On Wed, May 10, 2023 at 1:11 PM Zhang, Cathy <cathy.zhang@xxxxxxxxx> wrote:
>
> Hi Shakeel, Eric and all,
>
> How about adding memory pressure checking in sk_mem_uncharge()
> to decide if keep part of memory or not, which can help avoid the issue
> you fixed and the problem we find on the system with more CPUs.
>
> The code draft is like this:
>
> static inline void sk_mem_uncharge(struct sock *sk, int size)
> {
> int reclaimable;
> int reclaim_threshold = SK_RECLAIM_THRESHOLD;
>
> if (!sk_has_account(sk))
> return;
> sk->sk_forward_alloc += size;
>
> if (mem_cgroup_sockets_enabled && sk->sk_memcg &&
> mem_cgroup_under_socket_pressure(sk->sk_memcg)) {
> sk_mem_reclaim(sk);
> return;
> }
>
> reclaimable = sk->sk_forward_alloc - sk_unused_reserved_mem(sk);
>
> if (reclaimable > reclaim_threshold) {
> reclaimable -= reclaim_threshold;
> __sk_mem_reclaim(sk, reclaimable);
> }
> }
>
> I've run a test with the new code, the result looks good, it does not introduce
> latency, RPS is the same.
>
It will not work for sockets that are idle, after a burst.
If we restore per socket caches, we will need a shrinker.
Trust me, we do not want that kind of big hammer, crushing latencies.
Have you tried to increase batch sizes ?
Any kind of cache (even per-cpu) might need some adjustment when core
count or expected traffic is increasing.
This was somehow hinted in
commit 1813e51eece0ad6f4aacaeb738e7cced46feb470
Author: Shakeel Butt <shakeelb@xxxxxxxxxx>
Date: Thu Aug 25 00:05:06 2022 +0000
memcg: increase MEMCG_CHARGE_BATCH to 64
diff --git a/include/linux/memcontrol.h b/include/linux/memcontrol.h
index 222d7370134c73e59fdbdf598ed8d66897dbbf1d..0418229d30c25d114132a1ed46ac01358cf21424
100644
--- a/include/linux/memcontrol.h
+++ b/include/linux/memcontrol.h
@@ -334,7 +334,7 @@ struct mem_cgroup {
* TODO: maybe necessary to use big numbers in big irons or dynamic
based of the
* workload.
*/
-#define MEMCG_CHARGE_BATCH 64U
+#define MEMCG_CHARGE_BATCH 128U
extern struct mem_cgroup *root_mem_cgroup;
diff --git a/include/net/sock.h b/include/net/sock.h
index 656ea89f60ff90d600d16f40302000db64057c64..82f6a288be650f886b207e6a5e62a1d5dda808b0
100644
--- a/include/net/sock.h
+++ b/include/net/sock.h
@@ -1433,8 +1433,8 @@ sk_memory_allocated(const struct sock *sk)
return proto_memory_allocated(sk->sk_prot);
}
-/* 1 MB per cpu, in page units */
-#define SK_MEMORY_PCPU_RESERVE (1 << (20 - PAGE_SHIFT))
+/* 2 MB per cpu, in page units */
+#define SK_MEMORY_PCPU_RESERVE (1 << (21 - PAGE_SHIFT))
static inline void
sk_memory_allocated_add(struct sock *sk, int amt)
> > -----Original Message-----
> > From: Shakeel Butt <shakeelb@xxxxxxxxxx>
> > Sent: Wednesday, May 10, 2023 12:10 AM
> > To: Eric Dumazet <edumazet@xxxxxxxxxx>; Linux MM <linux-
> > mm@xxxxxxxxx>; Cgroups <cgroups@xxxxxxxxxxxxxxx>
> > Cc: Zhang, Cathy <cathy.zhang@xxxxxxxxx>; Paolo Abeni
> > <pabeni@xxxxxxxxxx>; davem@xxxxxxxxxxxxx; kuba@xxxxxxxxxx;
> > Brandeburg, Jesse <jesse.brandeburg@xxxxxxxxx>; Srinivas, Suresh
> > <suresh.srinivas@xxxxxxxxx>; Chen, Tim C <tim.c.chen@xxxxxxxxx>; You,
> > Lizhen <lizhen.you@xxxxxxxxx>; eric.dumazet@xxxxxxxxx;
> > netdev@xxxxxxxxxxxxxxx
> > Subject: Re: [PATCH net-next 1/2] net: Keep sk->sk_forward_alloc as a proper
> > size
> >
> > +linux-mm & cgroup
> >
> > Thread: https://lore.kernel.org/all/20230508020801.10702-1-
> > cathy.zhang@xxxxxxxxx/
> >
> > On Tue, May 9, 2023 at 8:43 AM Eric Dumazet <edumazet@xxxxxxxxxx>
> > wrote:
> > >
> > [...]
> > > Some mm experts should chime in, this is not a networking issue.
> >
> > Most of the MM folks are busy in LSFMM this week. I will take a look at this
> > soon.
