In GetCurrentNS, If there is a context-switch between
getCurrentThreadNetNSPath and GetNS, another goroutine may execute in
the original thread and change its network namespace, then the original
goroutine would get the updated network namespace, which could lead to
unexpected behavior, especially when GetCurrentNS is used to get the
host network namespace in netNS.Do.
The added test has a chance to reproduce it with "-count=50".
The patch fixes it by locking the thread in GetCurrentNS.
Signed-off-by: Quan Tian <qtian@vmware.com>
if the runtime is not passing portMappings in the runtimeConfig,
then DEL is a noop.
This solves performance issues, when the portmap plugin is
executed multiple times, holding the iptables lock, despite
it does not have anything to delete.
Signed-off-by: Antonio Ojea <aojea@redhat.com>
It may happen that you want to map a port only in one IP family.
It can be achieved using the unspecified IP address of the
corresponding IP family as HostIP i.e.:
podman run --rm --name some-nginx -d -p 0.0.0.0:8080:80 nginx
The problem is that current implementation considers the
unspecified address valid and appends it to the iptables rule:
-A CNI-DN-60380cb3197c5457ed6ba -s 10.88.0.0/16
-d 0.0.0.0/32 -p tcp -m tcp --dport 8080 -j CNI-HOSTPORT-SETMARK
This rule is not forwarding the traffic to the mapped port.
We should use the unspecified address only to discriminate the IP
family of the port mapping, but not use it to filter the dst.
Signed-off-by: Antonio Ojea <antonio.ojea.garcia@gmail.com>
This change sets the mac address if specified during the creation of the
macvlan interface. This is superior to setting it via the tuning plugin
because this ensures the mac address is set before an IP is set,
allowing a container to get a reserved IP address from DHCP.
Related #450
Signed-off-by: Clint Armstrong <clint@clintarmstrong.net>
When trying to move a master and slave interface into a container it is not
possible without first bringing the interfaces down. This change ensures
that the interface is set to down prior to trying to move the interface
into the container. This matches the behaviour on moving an interface out
of the container.
Signed-off-by: cns <christopher.swindle@metaswitch.com>
A /64 mask was used which routed an entire cidr based on source,
not only the bound address.
Fixes #478
Signed-off-by: Lars Ekman <lars.g.ekman@est.tech>
The DNAT hairpin rule only allow the container itself to access the
ports it is exposing thru the host IP. Other containers in the same
subnet might also want to access this service via the host IP, so
apply this rule to the whole subnet instead of just for the container.
This is particularly useful with setups using a reverse proxy for
https. With such a setup connections between containers (for ex.
oauth2) have to downgrade to http, or need complex dns setup to make
use of the internal IP of the reverse proxy. On the other hand going
thru the host IP is easy as that is probably what the service name
already resolve to.
Signed-off-by: Alban Bedel <albeu@free.fr>
--
v2: Fixed the tests
v3: Updated iptables rules documentation in README.md
v4: Fixed the network addresses in README.md to match iptables output
If the pluging receives portMappings in runtimeConfig, the pluing will add a NAT policy for each port mapping on the generated endpoints.
It enables HostPort usage on Windows with win-bridge.
Signed-off-by: Vincent Boulineau <vincent.boulineau@datadoghq.com>
fix #463
link host veth pair to bridge, the Initial state
of port is BR_STATE_DISABLED and change to
BR_STATE_FORWARDING async.
Signed-off-by: honglichang <honglichang@tencent.com>
The current ns package code is very careful about not leaving the calling
thread with the overridden namespace set, for example when origns.Set() fails.
This is achieved by starting a new green thread, locking its OS thread, and
never unlocking it. Which makes golang runtime to scrap the OS thread backing
the green thread after the go routine exits.
While this works, it's probably not as optimal: stopping and starting a new OS
thread is expensive and may be avoided if we unlock the thread after resetting
network namespace to the original. On the other hand, if resetting fails, it's
better to leave the thread locked and die.
While it won't work in all cases, we can still make an attempt to reuse the OS
thread when resetting the namespace succeeds. This can be achieved by unlocking
the thread conditionally to the namespace reset success.
Signed-off-by: Ihar Hrachyshka <ihrachys@redhat.com>
