Re: [PATCH net-next v8 08/15] quic: add path management

On Thu, Jan 29, 2026 at 11:20 AM Paolo Abeni <[email protected]> wrote:
>
> On 1/26/26 3:51 PM, Xin Long wrote:
> > This patch introduces 'quic_path_group' for managing paths, represented
> > by 'struct quic_path'. A connection may use two paths simultaneously
> > for connection migration.
> >
> > Each path is associated with a UDP tunnel socket (sk), and a single
> > UDP tunnel socket can be related to multiple paths from different sockets.
> > These UDP tunnel sockets are wrapped in 'quic_udp_sock' structures and
> > stored in a hash table.
> >
> > It includes mechanisms to bind and unbind paths, detect alternative paths
> > for migration, and swap paths to support seamless transition between
> > networks.
> >
> > - quic_path_bind(): Bind a path to a port and associate it with a UDP sk.
> >
> > - quic_path_unbind(): Unbind a path from a port and disassociate it from a
> >   UDP sk.
> >
> > - quic_path_swap(): Swap two paths to facilitate connection migration.
> >
> > - quic_path_detect_alt(): Determine if a packet is using an alternative
> >   path, used for connection migration.
> >
> >  It also integrates basic support for Packetization Layer Path MTU
> > Discovery (PLPMTUD), using PING frames and ICMP feedback to adjust path
> > MTU and handle probe confirmation or resets during routing changes.
> >
> > - quic_path_pl_recv(): state transition and pmtu update after the probe
> >   packet is acked.
> >
> > - quic_path_pl_toobig(): state transition and pmtu update after
> >   receiving a toobig or needfrag icmp packet.
> >
> > - quic_path_pl_send(): state transition and pmtu update after sending a
> >   probe packet.
> >
> > - quic_path_pl_reset(): restart the probing when path routing changes.
> >
> > - quic_path_pl_confirm(): check if probe packet gets acked.
> >
> > Signed-off-by: Tyler Fanelli <[email protected]>
> > Signed-off-by: Xin Long <[email protected]>
> > ---
> > v3:
> >   - Fix annotation in quic_udp_sock_lookup() (noted by Paolo).
> >   - Use inet_sk_get_local_port_range() instead of
> >     inet_get_local_port_range() (suggested by Paolo).
> >   - Adjust global UDP tunnel socket hashtable operations for the new
> >     hashtable type.
> >   - Delete quic_workqueue; use system_wq for UDP tunnel socket destroy.
> > v4:
> >   - Cache UDP tunnel socket pointer and its source address in struct
> >     quic_path for RCU-protected lookup/access.
> >   - Return -EAGAIN instead of -EINVAL in quic_path_bind() when UDP
> >     socket is being released in workqueue.
> >   - Move udp_tunnel_sock_release() out of the mutex_lock to avoid a
> >     warning of lockdep in quic_udp_sock_put_work().
> >   - Introduce quic_wq for UDP socket release work, so all pending works
> >     can be flushed before destroying the hashtable in quic_exit().
> > v5:
> >   - Rename quic_path_free() to quic_path_unbind() (suggested by Paolo).
> >   - Remove the 'serv' member from struct quic_path_group, since
> >     quic_is_serv() defined in a previous patch now uses
> >     sk->sk_max_ack_backlog for server-side detection.
> >   - Use quic_ktime_get_us() to set skb_cb->time, as RTT is measured
> >     in microseconds and jiffies_to_usecs() is not accurate enough.
> > v6:
> >   - Do not reset transport_header for QUIC in quic_udp_rcv(), allowing
> >     removal of udph_offset and enabling access to the UDP header via
> >     udp_hdr(); Pull skb->data in quic_udp_rcv() to allow access to the
> >     QUIC header via skb->data.
> > v7:
> >   - Pass udp sk to quic_path_rcv() and move the call to skb_linearize()
> >     and skb_set_owner_sk_safe() to .quic_path_rcv().
> >   - Delete the call to skb_linearize() and skb_set_owner_sk_safe() from
> >     quic_udp_err(), as it should not change skb in .encap_err_lookup()
> >     (noted by AI review).
> > v8:
> >   - Remove indirect quic_path_rcv and late call quic_packet_rcv()
> >     directly via extern (noted by Paolo).
> >   - Add a comment in quic_udp_rcv() clarifying it must return 0.
> >   - Add a comment in quic_udp_sock_put() clarifying the UDP socket
> >     may be freed in atomic RX context during connection migration.
> >   - Reorder some quic_path_group members to reduce struct size.
> > ---
> >  net/quic/Makefile   |   2 +-
> >  net/quic/path.c     | 520 ++++++++++++++++++++++++++++++++++++++++++++
> >  net/quic/path.h     | 170 +++++++++++++++
> >  net/quic/protocol.c |  11 +
> >  net/quic/socket.c   |   3 +
> >  net/quic/socket.h   |   7 +
> >  6 files changed, 712 insertions(+), 1 deletion(-)
> >  create mode 100644 net/quic/path.c
> >  create mode 100644 net/quic/path.h
> >
> > diff --git a/net/quic/Makefile b/net/quic/Makefile
> > index eee7501588d3..1565fb5cef9d 100644
> > --- a/net/quic/Makefile
> > +++ b/net/quic/Makefile
> > @@ -5,4 +5,4 @@
> >
> >  obj-$(CONFIG_IP_QUIC) += quic.o
> >
> > -quic-y := common.o family.o protocol.o socket.o stream.o connid.o
> > +quic-y := common.o family.o protocol.o socket.o stream.o connid.o path.o
> > diff --git a/net/quic/path.c b/net/quic/path.c
> > new file mode 100644
> > index 000000000000..9556607a009e
> > --- /dev/null
> > +++ b/net/quic/path.c
> > @@ -0,0 +1,520 @@
> > +// SPDX-License-Identifier: GPL-2.0-or-later
> > +/* QUIC kernel implementation
> > + * (C) Copyright Red Hat Corp. 2023
> > + *
> > + * This file is part of the QUIC kernel implementation
> > + *
> > + * Initialization/cleanup for QUIC protocol support.
> > + *
> > + * Written or modified by:
> > + *    Xin Long <[email protected]>
> > + */
> > +
> > +#include <net/udp_tunnel.h>
> > +#include <linux/quic.h>
> > +
> > +#include "common.h"
> > +#include "family.h"
> > +#include "path.h"
> > +
> > +static int quic_udp_rcv(struct sock *sk, struct sk_buff *skb)
> > +{
> > +     memset(skb->cb, 0, sizeof(skb->cb));
> > +     QUIC_SKB_CB(skb)->seqno = -1;
> > +     QUIC_SKB_CB(skb)->time = quic_ktime_get_us();
> > +
> > +     skb_pull(skb, sizeof(struct udphdr));
> > +     skb_dst_force(skb);
> > +     kfree_skb(skb);
> > +     return 0; /* .encap_rcv must return 0 if skb was either consumed or dropped. */
> > +}
> > +
> > +static int quic_udp_err(struct sock *sk, struct sk_buff *skb)
> > +{
> > +     return 0;
> > +}
> > +
> > +static void quic_udp_sock_put_work(struct work_struct *work)
> > +{
> > +     struct quic_udp_sock *us = container_of(work, struct quic_udp_sock, work);
> > +     struct quic_uhash_head *head;
> > +     struct sock *sk = us->sk;
> > +
> > +     /* Hold the sock to safely access it in quic_udp_sock_lookup() even after
> > +      * udp_tunnel_sock_release(). The release must occur before __hlist_del()
> > +      * so a new UDP tunnel socket can be created for the same address and port
> > +      * if quic_udp_sock_lookup() fails to find one.
> > +      *
> > +      * Note: udp_tunnel_sock_release() cannot be called under the mutex due to
> > +      * some lockdep warnings.
> > +      */
> > +     sock_hold(sk);
> > +     udp_tunnel_sock_release(sk->sk_socket);
> > +
> > +     head = quic_udp_sock_head(sock_net(sk), ntohs(us->addr.v4.sin_port));
> > +     mutex_lock(&head->lock);
> > +     __hlist_del(&us->node);
> > +     mutex_unlock(&head->lock);
> > +
> > +     sock_put(sk);
> > +     kfree(us);
> > +}
> > +
> > +static struct quic_udp_sock *quic_udp_sock_create(struct sock *sk, union quic_addr *a)
> > +{
> > +     struct udp_tunnel_sock_cfg tuncfg = {};
> > +     struct udp_port_cfg udp_conf = {};
> > +     struct net *net = sock_net(sk);
> > +     struct quic_uhash_head *head;
> > +     struct quic_udp_sock *us;
> > +     struct socket *sock;
> > +
> > +     us = kzalloc(sizeof(*us), GFP_KERNEL);
> > +     if (!us)
> > +             return NULL;
> > +
> > +     quic_udp_conf_init(sk, &udp_conf, a);
> > +     if (udp_sock_create(net, &udp_conf, &sock)) {
> > +             pr_debug("%s: failed to create udp sock\n", __func__);
> > +             kfree(us);
> > +             return NULL;
> > +     }
> > +
> > +     tuncfg.encap_type = 1;
> > +     tuncfg.encap_rcv = quic_udp_rcv;
> > +     tuncfg.encap_err_lookup = quic_udp_err;
> > +     setup_udp_tunnel_sock(net, sock, &tuncfg);
>
> Possibly you need to adjust UDP_MAX_TUNNEL_TYPES in udp_offload.c. You
> could check running a kernel with QUIC enabled and geneve, vxlan, FOU
> and xfrm disabled.
It does not currently implement gro_receive, so there is no need to touch
UDP_MAX_TUNNEL_TYPES. Adding .gro_receive support for QUIC can be
considered as a future enhancement.

>
> > +
> > +     refcount_set(&us->refcnt, 1);
> > +     us->sk = sock->sk;
> > +     memcpy(&us->addr, a, sizeof(*a));
> > +     us->bind_ifindex = sk->sk_bound_dev_if;
> > +
> > +     head = quic_udp_sock_head(net, ntohs(a->v4.sin_port));
> > +     hlist_add_head(&us->node, &head->head);
> > +     INIT_WORK(&us->work, quic_udp_sock_put_work);
>
> Is unclear to me if quick udp socket lookup be done locklessy with
> future series?
>
No, quic udp socket lookup will not be done locklessy.
As this is not on the DATA path, I didn't intend to use RCU things on it.