11561: Add lock_count to containers, cancel container on unlock exceeded
[arvados.git] / services / api / app / models / container.rb
1 # Copyright (C) The Arvados Authors. All rights reserved.
2 #
3 # SPDX-License-Identifier: AGPL-3.0
4
5 require 'log_reuse_info'
6 require 'whitelist_update'
7 require 'safe_json'
8 require 'update_priority'
9
10 class Container < ArvadosModel
11   include ArvadosModelUpdates
12   include HasUuid
13   include KindAndEtag
14   include CommonApiTemplate
15   include WhitelistUpdate
16   extend CurrentApiClient
17   extend DbCurrentTime
18   extend LogReuseInfo
19
20   serialize :environment, Hash
21   serialize :mounts, Hash
22   serialize :runtime_constraints, Hash
23   serialize :command, Array
24   serialize :scheduling_parameters, Hash
25   serialize :secret_mounts, Hash
26   serialize :runtime_status, Hash
27
28   before_validation :fill_field_defaults, :if => :new_record?
29   before_validation :set_timestamps
30   validates :command, :container_image, :output_path, :cwd, :priority, { presence: true }
31   validates :priority, numericality: { only_integer: true, greater_than_or_equal_to: 0 }
32   validate :validate_runtime_status
33   validate :validate_state_change
34   validate :validate_change
35   validate :validate_lock
36   validate :validate_output
37   after_validation :assign_auth
38   before_save :sort_serialized_attrs
39   before_save :update_secret_mounts_md5
40   before_save :scrub_secrets
41   before_save :clear_runtime_status_when_queued
42   after_save :update_cr_logs
43   after_save :handle_completed
44   after_save :propagate_priority
45   after_commit { UpdatePriority.run_update_thread }
46
47   has_many :container_requests, :foreign_key => :container_uuid, :class_name => 'ContainerRequest', :primary_key => :uuid
48   belongs_to :auth, :class_name => 'ApiClientAuthorization', :foreign_key => :auth_uuid, :primary_key => :uuid
49
50   api_accessible :user, extend: :common do |t|
51     t.add :command
52     t.add :container_image
53     t.add :cwd
54     t.add :environment
55     t.add :exit_code
56     t.add :finished_at
57     t.add :locked_by_uuid
58     t.add :log
59     t.add :mounts
60     t.add :output
61     t.add :output_path
62     t.add :priority
63     t.add :progress
64     t.add :runtime_constraints
65     t.add :runtime_status
66     t.add :started_at
67     t.add :state
68     t.add :auth_uuid
69     t.add :scheduling_parameters
70     t.add :runtime_user_uuid
71     t.add :runtime_auth_scopes
72   end
73
74   # Supported states for a container
75   States =
76     [
77      (Queued = 'Queued'),
78      (Locked = 'Locked'),
79      (Running = 'Running'),
80      (Complete = 'Complete'),
81      (Cancelled = 'Cancelled')
82     ]
83
84   State_transitions = {
85     nil => [Queued],
86     Queued => [Locked, Cancelled],
87     Locked => [Queued, Running, Cancelled],
88     Running => [Complete, Cancelled]
89   }
90
91   def self.limit_index_columns_read
92     ["mounts"]
93   end
94
95   def self.full_text_searchable_columns
96     super - ["secret_mounts", "secret_mounts_md5", "runtime_token"]
97   end
98
99   def self.searchable_columns *args
100     super - ["secret_mounts_md5", "runtime_token"]
101   end
102
103   def logged_attributes
104     super.except('secret_mounts', 'runtime_token')
105   end
106
107   def state_transitions
108     State_transitions
109   end
110
111   # Container priority is the highest "computed priority" of any
112   # matching request. The computed priority of a container-submitted
113   # request is the priority of the submitting container. The computed
114   # priority of a user-submitted request is a function of
115   # user-assigned priority and request creation time.
116   def update_priority!
117     return if ![Queued, Locked, Running].include?(state)
118     p = ContainerRequest.
119         where('container_uuid=? and priority>0', uuid).
120         includes(:requesting_container).
121         lock(true).
122         map do |cr|
123       if cr.requesting_container
124         cr.requesting_container.priority
125       else
126         (cr.priority << 50) - (cr.created_at.to_time.to_f * 1000).to_i
127       end
128     end.max || 0
129     update_attributes!(priority: p)
130   end
131
132   def propagate_priority
133     return true unless priority_changed?
134     act_as_system_user do
135       # Update the priority of child container requests to match new
136       # priority of the parent container (ignoring requests with no
137       # container assigned, because their priority doesn't matter).
138       ContainerRequest.
139         where(requesting_container_uuid: self.uuid,
140               state: ContainerRequest::Committed).
141         where('container_uuid is not null').
142         includes(:container).
143         map(&:container).
144         map(&:update_priority!)
145     end
146   end
147
148   # Create a new container (or find an existing one) to satisfy the
149   # given container request.
150   def self.resolve(req)
151     if req.runtime_token.nil?
152       runtime_user = if req.modified_by_user_uuid.nil?
153                        current_user
154                      else
155                        User.find_by_uuid(req.modified_by_user_uuid)
156                      end
157       runtime_auth_scopes = ["all"]
158     else
159       auth = ApiClientAuthorization.validate(token: req.runtime_token)
160       if auth.nil?
161         raise ArgumentError.new "Invalid runtime token"
162       end
163       runtime_user = User.find_by_id(auth.user_id)
164       runtime_auth_scopes = auth.scopes
165     end
166     c_attrs = act_as_user runtime_user do
167       {
168         command: req.command,
169         cwd: req.cwd,
170         environment: req.environment,
171         output_path: req.output_path,
172         container_image: resolve_container_image(req.container_image),
173         mounts: resolve_mounts(req.mounts),
174         runtime_constraints: resolve_runtime_constraints(req.runtime_constraints),
175         scheduling_parameters: req.scheduling_parameters,
176         secret_mounts: req.secret_mounts,
177         runtime_token: req.runtime_token,
178         runtime_user_uuid: runtime_user.uuid,
179         runtime_auth_scopes: runtime_auth_scopes
180       }
181     end
182     act_as_system_user do
183       if req.use_existing && (reusable = find_reusable(c_attrs))
184         reusable
185       else
186         Container.create!(c_attrs)
187       end
188     end
189   end
190
191   # Return a runtime_constraints hash that complies with requested but
192   # is suitable for saving in a container record, i.e., has specific
193   # values instead of ranges.
194   #
195   # Doing this as a step separate from other resolutions, like "git
196   # revision range to commit hash", makes sense only when there is no
197   # opportunity to reuse an existing container (e.g., container reuse
198   # is not implemented yet, or we have already found that no existing
199   # containers are suitable).
200   def self.resolve_runtime_constraints(runtime_constraints)
201     rc = {}
202     defaults = {
203       'keep_cache_ram' =>
204       Rails.configuration.container_default_keep_cache_ram,
205     }
206     defaults.merge(runtime_constraints).each do |k, v|
207       if v.is_a? Array
208         rc[k] = v[0]
209       else
210         rc[k] = v
211       end
212     end
213     rc
214   end
215
216   # Return a mounts hash suitable for a Container, i.e., with every
217   # readonly collection UUID resolved to a PDH.
218   def self.resolve_mounts(mounts)
219     c_mounts = {}
220     mounts.each do |k, mount|
221       mount = mount.dup
222       c_mounts[k] = mount
223       if mount['kind'] != 'collection'
224         next
225       end
226
227       uuid = mount.delete 'uuid'
228
229       if mount['portable_data_hash'].nil? and !uuid.nil?
230         # PDH not supplied, try by UUID
231         c = Collection.
232           readable_by(current_user).
233           where(uuid: uuid).
234           select(:portable_data_hash).
235           first
236         if !c
237           raise ArvadosModel::UnresolvableContainerError.new "cannot mount collection #{uuid.inspect}: not found"
238         end
239         mount['portable_data_hash'] = c.portable_data_hash
240       end
241     end
242     return c_mounts
243   end
244
245   # Return a container_image PDH suitable for a Container.
246   def self.resolve_container_image(container_image)
247     coll = Collection.for_latest_docker_image(container_image)
248     if !coll
249       raise ArvadosModel::UnresolvableContainerError.new "docker image #{container_image.inspect} not found"
250     end
251     coll.portable_data_hash
252   end
253
254   def self.find_reusable(attrs)
255     log_reuse_info { "starting with #{Container.all.count} container records in database" }
256     candidates = Container.where_serialized(:command, attrs[:command], md5: true)
257     log_reuse_info(candidates) { "after filtering on command #{attrs[:command].inspect}" }
258
259     candidates = candidates.where('cwd = ?', attrs[:cwd])
260     log_reuse_info(candidates) { "after filtering on cwd #{attrs[:cwd].inspect}" }
261
262     candidates = candidates.where_serialized(:environment, attrs[:environment], md5: true)
263     log_reuse_info(candidates) { "after filtering on environment #{attrs[:environment].inspect}" }
264
265     candidates = candidates.where('output_path = ?', attrs[:output_path])
266     log_reuse_info(candidates) { "after filtering on output_path #{attrs[:output_path].inspect}" }
267
268     image = resolve_container_image(attrs[:container_image])
269     candidates = candidates.where('container_image = ?', image)
270     log_reuse_info(candidates) { "after filtering on container_image #{image.inspect} (resolved from #{attrs[:container_image].inspect})" }
271
272     candidates = candidates.where_serialized(:mounts, resolve_mounts(attrs[:mounts]), md5: true)
273     log_reuse_info(candidates) { "after filtering on mounts #{attrs[:mounts].inspect}" }
274
275     secret_mounts_md5 = Digest::MD5.hexdigest(SafeJSON.dump(self.deep_sort_hash(attrs[:secret_mounts])))
276     candidates = candidates.where('secret_mounts_md5 = ?', secret_mounts_md5)
277     log_reuse_info(candidates) { "after filtering on secret_mounts_md5 #{secret_mounts_md5.inspect}" }
278
279     candidates = candidates.where_serialized(:runtime_constraints, resolve_runtime_constraints(attrs[:runtime_constraints]), md5: true)
280     log_reuse_info(candidates) { "after filtering on runtime_constraints #{attrs[:runtime_constraints].inspect}" }
281
282     log_reuse_info { "checking for state=Complete with readable output and log..." }
283
284     select_readable_pdh = Collection.
285       readable_by(current_user).
286       select(:portable_data_hash).
287       to_sql
288
289     usable = candidates.where(state: Complete, exit_code: 0)
290     log_reuse_info(usable) { "with state=Complete, exit_code=0" }
291
292     usable = usable.where("log IN (#{select_readable_pdh})")
293     log_reuse_info(usable) { "with readable log" }
294
295     usable = usable.where("output IN (#{select_readable_pdh})")
296     log_reuse_info(usable) { "with readable output" }
297
298     usable = usable.order('finished_at ASC').limit(1).first
299     if usable
300       log_reuse_info { "done, reusing container #{usable.uuid} with state=Complete" }
301       return usable
302     end
303
304     # Check for non-failing Running candidates and return the most likely to finish sooner.
305     log_reuse_info { "checking for state=Running..." }
306     running = candidates.where(state: Running).
307               where("(runtime_status->'error') is null").
308               order('progress desc, started_at asc').
309               limit(1).first
310     if running
311       log_reuse_info { "done, reusing container #{running.uuid} with state=Running" }
312       return running
313     else
314       log_reuse_info { "have no containers in Running state" }
315     end
316
317     # Check for Locked or Queued ones and return the most likely to start first.
318     locked_or_queued = candidates.
319                        where("state IN (?)", [Locked, Queued]).
320                        order('state asc, priority desc, created_at asc').
321                        limit(1).first
322     if locked_or_queued
323       log_reuse_info { "done, reusing container #{locked_or_queued.uuid} with state=#{locked_or_queued.state}" }
324       return locked_or_queued
325     else
326       log_reuse_info { "have no containers in Locked or Queued state" }
327     end
328
329     log_reuse_info { "done, no reusable container found" }
330     nil
331   end
332
333   def check_lock_fail
334     if self.state != Queued
335       raise LockFailedError.new("cannot lock when #{self.state}")
336     elsif self.priority <= 0
337       raise LockFailedError.new("cannot lock when priority<=0")
338     end
339   end
340
341   def lock
342     # Check invalid state transitions once before getting the lock
343     # (because it's cheaper that way) and once after getting the lock
344     # (because state might have changed while acquiring the lock).
345     check_lock_fail
346     transaction do
347       reload
348       check_lock_fail
349       update_attributes!(state: Locked, lock_count: self.lock_count+1)
350     end
351   end
352
353   def check_unlock_fail
354     if self.state != Locked
355       raise InvalidStateTransitionError.new("cannot unlock when #{self.state}")
356     elsif self.locked_by_uuid != current_api_client_authorization.uuid
357       raise InvalidStateTransitionError.new("locked by a different token")
358     end
359   end
360
361   def unlock
362     # Check invalid state transitions twice (see lock)
363     check_unlock_fail
364     transaction do
365       reload(lock: 'FOR UPDATE')
366       check_unlock_fail
367       if self.lock_count < Rails.configuration.max_container_dispatch_attempts
368         update_attributes!(state: Queued)
369       else
370         update_attributes!(state: Cancelled,
371                            runtime_status: {
372                              error: "Container exceeded 'max_container_dispatch_attempts' (lock_count=#{self.lock_count}."
373                            })
374       end
375     end
376   end
377
378   def self.readable_by(*users_list)
379     # Load optional keyword arguments, if they exist.
380     if users_list.last.is_a? Hash
381       kwargs = users_list.pop
382     else
383       kwargs = {}
384     end
385     Container.where(ContainerRequest.readable_by(*users_list).where("containers.uuid = container_requests.container_uuid").exists)
386   end
387
388   def final?
389     [Complete, Cancelled].include?(self.state)
390   end
391
392   def self.for_current_token
393     return if !current_api_client_authorization
394     _, _, _, container_uuid = Thread.current[:token].split('/')
395     if container_uuid.nil?
396       Container.where(auth_uuid: current_api_client_authorization.uuid).first
397     else
398       Container.where('auth_uuid=? or (uuid=? and runtime_token=?)',
399                       current_api_client_authorization.uuid,
400                       container_uuid,
401                       current_api_client_authorization.token).first
402     end
403   end
404
405   protected
406
407   def fill_field_defaults
408     self.state ||= Queued
409     self.environment ||= {}
410     self.runtime_constraints ||= {}
411     self.mounts ||= {}
412     self.cwd ||= "."
413     self.priority ||= 0
414     self.scheduling_parameters ||= {}
415   end
416
417   def permission_to_create
418     current_user.andand.is_admin
419   end
420
421   def ensure_owner_uuid_is_permitted
422     # validate_change ensures owner_uuid can't be changed at all --
423     # except during create, which requires admin privileges. Checking
424     # permission here would be superfluous.
425     true
426   end
427
428   def set_timestamps
429     if self.state_changed? and self.state == Running
430       self.started_at ||= db_current_time
431     end
432
433     if self.state_changed? and [Complete, Cancelled].include? self.state
434       self.finished_at ||= db_current_time
435     end
436   end
437
438   # Check that well-known runtime status keys have desired data types
439   def validate_runtime_status
440     [
441       'error', 'errorDetail', 'warning', 'warningDetail', 'activity'
442     ].each do |k|
443       if self.runtime_status.andand.include?(k) && !self.runtime_status[k].is_a?(String)
444         errors.add(:runtime_status, "'#{k}' value must be a string")
445       end
446     end
447   end
448
449   def validate_change
450     permitted = [:state]
451     progress_attrs = [:progress, :runtime_status, :log, :output]
452     final_attrs = [:exit_code, :finished_at]
453
454     if self.new_record?
455       permitted.push(:owner_uuid, :command, :container_image, :cwd,
456                      :environment, :mounts, :output_path, :priority,
457                      :runtime_constraints, :scheduling_parameters,
458                      :secret_mounts, :runtime_token,
459                      :runtime_user_uuid, :runtime_auth_scopes)
460     end
461
462     case self.state
463     when Locked
464       permitted.push :priority, :runtime_status, :log, :lock_count
465
466     when Queued
467       permitted.push :priority
468
469     when Running
470       permitted.push :priority, *progress_attrs
471       if self.state_changed?
472         permitted.push :started_at
473       end
474
475     when Complete
476       if self.state_was == Running
477         permitted.push *final_attrs, *progress_attrs
478       end
479
480     when Cancelled
481       case self.state_was
482       when Running
483         permitted.push :finished_at, *progress_attrs
484       when Queued, Locked
485         permitted.push :finished_at, :log, :runtime_status
486       end
487
488     else
489       # The state_transitions check will add an error message for this
490       return false
491     end
492
493     if self.state == Running &&
494        !current_api_client_authorization.nil? &&
495        (current_api_client_authorization.uuid == self.auth_uuid ||
496         current_api_client_authorization.token == self.runtime_token)
497       # The contained process itself can write final attrs but can't
498       # change priority or log.
499       permitted.push *final_attrs
500       permitted = permitted - [:log, :priority]
501     elsif self.locked_by_uuid && self.locked_by_uuid != current_api_client_authorization.andand.uuid
502       # When locked, progress fields cannot be updated by the wrong
503       # dispatcher, even though it has admin privileges.
504       permitted = permitted - progress_attrs
505     end
506     check_update_whitelist permitted
507   end
508
509   def validate_lock
510     if [Locked, Running].include? self.state
511       # If the Container was already locked, locked_by_uuid must not
512       # changes. Otherwise, the current auth gets the lock.
513       need_lock = locked_by_uuid_was || current_api_client_authorization.andand.uuid
514     else
515       need_lock = nil
516     end
517
518     # The caller can provide a new value for locked_by_uuid, but only
519     # if it's exactly what we expect. This allows a caller to perform
520     # an update like {"state":"Unlocked","locked_by_uuid":null}.
521     if self.locked_by_uuid_changed?
522       if self.locked_by_uuid != need_lock
523         return errors.add :locked_by_uuid, "can only change to #{need_lock}"
524       end
525     end
526     self.locked_by_uuid = need_lock
527   end
528
529   def validate_output
530     # Output must exist and be readable by the current user.  This is so
531     # that a container cannot "claim" a collection that it doesn't otherwise
532     # have access to just by setting the output field to the collection PDH.
533     if output_changed?
534       c = Collection.
535             readable_by(current_user, {include_trash: true}).
536             where(portable_data_hash: self.output).
537             first
538       if !c
539         errors.add :output, "collection must exist and be readable by current user."
540       end
541     end
542   end
543
544   def update_cr_logs
545     # If self.final?, this update is superfluous: the final log/output
546     # update will be done when handle_completed calls finalize! on
547     # each requesting CR.
548     return if self.final? || !self.log_changed?
549     leave_modified_by_user_alone do
550       ContainerRequest.where(container_uuid: self.uuid).each do |cr|
551         cr.update_collections(container: self, collections: ['log'])
552         cr.save!
553       end
554     end
555   end
556
557   def assign_auth
558     if self.auth_uuid_changed?
559          return errors.add :auth_uuid, 'is readonly'
560     end
561     if not [Locked, Running].include? self.state
562       # don't need one
563       self.auth.andand.update_attributes(expires_at: db_current_time)
564       self.auth = nil
565       return
566     elsif self.auth
567       # already have one
568       return
569     end
570     if self.runtime_token.nil?
571       if self.runtime_user_uuid.nil?
572         # legacy behavior, we don't have a runtime_user_uuid so get
573         # the user from the highest priority container request, needed
574         # when performing an upgrade and there are queued containers,
575         # and some tests.
576         cr = ContainerRequest.
577                where('container_uuid=? and priority>0', self.uuid).
578                order('priority desc').
579                first
580         if !cr
581           return errors.add :auth_uuid, "cannot be assigned because priority <= 0"
582         end
583         self.runtime_user_uuid = cr.modified_by_user_uuid
584         self.runtime_auth_scopes = ["all"]
585       end
586
587       # generate a new token
588       self.auth = ApiClientAuthorization.
589                     create!(user_id: User.find_by_uuid(self.runtime_user_uuid).id,
590                             api_client_id: 0,
591                             scopes: self.runtime_auth_scopes)
592     end
593   end
594
595   def sort_serialized_attrs
596     if self.environment_changed?
597       self.environment = self.class.deep_sort_hash(self.environment)
598     end
599     if self.mounts_changed?
600       self.mounts = self.class.deep_sort_hash(self.mounts)
601     end
602     if self.runtime_constraints_changed?
603       self.runtime_constraints = self.class.deep_sort_hash(self.runtime_constraints)
604     end
605     if self.scheduling_parameters_changed?
606       self.scheduling_parameters = self.class.deep_sort_hash(self.scheduling_parameters)
607     end
608     if self.runtime_auth_scopes_changed?
609       self.runtime_auth_scopes = self.runtime_auth_scopes.sort
610     end
611   end
612
613   def update_secret_mounts_md5
614     if self.secret_mounts_changed?
615       self.secret_mounts_md5 = Digest::MD5.hexdigest(
616         SafeJSON.dump(self.class.deep_sort_hash(self.secret_mounts)))
617     end
618   end
619
620   def scrub_secrets
621     # this runs after update_secret_mounts_md5, so the
622     # secret_mounts_md5 will still reflect the secrets that are being
623     # scrubbed here.
624     if self.state_changed? && self.final?
625       self.secret_mounts = {}
626       self.runtime_token = nil
627     end
628   end
629
630   def clear_runtime_status_when_queued
631     # Avoid leaking status messages between different dispatch attempts
632     if self.state_was == Locked && self.state == Queued
633       self.runtime_status = {}
634     end
635   end
636
637   def handle_completed
638     # This container is finished so finalize any associated container requests
639     # that are associated with this container.
640     if self.state_changed? and self.final?
641       act_as_system_user do
642
643         if self.state == Cancelled
644           retryable_requests = ContainerRequest.where("container_uuid = ? and priority > 0 and state = 'Committed' and container_count < container_count_max", uuid)
645         else
646           retryable_requests = []
647         end
648
649         if retryable_requests.any?
650           c_attrs = {
651             command: self.command,
652             cwd: self.cwd,
653             environment: self.environment,
654             output_path: self.output_path,
655             container_image: self.container_image,
656             mounts: self.mounts,
657             runtime_constraints: self.runtime_constraints,
658             scheduling_parameters: self.scheduling_parameters,
659             secret_mounts: self.secret_mounts_was,
660             runtime_token: self.runtime_token_was,
661             runtime_user_uuid: self.runtime_user_uuid,
662             runtime_auth_scopes: self.runtime_auth_scopes
663           }
664           c = Container.create! c_attrs
665           retryable_requests.each do |cr|
666             cr.with_lock do
667               leave_modified_by_user_alone do
668                 # Use row locking because this increments container_count
669                 cr.container_uuid = c.uuid
670                 cr.save!
671               end
672             end
673           end
674         end
675
676         # Notify container requests associated with this container
677         ContainerRequest.where(container_uuid: uuid,
678                                state: ContainerRequest::Committed).each do |cr|
679           leave_modified_by_user_alone do
680             cr.finalize!
681           end
682         end
683
684         # Cancel outstanding container requests made by this container.
685         ContainerRequest.
686           includes(:container).
687           where(requesting_container_uuid: uuid,
688                 state: ContainerRequest::Committed).each do |cr|
689           leave_modified_by_user_alone do
690             cr.update_attributes!(priority: 0)
691             cr.container.reload
692             if cr.container.state == Container::Queued || cr.container.state == Container::Locked
693               # If the child container hasn't started yet, finalize the
694               # child CR now instead of leaving it "on hold", i.e.,
695               # Queued with priority 0.  (OTOH, if the child is already
696               # running, leave it alone so it can get cancelled the
697               # usual way, get a copy of the log collection, etc.)
698               cr.update_attributes!(state: ContainerRequest::Final)
699             end
700           end
701         end
702       end
703     end
704   end
705 end