Kuniyuki Iwashima [off-list ref] writes:
Some protocols (e.g., TCP, UDP) implement memory accounting for socket
buffers and charge memory to per-protocol global counters pointed to by
sk->sk_proto->memory_allocated.
When running under a non-root cgroup, this memory is also charged to the
memcg as "sock" in memory.stat.
Even when a memcg controls memory usage, sockets of such protocols are
still subject to global limits (e.g., /proc/sys/net/ipv4/tcp_mem).
This makes it difficult to accurately estimate and configure appropriate
global limits, especially in multi-tenant environments.
If all workloads were guaranteed to be controlled under memcg, the issue
could be worked around by setting tcp_mem[0~2] to UINT_MAX.
In reality, this assumption does not always hold, and processes that
belong to the root cgroup or opt out of memcg can consume memory up to
the global limit, becoming a noisy neighbour.
Let's decouple memcg from the global per-protocol memory accounting if
it has a finite memory.max (!= "max").
I think you can't make the new behavior as the new default, simple because
it might break existing setups. Basically anyone who is using memory.max
will suddenly have their processes being opted out of the net memory
accounting. Idk how many users are actually relying on the network
memory accounting, but I believe way more than 0.
So I guess a net sysctl/some other knob is warranted here, with the old
behavior being the default.
Thanks
On Tue, Aug 12, 2025 at 6:58 PM Roman Gushchin [off-list ref] wrote:
Kuniyuki Iwashima [off-list ref] writes:
quoted
Some protocols (e.g., TCP, UDP) implement memory accounting for socket
buffers and charge memory to per-protocol global counters pointed to by
sk->sk_proto->memory_allocated.
When running under a non-root cgroup, this memory is also charged to the
memcg as "sock" in memory.stat.
Even when a memcg controls memory usage, sockets of such protocols are
still subject to global limits (e.g., /proc/sys/net/ipv4/tcp_mem).
This makes it difficult to accurately estimate and configure appropriate
global limits, especially in multi-tenant environments.
If all workloads were guaranteed to be controlled under memcg, the issue
could be worked around by setting tcp_mem[0~2] to UINT_MAX.
In reality, this assumption does not always hold, and processes that
belong to the root cgroup or opt out of memcg can consume memory up to
the global limit, becoming a noisy neighbour.
Let's decouple memcg from the global per-protocol memory accounting if
it has a finite memory.max (!= "max").
I think you can't make the new behavior as the new default, simple because
it might break existing setups. Basically anyone who is using memory.max
will suddenly have their processes being opted out of the net memory
accounting. Idk how many users are actually relying on the network
memory accounting, but I believe way more than 0.
So I guess a net sysctl/some other knob is warranted here, with the old
behavior being the default.
I think we don't need a knob to change the behaviour because
the affected case must have a broken assumption.
There are 3 possible cases below.
1) memory.max == "max"
2) memory.max != "max" and tcp_mem does not suppress
memory allocation
3) memory.max != "max" and tcp_mem suppresses memory
allocation
1) is not affected, and 2) is not affected too because decoupling
does not change the situation.
Then, for 3), this change will allow more memory than ever,
but it's still limited by memory.max, which is configured by
the sys admin.
If this could be a problem, then the total amount of all memcg's
memory.max should exceed the amount of system memory,
which is unlikely if configured properly.
Also, in the 3) case, TCP has quite bad performance and the
sys admin should have raised the tcp_mem limit and moved
to 2) like our unlimited tcp_mem setting.
What do you think ?