This patch series contains the set of changes to correctly set up
the infra for PF_RDS sockets that use TCP as the transport in multiple
network namespaces.
Patch 1 in the series is the minimal set of changes to allow
a single instance of RDS-TCP to run in any (i.e init_net or other)
namespace. The changes in this patch set ensure that the
execution of 'modprobe [-r] rds_tcp' correctly sets up the kernel
TCP sockets relative to the current netns.
Patch 2 of the series further allows multiple RDS-TCP instances,
one per network namespace. The changes in this patch allows dynamic
creation/tear-down of RDS-TCP client and server sockets across all
current and future namespaces.
Comments are specifically invited about the following:
There is some question in my mind as to whether Patch 2 should
use register_pernet_subsys() or register_pernet_device(): due
to the nature of the architecture, RDS/TCP is not a network device,
but more accurately a subsystem that encapsulates an RDS packet into
a TCP/IP header at the ksocket layer. However, the listen socket
is created as part of the ->init in the pernet_operations, and the
connect/accept sockets get created in the kernel dynamically, with the
intention that all of these sockets should be cleaned as part of ->exit.
Based on the comments in net_namespace.h, sockets would need
to be cleaned up as part of a pernet operation, else they would
hold up lo cleanup. In the current version of patch2, that cleanup is
achieved after the ethernet devices, by the socket keepalive timeout,
after which the ->exit will get called. I'm not sure there is a clean
way to avoid this. As thing stand, doing "ip netns delete <name>"
would result in syslogd messages about "unregister_netdevice: waiting
for lo to become free. Usage count .." being seen in the interval between
ethernet device migration to init_net and the keepalive timeout
Patch 3 in this set is independant of the above two changes, and is
a bugfix/follow up to eeb1bd5c encountered while testing the above.
Sowmini Varadhan (3):
Make RDS-TCP work correctly when it is set up in a netns other than
init_net
Support multiple RDS-TCP listen endpoints, one per netns.
sk_clone_lock() should only do get_net() if the parent is not a
kernel socket
net/core/sock.c | 3 +-
net/rds/bind.c | 3 +-
net/rds/connection.c | 16 ++++---
net/rds/ib.c | 2 +-
net/rds/ib_cm.c | 4 +-
net/rds/iw.c | 2 +-
net/rds/iw_cm.c | 4 +-
net/rds/rds.h | 11 +++--
net/rds/send.c | 3 +-
net/rds/tcp.c | 116 ++++++++++++++++++++++++++++++++++++++++++-------
net/rds/tcp.h | 7 ++-
net/rds/tcp_connect.c | 9 +++-
net/rds/tcp_listen.c | 40 ++++++-----------
net/rds/transport.c | 4 +-
14 files changed, 155 insertions(+), 69 deletions(-)
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
