services/api/app/models/container.rb

   1 require 'whitelist_update'
   2 require 'safe_json'
   3
   4 class Container < ArvadosModel
   5   include HasUuid
   6   include KindAndEtag
   7   include CommonApiTemplate
   8   include WhitelistUpdate
   9   extend CurrentApiClient
  10   extend DbCurrentTime
  11
  12   serialize :environment, Hash
  13   serialize :mounts, Hash
  14   serialize :runtime_constraints, Hash
  15   serialize :command, Array
  16   serialize :scheduling_parameters, Hash
  17
  18   before_validation :fill_field_defaults, :if => :new_record?
  19   before_validation :set_timestamps
  20   validates :command, :container_image, :output_path, :cwd, :priority, :presence => true
  21   validate :validate_state_change
  22   validate :validate_change
  23   validate :validate_lock
  24   validate :validate_output
  25   after_validation :assign_auth
  26   before_save :sort_serialized_attrs
  27   after_save :handle_completed
  28
  29   has_many :container_requests, :foreign_key => :container_uuid, :class_name => 'ContainerRequest', :primary_key => :uuid
  30   belongs_to :auth, :class_name => 'ApiClientAuthorization', :foreign_key => :auth_uuid, :primary_key => :uuid
  31
  32   api_accessible :user, extend: :common do |t|
  33     t.add :command
  34     t.add :container_image
  35     t.add :cwd
  36     t.add :environment
  37     t.add :exit_code
  38     t.add :finished_at
  39     t.add :locked_by_uuid
  40     t.add :log
  41     t.add :mounts
  42     t.add :output
  43     t.add :output_path
  44     t.add :priority
  45     t.add :progress
  46     t.add :runtime_constraints
  47     t.add :started_at
  48     t.add :state
  49     t.add :auth_uuid
  50     t.add :scheduling_parameters
  51   end
  52
  53   # Supported states for a container
  54   States =
  55     [
  56      (Queued = 'Queued'),
  57      (Locked = 'Locked'),
  58      (Running = 'Running'),
  59      (Complete = 'Complete'),
  60      (Cancelled = 'Cancelled')
  61     ]
  62
  63   State_transitions = {
  64     nil => [Queued],
  65     Queued => [Locked, Cancelled],
  66     Locked => [Queued, Running, Cancelled],
  67     Running => [Complete, Cancelled]
  68   }
  69
  70   def self.limit_index_columns_read
  71     ["mounts"]
  72   end
  73
  74   def state_transitions
  75     State_transitions
  76   end
  77
  78   def update_priority!
  79     if [Queued, Locked, Running].include? self.state
  80       # Update the priority of this container to the maximum priority of any of
  81       # its committed container requests and save the record.
  82       self.priority = ContainerRequest.
  83         where(container_uuid: uuid,
  84               state: ContainerRequest::Committed).
  85         maximum('priority')
  86       self.save!
  87     end
  88   end
  89
  90   # Create a new container (or find an existing one) to satisfy the
  91   # given container request.
  92   def self.resolve(req)
  93     c_attrs = {
  94       command: req.command,
  95       cwd: req.cwd,
  96       environment: req.environment,
  97       output_path: req.output_path,
  98       container_image: resolve_container_image(req.container_image),
  99       mounts: resolve_mounts(req.mounts),
 100       runtime_constraints: resolve_runtime_constraints(req.runtime_constraints),
 101       scheduling_parameters: req.scheduling_parameters,
 102     }
 103     act_as_system_user do
 104       if req.use_existing && (reusable = find_reusable(c_attrs))
 105         reusable
 106       else
 107         Container.create!(c_attrs)
 108       end
 109     end
 110   end
 111
 112   # Return a runtime_constraints hash that complies with requested but
 113   # is suitable for saving in a container record, i.e., has specific
 114   # values instead of ranges.
 115   #
 116   # Doing this as a step separate from other resolutions, like "git
 117   # revision range to commit hash", makes sense only when there is no
 118   # opportunity to reuse an existing container (e.g., container reuse
 119   # is not implemented yet, or we have already found that no existing
 120   # containers are suitable).
 121   def self.resolve_runtime_constraints(runtime_constraints)
 122     rc = {}
 123     defaults = {
 124       'keep_cache_ram' =>
 125       Rails.configuration.container_default_keep_cache_ram,
 126     }
 127     defaults.merge(runtime_constraints).each do |k, v|
 128       if v.is_a? Array
 129         rc[k] = v[0]
 130       else
 131         rc[k] = v
 132       end
 133     end
 134     rc
 135   end
 136
 137   # Return a mounts hash suitable for a Container, i.e., with every
 138   # readonly collection UUID resolved to a PDH.
 139   def self.resolve_mounts(mounts)
 140     c_mounts = {}
 141     mounts.each do |k, mount|
 142       mount = mount.dup
 143       c_mounts[k] = mount
 144       if mount['kind'] != 'collection'
 145         next
 146       end
 147       if (uuid = mount.delete 'uuid')
 148         c = Collection.
 149           readable_by(current_user).
 150           where(uuid: uuid).
 151           select(:portable_data_hash).
 152           first
 153         if !c
 154           raise ArvadosModel::UnresolvableContainerError.new "cannot mount collection #{uuid.inspect}: not found"
 155         end
 156         if mount['portable_data_hash'].nil?
 157           # PDH not supplied by client
 158           mount['portable_data_hash'] = c.portable_data_hash
 159         elsif mount['portable_data_hash'] != c.portable_data_hash
 160           # UUID and PDH supplied by client, but they don't agree
 161           raise ArgumentError.new "cannot mount collection #{uuid.inspect}: current portable_data_hash #{c.portable_data_hash.inspect} does not match #{c['portable_data_hash'].inspect} in request"
 162         end
 163       end
 164     end
 165     return c_mounts
 166   end
 167
 168   # Return a container_image PDH suitable for a Container.
 169   def self.resolve_container_image(container_image)
 170     coll = Collection.for_latest_docker_image(container_image)
 171     if !coll
 172       raise ArvadosModel::UnresolvableContainerError.new "docker image #{container_image.inspect} not found"
 173     end
 174     coll.portable_data_hash
 175   end
 176
 177   def self.find_reusable(attrs)
 178     candidates = Container.
 179       where_serialized(:command, attrs[:command]).
 180       where('cwd = ?', attrs[:cwd]).
 181       where_serialized(:environment, attrs[:environment]).
 182       where('output_path = ?', attrs[:output_path]).
 183       where('container_image = ?', resolve_container_image(attrs[:container_image])).
 184       where_serialized(:mounts, resolve_mounts(attrs[:mounts])).
 185       where_serialized(:runtime_constraints, resolve_runtime_constraints(attrs[:runtime_constraints]))
 186
 187     # Check for Completed candidates whose output and log are both readable.
 188     select_readable_pdh = Collection.
 189       readable_by(current_user).
 190       select(:portable_data_hash).
 191       to_sql
 192     usable = candidates.
 193       where(state: Complete).
 194       where(exit_code: 0).
 195       where("log IN (#{select_readable_pdh})").
 196       where("output IN (#{select_readable_pdh})").
 197       order('finished_at ASC').
 198       limit(1).
 199       first
 200     return usable if usable
 201
 202     # Check for Running candidates and return the most likely to finish sooner.
 203     running = candidates.where(state: Running).
 204       order('progress desc, started_at asc').limit(1).first
 205     return running if not running.nil?
 206
 207     # Check for Locked or Queued ones and return the most likely to start first.
 208     locked_or_queued = candidates.where("state IN (?)", [Locked, Queued]).
 209       order('state asc, priority desc, created_at asc').limit(1).first
 210     return locked_or_queued if not locked_or_queued.nil?
 211
 212     # No suitable candidate found.
 213     nil
 214   end
 215
 216   def check_lock_fail
 217     if self.state != Queued
 218       raise LockFailedError.new("cannot lock when #{self.state}")
 219     elsif self.priority <= 0
 220       raise LockFailedError.new("cannot lock when priority<=0")
 221     end
 222   end
 223
 224   def lock
 225     # Check invalid state transitions once before getting the lock
 226     # (because it's cheaper that way) and once after getting the lock
 227     # (because state might have changed while acquiring the lock).
 228     check_lock_fail
 229     begin
 230       reload(lock: 'FOR UPDATE NOWAIT')
 231     rescue
 232       raise LockFailedError.new("cannot lock: other transaction in progress")
 233     end
 234     check_lock_fail
 235     update_attributes!(state: Locked)
 236   end
 237
 238   def check_unlock_fail
 239     if self.state != Locked
 240       raise InvalidStateTransitionError.new("cannot unlock when #{self.state}")
 241     elsif self.locked_by_uuid != current_api_client_authorization.uuid
 242       raise InvalidStateTransitionError.new("locked by a different token")
 243     end
 244   end
 245
 246   def unlock
 247     # Check invalid state transitions twice (see lock)
 248     check_unlock_fail
 249     reload(lock: 'FOR UPDATE')
 250     check_unlock_fail
 251     update_attributes!(state: Queued)
 252   end
 253
 254   def self.readable_by(*users_list)
 255     if users_list.select { |u| u.is_admin }.any?
 256       return self
 257     end
 258     user_uuids = users_list.map { |u| u.uuid }
 259     uuid_list = user_uuids + users_list.flat_map { |u| u.groups_i_can(:read) }
 260     uuid_list.uniq!
 261     permitted = "(SELECT head_uuid FROM links WHERE link_class='permission' AND tail_uuid IN (:uuids))"
 262     joins(:container_requests).
 263       where("container_requests.uuid IN #{permitted} OR "+
 264             "container_requests.owner_uuid IN (:uuids)",
 265             uuids: uuid_list)
 266   end
 267
 268   def final?
 269     [Complete, Cancelled].include?(self.state)
 270   end
 271
 272   protected
 273
 274   def fill_field_defaults
 275     self.state ||= Queued
 276     self.environment ||= {}
 277     self.runtime_constraints ||= {}
 278     self.mounts ||= {}
 279     self.cwd ||= "."
 280     self.priority ||= 1
 281     self.scheduling_parameters ||= {}
 282   end
 283
 284   def permission_to_create
 285     current_user.andand.is_admin
 286   end
 287
 288   def permission_to_update
 289     # Override base permission check to allow auth_uuid to set progress and
 290     # output (only).  Whether it is legal to set progress and output in the current
 291     # state has already been checked in validate_change.
 292     current_user.andand.is_admin ||
 293       (!current_api_client_authorization.nil? and
 294        [self.auth_uuid, self.locked_by_uuid].include? current_api_client_authorization.uuid)
 295   end
 296
 297   def ensure_owner_uuid_is_permitted
 298     # Override base permission check to allow auth_uuid to set progress and
 299     # output (only).  Whether it is legal to set progress and output in the current
 300     # state has already been checked in validate_change.
 301     if !current_api_client_authorization.nil? and self.auth_uuid == current_api_client_authorization.uuid
 302       check_update_whitelist [:progress, :output]
 303     else
 304       super
 305     end
 306   end
 307
 308   def set_timestamps
 309     if self.state_changed? and self.state == Running
 310       self.started_at ||= db_current_time
 311     end
 312
 313     if self.state_changed? and [Complete, Cancelled].include? self.state
 314       self.finished_at ||= db_current_time
 315     end
 316   end
 317
 318   def validate_change
 319     permitted = [:state]
 320
 321     if self.new_record?
 322       permitted.push(:owner_uuid, :command, :container_image, :cwd,
 323                      :environment, :mounts, :output_path, :priority,
 324                      :runtime_constraints, :scheduling_parameters)
 325     end
 326
 327     case self.state
 328     when Queued, Locked
 329       permitted.push :priority
 330
 331     when Running
 332       permitted.push :priority, :progress, :output
 333       if self.state_changed?
 334         permitted.push :started_at
 335       end
 336
 337     when Complete
 338       if self.state_was == Running
 339         permitted.push :finished_at, :output, :log, :exit_code
 340       end
 341
 342     when Cancelled
 343       case self.state_was
 344       when Running
 345         permitted.push :finished_at, :output, :log
 346       when Queued, Locked
 347         permitted.push :finished_at
 348       end
 349
 350     else
 351       # The state_transitions check will add an error message for this
 352       return false
 353     end
 354
 355     check_update_whitelist permitted
 356   end
 357
 358   def validate_lock
 359     if [Locked, Running].include? self.state
 360       # If the Container was already locked, locked_by_uuid must not
 361       # changes. Otherwise, the current auth gets the lock.
 362       need_lock = locked_by_uuid_was || current_api_client_authorization.andand.uuid
 363     else
 364       need_lock = nil
 365     end
 366
 367     # The caller can provide a new value for locked_by_uuid, but only
 368     # if it's exactly what we expect. This allows a caller to perform
 369     # an update like {"state":"Unlocked","locked_by_uuid":null}.
 370     if self.locked_by_uuid_changed?
 371       if self.locked_by_uuid != need_lock
 372         return errors.add :locked_by_uuid, "can only change to #{need_lock}"
 373       end
 374     end
 375     self.locked_by_uuid = need_lock
 376   end
 377
 378   def validate_output
 379     # Output must exist and be readable by the current user.  This is so
 380     # that a container cannot "claim" a collection that it doesn't otherwise
 381     # have access to just by setting the output field to the collection PDH.
 382     if output_changed?
 383       c = Collection.unscoped do
 384         Collection.
 385             readable_by(current_user).
 386             where(portable_data_hash: self.output).
 387             first
 388       end
 389       if !c
 390         errors.add :output, "collection must exist and be readable by current user."
 391       end
 392     end
 393   end
 394
 395   def assign_auth
 396     if self.auth_uuid_changed?
 397       return errors.add :auth_uuid, 'is readonly'
 398     end
 399     if not [Locked, Running].include? self.state
 400       # don't need one
 401       self.auth.andand.update_attributes(expires_at: db_current_time)
 402       self.auth = nil
 403       return
 404     elsif self.auth
 405       # already have one
 406       return
 407     end
 408     cr = ContainerRequest.
 409       where('container_uuid=? and priority>0', self.uuid).
 410       order('priority desc').
 411       first
 412     if !cr
 413       return errors.add :auth_uuid, "cannot be assigned because priority <= 0"
 414     end
 415     self.auth = ApiClientAuthorization.
 416       create!(user_id: User.find_by_uuid(cr.modified_by_user_uuid).id,
 417               api_client_id: 0)
 418   end
 419
 420   def sort_serialized_attrs
 421     if self.environment_changed?
 422       self.environment = self.class.deep_sort_hash(self.environment)
 423     end
 424     if self.mounts_changed?
 425       self.mounts = self.class.deep_sort_hash(self.mounts)
 426     end
 427     if self.runtime_constraints_changed?
 428       self.runtime_constraints = self.class.deep_sort_hash(self.runtime_constraints)
 429     end
 430     if self.scheduling_parameters_changed?
 431       self.scheduling_parameters = self.class.deep_sort_hash(self.scheduling_parameters)
 432     end
 433   end
 434
 435   def handle_completed
 436     # This container is finished so finalize any associated container requests
 437     # that are associated with this container.
 438     if self.state_changed? and self.final?
 439       act_as_system_user do
 440
 441         if self.state == Cancelled
 442           retryable_requests = ContainerRequest.where("container_uuid = ? and priority > 0 and state = 'Committed' and container_count < container_count_max", uuid)
 443         else
 444           retryable_requests = []
 445         end
 446
 447         if retryable_requests.any?
 448           c_attrs = {
 449             command: self.command,
 450             cwd: self.cwd,
 451             environment: self.environment,
 452             output_path: self.output_path,
 453             container_image: self.container_image,
 454             mounts: self.mounts,
 455             runtime_constraints: self.runtime_constraints,
 456             scheduling_parameters: self.scheduling_parameters
 457           }
 458           c = Container.create! c_attrs
 459           retryable_requests.each do |cr|
 460             cr.with_lock do
 461               # Use row locking because this increments container_count
 462               cr.container_uuid = c.uuid
 463               cr.save
 464             end
 465           end
 466         end
 467
 468         # Notify container requests associated with this container
 469         ContainerRequest.where(container_uuid: uuid,
 470                                state: ContainerRequest::Committed).each do |cr|
 471           cr.finalize!
 472         end
 473
 474         # Try to cancel any outstanding container requests made by this container.
 475         ContainerRequest.where(requesting_container_uuid: uuid,
 476                                state: ContainerRequest::Committed).each do |cr|
 477           cr.priority = 0
 478           cr.save
 479         end
 480       end
 481     end
 482   end
 483
 484 end