On Wed, Mar 4, 2020 at 5:36 PM Eric Dumazet <edumazet@xxxxxxxxxx> wrote:
>
> On Wed, Mar 4, 2020 at 3:39 PM Shakeel Butt <shakeelb@xxxxxxxxxx> wrote:
> >
> > If a TCP socket is allocated in IRQ context or cloned from unassociated
> > (i.e. not associated to a memcg) in IRQ context then it will remain
> > unassociated for its whole life. Almost half of the TCPs created on the
> > system are created in IRQ context, so, memory used by such sockets will
> > not be accounted by the memcg.
> >
> > This issue is more widespread in cgroup v1 where network memory
> > accounting is opt-in but it can happen in cgroup v2 if the source socket
> > for the cloning was created in root memcg.
> >
> > To fix the issue, just do the late association of the unassociated
> > sockets at accept() time in the process context and then force charge
> > the memory buffer already reserved by the socket.
> >
> > Signed-off-by: Shakeel Butt <shakeelb@xxxxxxxxxx>
> > ---
> > Changes since v1:
> > - added sk->sk_rmem_alloc to initial charging.
> > - added synchronization to get memory usage and set sk_memcg race-free.
> >
> > net/ipv4/inet_connection_sock.c | 19 +++++++++++++++++++
> > 1 file changed, 19 insertions(+)
> >
> > diff --git a/net/ipv4/inet_connection_sock.c b/net/ipv4/inet_connection_sock.c
> > index a4db79b1b643..7bcd657cd45e 100644
> > --- a/net/ipv4/inet_connection_sock.c
> > +++ b/net/ipv4/inet_connection_sock.c
> > @@ -482,6 +482,25 @@ struct sock *inet_csk_accept(struct sock *sk, int flags, int *err, bool kern)
> > }
> > spin_unlock_bh(&queue->fastopenq.lock);
> > }
> > +
> > + if (mem_cgroup_sockets_enabled && !newsk->sk_memcg) {
> > + int amt;
> > +
> > + /* atomically get the memory usage and set sk->sk_memcg. */
> > + lock_sock(newsk);
> > +
> > + /* The sk has not been accepted yet, no need to look at
> > + * sk->sk_wmem_queued.
> > + */
> > + amt = sk_mem_pages(newsk->sk_forward_alloc +
> > + atomic_read(&sk->sk_rmem_alloc));
> > + mem_cgroup_sk_alloc(newsk);
> > +
> > + release_sock(newsk);
> > +
> > + if (newsk->sk_memcg)
>
> Most sockets in accept queue should have amt == 0, so maybe avoid
> calling this thing only when amt == 0 ?
>
Thanks, will do in the next version. BTW I have tested with adding
mdelay() here and running iperf3 and I did see non-zero amt.
> Also I would release_sock(newsk) after this, otherwise incoming
> packets could mess with newsk->sk_forward_alloc
>
I think that is fine. Once sk->sk_memcg is set then
mem_cgroup_charge_skmem() will be called for new incoming packets.
Here we just need to call mem_cgroup_charge_skmem() with amt before
sk->sk_memcg was set.
> if (amt && newsk->sk_memcg)
> mem_cgroup_charge_skmem(newsk->sk_memcg, amt);
> release_sock(newsk);
>
> Also, I wonder if mem_cgroup_charge_skmem() has been used at all
> these last four years
> on arches with PAGE_SIZE != 4096
>
> ( SK_MEM_QUANTUM is not anymore PAGE_SIZE, but 4096)
>
Oh so sk_mem_pages() does not really give the number of pages. Yeah
this needs a fix for non-4906 page size architectures. Though I can
understand why this has not been caught yet. Network memory accounting
is opt-in in cgroup v1 and most of the users still use v1. In cgroup
v2, it is enabled and there is no way to opt-out. Facebook is a
well-known v2 user and it seems like they don't have non-4096 page
size arch systems.
