test_reassign_port_between_servers failing with tap device is busy errors in neutron xenial jobs since 7/28

Another instance:

http://logs.openstack.org/52/347352/8/check/gate-tempest-dsvm-neutron-full-ubuntu-xenial/1f82cb6/console.html

logstash shows it starting on 7/28: http://goo.gl/u4kqup

This change switched the neutron gate jobs over to xenial:

https://review.openstack.org/#/c/348078/

Looks like it might just be failing in this test:

tempest.api.compute.servers.test_attach_interfaces.AttachInterfacesTestJSON.test_reassign_port_between_servers

Looks like possibly an issue with glean?

http://logs.openstack.org/09/347509/4/gate/gate-tempest-dsvm-neutron-full-ubuntu-xenial/2051841/logs/screen-n-cpu.txt.gz#_2016-07-29_12_59_23_311

attach device xml:
<interface type="bridge">
  <mac address="fa:16:3e:79:df:76"/>
  <model type="virtio"/>
  <driver name="qemu"/>
  <source bridge="qbr953222b9-a8"/>
  <target dev="tap953222b9-a8"/>
</interface>

http://logs.openstack.org/09/347509/4/gate/gate-tempest-dsvm-neutron-full-ubuntu-xenial/2051841/logs/syslog.txt.gz#_Jul_29_12_59_21

Jul 29 12:59:21 ubuntu-xenial-osic-cloud1-3018328 glean.sh[373]: Traceback (most recent call last):
Jul 29 12:59:21 ubuntu-xenial-osic-cloud1-3018328 glean.sh[373]: File "/usr/local/bin/glean", line 11, in <module>
Jul 29 12:59:21 ubuntu-xenial-osic-cloud1-3018328 glean.sh[373]: sys.exit(main())
Jul 29 12:59:21 ubuntu-xenial-osic-cloud1-3018328 glean.sh[373]: File "/usr/local/lib/python2.7/dist-packages/glean/cmd.py", line 722, in main
Jul 29 12:59:21 ubuntu-xenial-osic-cloud1-3018328 glean.sh[373]: write_network_info_from_config_drive(args)
Jul 29 12:59:21 ubuntu-xenial-osic-cloud1-3018328 glean.sh[373]: File "/usr/local/lib/python2.7/dist-packages/glean/cmd.py", line 562, in write_network_info_from_config_drive
Jul 29 12:59:21 ubuntu-xenial-osic-cloud1-3018328 glean.sh[373]: sys_interfaces = get_sys_interfaces(args.interface, args)
Jul 29 12:59:21 ubuntu-xenial-osic-cloud1-3018328 glean.sh[373]: File "/usr/local/lib/python2.7/dist-packages/glean/cmd.py", line 508, in get_sys_interfaces
Jul 29 12:59:21 ubuntu-xenial-osic-cloud1-3018328 glean.sh[373]: '%s/%s/addr_assign_type' % (sys_root, iface), 'r').read().strip()
Jul 29 12:59:21 ubuntu-xenial-osic-cloud1-3018328 glean.sh[373]: IOError: [Errno 2] No such file or directory: '/sys/class/net/qbr953222b9/a8/addr_assign_type'

The device detach from the guest in libvirt is asynchronous, so this must be much slower in libvirt 1.3.1 on xenial nodes, so neutron is telling us that the port is detach (device_id is None on the port) after detach - which is what the tempest test is polling on - before the interface is actually detached from the guest.

So we probably need a retry loop in detach_interface in the libvirt driver (like we have for detach_volume) to retry until timeout for the interface to be gone from the guest and consider the detach successful.

However, detach_interface is a cast from compute API to the compute manager, so the Tempest test doesn't really have a way to poll that the interface is actually detached from the guest (beyond doing something like ssh'ing into the guest to verify the interface with the given mac is gone).

Nevermind, it's the compute manager in nova that's telling neutron that the port is no longer bound:

https://github.com/openstack/nova/blob/fdf3328107e53f1c5578c2e4dfbad78d832b01c6/nova/compute/manager.py#L4990

We first call the virt driver to detach the interface (which is async) and then update the port telling neutron that the device_id is '', which is what tempest is waiting for.

So if we add a poll / retry in the libvirt guest module for the detach, we can delay the port update which tempest is waiting for and we should be good.

Skipping the test in tempest for now: https://review.openstack.org/#/c/348955/

Fix proposed to branch: master
Review: https://review.openstack.org/349014

Not seen in the gate any more, the fixing patch is in merge conflict and really old

Change abandoned by Sean Dague (<email address hidden>) on branch: master
Review: https://review.openstack.org/349014
Reason: This review is > 6 weeks without comment, and failed Jenkins the last time it was checked. We are abandoning this for now. Feel free to reactivate the review by pressing the restore button and leaving a 'recheck' comment to get fresh test results.

Related fix proposed to branch: master
Review: https://review.openstack.org/431778

It's going to be hard getting this back to Newton because it depends on:

https://review.openstack.org/#/c/372243/

Which isn't in Newton. It's in stable/ocata though.

As noted in that patch, it's tests are broken too and being fixed here:

https://review.openstack.org/#/c/431778/

So to get this to Newton, we'd need to backport:

https://review.openstack.org/#/c/372243/

And then:

https://review.openstack.org/#/c/431778/

And then:

https://review.openstack.org/#/c/349014/

Reviewed: https://review.openstack.org/431778
Committed: https://git.openstack.org/cgit/openstack/nova/commit/?id=1ecd71b08d14450e475dc9512d40828da6fcfe15
Submitter: Jenkins
Branch: master

commit 1ecd71b08d14450e475dc9512d40828da6fcfe15
Author: Matt Riedemann <email address hidden>
Date: Thu Feb 9 18:41:11 2017 -0500

    libvirt: fix and break up _test_attach_detach_interface

    The detach_interface flow in this test was broken because
    it wasn't mocking out domain.detachDeviceFlags so the xml
    it was expecting to be passed to that method wasn't actually
    being verified. The same thing is broken in test
    test_detach_interface_device_with_same_mac_address because
    it copies the other broken test code.

    This change breaks apart the monster attach/detach test method
    and converts the detach_interface portion to mock and fixes
    the broken assertion.

    test_detach_interface_device_with_same_mac_address is just
    fixed, not converted to mock.

    Change-Id: I6d9a975876c5652ef544c587f65b1bdd1543848b
    Related-Bug: #1607714

Related fix proposed to branch: stable/ocata
Review: https://review.openstack.org/448188

Fix proposed to branch: stable/ocata
Review: https://review.openstack.org/448189

Reviewed: https://review.openstack.org/349014
Committed: https://git.openstack.org/cgit/openstack/nova/commit/?id=a3b3e8d8314b0cedc2604be509f0f4d523a35ed5
Submitter: Jenkins
Branch: master

commit a3b3e8d8314b0cedc2604be509f0f4d523a35ed5
Author: Matt Riedemann <email address hidden>
Date: Thu Feb 9 15:54:41 2017 -0500

    libvirt: wait for interface detach from the guest

    The test_reassign_port_between_servers test in Tempest creates
    a port in neutron and two servers. It attaches the port to the
    first server and then quickly detaches it and waits for the
    port.device_id to be unbound from the server. Then it repeats
    that for the second server.

    The interface detach from the guest is asynchronous and happens
    before nova unbinds the port, so there is a race where the port's
    device_id is unset but the interface is still on the first guest
    when we try to attach to the second guest, which fails.

    This is a latent bug, but apparently has been tickled by the
    move to our neutron CI jobs to use ubuntu xenial nodes.

    The fix is to add a detach and retry loop on the interface detach
    on the guest so that we can wait until the interface is gone
    from the guest before nova unbinds the port in neutron, which is
    what the Tempest test is waiting for. Then the device should be
    available for attaching to the second guest.

    This is similar to what we do with detaching volumes.

    Closes-Bug: #1607714

    Change-Id: Ic04aad8923ea2edf1d16e32c208cd41fdf898834

Related fix proposed to branch: stable/newton
Review: https://review.openstack.org/470348

Fix proposed to branch: stable/newton
Review: https://review.openstack.org/470349

This issue was fixed in the openstack/nova 16.0.0.0b2 development milestone.

Reviewed: https://review.openstack.org/448188
Committed: https://git.openstack.org/cgit/openstack/nova/commit/?id=c0820944ea8554f5b1db0d538781caf0e75b6b0e
Submitter: Jenkins
Branch: stable/ocata

commit c0820944ea8554f5b1db0d538781caf0e75b6b0e
Author: Matt Riedemann <email address hidden>
Date: Thu Feb 9 18:41:11 2017 -0500

    libvirt: fix and break up _test_attach_detach_interface

    The detach_interface flow in this test was broken because
    it wasn't mocking out domain.detachDeviceFlags so the xml
    it was expecting to be passed to that method wasn't actually
    being verified. The same thing is broken in test
    test_detach_interface_device_with_same_mac_address because
    it copies the other broken test code.

    This change breaks apart the monster attach/detach test method
    and converts the detach_interface portion to mock and fixes
    the broken assertion.

    test_detach_interface_device_with_same_mac_address is just
    fixed, not converted to mock.

    Change-Id: I6d9a975876c5652ef544c587f65b1bdd1543848b
    Related-Bug: #1607714
    (cherry picked from commit 1ecd71b08d14450e475dc9512d40828da6fcfe15)

Reviewed: https://review.openstack.org/448189
Committed: https://git.openstack.org/cgit/openstack/nova/commit/?id=02ad4f862a7c5b51100288b6b22f15087788d8d7
Submitter: Jenkins
Branch: stable/ocata

commit 02ad4f862a7c5b51100288b6b22f15087788d8d7
Author: Matt Riedemann <email address hidden>
Date: Thu Feb 9 15:54:41 2017 -0500

    libvirt: wait for interface detach from the guest

    The test_reassign_port_between_servers test in Tempest creates
    a port in neutron and two servers. It attaches the port to the
    first server and then quickly detaches it and waits for the
    port.device_id to be unbound from the server. Then it repeats
    that for the second server.

    The interface detach from the guest is asynchronous and happens
    before nova unbinds the port, so there is a race where the port's
    device_id is unset but the interface is still on the first guest
    when we try to attach to the second guest, which fails.

    This is a latent bug, but apparently has been tickled by the
    move to our neutron CI jobs to use ubuntu xenial nodes.

    The fix is to add a detach and retry loop on the interface detach
    on the guest so that we can wait until the interface is gone
    from the guest before nova unbinds the port in neutron, which is
    what the Tempest test is waiting for. Then the device should be
    available for attaching to the second guest.

    This is similar to what we do with detaching volumes.

    Closes-Bug: #1607714

    Change-Id: Ic04aad8923ea2edf1d16e32c208cd41fdf898834
    (cherry picked from commit a3b3e8d8314b0cedc2604be509f0f4d523a35ed5)

Reviewed: https://review.openstack.org/470348
Committed: https://git.openstack.org/cgit/openstack/nova/commit/?id=70c44831a5cdf605365a0b60994e2a458907d27f
Submitter: Jenkins
Branch: stable/newton

commit 70c44831a5cdf605365a0b60994e2a458907d27f
Author: Matt Riedemann <email address hidden>
Date: Thu Feb 9 18:41:11 2017 -0500

    libvirt: fix and break up _test_attach_detach_interface

    The detach_interface flow in this test was broken because
    it wasn't mocking out domain.detachDeviceFlags so the xml
    it was expecting to be passed to that method wasn't actually
    being verified. The same thing is broken in test
    test_detach_interface_device_with_same_mac_address because
    it copies the other broken test code.

    This change breaks apart the monster attach/detach test method
    and converts the detach_interface portion to mock and fixes
    the broken assertion.

    test_detach_interface_device_with_same_mac_address is just
    fixed, not converted to mock.

    Conflicts:
          nova/tests/unit/virt/libvirt/test_driver.py

    NOTE(mriedem): The conflict is due to change
    I5c461a8242c51994d12ce9c6774d5f956232f950 not being in Newton.

    Change-Id: I6d9a975876c5652ef544c587f65b1bdd1543848b
    Related-Bug: #1607714
    (cherry picked from commit 1ecd71b08d14450e475dc9512d40828da6fcfe15)
    (cherry picked from commit c0820944ea8554f5b1db0d538781caf0e75b6b0e)

Reviewed: https://review.openstack.org/470349
Committed: https://git.openstack.org/cgit/openstack/nova/commit/?id=1e66b034eb2c7176d9e163d7944b04d1928e148f
Submitter: Jenkins
Branch: stable/newton

commit 1e66b034eb2c7176d9e163d7944b04d1928e148f
Author: Matt Riedemann <email address hidden>
Date: Thu Feb 9 15:54:41 2017 -0500

    libvirt: wait for interface detach from the guest

    The test_reassign_port_between_servers test in Tempest creates
    a port in neutron and two servers. It attaches the port to the
    first server and then quickly detaches it and waits for the
    port.device_id to be unbound from the server. Then it repeats
    that for the second server.

    The interface detach from the guest is asynchronous and happens
    before nova unbinds the port, so there is a race where the port's
    device_id is unset but the interface is still on the first guest
    when we try to attach to the second guest, which fails.

    This is a latent bug, but apparently has been tickled by the
    move to our neutron CI jobs to use ubuntu xenial nodes.

    The fix is to add a detach and retry loop on the interface detach
    on the guest so that we can wait until the interface is gone
    from the guest before nova unbinds the port in neutron, which is
    what the Tempest test is waiting for. Then the device should be
    available for attaching to the second guest.

    This is similar to what we do with detaching volumes.

    Closes-Bug: #1607714

    Conflicts:
          nova/tests/unit/virt/libvirt/test_driver.py

    NOTE(mriedem): The conflict is due to change
    I5c461a8242c51994d12ce9c6774d5f956232f950 not being in Newton.

    Change-Id: Ic04aad8923ea2edf1d16e32c208cd41fdf898834
    (cherry picked from commit a3b3e8d8314b0cedc2604be509f0f4d523a35ed5)
    (cherry picked from commit ca0a46e36615f227f91f92d746916bbf17d1143c)

This issue was fixed in the openstack/nova 15.0.7 release.

This issue was fixed in the openstack/nova 14.0.8 release.