Re: [PATCH] net: limit a number of namespaces which can be cleaned up concurrently

On Sat, Oct 15, 2016 at 9:36 AM, Eric W. Biederman
[off-list ref] wrote:

Andrei Vagin [off-list ref] writes:

quoted

On Thu, Oct 13, 2016 at 10:06:28PM -0500, Eric W. Biederman wrote:

quoted

Andrei Vagin [off-list ref] writes:

quoted

On Thu, Oct 13, 2016 at 10:49:38AM -0500, Eric W. Biederman wrote:

quoted

Andrei Vagin [off-list ref] writes:

quoted

From: Andrey Vagin <redacted>

The operation of destroying netns is heavy and it is executed under
net_mutex. If many namespaces are destroyed concurrently, net_mutex can
be locked for a long time. It is impossible to create a new netns during
this period of time.

This may be the right approach or at least the right approach to bound
net_mutex hold times but I have to take exception to calling network
namespace cleanup heavy.

The only particularly time consuming operation I have ever found are calls to
synchronize_rcu/sycrhonize_sched/synchronize_net.

I booted the kernel with maxcpus=1, in this case these functions work
very fast and the problem is there any way.

Accoding to perf, we spend a lot of time in kobject_uevent:

-   99.96%     0.00%  kworker/u4:1     [kernel.kallsyms]  [k] unregister_netdevice_many
   - unregister_netdevice_many
      - 99.95% rollback_registered_many
         - 99.64% netdev_unregister_kobject
            - 33.43% netdev_queue_update_kobjects
               - 33.40% kobject_put
                  - kobject_release
                     + 33.37% kobject_uevent
                     + 0.03% kobject_del
               + 0.03% sysfs_remove_group
            - 33.13% net_rx_queue_update_kobjects
               - kobject_put
               - kobject_release
                  + 33.11% kobject_uevent
                  + 0.01% kobject_del
                    0.00% rx_queue_release
            - 33.08% device_del
               + 32.75% kobject_uevent
               + 0.17% device_remove_attrs
               + 0.07% dpm_sysfs_remove
               + 0.04% device_remove_class_symlinks
               + 0.01% kobject_del
               + 0.01% device_pm_remove
               + 0.01% sysfs_remove_file_ns
               + 0.00% klist_del
               + 0.00% driver_deferred_probe_del
                 0.00% cleanup_glue_dir.isra.14.part.15
                 0.00% to_acpi_device_node
                 0.00% sysfs_remove_group
              0.00% klist_del
              0.00% device_remove_attrs
         + 0.26% call_netdevice_notifiers_info
         + 0.04% rtmsg_ifinfo_build_skb
         + 0.01% rtmsg_ifinfo_send
        0.00% dev_uc_flush
        0.00% netif_reset_xps_queues_gt

Someone can listen these uevents, so we can't stop sending them without
breaking backward compatibility. We can try to optimize
kobject_uevent...

Oh that is a surprise.  We can definitely skip genenerating uevents for
network namespaces that are exiting because by definition no one can see
those network namespaces.  If a socket existed that could see those
uevents it would hold a reference to the network namespace and as such
the network namespace could not exit.

That sounds like it is worth investigating a little more deeply.

I am surprised that allocation and freeing is so heavy we are spending
lots of time doing that.  On the other hand kobj_bcast_filter is very
dumb and very late so I expect something can be moved earlier and make
that code cheaper with the tiniest bit of work.

I'm sorry, I've collected this data for a kernel with debug options
(DEBUG_SPINLOCK, PROVE_LOCKING, DEBUG_LIST, etc). If a kernel is
compiled without debug options, kobject_uevent becomes less expensive,
but still expensive.

-   98.64%     0.00%  kworker/u4:2  [kernel.kallsyms]    [k] cleanup_net
   - cleanup_net
      - 98.54% ops_exit_list.isra.4
         - 60.48% default_device_exit_batch
            - 60.40% unregister_netdevice_many
               - rollback_registered_many
                  - 59.82% netdev_unregister_kobject
                     - 20.10% device_del
                        + 19.44% kobject_uevent
                        + 0.40% device_remove_attrs
                        + 0.17% dpm_sysfs_remove
                        + 0.04% device_remove_class_symlinks
                        + 0.04% kobject_del
                        + 0.01% device_pm_remove
                        + 0.01% sysfs_remove_file_ns
                     - 19.89% netdev_queue_update_kobjects
                        + 19.81% kobject_put
                        + 0.07% sysfs_remove_group
                     - 19.79% net_rx_queue_update_kobjects
                          kobject_put
                        - kobject_release
                           + 19.77% kobject_uevent
                           + 0.02% kobject_del
                             0.01% rx_queue_release
                     + 0.02% kset_unregister
                       0.01% pm_runtime_set_memalloc_noio
                       0.01% bus_remove_device
                  + 0.45% call_netdevice_notifiers_info
                  + 0.07% rtmsg_ifinfo_build_skb
                  + 0.04% rtmsg_ifinfo_send
                    0.01% kset_unregister
            + 0.07% rtnl_unlock
         + 19.27% rpcsec_gss_exit_net
         + 5.45% tcp_net_metrics_exit
         + 5.31% sunrpc_exit_net
         + 3.18% ip6addrlbl_net_exit


So after removing kobject_uevent, cleanup_net becomes more than two times faster:

1000 namespaces are cleaned up for 2.8 seconds with uevents, and 1.2 senconds
without uevents. I do this experiments with max_cpus=1 to exclude synchronize_rcu.

As a summary we can skip generating uevents, but it doesn't solve the original
problem. If we want to avoid the limit introduced in this patch, we have
to reduce the time for destroing net namespace in dozen times, don't
we?

It definitely looks like optimizing kobject_uevent for this case is
worth while.

I would not mind getting the raw cost of network namespace cleanups
below 2.8ms or with uevent cleanups 1.2ms.  There is just a lot going on
for a lot of good reasons in the networking stack so that can be tricky.

The larger issue is that there is a trade off between latency and
throughput in network namespace destruction.  Consider the case of
vsftpd.  Which creates a new network namespace for every connection.
Something like that can wind up with a huge backlog of network
namespaces to clean up while continually creating more.  The system will
go OOM if we don't stop and cleanup what we have.

And the batching is very very important for throughput.  So the smallest
batch size we could really accept is a batch size that does not hurt
throughput when destroying network namespaces.  Otherwise we will have a
growing backlog of network namespaces to cleanup and a system that
eventuallys stops being usable at all.  In that context I think a long
hold time on net_mutex is preferable to a system that does not work at
all.

Now I would love to make both the throughput and the latency better I
would be all in favor of that, but that requires some deep changes to
the network namespace initialization and cleanup.  Unfortunately I
haven't stared at the problem enough to know what those changes would
need to be.  But something where we would not need to serialize network
namespace cleanup between different network namespaces.  And ideally
something we could implement incrementally as there is so much
networking code I don't expect we could verify and change verything
overnight.

Eric, I get your point. All these sounds reasonable.

And here is another idea about net_mutex.

The longer I look at net_mutex the more it looks like that it can be replaced
on a read-write lock. It protects per-namespace lists of operations, which are
modified only when modules are loaded or unloaded. And the kernel reads
these lists to create or destroy a new network namespace.

Eric and David, what do you think about this idea? Do you have any ideas
why it will not work.

I don't know this code so well to not skip something obvious. The attached
patch shows how it looks like.

If it will works we will be able to create and destroy net namespaces
concurrently.
And even call cleanup_net() from a few threads if we have a big
backlog. It is only
one of steps which may be useful to fix this problem.

That plus in practice the bottleneck has always been the synchronize_rcu
calls which tend to take at least a millisecond a piece.  Being able
overlap those synchronize_rcu calls in the common case has reduced
the time to run the network stack cleanup code by very dramatic amounts.

Right now I am very happy that the network namespace cleanup code is
working properly.  When I started the network stack cleanup code to
cleanup network namespaces I found actual functional bugs.  I will be
even happier if we can figure out how to make it all run fast.

But ultimately we have the net_mutex and the rtnl_lock that serialize
things on the setup and cleanup paths and to allow creation to proceed
while cleanup is ongoing we need to find a way to avoid serialization by
either of those, and I have honestly drawn a blank.

So right now my best suggestion for making things better is to find and
fix each little piece we can fix.  Until the things are working as best
we can make them work.  It is not sexy or glamorous or fast but it makes
things better and is the best that I can see to do.

Eric

quoted

Here is a perf report after skipping generating uevents:
-   93.27%     0.00%  kworker/u4:1  [kernel.kallsyms]   [k] cleanup_net
   - cleanup_net
      - 92.97% ops_exit_list.isra.4
         - 35.14% rpcsec_gss_exit_net
            - gss_svc_shutdown_net
               - 17.40% rsc_cache_destroy_net
                  + 8.64% cache_unregister_net
                  + 8.52% cache_purge
                  + 0.22% cache_destroy_net
               + 9.00% cache_unregister_net
               + 8.49% cache_purge
               + 0.15% destroy_use_gss_proxy_proc_entry
               + 0.10% cache_destroy_net
         - 14.35% tcp_net_metrics_exit
            - 7.32% tcp_metrics_flush_all
               + 4.86% _raw_spin_unlock_bh
                 0.59% __local_bh_enable_ip
              6.12% _raw_spin_lock_bh
              0.90% _raw_spin_unlock_bh
         - 13.08% sunrpc_exit_net
            - 6.91% ip_map_cache_destroy
               + 3.90% cache_unregister_net
               + 2.86% cache_purge
               + 0.15% cache_destroy_net
            + 5.95% unix_gid_cache_destroy
            + 0.12% rpc_pipefs_exit_net
            + 0.10% rpc_proc_exit
         - 7.35% ip6addrlbl_net_exit
            + call_rcu_sched
         + 3.34% xfrm_net_exit
         + 1.22% ipv6_frags_exit_net
         + 1.17% ipv4_frags_exit_net
         + 0.78% fib_net_exit
         + 0.76% inet6_net_exit
         + 0.76% devinet_exit_net
         + 0.68% addrconf_exit_net
         + 0.63% igmp6_net_exit
         + 0.59% ipv4_mib_exit_net
         + 0.59% uevent_net_exit

quoted

Eric

_______________________________________________
Containers mailing list
Containers-cunTk1MwBs9QetFLy7KEm3xJsTq8ys+cHZ5vskTnxNA@public.gmane.org
https://lists.linuxfoundation.org/mailman/listinfo/containers