Re: [PATCH v2 net-next 1/2] tcp: account for memory pressure signaled by cgroup

[Shakeel Butt <shakeel.butt@linux.dev>]

On Mon, Jul 14, 2025 at 04:36:12PM +0200, Daniel Sedlak wrote:
> This patch is a result of our long-standing debug sessions, where it all
> started as "networking is slow", and TCP network throughput suddenly
> dropped from tens of Gbps to few Mbps, and we could not see anything in
> the kernel log or netstat counters.
> 
> Currently, we have two memory pressure counters for TCP sockets [1],
> which we manipulate only when the memory pressure is signalled through
> the proto struct [2]. However, the memory pressure can also be signaled
> through the cgroup memory subsystem, which we do not reflect in the
> netstat counters. In the end, when the cgroup memory subsystem signals
> that it is under pressure, we silently reduce the advertised TCP window
> with tcp_adjust_rcv_ssthresh() to 4*advmss, which causes a significant
> throughput reduction.
> 
> So this patch adds a new counter to account for memory pressure
> signaled by the memory cgroup, so it is much easier to spot.
> 
> Link: https://elixir.bootlin.com/linux/v6.15.4/source/include/uapi/linux/snmp.h#L231-L232 [1]
> Link: https://elixir.bootlin.com/linux/v6.15.4/source/include/net/sock.h#L1300-L1301 [2]
> Co-developed-by: Matyas Hurtik <matyas.hurtik@cdn77.com>
> Signed-off-by: Matyas Hurtik <matyas.hurtik@cdn77.com>
> Signed-off-by: Daniel Sedlak <daniel.sedlak@cdn77.com>
> ---
>  Documentation/networking/net_cachelines/snmp.rst |  1 +
>  include/net/tcp.h                                | 14 ++++++++------
>  include/uapi/linux/snmp.h                        |  1 +
>  net/ipv4/proc.c                                  |  1 +
>  4 files changed, 11 insertions(+), 6 deletions(-)
> 
> diff --git a/Documentation/networking/net_cachelines/snmp.rst b/Documentation/networking/net_cachelines/snmp.rst
> index bd44b3eebbef..ed17ff84e39c 100644
> --- a/Documentation/networking/net_cachelines/snmp.rst
> +++ b/Documentation/networking/net_cachelines/snmp.rst
> @@ -76,6 +76,7 @@ unsigned_long  LINUX_MIB_TCPABORTONLINGER
>  unsigned_long  LINUX_MIB_TCPABORTFAILED
>  unsigned_long  LINUX_MIB_TCPMEMORYPRESSURES
>  unsigned_long  LINUX_MIB_TCPMEMORYPRESSURESCHRONO
> +unsigned_long  LINUX_MIB_TCPCGROUPSOCKETPRESSURE
>  unsigned_long  LINUX_MIB_TCPSACKDISCARD
>  unsigned_long  LINUX_MIB_TCPDSACKIGNOREDOLD
>  unsigned_long  LINUX_MIB_TCPDSACKIGNOREDNOUNDO
> diff --git a/include/net/tcp.h b/include/net/tcp.h
> index 761c4a0ad386..aae3efe24282 100644
> --- a/include/net/tcp.h
> +++ b/include/net/tcp.h
> @@ -267,6 +267,11 @@ extern long sysctl_tcp_mem[3];
>  #define TCP_RACK_STATIC_REO_WND  0x2 /* Use static RACK reo wnd */
>  #define TCP_RACK_NO_DUPTHRESH    0x4 /* Do not use DUPACK threshold in RACK */
>  
> +#define TCP_INC_STATS(net, field)	SNMP_INC_STATS((net)->mib.tcp_statistics, field)
> +#define __TCP_INC_STATS(net, field)	__SNMP_INC_STATS((net)->mib.tcp_statistics, field)
> +#define TCP_DEC_STATS(net, field)	SNMP_DEC_STATS((net)->mib.tcp_statistics, field)
> +#define TCP_ADD_STATS(net, field, val)	SNMP_ADD_STATS((net)->mib.tcp_statistics, field, val)
> +
>  extern atomic_long_t tcp_memory_allocated;
>  DECLARE_PER_CPU(int, tcp_memory_per_cpu_fw_alloc);
>  
> @@ -277,8 +282,10 @@ extern unsigned long tcp_memory_pressure;
>  static inline bool tcp_under_memory_pressure(const struct sock *sk)
>  {
>  	if (mem_cgroup_sockets_enabled && sk->sk_memcg &&
> -	    mem_cgroup_under_socket_pressure(sk->sk_memcg))
> +	    mem_cgroup_under_socket_pressure(sk->sk_memcg)) {
> +		TCP_INC_STATS(sock_net(sk), LINUX_MIB_TCPCGROUPSOCKETPRESSURE);
>  		return true;

Incrementing it here will give a very different semantic to this stat
compared to LINUX_MIB_TCPMEMORYPRESSURES. Here the increments mean the
number of times the kernel check if a given socket is under memcg
pressure for a net namespace. Is that what we want?