netdev - Re: [PATCH net-next 09/10] udp: make busylock per socket

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [day] [month] [year] [list]
Message-ID: <willemdebruijn.kernel.17894869e21ae@gmail.com>
Date: Tue, 16 Sep 2025 15:15:00 -0400
From: Willem de Bruijn <willemdebruijn.kernel@...il.com>
To: Eric Dumazet <edumazet@...gle.com>, 
 Willem de Bruijn <willemdebruijn.kernel@...il.com>
Cc: "David S . Miller" <davem@...emloft.net>, 
 Jakub Kicinski <kuba@...nel.org>, 
 Paolo Abeni <pabeni@...hat.com>, 
 Simon Horman <horms@...nel.org>, 
 Willem de Bruijn <willemb@...gle.com>, 
 Kuniyuki Iwashima <kuniyu@...gle.com>, 
 David Ahern <dsahern@...nel.org>, 
 netdev@...r.kernel.org, 
 eric.dumazet@...il.com
Subject: Re: [PATCH net-next 09/10] udp: make busylock per socket

Eric Dumazet wrote:
> On Tue, Sep 16, 2025 at 9:31 AM Willem de Bruijn
> <willemdebruijn.kernel@...il.com> wrote:
> >
> > Eric Dumazet wrote:
> > > While having all spinlocks packed into an array was a space saver,
> > > this also caused NUMA imbalance and hash collisions.
> > >
> > > UDPv6 socket size becomes 1600 after this patch.
> > >
> > > Signed-off-by: Eric Dumazet <edumazet@...gle.com>
> > > ---
> > >  include/linux/udp.h |  1 +
> > >  include/net/udp.h   |  1 +
> > >  net/ipv4/udp.c      | 20 ++------------------
> > >  3 files changed, 4 insertions(+), 18 deletions(-)
> > >
> > > diff --git a/include/linux/udp.h b/include/linux/udp.h
> > > index 6ed008ab166557e868c1918daaaa5d551b7989a7..e554890c4415b411f35007d3ece9e6042db7a544 100644
> > > --- a/include/linux/udp.h
> > > +++ b/include/linux/udp.h
> > > @@ -109,6 +109,7 @@ struct udp_sock {
> > >        */
> > >       struct hlist_node       tunnel_list;
> > >       struct numa_drop_counters drop_counters;
> > > +     spinlock_t              busylock ____cacheline_aligned_in_smp;
> > >  };
> > >
> > >  #define udp_test_bit(nr, sk)                 \
> > > diff --git a/include/net/udp.h b/include/net/udp.h
> > > index a08822e294b038c0d00d4a5f5cac62286a207926..eecd64097f91196897f45530540b9c9b68c5ba4e 100644
> > > --- a/include/net/udp.h
> > > +++ b/include/net/udp.h
> > > @@ -289,6 +289,7 @@ static inline void udp_lib_init_sock(struct sock *sk)
> > >       struct udp_sock *up = udp_sk(sk);
> > >
> > >       sk->sk_drop_counters = &up->drop_counters;
> > > +     spin_lock_init(&up->busylock);
> > >       skb_queue_head_init(&up->reader_queue);
> > >       INIT_HLIST_NODE(&up->tunnel_list);
> > >       up->forward_threshold = sk->sk_rcvbuf >> 2;
> > > diff --git a/net/ipv4/udp.c b/net/ipv4/udp.c
> > > index 25143f932447df2a84dd113ca33e1ccf15b3503c..7d1444821ee51a19cd5fd0dd5b8d096104c9283c 100644
> > > --- a/net/ipv4/udp.c
> > > +++ b/net/ipv4/udp.c
> > > @@ -1689,17 +1689,11 @@ static void udp_skb_dtor_locked(struct sock *sk, struct sk_buff *skb)
> > >   * to relieve pressure on the receive_queue spinlock shared by consumer.
> > >   * Under flood, this means that only one producer can be in line
> > >   * trying to acquire the receive_queue spinlock.
> > > - * These busylock can be allocated on a per cpu manner, instead of a
> > > - * per socket one (that would consume a cache line per socket)
> > >   */
> > > -static int udp_busylocks_log __read_mostly;
> > > -static spinlock_t *udp_busylocks __read_mostly;
> > > -
> > > -static spinlock_t *busylock_acquire(void *ptr)
> > > +static spinlock_t *busylock_acquire(struct sock *sk)
> > >  {
> > > -     spinlock_t *busy;
> > > +     spinlock_t *busy = &udp_sk(sk)->busylock;
> > >
> > > -     busy = udp_busylocks + hash_ptr(ptr, udp_busylocks_log);
> > >       spin_lock(busy);
> > >       return busy;
> > >  }
> > > @@ -3997,7 +3991,6 @@ static void __init bpf_iter_register(void)
> > >  void __init udp_init(void)
> > >  {
> > >       unsigned long limit;
> > > -     unsigned int i;
> > >
> > >       udp_table_init(&udp_table, "UDP");
> > >       limit = nr_free_buffer_pages() / 8;
> > > @@ -4006,15 +3999,6 @@ void __init udp_init(void)
> > >       sysctl_udp_mem[1] = limit;
> > >       sysctl_udp_mem[2] = sysctl_udp_mem[0] * 2;
> > >
> > > -     /* 16 spinlocks per cpu */
> > > -     udp_busylocks_log = ilog2(nr_cpu_ids) + 4;
> > > -     udp_busylocks = kmalloc(sizeof(spinlock_t) << udp_busylocks_log,
> > > -                             GFP_KERNEL);
> >
> > A per sock busylock is preferable over increasing this array to be
> > full percpu (and converting percpu to avoid false sharing)?
> >
> > Because that would take a lot of space on modern server platforms?
> > Just trying to understand the trade-off made.
> 
> The goal of the busylock is to have a single gate before sk->sk_receive_queue.
> 
> Having per-cpu spinlocks will not fit the need ?

Oh of course. For high rate UDP servers I was mistakenly immediately
thinking of SO_REUSEPORT and CPU pinning.

And thus different sockets that may accidentally share a hashed lock
or hit cacheline false sharing.

But this is a single socket being hit with traffic from multiple
producers.
 
> Note that having per-NUMA receive queues is on my plate, but not finished yet.

Interesting!

> I tried to remove the busylock (because modern UDP has a second queue
> (up->reader_queue),
> so __skb_recv_udp() splices things in batches), but busylock was still
> beneficial.

Good to know, thanks.