openstack
/
nova
Code Issues Proposed changes

Merge "Functional test with pGPUs"

This commit is contained in:
Zuul 2020-04-09 23:29:21 +00:00 committed by Gerrit Code Review
parent 77482b05af 9dcc0941f1
commit d695b7dfbc
4 changed files with 382 additions and 31 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

47
nova/tests/functional/integrated_helpers.py
View File

@ -449,7 +449,54 @@ class _IntegratedTestBase(test.TestCase, InstanceHelperMixin):
("The expected wsgi middlewares %s are not "
"existed") % expected_middleware)
# TODO(sbauza): Drop this method once test classes inherit from a mixin
def _get_provider_uuid_by_name(self, name):
return self.placement_api.get(
'/resource_providers?name=%s' % name).body[
'resource_providers'][0]['uuid']
# TODO(sbauza): Drop this method once test classes inherit from a mixin
def _get_all_rp_uuids_in_a_tree(self, in_tree_rp_uuid):
rps = self.placement_api.get(
'/resource_providers?in_tree=%s' % in_tree_rp_uuid,
version='1.20').body['resource_providers']
return [rp['uuid'] for rp in rps]
# TODO(sbauza): Drop this method once test classes inherit from a mixin
def _get_provider_inventory(self, rp_uuid):
return self.placement_api.get(
'/resource_providers/%s/inventories' % rp_uuid).body['inventories']
# TODO(sbauza): Drop this method once test classes inherit from a mixin
def _get_provider_usages(self, provider_uuid):
return self.placement_api.get(
'/resource_providers/%s/usages' % provider_uuid).body['usages']
# TODO(sbauza): Drop this method once test classes inherit from a mixin
def _create_trait(self, trait):
return self.placement_api.put('/traits/%s' % trait, {}, version='1.6')
# TODO(sbauza): Drop this method once test classes inherit from a mixin
def _set_provider_traits(self, rp_uuid, traits):
"""This will overwrite any existing traits.
:param rp_uuid: UUID of the resource provider to update
:param traits: list of trait strings to set on the provider
:returns: APIResponse object with the results
"""
provider = self.placement_api.get(
'/resource_providers/%s' % rp_uuid).body
put_traits_req = {
'resource_provider_generation': provider['generation'],
'traits': traits
}
return self.placement_api.put(
'/resource_providers/%s/traits' % rp_uuid,
put_traits_req, version='1.6')
# FIXME(sbauza): There is little value to have this be a whole base testclass
# instead of a mixin only providing methods for accessing Placement endpoint.
class ProviderUsageBaseTestCase(test.TestCase, InstanceHelperMixin):
"""Base test class for functional tests that check provider usage
and consumer allocations in Placement during various operations.

20
nova/tests/functional/libvirt/test_reshape.py
View File

@ -59,10 +59,28 @@ class VGPUReshapeTests(base.ServersTestBase):
# the old tree as that would be a bad time for reshape. Later when the
# compute service is restarted the driver will do the reshape.
mdevs = {
'mdev_4b20d080_1b54_4048_85b3_a6a62d165c01':
fakelibvirt.FakeMdevDevice(
dev_name='mdev_4b20d080_1b54_4048_85b3_a6a62d165c01',
type_id=fakelibvirt.NVIDIA_11_VGPU_TYPE,
parent=fakelibvirt.PGPU1_PCI_ADDR),
'mdev_4b20d080_1b54_4048_85b3_a6a62d165c02':
fakelibvirt.FakeMdevDevice(
dev_name='mdev_4b20d080_1b54_4048_85b3_a6a62d165c02',
type_id=fakelibvirt.NVIDIA_11_VGPU_TYPE,
parent=fakelibvirt.PGPU2_PCI_ADDR),
'mdev_4b20d080_1b54_4048_85b3_a6a62d165c03':
fakelibvirt.FakeMdevDevice(
dev_name='mdev_4b20d080_1b54_4048_85b3_a6a62d165c03',
type_id=fakelibvirt.NVIDIA_11_VGPU_TYPE,
parent=fakelibvirt.PGPU3_PCI_ADDR),
}
fake_connection = self._get_connection(
# We need more RAM or the 3rd server won't be created
host_info=fakelibvirt.HostInfo(kB_mem=8192),
mdev_info=fakelibvirt.HostMdevDevicesInfo())
mdev_info=fakelibvirt.HostMdevDevicesInfo(devices=mdevs))
self.mock_conn.return_value = fake_connection
# start a compute with vgpu support disabled so the driver will

234
nova/tests/functional/libvirt/test_vgpu.py Normal file
View File

@ -0,0 +1,234 @@
#
# Licensed under the Apache License, Version 2.0 (the "License"); you may
# not use this file except in compliance with the License. You may obtain
# a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
# License for the specific language governing permissions and limitations
# under the License.
import fixtures
import re
import os_resource_classes as orc
from oslo_config import cfg
from oslo_log import log as logging
from oslo_utils import uuidutils
import nova.conf
from nova import context
from nova import objects
from nova.tests.functional.libvirt import base
from nova.tests.unit.virt.libvirt import fakelibvirt
from nova.virt.libvirt import utils as libvirt_utils
CONF = cfg.CONF
LOG = logging.getLogger(__name__)
class VGPUTestBase(base.ServersTestBase):
FAKE_LIBVIRT_VERSION = 5000000
FAKE_QEMU_VERSION = 3001000
def setUp(self):
super(VGPUTestBase, self).setUp()
self.useFixture(fixtures.MockPatch(
'nova.virt.libvirt.LibvirtDriver._get_local_gb_info',
return_value={'total': 128,
'used': 44,
'free': 84}))
self.useFixture(fixtures.MockPatch(
'nova.privsep.libvirt.create_mdev',
side_effect=self._create_mdev))
self.context = context.get_admin_context()
def pci2libvirt_address(self, address):
return "pci_{}_{}_{}_{}".format(*re.split("[.:]", address))
def libvirt2pci_address(self, dev_name):
return "{}:{}:{}.{}".format(*dev_name[4:].split('_'))
def _create_mdev(self, physical_device, mdev_type, uuid=None):
# We need to fake the newly created sysfs object by adding a new
# FakeMdevDevice in the existing persisted Connection object so
# when asking to get the existing mdevs, we would see it.
if not uuid:
uuid = uuidutils.generate_uuid()
mdev_name = libvirt_utils.mdev_uuid2name(uuid)
libvirt_parent = self.pci2libvirt_address(physical_device)
self.fake_connection.mdev_info.devices.update(
{mdev_name: fakelibvirt.FakeMdevDevice(dev_name=mdev_name,
type_id=mdev_type,
parent=libvirt_parent)})
return uuid
def _start_compute_service(self, hostname):
self.fake_connection = self._get_connection(
host_info=fakelibvirt.HostInfo(cpu_nodes=2, kB_mem=8192),
# We want to create two pGPUs but no other PCI devices
pci_info=fakelibvirt.HostPCIDevicesInfo(num_pci=0,
num_pfs=0,
num_vfs=0,
num_mdevcap=2),
hostname=hostname)
self.mock_conn.return_value = self.fake_connection
compute = self.start_service('compute', host=hostname)
rp_uuid = self._get_provider_uuid_by_name(hostname)
rp_uuids = self._get_all_rp_uuids_in_a_tree(rp_uuid)
for rp in rp_uuids:
inventory = self._get_provider_inventory(rp)
if orc.VGPU in inventory:
usage = self._get_provider_usages(rp)
self.assertEqual(16, inventory[orc.VGPU]['total'])
self.assertEqual(0, usage[orc.VGPU])
# Since we haven't created any mdevs yet, we shouldn't find them
self.assertEqual([], compute.driver._get_mediated_devices())
return compute
class VGPUTests(VGPUTestBase):
def setUp(self):
super(VGPUTests, self).setUp()
extra_spec = {"resources:VGPU": "1"}
self.flavor = self._create_flavor(extra_spec=extra_spec)
# Start compute1 supporting only nvidia-11
self.flags(
enabled_vgpu_types=fakelibvirt.NVIDIA_11_VGPU_TYPE,
group='devices')
self.compute1 = self._start_compute_service('host1')
def test_create_servers_with_vgpu(self):
self._create_server(
image_uuid='155d900f-4e14-4e4c-a73d-069cbf4541e6',
flavor_id=self.flavor, host=self.compute1.host,
expected_state='ACTIVE')
# Now we should find a new mdev
mdevs = self.compute1.driver._get_mediated_devices()
self.assertEqual(1, len(mdevs))
# Checking also the allocations for the parent pGPU
parent_name = mdevs[0]['parent']
parent_rp_name = self.compute1.host + '_' + parent_name
parent_rp_uuid = self._get_provider_uuid_by_name(parent_rp_name)
usage = self._get_provider_usages(parent_rp_uuid)
self.assertEqual(1, usage[orc.VGPU])
class VGPUMultipleTypesTests(VGPUTestBase):
def setUp(self):
super(VGPUMultipleTypesTests, self).setUp()
extra_spec = {"resources:VGPU": "1"}
self.flavor = self._create_flavor(extra_spec=extra_spec)
self.flags(
enabled_vgpu_types=[fakelibvirt.NVIDIA_11_VGPU_TYPE,
fakelibvirt.NVIDIA_12_VGPU_TYPE],
group='devices')
# we need to call the below again to ensure the updated
# 'device_addresses' value is read and the new groups created
nova.conf.devices.register_dynamic_opts(CONF)
# host1 will have 2 physical GPUs :
# - 0000:81:00.0 will only support nvidia-11
# - 0000:81:01.0 will only support nvidia-12
pgpu1_pci_addr = self.libvirt2pci_address(fakelibvirt.PGPU1_PCI_ADDR)
pgpu2_pci_addr = self.libvirt2pci_address(fakelibvirt.PGPU2_PCI_ADDR)
self.flags(device_addresses=[pgpu1_pci_addr], group='vgpu_nvidia-11')
self.flags(device_addresses=[pgpu2_pci_addr], group='vgpu_nvidia-12')
# Prepare traits for later on
self._create_trait('CUSTOM_NVIDIA_11')
self._create_trait('CUSTOM_NVIDIA_12')
self.compute1 = self._start_compute_service('host1')
def test_create_servers_with_vgpu(self):
self._create_server(
image_uuid='155d900f-4e14-4e4c-a73d-069cbf4541e6',
flavor_id=self.flavor, host=self.compute1.host,
expected_state='ACTIVE')
mdevs = self.compute1.driver._get_mediated_devices()
self.assertEqual(1, len(mdevs))
# We can be deterministic : since 0000:81:01.0 is asked to only support
# nvidia-12 *BUT* doesn't actually have this type as a PCI capability,
# we are sure that only 0000:81:00.0 is used.
parent_name = mdevs[0]['parent']
self.assertEqual(fakelibvirt.PGPU1_PCI_ADDR, parent_name)
# We are also sure that there is no RP for 0000:81:01.0 since there
# is no inventory for nvidia-12
root_rp_uuid = self._get_provider_uuid_by_name(self.compute1.host)
rp_uuids = self._get_all_rp_uuids_in_a_tree(root_rp_uuid)
# We only have 2 RPs : the root RP and only the pGPU1 RP...
self.assertEqual(2, len(rp_uuids))
# ... but we double-check by asking the RP by its expected name
expected_pgpu2_rp_name = (self.compute1.host + '_' +
fakelibvirt.PGPU2_PCI_ADDR)
pgpu2_rp = self.placement_api.get(
'/resource_providers?name=' + expected_pgpu2_rp_name).body[
'resource_providers']
# See, Placement API returned no RP for this name as it doesn't exist.
self.assertEqual([], pgpu2_rp)
def test_create_servers_with_specific_type(self):
# Regenerate the PCI addresses so both pGPUs now support nvidia-12
self.fake_connection.pci_info = fakelibvirt.HostPCIDevicesInfo(
num_pci=0, num_pfs=0, num_vfs=0, num_mdevcap=2,
multiple_gpu_types=True)
# Make a restart to update the Resource Providers
self.compute1 = self.restart_compute_service(self.compute1)
pgpu1_rp_uuid = self._get_provider_uuid_by_name(
self.compute1.host + '_' + fakelibvirt.PGPU1_PCI_ADDR)
pgpu2_rp_uuid = self._get_provider_uuid_by_name(
self.compute1.host + '_' + fakelibvirt.PGPU2_PCI_ADDR)
pgpu1_inventory = self._get_provider_inventory(pgpu1_rp_uuid)
self.assertEqual(16, pgpu1_inventory[orc.VGPU]['total'])
pgpu2_inventory = self._get_provider_inventory(pgpu2_rp_uuid)
self.assertEqual(8, pgpu2_inventory[orc.VGPU]['total'])
# Attach traits to the pGPU RPs
self._set_provider_traits(pgpu1_rp_uuid, ['CUSTOM_NVIDIA_11'])
self._set_provider_traits(pgpu2_rp_uuid, ['CUSTOM_NVIDIA_12'])
expected = {'CUSTOM_NVIDIA_11': fakelibvirt.PGPU1_PCI_ADDR,
'CUSTOM_NVIDIA_12': fakelibvirt.PGPU2_PCI_ADDR}
for trait in expected.keys():
# Add a trait to the flavor
extra_spec = {"resources:VGPU": "1",
"trait:%s" % trait: "required"}
flavor = self._create_flavor(extra_spec=extra_spec)
# Use the new flavor for booting
server = self._create_server(
image_uuid='155d900f-4e14-4e4c-a73d-069cbf4541e6',
flavor_id=flavor, host=self.compute1.host,
expected_state='ACTIVE')
# Get the instance we just created
inst = objects.Instance.get_by_uuid(self.context, server['id'])
# Get the mdevs that were allocated for this instance, we should
# only have one
mdevs = self.compute1.driver._get_all_assigned_mediated_devices(
inst)
self.assertEqual(1, len(mdevs))
# It's a dict of mdev_uuid/instance_uuid pairs, we only care about
# the keys
mdevs = list(mdevs.keys())
# Now get the detailed information about this single mdev
mdev_info = self.compute1.driver._get_mediated_device_information(
libvirt_utils.mdev_uuid2name(mdevs[0]))
# We can be deterministic : since we asked for a specific type,
# we know which pGPU we landed.
self.assertEqual(expected[trait], mdev_info['parent'])

112
nova/tests/unit/virt/libvirt/fakelibvirt.py
View File

@ -195,10 +195,18 @@ VF_PROD_NAME = 'X540 Ethernet Controller Virtual Function'
VF_DRIVER_NAME = 'ixgbevf'
VF_CAP_TYPE = 'phys_function'
MDEV_CAPABLE_VEND_ID = '10DE'
MDEV_CAPABLE_VEND_NAME = 'Nvidia'
MDEV_CAPABLE_PROD_ID = '0FFE'
MDEV_CAPABLE_PROD_NAME = 'GRID M60-0B'
MDEV_CAPABLE_DRIVER_NAME = 'nvidia'
MDEV_CAPABLE_CAP_TYPE = 'mdev_types'
NVIDIA_11_VGPU_TYPE = 'nvidia-11'
PGPU1_PCI_ADDR = 'pci_0000_06_00_0'
PGPU2_PCI_ADDR = 'pci_0000_07_00_0'
PGPU3_PCI_ADDR = 'pci_0000_08_00_0'
NVIDIA_12_VGPU_TYPE = 'nvidia-12'
PGPU1_PCI_ADDR = 'pci_0000_81_00_0'
PGPU2_PCI_ADDR = 'pci_0000_81_01_0'
PGPU3_PCI_ADDR = 'pci_0000_81_02_0'
class FakePCIDevice(object):
@ -240,9 +248,16 @@ class FakePCIDevice(object):
</device>""".strip()) # noqa
cap_templ = "<capability type='%(cap_type)s'>%(addresses)s</capability>"
addr_templ = "<address domain='0x0000' bus='0x81' slot='%(slot)#02x' function='%(function)#02x'/>" # noqa
mdevtypes_templ = textwrap.dedent("""
<type id='%(type_id)s'>
<name>GRID M60-0B</name><deviceAPI>vfio-pci</deviceAPI>
<availableInstances>%(instances)s</availableInstances>
</type>""".strip()) # noqa
is_capable_of_mdevs = False
def __init__(self, dev_type, slot, function, iommu_group, numa_node,
vf_ratio=None):
vf_ratio=None, multiple_gpu_types=False):
"""Populate pci devices
:param dev_type: (string) Indicates the type of the device (PCI, PF,
@ -253,8 +268,11 @@ class FakePCIDevice(object):
:param numa_node: (int) NUMA node of the device.
:param vf_ratio: (int) Ratio of Virtual Functions on Physical. Only
applicable if ``dev_type`` is one of: ``PF``, ``VF``.
:param multiple_gpu_types: (bool) Supports different vGPU types
"""
vend_id = PCI_VEND_ID
vend_name = PCI_VEND_NAME
if dev_type == 'PCI':
if vf_ratio:
raise ValueError('vf_ratio does not apply for PCI devices')
@ -295,14 +313,34 @@ class FakePCIDevice(object):
'function': 0,
}
}
elif dev_type == 'MDEV_TYPES':
prod_id = MDEV_CAPABLE_PROD_ID
prod_name = MDEV_CAPABLE_PROD_NAME
driver = MDEV_CAPABLE_DRIVER_NAME
vend_id = MDEV_CAPABLE_VEND_ID
vend_name = MDEV_CAPABLE_VEND_NAME
types = [self.mdevtypes_templ % {
'type_id': NVIDIA_11_VGPU_TYPE,
'instances': 16,
}]
if multiple_gpu_types:
types.append(self.mdevtypes_templ % {
'type_id': NVIDIA_12_VGPU_TYPE,
'instances': 8,
})
capability = self.cap_templ % {
'cap_type': MDEV_CAPABLE_CAP_TYPE,
'addresses': '\n'.join(types)
}
self.is_capable_of_mdevs = True
else:
raise ValueError('Expected one of: PCI, VF, PCI')
self.pci_device = self.pci_device_template % {
'slot': slot,
'function': function,
'vend_id': PCI_VEND_ID,
'vend_name': PCI_VEND_NAME,
'vend_id': vend_id,
'vend_name': vend_name,
'prod_id': prod_id,
'prod_name': prod_name,
'driver': driver,
@ -326,26 +364,31 @@ class HostPCIDevicesInfo(object):
TOTAL_NUMA_NODES = 2
pci_devname_template = 'pci_0000_81_%(slot)02x_%(function)d'
def __init__(self, num_pci=0, num_pfs=2, num_vfs=8, numa_node=None):
def __init__(self, num_pci=0, num_pfs=2, num_vfs=8, num_mdevcap=0,
numa_node=None, multiple_gpu_types=False):
"""Create a new HostPCIDevicesInfo object.
:param num_pci: (int) The number of (non-SR-IOV) PCI devices.
:param num_pci: (int) The number of (non-SR-IOV) and (non-MDEV capable)
PCI devices.
:param num_pfs: (int) The number of PCI SR-IOV Physical Functions.
:param num_vfs: (int) The number of PCI SR-IOV Virtual Functions.
:param num_mdevcap: (int) The number of PCI devices capable of creating
mediated devices.
:param iommu_group: (int) Initial IOMMU group ID.
:param numa_node: (int) NUMA node of the device; if set all of the
devices will be assigned to the specified node else they will be
split between ``$TOTAL_NUMA_NODES`` nodes.
:param multiple_gpu_types: (bool) Supports different vGPU types
"""
self.devices = {}
if not (num_vfs or num_pfs):
if not (num_vfs or num_pfs) and not num_mdevcap:
return
if num_vfs and not num_pfs:
raise ValueError('Cannot create VFs without PFs')
if num_vfs % num_pfs:
if num_pfs and num_vfs % num_pfs:
raise ValueError('num_vfs must be a factor of num_pfs')
slot = 0
@ -369,6 +412,24 @@ class HostPCIDevicesInfo(object):
slot += 1
iommu_group += 1
# Generate MDEV capable devs
for dev in range(num_mdevcap):
pci_dev_name = self.pci_devname_template % {
'slot': slot, 'function': function}
LOG.info('Generating MDEV capable device %r', pci_dev_name)
self.devices[pci_dev_name] = FakePCIDevice(
dev_type='MDEV_TYPES',
slot=slot,
function=function,
iommu_group=iommu_group,
numa_node=self._calc_numa_node(dev, numa_node),
multiple_gpu_types=multiple_gpu_types)
slot += 1
iommu_group += 1
vf_ratio = num_vfs // num_pfs if num_pfs else 0
# Generate PFs
@ -425,6 +486,10 @@ class HostPCIDevicesInfo(object):
pci_dev = self.devices.get(device_name)
return pci_dev
def get_all_mdev_capable_devices(self):
return [dev for dev in self.devices
if self.devices[dev].is_capable_of_mdevs]
class FakeMdevDevice(object):
template = """
@ -453,21 +518,11 @@ class FakeMdevDevice(object):
class HostMdevDevicesInfo(object):
def __init__(self):
self.devices = {
'mdev_4b20d080_1b54_4048_85b3_a6a62d165c01':
FakeMdevDevice(
dev_name='mdev_4b20d080_1b54_4048_85b3_a6a62d165c01',
type_id=NVIDIA_11_VGPU_TYPE, parent=PGPU1_PCI_ADDR),
'mdev_4b20d080_1b54_4048_85b3_a6a62d165c02':
FakeMdevDevice(
dev_name='mdev_4b20d080_1b54_4048_85b3_a6a62d165c02',
type_id=NVIDIA_11_VGPU_TYPE, parent=PGPU2_PCI_ADDR),
'mdev_4b20d080_1b54_4048_85b3_a6a62d165c03':
FakeMdevDevice(
dev_name='mdev_4b20d080_1b54_4048_85b3_a6a62d165c03',
type_id=NVIDIA_11_VGPU_TYPE, parent=PGPU3_PCI_ADDR),
}
def __init__(self, devices=None):
if devices is not None:
self.devices = devices
else:
self.devices = {}
def get_all_devices(self):
return self.devices.keys()
@ -1268,7 +1323,7 @@ class Connection(object):
self.pci_info = pci_info or HostPCIDevicesInfo(num_pci=0,
num_pfs=0,
num_vfs=0)
self.mdev_info = mdev_info or []
self.mdev_info = mdev_info or HostMdevDevicesInfo(devices={})
self.hostname = hostname or 'compute1'
def _add_filter(self, nwfilter):
@ -1573,10 +1628,7 @@ class Connection(object):
if cap == 'mdev':
return self.mdev_info.get_all_devices()
if cap == 'mdev_types':
# TODO(gibi): We should return something like
# https://libvirt.org/drvnodedev.html#MDEVCap but I tried and it
# did not work for me.
return None
return self.pci_info.get_all_mdev_capable_devices()
else:
raise ValueError('Capability "%s" is not supported' % cap)