Re: [PATCH 3/5] net: add udp limit to net cgroup

On Wed, Aug 10, 2016 at 05:53:16PM -0700, Anoop Naravaram wrote:

udp port limit
--------------
This property controls the limit of udp ports that can be used by the
processes in a cgroup. The controller manages udp statistics (usage,
limit, etc) for each cgroup. Every cgroup also keeps track of the udp
ports acquired by its descendants. If a process tries to acquire a port
when its cgroup has already reached its limit, it will fail with error
EACCES. It will also fail if one of the cgroup's ancestors has reached its
limit. There are 5 files exposed to userspace to configure this property:

* 'net.udp_usage': Reading this file gives the number of udp ports used by
processes in this cgroup and all its descendants.
* 'net.udp_limit': Writing this file sets the total number of udp ports
that can be used by processes in this cgroup and all
its descendants. This file can also be read.
* 'net.udp_maxusage': Reading this file gives the highest value of
net.udp_usage that has been seen for this cgroup.
* 'net.udp_failcnt': Reading this file gives the number of times a
process in this cgroup or one of its descendants has attempted to acquire
a udp port but failed because the limit of this cgroup was reached.
* 'net.udp_underflowcnt': Reading this file gives the number of times a
process in this cgroup or one of its descendants released a udp port when
the usage value of this cgroup was 0.

I have similar concern here. I don't think we should bloat the kernel by
trying implement every possible port restriction combination and count.
For the same reasons we don't do all possible tcp stats that research
community finds useful. This additional code is a maintenance headache.

On Fri, Aug 12, 2016 at 9:35 PM, Alexei Starovoitov
[off-list ref] wrote:
[...]

quoted

* 'net.udp_usage': Reading this file gives the number of udp ports used by
processes in this cgroup and all its descendants.
* 'net.udp_limit': Writing this file sets the total number of udp ports
that can be used by processes in this cgroup and all
its descendants. This file can also be read.
* 'net.udp_maxusage': Reading this file gives the highest value of
net.udp_usage that has been seen for this cgroup.
* 'net.udp_failcnt': Reading this file gives the number of times a
process in this cgroup or one of its descendants has attempted to acquire
a udp port but failed because the limit of this cgroup was reached.
* 'net.udp_underflowcnt': Reading this file gives the number of times a
process in this cgroup or one of its descendants released a udp port when
the usage value of this cgroup was 0.

I have similar concern here. I don't think we should bloat the kernel by
trying implement every possible port restriction combination and count.
For the same reasons we don't do all possible tcp stats that research
community finds useful. This additional code is a maintenance headache.

At this moment if a rampant process decides to just open a datagram
socket and just binds (does not listen) and does this for all
available ports, you are guaranteed to have machine becoming useless
for some (totally unrelated) legitimate process trying to do useful
things.

From kernel we can't control such user-space behavior, but at least

restrict the damage to "a specific set of processes". This code
provides that feature to restrict the damage.

Additional code is always a maintenance headache but that doesn't stop
us from adding new code, isn't it? :)