sql_conds = nil
user_uuids = users_list.map { |u| u.uuid }
+ all_user_uuids = []
# For details on how the trashed_groups table is constructed, see
# see db/migrate/20200501150153_permission_table.rb
exclude_trashed_records = "AND (#{sql_table}.trash_at is NULL or #{sql_table}.trash_at > statement_timestamp())"
end
+ trashed_check = ""
+ if !include_trash && sql_table != "api_client_authorizations"
+ trashed_check = "#{sql_table}.owner_uuid NOT IN (SELECT group_uuid FROM #{TRASHED_GROUPS} " +
+ "where trash_at <= statement_timestamp()) #{exclude_trashed_records}"
+ end
+
if users_list.select { |u| u.is_admin }.any?
# Admin skips most permission checks, but still want to filter on trashed items.
- if !include_trash
- if sql_table != "api_client_authorizations"
- # Only include records where the owner is not trashed
- sql_conds = "#{sql_table}.owner_uuid NOT IN (SELECT group_uuid FROM #{TRASHED_GROUPS} "+
- "where trash_at <= statement_timestamp()) #{exclude_trashed_records}"
- end
+ if !include_trash && sql_table != "api_client_authorizations"
+ # Only include records where the owner is not trashed
+ sql_conds = trashed_check
end
else
- trashed_check = ""
- if !include_trash then
- trashed_check = "AND target_uuid NOT IN (SELECT group_uuid FROM #{TRASHED_GROUPS} where trash_at <= statement_timestamp())"
- end
-
# The core of the permission check is a join against the
# materialized_permissions table to determine if the user has at
# least read permission to either the object itself or its
# A user can have can_manage access to another user, this grants
# full access to all that user's stuff. To implement that we
# need to include those other users in the permission query.
- user_uuids_subquery = USER_UUIDS_SUBQUERY_TEMPLATE % {user: ":user_uuids", perm_level: 1}
+
+ # This was previously implemented by embedding the subquery
+ # directly into the query, but it was discovered later that this
+ # causes the Postgres query planner to do silly things because
+ # the query heuristics assumed the subquery would have a lot
+ # more rows that it does, and choose a bad merge strategy. By
+ # doing the query here and embedding the result as a constant,
+ # Postgres also knows exactly how many items there are and can
+ # choose the right query strategy.
+ #
+ # (note: you could also do this with a temporary table, but that
+ # would require all every request be wrapped in a transaction,
+ # which is not currently the case).
+
+ all_user_uuids = ActiveRecord::Base.connection.exec_query %{
+#{USER_UUIDS_SUBQUERY_TEMPLATE % {user: "'#{user_uuids.join "', '"}'", perm_level: 1}}
+},
+ 'readable_by.user_uuids'
+
+ user_uuids_subquery = ":user_uuids"
# Note: it is possible to combine the direct_check and
- # owner_check into a single EXISTS() clause, however it turns
+ # owner_check into a single IN (SELECT) clause, however it turns
# out query optimizer doesn't like it and forces a sequential
- # table scan. Constructing the query with separate EXISTS()
+ # table scan. Constructing the query with separate IN (SELECT)
# clauses enables it to use the index.
#
# see issue 13208 for details.
# Match a direct read permission link from the user to the record uuid
direct_check = "#{sql_table}.uuid IN (SELECT target_uuid FROM #{PERMISSION_VIEW} "+
- "WHERE user_uuid IN (#{user_uuids_subquery}) AND perm_level >= 1 #{trashed_check})"
+ "WHERE user_uuid IN (#{user_uuids_subquery}) AND perm_level >= 1)"
# Match a read permission for the user to the record's
# owner_uuid. This is so we can have a permissions table that
# other user owns.
owner_check = ""
if sql_table != "api_client_authorizations" and sql_table != "groups" then
- owner_check = "OR #{sql_table}.owner_uuid IN (SELECT target_uuid FROM #{PERMISSION_VIEW} "+
- "WHERE user_uuid IN (#{user_uuids_subquery}) AND perm_level >= 1 #{trashed_check} AND traverse_owned) "
+ owner_check = "#{sql_table}.owner_uuid IN (SELECT target_uuid FROM #{PERMISSION_VIEW} "+
+ "WHERE user_uuid IN (#{user_uuids_subquery}) AND perm_level >= 1 AND traverse_owned) "
+
+ # We want to do owner_check before direct_check in the OR
+ # clause. The order of the OR clause isn't supposed to
+ # matter, but in practice, it does -- apparently in the
+ # absence of other hints, it uses the ordering from the query.
+ # For certain types of queries (like filtering on owner_uuid),
+ # every item will match the owner_check clause, so then
+ # Postgres will optimize out the direct_check entirely.
+ direct_check = " OR " + direct_check
end
links_cond = ""
if sql_table == "links"
- # Match any permission link that gives one of the authorized
- # users some permission _or_ gives anyone else permission to
- # view one of the authorized users.
+ # 1) Match permission links incoming or outgoing on the
+ # user, i.e. granting permission on the user, or granting
+ # permission to the user.
+ #
+ # 2) Match permission links which grant permission on an
+ # object that this user can_manage.
+ #
links_cond = "OR (#{sql_table}.link_class IN (:permission_link_classes) AND "+
- "(#{sql_table}.head_uuid IN (#{user_uuids_subquery}) OR #{sql_table}.tail_uuid IN (#{user_uuids_subquery})))"
+ " ((#{sql_table}.head_uuid IN (#{user_uuids_subquery}) OR #{sql_table}.tail_uuid IN (#{user_uuids_subquery})) OR " +
+ " #{sql_table}.head_uuid IN (SELECT target_uuid FROM #{PERMISSION_VIEW} "+
+ " WHERE user_uuid IN (#{user_uuids_subquery}) AND perm_level >= 3))) "
end
- sql_conds = "(#{direct_check} #{owner_check} #{links_cond}) #{exclude_trashed_records}"
+ sql_conds = "(#{owner_check} #{direct_check} #{links_cond}) #{trashed_check.empty? ? "" : "AND"} #{trashed_check}"
end
end
self.where(sql_conds,
- user_uuids: user_uuids,
- permission_link_classes: ['permission', 'resources'])
+ user_uuids: all_user_uuids.collect{|c| c["target_uuid"]},
+ permission_link_classes: ['permission'])
end
def save_with_unique_name!
nil
end
+ # Fill in implied zero/false values in database records that were
+ # created before #17014 made them explicit, and reset the Rails
+ # "changed" state so the record doesn't appear to have been modified
+ # after loading.
+ #
+ # Invoked by Container and ContainerRequest models as an after_find
+ # hook.
+ def fill_container_defaults_after_find
+ fill_container_defaults
+ set_attribute_was('runtime_constraints', runtime_constraints)
+ set_attribute_was('scheduling_parameters', scheduling_parameters)
+ clear_changes_information
+ end
+
+ # Fill in implied zero/false values. Invoked by ContainerRequest as
+ # a before_validation hook in order to (a) ensure every key has a
+ # value in the updated database record and (b) ensure the attribute
+ # whitelist doesn't reject a change from an explicit zero/false
+ # value in the database to an implicit zero/false value in an update
+ # request.
+ def fill_container_defaults
+ self.runtime_constraints = {
+ 'API' => false,
+ 'cuda' => {
+ 'device_count' => 0,
+ 'driver_version' => '',
+ 'hardware_capability' => '',
+ },
+ 'keep_cache_ram' => 0,
+ 'ram' => 0,
+ 'vcpus' => 0,
+ }.merge(attributes['runtime_constraints'] || {})
+ self.scheduling_parameters = {
+ 'max_run_time' => 0,
+ 'partitions' => [],
+ 'preemptible' => false,
+ }.merge(attributes['scheduling_parameters'] || {})
+ end
+
# ArvadosModel.find_by_uuid needs extra magic to allow it to return
# an object in any class.
def self.find_by_uuid uuid
projects / arvados.git / blobdiff