A socket is pressure-aware when its protocol has pressure defined, that
is sk_has_memory_pressure(sk) != NULL, e.g. TCP. These protocols might
want to limit the usage of socket memory depending on both the state of
global & memcg pressure through sk_under_memory_pressure(sk).
While for allocation, memcg pressure will be simply ignored when usage
is under global limit (sysctl_mem[0]). This behavior has different impacts
on different cgroup modes. In cgroupv2 socket and other purposes share a
same memory limit, thus allowing sockmem to burst under memcg reclaiming
pressure could lead to longer stall, sometimes even OOM. While cgroupv1
has no such worries.
As a cloud service provider, we encountered a problem in our production
environment during the transition from cgroup v1 to v2 (partly due to the
heavy taxes of accounting socket memory in v1). Say one workload behaves
fine in cgroupv1 with memcg limit configured to 10GB memory and another
1GB tcpmem, but will suck (or even be OOM-killed) in v2 with 11GB memory
due to burst memory usage on socket, since there is no specific limit for
socket memory in cgroupv2 and relies largely on workloads doing traffic
control themselves.
It's rational for the workloads to build some traffic control to better
utilize the resources they bought, but from kernel's point of view it's
also reasonable to suppress the allocation of socket memory once there is
a shortage of free memory, given that performance degradation is better
than failure.
As per the above, this patch aims to be more conservative on allocation
for the pressure-aware sockets under global and/or memcg pressure. While
OTOH throttling on incoming traffic could hurt latency badly possibly
due to SACKed segs get dropped from the OFO queue. See a related commit
720ca52bcef22 ("net-memcg: avoid stalls when under memory pressure").
This patch preserves this decision by throttling RX allocation only at
critical pressure level when it hardly makes sense to continue receive
data.
No functional change intended for pressure-unaware protocols.
Signed-off-by: Abel Wu <wuyun.abel@xxxxxxxxxxxxx>
---
net/core/sock.c | 29 +++++++++++++++++++++++++++--
1 file changed, 27 insertions(+), 2 deletions(-)
diff --git a/net/core/sock.c b/net/core/sock.c
index af778fc60a4d..6c1d13547f1b 100644
--- a/net/core/sock.c
+++ b/net/core/sock.c
@@ -3041,6 +3041,7 @@ EXPORT_SYMBOL(sk_wait_data);
int __sk_mem_raise_allocated(struct sock *sk, int size, int amt, int kind)
{
struct mem_cgroup *memcg = mem_cgroup_sockets_enabled ? sk->sk_memcg : NULL;
+ bool under_memcg_pressure = false;
struct proto *prot = sk->sk_prot;
bool charged = false;
long allocated;
@@ -3051,13 +3052,25 @@ int __sk_mem_raise_allocated(struct sock *sk, int size, int amt, int kind)
if (memcg) {
if (!mem_cgroup_charge_skmem(memcg, amt, gfp_memcg_charge()))
goto suppress_allocation;
+
+ /* Get pressure info from net-memcg. But consider the memcg
+ * to be under pressure for incoming traffic iff at 'critical'
+ * level, see commit 720ca52bcef22 ("net-memcg: avoid stalls
+ * when under memory pressure").
+ */
+ if (sk_has_memory_pressure(sk) &&
+ mem_cgroup_under_socket_pressure(memcg, &under_memcg_pressure) &&
+ !in_softirq())
+ under_memcg_pressure = true;
+
charged = true;
}
/* Under limit. */
if (allocated <= sk_prot_mem_limits(sk, 0)) {
sk_leave_memory_pressure(sk);
- return 1;
+ if (!under_memcg_pressure)
+ return 1;
}
/* Under pressure. */
@@ -3087,8 +3100,20 @@ int __sk_mem_raise_allocated(struct sock *sk, int size, int amt, int kind)
if (sk_has_memory_pressure(sk)) {
u64 alloc;
- if (!sk_under_memory_pressure(sk))
+ /* Be more conservative if the socket's memcg (or its
+ * parents) is under reclaim pressure, try to possibly
+ * avoid further memstall.
+ */
+ if (under_memcg_pressure)
+ goto suppress_allocation;
+
+ if (!sk_under_global_memory_pressure(sk))
return 1;
+
+ /* Trying to be fair among all the sockets of same
+ * protocal under global memory pressure, by allowing
+ * the ones that under average usage to raise.
+ */
alloc = sk_sockets_allocated_read_positive(sk);
if (sk_prot_mem_limits(sk, 2) > alloc *
sk_mem_pages(sk->sk_wmem_queued +
--
2.37.3
