Currently updates to security group rules or membership
are handled by immediately triggering a call to refresh_firewall.
This call is quite expensive, and it is often executed with a
very high frequency.
With this patch, the notification handler simply adds devices for
which the firewall should be refreshed to a set, which will then
be processed in another routine. The latter is supposed to
be called in the main agent loop.
This patch for 'provider updates' simply sets a flag for refreshing
the firewall for all devices.
In order to avoid breaking other agents leveraging the security
group RPC mixin, the reactive behaviour is still available, and is
still the default way of handling security group updates.
NVP plugin:fix delete sec group when backend is out of sync
If a security group does not exist on the NVP backend, an error
should not be raised on deletion of the security group.
This patch changes the plugin behavior by deleting the record
from the database and just logging that the security group
was not found on the NVP backend.
Fawad Khaliq [Wed, 5 Feb 2014 18:15:13 +0000 (10:15 -0800)]
Fix error message typo
* Fix error message typo in "_network_admin_state"
function where "Network Admin State Validation Falied"
should be changed to "Network Admin State Validation Failed"
Change the behaviour of the L3 agent in order to set the IP addresses
for the floating IPs on the external gateway interface after the
relevant NAT rules have been applied.
This will avoid a transitory period in which the floating IP exists
and is reachable but it not yet wired to the actual target.
In the majority of cases a port_update notification pertains
a change in the properties affecting port filter, and does
not affect port wiring, ie: the local vlan tag.
This patch simply avoids doing port wiring/unwiring if the
local vlan tag did not change.
The extra overhead for the ovs-db get operation is offset
by the fact that get commands are generally faster than
set commands, and by avoiding executing the ovs-ofctl operation.
Process port_update notifications in the main agent loop
Instead of processing a port update notification directly in
the RPC call, the actual processing is moved into the main
rpc loop, whereas the RPC call just adds the updated port
identifier to a set of updated ports.
In this way, a port_update notification won't compete with the
main rpc loop, causing long delays into its completion under
heavy load. Also, repeated port_update notifications received
within a single iteration of the main agent loop will be
coalesced and processed only once.
This will also avoid the risk of processing notifications out
of order thus ending up with an actual configuration which
differs from the desired one.
This patch still performs L2 wiring for updated ports even if
it is necessary only when the administrative state of a port
changes.
The update_ports method has been renamed to scan_ports as the latter
name appears to be more in line with what the method actually does.
Brian Haley [Thu, 30 Jan 2014 20:05:49 +0000 (15:05 -0500)]
Change metadata-agent to have a configurable backlog
The metadata agent currently runs with a default socket backlog
of 128. This isn't enough on a busy network node, even when
spawning multiple worker processes.
This change addes a new "metadata_backlog = XX" to the ini file
to support a configurable value to help improve performance.
Brian Haley [Thu, 30 Jan 2014 19:39:47 +0000 (14:39 -0500)]
Change metadata-agent to spawn multiple workers
There is currently only one metadata-agent per network node,
which could be handling connections from hundreds or thousands
of metadata-namespace-proxy processes.
This change addes a new "metadata_workers = XX" to the ini file
to support creating more workers to help improve performance.
Tweak version nvp/nsx version validation logic for router operations
This patch improves how the nsx/nvp controller version is validated
prior to some router operations. This is done by greatly simplifying
the boolean condition.
Missing unit tests are also added for increased coverage.
Carl Baldwin [Mon, 18 Nov 2013 23:32:19 +0000 (23:32 +0000)]
Simplify ip allocation/recycling to relieve db pressure
I found that multiple calls to delete_port can pile up on the
_recycle_ip operation. This patch simplifies this operation. It
reduces the _recycle_ip operation to a single row delete in the ip
allocations table and doesn't touch the availability table.
To acheive the recycling of ips in a pool, this code runs a more
complex operation of rebuilding the availability table when it is
exhausted. Only one API process will perform this more expensive
operation and others waiting for allocation will immediately benefit.
The amortized cost of this operation is much less than the cumulative
cost of running the more expensive _recycle_ip operation for every
port delete.
IP allocation behaves a bit differently with this patch. Instead of
giving out the first IP available in a pool, the entire pool will be
allocated before wrapping around and recycling ip addresses that have
been released. This is a desirable feature as it puts ip addresses in
a sort of quarantine after they are released. It is easier to
distinguish newly allocated ips from old ones.
Carl Baldwin [Fri, 24 Jan 2014 22:35:48 +0000 (22:35 +0000)]
Reduce severity of log messages in validation methods
I noticed this while reviewing Ic2c87174. When I read through log
files, I don't want to see errors like this that come from validating
bad user input. Info severity is more appropriate.
Maru Newby [Mon, 20 Jan 2014 19:28:03 +0000 (19:28 +0000)]
Minimize the cost of checking for api worker exit
A recent change to oslo allows the configuration of the interval
that ProcessLauncher waits between checks of child exit. The
default interval of 0.01s resulted in the neutron service consuming
unnecessary cpu cycles checking whether api workers had exited (5%
cpu on idle in a VM). This patch extends the interval to 1s to
minimize the cost of the checks.
Aaron Rosen [Mon, 13 Jan 2014 21:57:04 +0000 (13:57 -0800)]
Remove and recreate interface if already exists
If the dhcp-agent machine restarts when openvswitch comes up it logs the
following warning messages for all tap interfaces that do not exist:
bridge|WARN|could not open network device tap2cf7dbad-9d (No such device)
Once the dhcp-agent starts it recreates the interfaces and re-adds them to the
ovs-bridge. Unfortunately, ovs does not reinitialize the interfaces as they
are already in ovsdb and does not assign them a ofport number.
This situation corrects itself though the next time a port is added to the
ovs-bridge which is why no one has probably noticed this issue till now.
In order to correct this we should first remove interface that exist and
then readd them.
Carl Baldwin [Fri, 17 Jan 2014 19:28:10 +0000 (19:28 +0000)]
Use an independent iptables lock per namespace
Since iptables is independent from namespace to namespace, it makes
sense to use an independent lock per namespace. This improvement is
aimed at improving the parallel performance in the L3 agent.
In the NVP plugin, metadata processing performs several plugin operations
with the same context (and db session). The first operation might leave
persisted objects in the session instance which then conflict with objects
created in the second operation.
Roman Podoliaka [Thu, 12 Dec 2013 06:20:15 +0000 (08:20 +0200)]
Fix the migration adding a UC to agents table
The migration script mistakenly assumes that all core
plugins use agents extension, which is not true (e.g.
plumgrid and bigswitch don't).
Apply this migration script only for plugins that are
stated in the original migration script adding agents
table (511471cc46b_agent_ext_model_supp.py).
Eugene Nikanorov [Mon, 13 Jan 2014 14:58:59 +0000 (18:58 +0400)]
Fix race condition in delete_port method. Fix update_port method
Port can be gone between
l3plugin.prevent_l3_port_deletion(context, id)
and port query. Need to handle it properly.
ALso need to handle non-existing port in update_port properly.
Carl Baldwin [Tue, 12 Nov 2013 22:52:47 +0000 (22:52 +0000)]
Use information from the dnsmasq hosts file to call dhcp_release
Certain situations can cause the DHCP agent's local cache to get out
of sync with the leases held internally by dnsmasq. This method of
detecting when to call dhcp_release is idempotent and not dependent on
the cache. It is more robust.
Fix pip install failure due to missing nvp.ini file
It looks like sdist does not support symlinks, therefore
letting nvp.ini point to nsx.ini is not a good solution.
Since nvp.ini is going away, leave a copy for now, but
add a warning so that users are aware of the switch,
whilst preserving full backward-compatibility.
This patch adds a new configuration variable for the timeout on
ovs-vsctl commands, and sets the default timeout to 10 seconds.
This is aimed at allowing users to tune the agents in order to avoid
timeout errors on their deployments.
Fix race condition on ml2 delete and update port methods
Synchronize access to ports table when deleting and updating
a port. Otherwise concurrent update/delete request for the same port
may cause neutron server to throw an exception and return
'500 Internal server error' for such requests.
When the rename of quantum->neutron occurred here ee3fe4e8 it also renamed
the the table creation from quantum_nvp_port_mapping to
neutron_nvp_port_mapping. This went undetected for a long time because
when neutron-server starts up it pushes down the scheme for tables that
are not there so the table would be created.
Because of this the following migration 50e86cb2637a called
op.rename_table('neutron_nvp_port_mapping', 'neutron_nsx_port_mappings')
though the table name being used was quantum_nvp_port_mapping. Because of this
the quantum_id->nvp_id mapping was never migrated over to the new table and
you would be left with a quantum_nvp_port_mapping table hanging around.
In addition, the downgrade would rename the table to neutron_nvp_port_mapping
instead of quantum_nvp_port_mapping. This patch addresses this issues.
Ann Kamyshnikova [Fri, 20 Sep 2013 11:48:37 +0000 (15:48 +0400)]
Update lockutils and fixture in openstack.common
lockutils: included commits: 8b2b0b7 Use hacking import_exceptions for gettextutils._ 6d0a6c3 Correct invalid docstrings 12bcdb7 Remove vim header 79e6bc6 fix lockutils.lock() to make it thread-safe ace5120 Add main() to lockutils that creates temp dir for locks 537d8e2 Allow lockutils to get lock_path conf from envvar 371fa42 Move LockFixture into a fixtures module d498c42 Fix to properly log when we release a semaphore 29d387c Add LockFixture to lockutils 3e3ac0c Modify lockutils.py due to dispose of eventlet 90b6a65 Fix locking bug 27d4b41 Move synchronized body to a first-class function 15c17fb Make lock_file_prefix optional 1a2df89 Enable H302 hacking check
fixture: created, included commits: 45658e2 Fix violations of H302:import only modules 12bcdb7 Remove vim header 3970d46 Fix typos in oslo 371fa42 Move LockFixture into a fixtures module f4a4855 Consolidate the use of stubs 6111131 Make openstack.common.fixture.config Py3 compliant 3906979 Add a fixture for dealing with config d332cca Add a fixture for dealing with mock patching. 1bc3ecf Start adding reusable test fixtures.
Also tox.ini was corrected to let lockutils work in tests.
This change is needed for work on bp: db-sync-models-with-migrations
Aaron Rosen [Wed, 8 Jan 2014 21:03:29 +0000 (13:03 -0800)]
Add test to port_security to test with security_groups
This patch adds a missing testcase to the port_security tests to test
for creating a port with port_security_enabled=False and passing in
a security group.
Code Review / openstack-build / neutron-build.git / log