE Anthony's avatar
BAIRE Anthony committed
1192

1193
1194
1195
1196
1197
        if state == JobState.WAITING:
            # job is not yet started

            # pull the image to the swarm
            if info.ver_id is not None:
BAIRE Anthony's avatar
BAIRE Anthony committed
1198
1199
                # NOTE: race condition: will fail if ver.state==sandbox
                #  jobs must be submitted after the image is committed
1200
1201
                yield from ctrl.image_manager.pull(info.ver_id, swarm=True)

1202
1203
1204
            # request a slot from the shared swarm
            with info.client.request_slot(info.ctr_name, info.cpu or 0, info.mem or 0):
                info.node_id = yield from info.client.wait_slot(info.ctr_name)
1205
                yield from self.run_in_executor(self._create_job, info, lock=False)
1206
            yield from self._finish_job(info)
1207

BAIRE Anthony's avatar
BAIRE Anthony committed
1208
        elif state == JobState.RUNNING: # pragma: nobranch
1209
1210
            # the job is already running
            # -> wait for its termination
1211
            yield from self._finish_job(info)
1212

BAIRE Anthony's avatar<