// This test verifies that persistent volumes are unmounted properly
// after a checkpointed framework disappears and the slave restarts.
//
// TODO(jieyu): Even though the command task specifies a new
// filesystem root, the executor (command executor) itself does not
// change filesystem root (uses the host filesystem). We need to add a
// test to test the scenario that the executor itself changes rootfs.
TEST_F(LinuxFilesystemIsolatorMesosTest,
ROOT_RecoverOrphanedPersistentVolume)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
string registry = path::join(sandbox.get(), "registry");
AWAIT_READY(DockerArchive::create(registry, "test_image"));
slave::Flags flags = CreateSlaveFlags();
flags.resources = "cpus:2;mem:1024;disk(role1):1024";
flags.isolation = "filesystem/linux,docker/runtime";
flags.docker_registry = registry;
flags.docker_store_dir = path::join(sandbox.get(), "store");
flags.image_providers = "docker";
Fetcher fetcher(flags);
Try<MesosContainerizer*> create =
MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(create);
Owned<Containerizer> containerizer(create.get());
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(
detector.get(),
containerizer.get(),
flags);
ASSERT_SOME(slave);
MockScheduler sched;
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_roles(0, "role1");
frameworkInfo.set_checkpoint(true);
MesosSchedulerDriver driver(
&sched,
frameworkInfo,
master.get()->pid,
DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
Offer offer = offers.get()[0];
string dir1 = path::join(sandbox.get(), "dir1");
ASSERT_SOME(os::mkdir(dir1));
Resource persistentVolume = createPersistentVolume(
Megabytes(64),
"role1",
"id1",
"path1",
None(),
None(),
frameworkInfo.principal());
// Create a task that does nothing for a long time.
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:512").get() + persistentVolume,
"sleep 1000");
task.mutable_container()->CopyFrom(createContainerInfo(
"test_image",
{createVolumeHostPath("/tmp", dir1, Volume::RW)}));
Future<TaskStatus> statusStarting;
Future<TaskStatus> statusRunning;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&statusStarting))
.WillOnce(FutureArg<1>(&statusRunning))
.WillRepeatedly(DoDefault());
Future<Nothing> ack =
FUTURE_DISPATCH(_, &Slave::_statusUpdateAcknowledgement);
// Create the persistent volumes and launch task via `acceptOffers`.
driver.acceptOffers(
{offer.id()},
{CREATE(persistentVolume), LAUNCH({task})});
AWAIT_READY(statusStarting);
EXPECT_EQ(TASK_STARTING, statusStarting->state());
AWAIT_READY(statusRunning);
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
// Wait for the ACK to be checkpointed.
AWAIT_READY(ack);
Future<hashset<ContainerID>> containers = containerizer->containers();
AWAIT_READY(containers);
ASSERT_EQ(1u, containers->size());
ContainerID containerId = *containers->begin();
// Restart the slave.
slave.get()->terminate();
// Wipe the slave meta directory so that the slave will treat the
// above running task as an orphan.
ASSERT_SOME(os::rmdir(slave::paths::getMetaRootDir(flags.work_dir)));
Future<Nothing> _recover = FUTURE_DISPATCH(_, &Slave::_recover);
// Recreate the containerizer using the same helper as above.
containerizer.reset();
create = MesosContainerizer::create(flags, true, &fetcher);
ASSERT_SOME(create);
containerizer.reset(create.get());
slave = StartSlave(detector.get(), containerizer.get(), flags);
ASSERT_SOME(slave);
// Wait until slave recovery is complete.
AWAIT_READY(_recover);
// Wait until the orphan containers are cleaned up.
AWAIT_READY(containerizer->wait(containerId));
Try<fs::MountInfoTable> table = fs::MountInfoTable::read();
ASSERT_SOME(table);
// All mount targets should be under this directory.
string directory = slave::paths::getSandboxRootDir(flags.work_dir);
// Verify that the orphaned container's persistent volume and
// the rootfs are unmounted.
foreach (const fs::MountInfoTable::Entry& entry, table->entries) {
EXPECT_FALSE(strings::contains(entry.target, directory))
<< "Target was not unmounted: " << entry.target;
}
driver.stop();
driver.join();
}
// This test verifies that, after a master failover, reconciliation of an
// operation that is still pending on an agent results in `OPERATION_PENDING`.
TEST_P(OperationReconciliationTest, AgentPendingOperationAfterMasterFailover)
{
Clock::pause();
mesos::internal::master::Flags masterFlags = CreateMasterFlags();
Try<Owned<cluster::Master>> master = StartMaster(masterFlags);
ASSERT_SOME(master);
Future<UpdateSlaveMessage> updateSlaveMessage =
FUTURE_PROTOBUF(UpdateSlaveMessage(), _, _);
auto detector = std::make_shared<StandaloneMasterDetector>(master.get()->pid);
mesos::internal::slave::Flags slaveFlags = CreateSlaveFlags();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get(), slaveFlags);
ASSERT_SOME(slave);
// Advance the clock to trigger agent registration.
Clock::advance(slaveFlags.registration_backoff_factor);
// Wait for the agent to register.
AWAIT_READY(updateSlaveMessage);
// Start and register a resource provider.
ResourceProviderInfo resourceProviderInfo;
resourceProviderInfo.set_type("org.apache.mesos.rp.test");
resourceProviderInfo.set_name("test");
Resource disk = createDiskResource(
"200", "*", None(), None(), createDiskSourceRaw(None(), "profile"));
Owned<MockResourceProvider> resourceProvider(
new MockResourceProvider(
resourceProviderInfo,
Resources(disk)));
// We override the mock resource provider's default action, so the operation
// will stay in `OPERATION_PENDING`.
Future<resource_provider::Event::ApplyOperation> applyOperation;
EXPECT_CALL(*resourceProvider, applyOperation(_))
.WillOnce(FutureArg<0>(&applyOperation));
Owned<EndpointDetector> endpointDetector(
mesos::internal::tests::resource_provider::createEndpointDetector(
slave.get()->pid));
updateSlaveMessage = FUTURE_PROTOBUF(UpdateSlaveMessage(), _, _);
// NOTE: We need to resume the clock so that the resource provider can
// fully register.
Clock::resume();
ContentType contentType = GetParam();
resourceProvider->start(endpointDetector, contentType);
// Wait until the agent's resources have been updated to include the
// resource provider resources.
AWAIT_READY(updateSlaveMessage);
ASSERT_TRUE(updateSlaveMessage->has_resource_providers());
ASSERT_EQ(1, updateSlaveMessage->resource_providers().providers_size());
Clock::pause();
// Start a v1 framework.
auto scheduler = std::make_shared<MockHTTPScheduler>();
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_roles(0, DEFAULT_TEST_ROLE);
EXPECT_CALL(*scheduler, connected(_))
.WillOnce(scheduler::SendSubscribe(frameworkInfo));
Future<scheduler::Event::Subscribed> subscribed;
EXPECT_CALL(*scheduler, subscribed(_, _))
.WillOnce(FutureArg<1>(&subscribed));
// Ignore heartbeats.
EXPECT_CALL(*scheduler, heartbeat(_))
.WillRepeatedly(Return());
// Decline offers that do not contain wanted resources.
EXPECT_CALL(*scheduler, offers(_, _))
.WillRepeatedly(scheduler::DeclineOffers());
Future<scheduler::Event::Offers> offers;
auto isRaw = [](const Resource& r) {
return r.has_disk() &&
r.disk().has_source() &&
r.disk().source().type() == Resource::DiskInfo::Source::RAW;
};
EXPECT_CALL(*scheduler, offers(_, scheduler::OffersHaveAnyResource(
std::bind(isRaw, lambda::_1))))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(scheduler::DeclineOffers()); // Decline successive offers.
scheduler::TestMesos mesos(
master.get()->pid, contentType, scheduler, detector);
AWAIT_READY(subscribed);
FrameworkID frameworkId(subscribed->framework_id());
// NOTE: If the framework has not declined an unwanted offer yet when
// the master updates the agent with the RAW disk resource, the new
// allocation triggered by this update won't generate an allocatable
// offer due to no CPU and memory resources. So here we first settle
// the clock to ensure that the unwanted offer has been declined, then
// advance the clock to trigger another allocation.
Clock::settle();
Clock::advance(masterFlags.allocation_interval);
AWAIT_READY(offers);
ASSERT_FALSE(offers->offers().empty());
const Offer& offer = offers->offers(0);
const AgentID& agentId = offer.agent_id();
Option<Resource> source;
Option<ResourceProviderID> resourceProviderId;
foreach (const Resource& resource, offer.resources()) {
if (isRaw(resource)) {
source = resource;
ASSERT_TRUE(resource.has_provider_id());
resourceProviderId = resource.provider_id();
break;
}
}
ASSERT_SOME(source);
ASSERT_SOME(resourceProviderId);
OperationID operationId;
operationId.set_value("operation");
mesos.send(createCallAccept(
frameworkId,
offer,
{CREATE_DISK(
source.get(),
Resource::DiskInfo::Source::MOUNT,
None(),
operationId)}));
AWAIT_READY(applyOperation);
// Simulate master failover.
EXPECT_CALL(*scheduler, disconnected(_));
detector->appoint(None());
master->reset();
master = StartMaster();
ASSERT_SOME(master);
// Settle the clock to ensure the master finishes recovering the registry.
Clock::settle();
Future<SlaveReregisteredMessage> slaveReregistered = FUTURE_PROTOBUF(
SlaveReregisteredMessage(), master.get()->pid, slave.get()->pid);
updateSlaveMessage = FUTURE_PROTOBUF(UpdateSlaveMessage(), _, _);
EXPECT_CALL(*scheduler, connected(_))
.WillOnce(scheduler::SendSubscribe(frameworkInfo, frameworkId));
Future<scheduler::Event::Subscribed> frameworkResubscribed;
EXPECT_CALL(*scheduler, subscribed(_, _))
.WillOnce(FutureArg<1>(&frameworkResubscribed));
// Simulate a new master detected event to the agent and the scheduler.
detector->appoint(master.get()->pid);
// Advance the clock, so that the agent re-registers.
Clock::advance(slaveFlags.registration_backoff_factor);
// Resume the clock to avoid deadlocks related to agent registration.
// See MESOS-8828.
Clock::resume();
// Wait for the framework and agent to re-register.
AWAIT_READY(slaveReregistered);
AWAIT_READY(updateSlaveMessage);
AWAIT_READY(frameworkResubscribed);
Clock::pause();
// Test explicit reconciliation
{
scheduler::Call::ReconcileOperations::Operation operation;
operation.mutable_operation_id()->CopyFrom(operationId);
operation.mutable_agent_id()->CopyFrom(agentId);
const Future<scheduler::APIResult> result =
mesos.call({createCallReconcileOperations(frameworkId, {operation})});
AWAIT_READY(result);
// The master should respond with '200 OK' and with a `scheduler::Response`.
ASSERT_EQ(process::http::Status::OK, result->status_code());
ASSERT_TRUE(result->has_response());
const scheduler::Response response = result->response();
ASSERT_EQ(scheduler::Response::RECONCILE_OPERATIONS, response.type());
ASSERT_TRUE(response.has_reconcile_operations());
const scheduler::Response::ReconcileOperations& reconcile =
response.reconcile_operations();
ASSERT_EQ(1, reconcile.operation_statuses_size());
const OperationStatus& operationStatus = reconcile.operation_statuses(0);
EXPECT_EQ(operationId, operationStatus.operation_id());
EXPECT_EQ(OPERATION_PENDING, operationStatus.state());
EXPECT_FALSE(operationStatus.has_uuid());
}
// Test implicit reconciliation
{
const Future<scheduler::APIResult> result =
mesos.call({createCallReconcileOperations(frameworkId, {})});
AWAIT_READY(result);
// The master should respond with '200 OK' and with a `scheduler::Response`.
ASSERT_EQ(process::http::Status::OK, result->status_code());
ASSERT_TRUE(result->has_response());
const scheduler::Response response = result->response();
ASSERT_EQ(scheduler::Response::RECONCILE_OPERATIONS, response.type());
ASSERT_TRUE(response.has_reconcile_operations());
const scheduler::Response::ReconcileOperations& reconcile =
response.reconcile_operations();
ASSERT_EQ(1, reconcile.operation_statuses_size());
const OperationStatus& operationStatus = reconcile.operation_statuses(0);
EXPECT_EQ(operationId, operationStatus.operation_id());
EXPECT_EQ(OPERATION_PENDING, operationStatus.state());
EXPECT_FALSE(operationStatus.has_uuid());
}
}
// This test verifies that the framework can launch a command task
// that specifies both container image and persistent volumes.
TEST_F(LinuxFilesystemIsolatorMesosTest,
ROOT_ChangeRootFilesystemCommandExecutorPersistentVolume)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
string registry = path::join(sandbox.get(), "registry");
AWAIT_READY(DockerArchive::create(registry, "test_image"));
slave::Flags flags = CreateSlaveFlags();
flags.resources = "cpus:2;mem:1024;disk(role1):1024";
flags.isolation = "filesystem/linux,docker/runtime";
flags.docker_registry = registry;
flags.docker_store_dir = path::join(sandbox.get(), "store");
flags.image_providers = "docker";
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get(), flags);
ASSERT_SOME(slave);
MockScheduler sched;
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_roles(0, "role1");
MesosSchedulerDriver driver(
&sched,
frameworkInfo,
master.get()->pid,
DEFAULT_CREDENTIAL);
Future<FrameworkID> frameworkId;
EXPECT_CALL(sched, registered(&driver, _, _))
.WillOnce(FutureArg<1>(&frameworkId));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(frameworkId);
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
Offer offer = offers.get()[0];
string dir1 = path::join(sandbox.get(), "dir1");
ASSERT_SOME(os::mkdir(dir1));
Resource persistentVolume = createPersistentVolume(
Megabytes(64),
"role1",
"id1",
"path1",
None(),
None(),
frameworkInfo.principal());
// We use the filter explicitly here so that the resources will not
// be filtered for 5 seconds (the default).
Filters filters;
filters.set_refuse_seconds(0);
TaskInfo task = createTask(
offer.slave_id(),
Resources::parse("cpus:1;mem:512").get() + persistentVolume,
"echo abc > path1/file");
task.mutable_container()->CopyFrom(createContainerInfo(
"test_image",
{createVolumeHostPath("/tmp", dir1, Volume::RW)}));
// Create the persistent volumes and launch task via `acceptOffers`.
driver.acceptOffers(
{offer.id()},
{CREATE(persistentVolume), LAUNCH({task})},
filters);
Future<TaskStatus> statusStarting;
Future<TaskStatus> statusRunning;
Future<TaskStatus> statusFinished;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&statusStarting))
.WillOnce(FutureArg<1>(&statusRunning))
.WillOnce(FutureArg<1>(&statusFinished));
AWAIT_READY(statusStarting);
EXPECT_EQ(TASK_STARTING, statusStarting->state());
AWAIT_READY(statusRunning);
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
AWAIT_READY(statusFinished);
EXPECT_EQ(TASK_FINISHED, statusFinished->state());
// NOTE: The command executor's id is the same as the task id.
ExecutorID executorId;
executorId.set_value(task.task_id().value());
string directory = slave::paths::getExecutorLatestRunPath(
flags.work_dir,
offer.slave_id(),
frameworkId.get(),
executorId);
EXPECT_FALSE(os::exists(path::join(directory, "path1")));
string volumePath = slave::paths::getPersistentVolumePath(
flags.work_dir,
"role1",
"id1");
EXPECT_SOME_EQ("abc\n", os::read(path::join(volumePath, "file")));
driver.stop();
driver.join();
}
// Tests that the task fails when it attempts to write to a persistent volume
// mounted as read-only. Note that although we use a shared persistent volume,
// the behavior is the same for non-shared persistent volumes.
TEST_F(LinuxFilesystemIsolatorMesosTest,
ROOT_WriteAccessSharedPersistentVolumeReadOnlyMode)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
string registry = path::join(sandbox.get(), "registry");
AWAIT_READY(DockerArchive::create(registry, "test_image"));
slave::Flags flags = CreateSlaveFlags();
flags.resources = "cpus:2;mem:128;disk(role1):128";
flags.isolation = "filesystem/linux,docker/runtime";
flags.docker_registry = registry;
flags.docker_store_dir = path::join(sandbox.get(), "store");
flags.image_providers = "docker";
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get(), flags);
ASSERT_SOME(slave);
MockScheduler sched;
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_roles(0, "role1");
frameworkInfo.add_capabilities()->set_type(
FrameworkInfo::Capability::SHARED_RESOURCES);
MesosSchedulerDriver driver(
&sched,
frameworkInfo,
master.get()->pid,
DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
// We create a shared volume which shall be used by the task to
// write to that volume.
Resource volume = createPersistentVolume(
Megabytes(4),
"role1",
"id1",
"volume_path",
None(),
None(),
frameworkInfo.principal(),
true); // Shared volume.
// The task uses the shared volume as read-only.
Resource roVolume = volume;
roVolume.mutable_disk()->mutable_volume()->set_mode(Volume::RO);
Resources taskResources =
Resources::parse("cpus:1;mem:64;disk(role1):1").get() + roVolume;
TaskInfo task = createTask(
offers.get()[0].slave_id(),
taskResources,
"echo hello > volume_path/file");
// The task fails to write to the volume since the task's resources
// intends to use the volume as read-only.
Future<TaskStatus> statusStarting;
Future<TaskStatus> statusRunning;
Future<TaskStatus> statusFailed;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&statusStarting))
.WillOnce(FutureArg<1>(&statusRunning))
.WillOnce(FutureArg<1>(&statusFailed));
driver.acceptOffers(
{offers.get()[0].id()},
{CREATE(volume),
LAUNCH({task})});
AWAIT_READY(statusStarting);
EXPECT_EQ(task.task_id(), statusStarting->task_id());
EXPECT_EQ(TASK_STARTING, statusStarting->state());
AWAIT_READY(statusRunning);
EXPECT_EQ(task.task_id(), statusRunning->task_id());
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
AWAIT_READY(statusFailed);
EXPECT_EQ(task.task_id(), statusFailed->task_id());
EXPECT_EQ(TASK_FAILED, statusFailed->state());
driver.stop();
driver.join();
}
// This test verifies that the volume usage accounting for sandboxes
// with bind-mounted volumes (while linux filesystem isolator is used)
// works correctly by creating a file within the volume the size of
// which exceeds the sandbox quota.
TEST_F(LinuxFilesystemIsolatorMesosTest,
ROOT_VolumeUsageExceedsSandboxQuota)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
string registry = path::join(sandbox.get(), "registry");
AWAIT_READY(DockerArchive::create(registry, "test_image"));
slave::Flags flags = CreateSlaveFlags();
flags.resources = "cpus:2;mem:128;disk(role1):128";
flags.isolation = "disk/du,filesystem/linux,docker/runtime";
flags.docker_registry = registry;
flags.docker_store_dir = path::join(sandbox.get(), "store");
flags.image_providers = "docker";
// NOTE: We can't pause the clock because we need the reaper to reap
// the 'du' subprocess.
flags.container_disk_watch_interval = Milliseconds(1);
flags.enforce_container_disk_quota = true;
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = StartSlave(detector.get(), flags);
ASSERT_SOME(slave);
MockScheduler sched;
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_roles(0, "role1");
MesosSchedulerDriver driver(
&sched,
frameworkInfo,
master.get()->pid,
DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
ASSERT_FALSE(offers->empty());
// We request a sandbox (1MB) that is smaller than the persistent
// volume (4MB) and attempt to create a file in that volume that is
// twice the size of the sanbox (2MB).
Resources volume = createPersistentVolume(
Megabytes(4),
"role1",
"id1",
"volume_path",
None(),
None(),
frameworkInfo.principal());
Resources taskResources =
Resources::parse("cpus:1;mem:64;disk(role1):1").get() + volume;
// We sleep to give quota enforcement (du) a chance to kick in.
TaskInfo task = createTask(
offers.get()[0].slave_id(),
taskResources,
"dd if=/dev/zero of=volume_path/file bs=1048576 count=2 && sleep 1");
Future<TaskStatus> statusStarting;
Future<TaskStatus> statusRunning;
Future<TaskStatus> statusFinished;
EXPECT_CALL(sched, statusUpdate(&driver, _))
.WillOnce(FutureArg<1>(&statusStarting))
.WillOnce(FutureArg<1>(&statusRunning))
.WillOnce(FutureArg<1>(&statusFinished));
driver.acceptOffers(
{offers.get()[0].id()},
{CREATE(volume),
LAUNCH({task})});
AWAIT_READY(statusStarting);
EXPECT_EQ(task.task_id(), statusStarting->task_id());
EXPECT_EQ(TASK_STARTING, statusStarting->state());
AWAIT_READY(statusRunning);
EXPECT_EQ(task.task_id(), statusRunning->task_id());
EXPECT_EQ(TASK_RUNNING, statusRunning->state());
AWAIT_READY(statusFinished);
EXPECT_EQ(task.task_id(), statusFinished->task_id());
EXPECT_EQ(TASK_FINISHED, statusFinished->state());
driver.stop();
driver.join();
}
// This test verifies that authorization based endpoint filtering
// works correctly on the /state endpoint.
// Both default users are allowed to view high level frameworks, but only
// one is allowed to view the tasks.
// After launching a single task per each framework, one for role "superhero"
// and the other for role "muggle", this test verifies that each of two
// default users can view resource allocations and resource reservations for
// corresponding allowed roles only.
TYPED_TEST(SlaveAuthorizerTest, FilterStateEndpoint)
{
ACLs acls;
const string roleSuperhero = "superhero";
const string roleMuggle = "muggle";
{
// Default principal can see all frameworks.
mesos::ACL::ViewFramework* acl = acls.add_view_frameworks();
acl->mutable_principals()->add_values(DEFAULT_CREDENTIAL.principal());
acl->mutable_users()->set_type(ACL::Entity::ANY);
}
{
// Second default principal can see all frameworks.
mesos::ACL::ViewFramework* acl = acls.add_view_frameworks();
acl->mutable_principals()->add_values(DEFAULT_CREDENTIAL_2.principal());
acl->mutable_users()->set_type(ACL::Entity::ANY);
}
{
// No other principal can see frameworks running under any user.
ACL::ViewFramework* acl = acls.add_view_frameworks();
acl->mutable_principals()->set_type(ACL::Entity::ANY);
acl->mutable_users()->set_type(ACL::Entity::NONE);
}
{
// Default principal can see all executors.
mesos::ACL::ViewExecutor* acl = acls.add_view_executors();
acl->mutable_principals()->add_values(DEFAULT_CREDENTIAL.principal());
acl->mutable_users()->set_type(ACL::Entity::ANY);
}
{
// No other principal can see executors running under any user.
ACL::ViewExecutor* acl = acls.add_view_executors();
acl->mutable_principals()->set_type(ACL::Entity::ANY);
acl->mutable_users()->set_type(ACL::Entity::NONE);
}
{
// Default principal can see all tasks.
mesos::ACL::ViewTask* acl = acls.add_view_tasks();
acl->mutable_principals()->add_values(DEFAULT_CREDENTIAL.principal());
acl->mutable_users()->set_type(ACL::Entity::ANY);
}
{
// No other principal can see tasks running under any user.
ACL::ViewTask* acl = acls.add_view_tasks();
acl->mutable_principals()->set_type(ACL::Entity::ANY);
acl->mutable_users()->set_type(ACL::Entity::NONE);
}
{
// Default principal can view "superhero" role only.
ACL::ViewRole* acl = acls.add_view_roles();
acl->mutable_principals()->add_values(DEFAULT_CREDENTIAL.principal());
acl->mutable_roles()->add_values(roleSuperhero);
acl = acls.add_view_roles();
acl->mutable_principals()->add_values(DEFAULT_CREDENTIAL.principal());
acl->mutable_roles()->set_type(mesos::ACL::Entity::NONE);
}
{
// Second default principal can view "muggle" role only.
ACL::ViewRole* acl = acls.add_view_roles();
acl->mutable_principals()->add_values(DEFAULT_CREDENTIAL_2.principal());
acl->mutable_roles()->add_values(roleMuggle);
acl = acls.add_view_roles();
acl->mutable_principals()->add_values(DEFAULT_CREDENTIAL_2.principal());
acl->mutable_roles()->set_type(mesos::ACL::Entity::NONE);
}
// Create an `Authorizer` with the ACLs.
Try<Authorizer*> create = TypeParam::create(parameterize(acls));
ASSERT_SOME(create);
Owned<Authorizer> authorizer(create.get());
Try<Owned<cluster::Master>> master = this->StartMaster(authorizer.get());
ASSERT_SOME(master);
// Register framework with user "bar" and role "superhero".
FrameworkInfo frameworkSuperhero = DEFAULT_FRAMEWORK_INFO;
frameworkSuperhero.set_name("framework-" + roleSuperhero);
frameworkSuperhero.set_roles(0, roleSuperhero);
frameworkSuperhero.set_user("bar");
// Create an executor with user "bar".
ExecutorInfo executorSuperhero =
createExecutorInfo("test-executor-" + roleSuperhero, "sleep 2");
executorSuperhero.mutable_command()->set_user("bar");
MockExecutor execSuperhero(executorSuperhero.executor_id());
// Register framework with user "foo" and role "muggle".
FrameworkInfo frameworkMuggle = DEFAULT_FRAMEWORK_INFO;
frameworkMuggle.set_name("framework-" + roleMuggle);
frameworkMuggle.set_principal(DEFAULT_CREDENTIAL_2.principal());
frameworkMuggle.set_roles(0, roleMuggle);
frameworkMuggle.set_user("foo");
// Create an executor with user "foo".
ExecutorInfo executorMuggle =
createExecutorInfo("test-executor-" + roleMuggle, "sleep 2");
executorMuggle.mutable_command()->set_user("foo");
MockExecutor execMuggle(executorMuggle.executor_id());
TestContainerizer containerizer(
{{executorSuperhero.executor_id(), &execSuperhero},
{executorMuggle.executor_id(), &execMuggle}});
slave::Flags flags = this->CreateSlaveFlags();
// Statically reserve resources for each role.
flags.resources = "cpus(" + roleSuperhero + "):2;" + "cpus(" + roleMuggle +
"):3;mem(" + roleSuperhero + "):512;" + "mem(" + roleMuggle + "):1024;";
Owned<MasterDetector> detector = master.get()->createDetector();
Try<Owned<cluster::Slave>> slave = this->StartSlave(
detector.get(), &containerizer, authorizer.get(), flags);
ASSERT_SOME(slave);
MockScheduler schedSuperhero;
MesosSchedulerDriver driverSuperhero(
&schedSuperhero,
frameworkSuperhero,
master.get()->pid,
DEFAULT_CREDENTIAL);
EXPECT_CALL(execSuperhero, registered(_, _, _, _))
.Times(AtMost(1));
Future<FrameworkID> frameworkIdSuperhero;
EXPECT_CALL(schedSuperhero, registered(&driverSuperhero, _, _))
.WillOnce(FutureArg<1>(&frameworkIdSuperhero));
Future<vector<Offer>> offersSuperhero;
EXPECT_CALL(schedSuperhero, resourceOffers(&driverSuperhero, _))
.WillOnce(FutureArg<1>(&offersSuperhero))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driverSuperhero.start();
AWAIT_READY(frameworkIdSuperhero);
AWAIT_READY(offersSuperhero);
ASSERT_FALSE(offersSuperhero->empty());
// Define a task which will run on executorSuperhero of frameworkSuperhero.
TaskInfo taskSuperhero;
taskSuperhero.set_name("test-" + roleSuperhero);
taskSuperhero.mutable_task_id()->set_value("1");
taskSuperhero.mutable_slave_id()->MergeFrom(
offersSuperhero.get()[0].slave_id());
taskSuperhero.mutable_resources()->MergeFrom(
offersSuperhero.get()[0].resources());
taskSuperhero.mutable_executor()->MergeFrom(executorSuperhero);
EXPECT_CALL(execSuperhero, launchTask(_, _))
.WillOnce(SendStatusUpdateFromTask(TASK_RUNNING))
.WillRepeatedly(Return());
Future<TaskStatus> statusSuperhero;
EXPECT_CALL(schedSuperhero, statusUpdate(&driverSuperhero, _))
.WillOnce(FutureArg<1>(&statusSuperhero));
driverSuperhero.launchTasks(offersSuperhero.get()[0].id(), {taskSuperhero});
AWAIT_READY(statusSuperhero);
EXPECT_EQ(TASK_RUNNING, statusSuperhero->state());
MockScheduler schedMuggle;
MesosSchedulerDriver driverMuggle(
&schedMuggle,
frameworkMuggle,
master.get()->pid,
DEFAULT_CREDENTIAL_2);
EXPECT_CALL(execMuggle, registered(_, _, _, _))
.Times(AtMost(1));
Future<FrameworkID> frameworkIdMuggle;
EXPECT_CALL(schedMuggle, registered(&driverMuggle, _, _))
.WillOnce(FutureArg<1>(&frameworkIdMuggle));
Future<vector<Offer>> offersMuggle;
EXPECT_CALL(schedMuggle, resourceOffers(&driverMuggle, _))
.WillOnce(FutureArg<1>(&offersMuggle))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driverMuggle.start();
AWAIT_READY(frameworkIdMuggle);
AWAIT_READY(offersMuggle);
ASSERT_FALSE(offersMuggle->empty());
// Define a task which will run on executorMuggle of frameworkMuggle.
TaskInfo taskMuggle;
taskMuggle.set_name("test-" + roleMuggle);
taskMuggle.mutable_task_id()->set_value("2");
taskMuggle.mutable_slave_id()->MergeFrom(
offersMuggle.get()[0].slave_id());
taskMuggle.mutable_resources()->MergeFrom(
offersMuggle.get()[0].resources());
taskMuggle.mutable_executor()->MergeFrom(executorMuggle);
EXPECT_CALL(execMuggle, launchTask(_, _))
.WillOnce(SendStatusUpdateFromTask(TASK_RUNNING))
.WillRepeatedly(Return());
Future<TaskStatus> statusMuggle;
EXPECT_CALL(schedMuggle, statusUpdate(&driverMuggle, _))
.WillOnce(FutureArg<1>(&statusMuggle));
driverMuggle.launchTasks(offersMuggle.get()[0].id(), {taskMuggle});
AWAIT_READY(statusMuggle);
ASSERT_EQ(TASK_RUNNING, statusMuggle->state());
// Retrieve endpoint with the user allowed to view the frameworks.
// The default user allowed to view role "superhero" only.
{
Future<Response> response = http::get(
slave.get()->pid,
"state",
None(),
createBasicAuthHeaders(DEFAULT_CREDENTIAL));
AWAIT_EXPECT_RESPONSE_STATUS_EQ(OK().status, response);
Try<JSON::Object> parse = JSON::parse<JSON::Object>(response->body);
ASSERT_SOME(parse);
JSON::Object state = parse.get();
ASSERT_TRUE(state.values["frameworks"].is<JSON::Array>());
JSON::Array frameworks = state.values["frameworks"].as<JSON::Array>();
EXPECT_EQ(2u, frameworks.values.size());
foreach (const JSON::Value& value, frameworks.values) {
JSON::Object framework = value.as<JSON::Object>();
EXPECT_FALSE(framework.values.empty());
ASSERT_TRUE(framework.values["executors"].is<JSON::Array>());
JSON::Array executors = framework.values["executors"].as<JSON::Array>();
EXPECT_EQ(1u, executors.values.size());
JSON::Object executor = executors.values.front().as<JSON::Object>();
EXPECT_EQ(1u, executor.values["tasks"].as<JSON::Array>().values.size());
}
ASSERT_TRUE(state.values["reserved_resources"].is<JSON::Object>());
JSON::Object reserved_resources =
state.values["reserved_resources"].as<JSON::Object>();
EXPECT_TRUE(reserved_resources.values[roleSuperhero].is<JSON::Object>());
EXPECT_FALSE(reserved_resources.values[roleMuggle].is<JSON::Object>());
ASSERT_TRUE(
state.values["reserved_resources_allocated"].is<JSON::Object>());
JSON::Object reserved_resources_allocated =
state.values["reserved_resources_allocated"].as<JSON::Object>();
EXPECT_TRUE(
reserved_resources_allocated.values[roleSuperhero].is<JSON::Object>());
EXPECT_FALSE(
reserved_resources_allocated.values[roleMuggle].is<JSON::Object>());
ASSERT_TRUE(state.values["reserved_resources_full"].is<JSON::Object>());
JSON::Object reserved_resources_full =
state.values["reserved_resources_full"].as<JSON::Object>();
EXPECT_TRUE(
reserved_resources_full.values[roleSuperhero].is<JSON::Array>());
EXPECT_FALSE(
reserved_resources_full.values[roleMuggle].is<JSON::Array>());
}
// Retrieve endpoint with the user allowed to view the frameworks,
// but not the executors.
// The second default user allowed to view role "muggle" only.
{
Future<Response> response = http::get(
slave.get()->pid,
"state",
None(),
createBasicAuthHeaders(DEFAULT_CREDENTIAL_2));
AWAIT_EXPECT_RESPONSE_STATUS_EQ(OK().status, response);
Try<JSON::Object> parse = JSON::parse<JSON::Object>(response->body);
ASSERT_SOME(parse);
JSON::Object state = parse.get();
ASSERT_TRUE(state.values["frameworks"].is<JSON::Array>());
JSON::Array frameworks = state.values["frameworks"].as<JSON::Array>();
EXPECT_EQ(2u, frameworks.values.size());
foreach (const JSON::Value& value, frameworks.values) {
JSON::Object framework = value.as<JSON::Object>();
EXPECT_FALSE(framework.values.empty());
EXPECT_TRUE(
framework.values["executors"].as<JSON::Array>().values.empty());
}
ASSERT_TRUE(state.values["reserved_resources"].is<JSON::Object>());
JSON::Object reserved_resources =
state.values["reserved_resources"].as<JSON::Object>();
EXPECT_TRUE(reserved_resources.values[roleMuggle].is<JSON::Object>());
EXPECT_FALSE(reserved_resources.values[roleSuperhero].is<JSON::Object>());
ASSERT_TRUE(
state.values["reserved_resources_allocated"].is<JSON::Object>());
JSON::Object reserved_resources_allocated =
state.values["reserved_resources_allocated"].as<JSON::Object>();
EXPECT_TRUE(
reserved_resources_allocated.values[roleMuggle].is<JSON::Object>());
EXPECT_FALSE(
reserved_resources_allocated.values[roleSuperhero].is<JSON::Object>());
ASSERT_TRUE(state.values["reserved_resources_full"].is<JSON::Object>());
JSON::Object reserved_resources_full =
state.values["reserved_resources_full"].as<JSON::Object>();
EXPECT_TRUE(
reserved_resources_full.values[roleMuggle].is<JSON::Array>());
EXPECT_FALSE(
reserved_resources_full.values[roleSuperhero].is<JSON::Array>());
}
EXPECT_CALL(execSuperhero, shutdown(_))
.Times(AtMost(1));
EXPECT_CALL(execMuggle, shutdown(_))
.Times(AtMost(1));
driverSuperhero.stop();
driverSuperhero.join();
driverMuggle.stop();
driverMuggle.join();
}
// This test verifies that the master reconciles operations that are missing
// from a reregistering slave. In this case, we drop the ApplyOperationMessage
// and expect the master to send a ReconcileOperationsMessage after the slave
// reregisters.
TEST_F(MasterSlaveReconciliationTest, ReconcileDroppedOperation)
{
Try<Owned<cluster::Master>> master = StartMaster();
ASSERT_SOME(master);
StandaloneMasterDetector detector(master.get()->pid);
Future<UpdateSlaveMessage> updateSlaveMessage =
FUTURE_PROTOBUF(UpdateSlaveMessage(), _, _);
Try<Owned<cluster::Slave>> slave = StartSlave(&detector);
ASSERT_SOME(slave);
// Since any out-of-sync operation state in `UpdateSlaveMessage` triggers a
// reconciliation, await the message from the initial agent registration
// sequence beforce continuing. Otherwise we risk the master reconciling with
// the agent before we fail over the master.
AWAIT_READY(updateSlaveMessage);
// Register the framework in a non-`*` role so it can reserve resources.
FrameworkInfo frameworkInfo = DEFAULT_FRAMEWORK_INFO;
frameworkInfo.set_roles(0, DEFAULT_TEST_ROLE);
MockScheduler sched;
MesosSchedulerDriver driver(
&sched, frameworkInfo, master.get()->pid, DEFAULT_CREDENTIAL);
EXPECT_CALL(sched, registered(&driver, _, _));
Future<vector<Offer>> offers;
EXPECT_CALL(sched, resourceOffers(&driver, _))
.WillOnce(FutureArg<1>(&offers))
.WillRepeatedly(Return()); // Ignore subsequent offers.
driver.start();
AWAIT_READY(offers);
// We prevent the operation from reaching the agent.
Future<ApplyOperationMessage> applyOperationMessage =
DROP_PROTOBUF(ApplyOperationMessage(), _, _);
// Perform a reserve operation on the offered resources.
// This will trigger an `ApplyOperationMessage`.
ASSERT_FALSE(offers->empty());
const Offer& offer = offers->at(0);
Resources reservedResources = offer.resources();
reservedResources =
reservedResources.pushReservation(createDynamicReservationInfo(
frameworkInfo.roles(0), frameworkInfo.principal()));
driver.acceptOffers({offer.id()}, {RESERVE(reservedResources)});
AWAIT_READY(applyOperationMessage);
// We expect the master to detect the missing operation when the
// slave reregisters and to reconcile the operations on that slave.
Future<ReconcileOperationsMessage> reconcileOperationsMessage =
FUTURE_PROTOBUF(ReconcileOperationsMessage(), _, _);
// Simulate a master failover to trigger slave reregistration.
detector.appoint(master.get()->pid);
AWAIT_READY(reconcileOperationsMessage);
ASSERT_EQ(1, reconcileOperationsMessage->operations_size());
EXPECT_EQ(
applyOperationMessage->operation_uuid(),
reconcileOperationsMessage->operations(0).operation_uuid());
}