Back in 2024 I reported a 7-12% regression on an iperf3 UDP loopback
thoughput test that we traced to the extra overhead of calling
compute_score on two places, introduced by commit f0ea27e7bf ("udp:
re-score reuseport groups when connected sockets are present"). At the
time, I pointed out the overhead was caused by the multiple calls,
associated with cpu-specific mitigations, and merged commit
50aee97d15 ("udp: Avoid call to compute_score on multiple sites") to
jump back explicitly, to force the rescore call in a single place.
Recently though, we got another regression report against a newer distro
version, which a team colleague traced back to the same root-cause.
Turns out that once we updated to gcc-13, the compiler got smart enough
to unroll the loop, undoing my previous mitigation. Let's bite the
bullet and __always_inline compute_score on both ipv4 and ipv6 to
prevent gcc from de-optimizing it again in the future. These functions
are only called in two places each, udpX_lib_lookup1 and
udpX_lib_lookup2, so the extra size shouldn't be a problem and it is hot
enough to be very visible in profilings. In fact, with gcc13, forcing
the inline will prevent gcc from unrolling the fix from commit
50aee97d15, so we don't end up increasing udpX_lib_lookup2 at all.
I haven't recollected the results myself, as I don't have access to the
machine at the moment. But the same colleague reported 4.67%
inprovement with this patch in the loopback benchmark, solving the
regression report within noise margins.
Eric Dumazet reported no size change to vmlinux when built with clang.
I report the same also with gcc-13:
scripts/bloat-o-meter vmlinux vmlinux-inline
add/remove: 0/2 grow/shrink: 4/0 up/down: 616/-416 (200)
Function old new delta
udp6_lib_lookup2 762 949 +187
__udp6_lib_lookup 810 975 +165
udp4_lib_lookup2 757 906 +149
__udp4_lib_lookup 871 986 +115
__pfx_compute_score 32 - -32
compute_score 384 - -384
Total: Before=35011784, After=35011984, chg +0.00%
Fixes: 50aee97d15 ("udp: Avoid call to compute_score on multiple sites")
Reviewed-by: Eric Dumazet <edumazet@google.com>
Acked-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Gabriel Krisman Bertazi <krisman@suse.de>
Link: https://patch.msgid.link/20260410155936.654915-1-krisman@suse.de
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Since UDP and UDP-Lite had dedicated socket hash tables for
each, we have had to pass the pointer down to many socket
lookup functions.
UDP-Lite gone, and we do not need to do that.
Let's fetch net->ipv4.udp_table only where needed in IPv6
stack: __udp6_lib_lookup() and __udp6_lib_mcast_deliver().
__udp6_lib_err() is renamed to udpv6_err() as its wrapper
is no longer needed.
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20260311052020.1213705-14-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
BPF socket lookup for SO_REUSEPORT does not support UDP-Lite.
In __udp4_lib_lookup() and __udp6_lib_lookup(), it checks if
the passed udptable pointer is the same as net->ipv4.udp_table,
which is only true for UDP.
Now, the condition is always true.
Let's remove the check.
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20260311052020.1213705-13-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Since UDP and UDP-Lite had dedicated socket hash tables for
each, we have had to fetch them from different pointers for
procfs or bpf iterator.
UDP always has its global or per-netns table in
net->ipv4.udp_table and struct udp_seq_afinfo.udp_table
is NULL.
OTOH, UDP-Lite had only one global table in the pointer.
We no longer use the field.
Let's remove it and udp_get_table_seq().
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20260311052020.1213705-12-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Since UDP and UDP-Lite had dedicated socket hash tables for
each, we have had to fetch them from different pointers.
UDP always has its global or per-netns table in
net->ipv4.udp_table and struct proto.h.udp_table is NULL.
OTOH, UDP-Lite had only one global table in the pointer.
We no longer use the field.
Let's remove it and udp_get_table_prot().
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20260311052020.1213705-11-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
UDP-Lite supports variable-length checksum and has two socket
options, UDPLITE_SEND_CSCOV and UDPLITE_RECV_CSCOV, to control
the checksum coverage.
Let's remove the support.
setsockopt(UDPLITE_SEND_CSCOV / UDPLITE_RECV_CSCOV) was only
available for UDP-Lite and returned -ENOPROTOOPT for UDP.
Now, the options are handled in ip_setsockopt() and
ipv6_setsockopt(), which still return the same error.
getsockopt(UDPLITE_SEND_CSCOV / UDPLITE_RECV_CSCOV) was available
for UDP and always returned 0, meaning full checksum, but now
-ENOPROTOOPT is returned.
Given that getsockopt() is meaningless for UDP and even the options
are not defined under include/uapi/, this should not be a problem.
$ man 7 udplite
...
BUGS
Where glibc support is missing, the following definitions
are needed:
#define IPPROTO_UDPLITE 136
#define UDPLITE_SEND_CSCOV 10
#define UDPLITE_RECV_CSCOV 11
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20260311052020.1213705-10-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
UDP TX paths also have some code for UDP-Lite partial
checksum:
* udplite_csum() in udp_send_skb() and udp_v6_send_skb()
* udplite_getfrag() in udp_sendmsg() and udpv6_sendmsg()
Let's remove such code.
Now, we can use IPPROTO_UDP directly instead of sk->sk_protocol
or fl6->flowi6_proto for csum_tcpudp_magic() and csum_ipv6_magic().
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20260311052020.1213705-9-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
UDP-Lite supports the partial checksum and the coverage is
stored in the position of the length field of struct udphdr.
In RX paths, udp4_csum_init() / udp6_csum_init() save the value
in UDP_SKB_CB(skb)->cscov and set UDP_SKB_CB(skb)->partial_cov
to 1 if the coverage is not full.
The subsequent processing diverges depending on the value,
but such paths are now dead.
Also, these functions have some code guarded for UDP:
* udp_unicast_rcv_skb / udp6_unicast_rcv_skb
* __udp4_lib_rcv() and __udp6_lib_rcv().
Let's remove the partial csum code and the unnecessary
guard for UDP-Lite in RX.
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20260311052020.1213705-8-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Since UDP and UDP-Lite shared most of the code, we have had
to check the protocol every time we increment SNMP stats.
Now that the UDP-Lite paths are dead, let's remove UDP-Lite
SNMP stats.
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20260311052020.1213705-6-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
As announced in commit be28c14ac8 ("udplite: Print deprecation
notice."), it's time to deprecate UDP-Lite.
As a first step, let's drop support for IPv6 UDP-Lite sockets.
We will remove the remaining dead code gradually.
Along with the removal of udplite.c, most of the functions exposed
via udp_impl.h are made static.
The prototypes of udpv6_sendmsg() and udpv6_recvmsg() are moved
to udp.h, but only udpv6_recvmsg() has INDIRECT_CALLABLE_DECLARE()
because udpv6_sendmsg() is exported for rxrpc since commit ed472b0c87
("rxrpc: Call udp_sendmsg() directly").
Also, udpv6_recvmsg() needs INDIRECT_CALLABLE_SCOPE for
CONFIG_MITIGATION_RETPOLINE=n.
Note that udplite.h is included temporarily for udplite_csum().
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20260311052020.1213705-3-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Since commit a3d2599b24 ("ipv{4,6}/udp{,lite}: simplify proc
registration"), udp4_seq_show() and udp6_seq_show() are not
used in net/ipv4/udplite.c and net/ipv6/udplite.c.
Instead, udp_seq_ops and udp6_seq_ops are exposed to UDP-Lite.
Let's make udp4_seq_show() and udp6_seq_show() static.
udp_seq_ops and udp6_seq_ops are moved to udp_impl.h so that
we can make them static when the header is removed.
Signed-off-by: Kuniyuki Iwashima <kuniyu@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20260311052020.1213705-2-kuniyu@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This function has a single caller in net/ipv6/udp.c.
Move it there so that the compiler can decide to (auto)inline
it if he prefers to. IBT glue is removed anyway.
With clang, we can see it was able to inline it and also
inlined one other helper at the same time.
UDPLITE removal will also help.
$ scripts/bloat-o-meter -t vmlinux.old vmlinux.new
add/remove: 0/2 grow/shrink: 1/0 up/down: 840/-785 (55)
Function old new delta
__udp6_lib_rcv 1247 2087 +840
__pfx_udp6_csum_init 16 - -16
udp6_csum_init 769 - -769
Total: Before=25074399, After=25074454, chg +0.00%
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20260223093445.3696368-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Convert struct proto pre_connect(), connect(), bind(), and bind_add()
callback function prototypes from struct sockaddr to struct sockaddr_unsized.
This does not change per-implementation use of sockaddr for passing around
an arbitrarily sized sockaddr struct. Those will be addressed in future
patches.
Additionally removes the no longer referenced struct sockaddr from
include/net/inet_common.h.
No binary changes expected.
Signed-off-by: Kees Cook <kees@kernel.org>
Link: https://patch.msgid.link/20251104002617.2752303-5-kees@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
busylock was protecting UDP sockets against packet floods,
but unfortunately was not protecting the host itself.
Under stress, many cpus could spin while acquiring the busylock,
and NIC had to drop packets. Or packets would be dropped
in cpu backlog if RPS/RFS were in place.
This patch replaces the busylock by intermediate
lockless queues. (One queue per NUMA node).
This means that fewer number of cpus have to acquire
the UDP receive queue lock.
Most of the cpus can either:
- immediately drop the packet.
- or queue it in their NUMA aware lockless queue.
Then one of the cpu is chosen to process this lockless queue
in a batch.
The batch only contains packets that were cooked on the same
NUMA node, thus with very limited latency impact.
Tested:
DDOS targeting a victim UDP socket, on a platform with 6 NUMA nodes
(Intel(R) Xeon(R) 6985P-C)
Before:
nstat -n ; sleep 1 ; nstat | grep Udp
Udp6InDatagrams 1004179 0.0
Udp6InErrors 3117 0.0
Udp6RcvbufErrors 3117 0.0
After:
nstat -n ; sleep 1 ; nstat | grep Udp
Udp6InDatagrams 1116633 0.0
Udp6InErrors 14197275 0.0
Udp6RcvbufErrors 14197275 0.0
We can see this host can now proces 14.2 M more packets per second
while under attack, and the victim socket can receive 11 % more
packets.
I used a small bpftrace program measuring time (in us) spent in
__udp_enqueue_schedule_skb().
Before:
@udp_enqueue_us[398]:
[0] 24901 |@@@ |
[1] 63512 |@@@@@@@@@ |
[2, 4) 344827 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[4, 8) 244673 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[8, 16) 54022 |@@@@@@@@ |
[16, 32) 222134 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[32, 64) 232042 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@ |
[64, 128) 4219 | |
[128, 256) 188 | |
After:
@udp_enqueue_us[398]:
[0] 5608855 |@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@@|
[1] 1111277 |@@@@@@@@@@ |
[2, 4) 501439 |@@@@ |
[4, 8) 102921 | |
[8, 16) 29895 | |
[16, 32) 43500 | |
[32, 64) 31552 | |
[64, 128) 979 | |
[128, 256) 13 | |
Note that the remaining bottleneck for this platform is in
udp_drops_inc() because we limited struct numa_drop_counters
to only two nodes so far.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20250922104240.2182559-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Generic sk_drops_inc() reads sk->sk_drop_counters.
We know the precise location for UDP sockets.
Move sk_drop_counters out of sock_read_rxtx
so that sock_write_rxtx starts at a cache line boundary.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20250916160951.541279-9-edumazet@google.com
Reviewed-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Add READ_ONCE() annotations because np->rxpmtu can be changed
while udpv6_recvmsg() and rawv6_recvmsg() read it.
Since this is a very rarely used feature, and that udpv6_recvmsg()
and rawv6_recvmsg() read np->rxopt anyway, change the test order
so that np->rxpmtu does not need to be in a hot cache line.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20250916160951.541279-4-edumazet@google.com
Reviewed-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
We want to split sk->sk_drops in the future to reduce
potential contention on this field.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20250826125031.1578842-2-edumazet@google.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Add a new SKB drop reason (SKB_DROP_REASON_PFMEMALLOC) to track packets
dropped due to memory pressure. In production environments, we've observed
memory exhaustion reported by memory layer stack traces, but these drops
were not properly tracked in the SKB drop reason infrastructure.
While most network code paths now properly report pfmemalloc drops, some
protocol-specific socket implementations still use sk_filter() without
drop reason tracking:
- Bluetooth L2CAP sockets
- CAIF sockets
- IUCV sockets
- Netlink sockets
- SCTP sockets
- Unix domain sockets
These remaining cases represent less common paths and could be converted
in a follow-up patch if needed. The current implementation provides
significantly improved observability into memory pressure events in the
network stack, especially for key protocols like TCP and UDP, helping to
diagnose problems in production environments.
Reported-by: Matt Fleming <mfleming@cloudflare.com>
Signed-off-by: Jesper Dangaard Brouer <hawk@kernel.org>
Link: https://patch.msgid.link/175268316579.2407873.11634752355644843509.stgit@firesoul
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
____cacheline_aligned_in_smp attribute only makes sure to align
a field to a cache line. It does not prevent the linker to use
the remaining of the cache line for other variables, causing
potential false sharing.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20250630093540.3052835-5-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
sk->sk_uid can be read while another thread changes its
value in sockfs_setattr().
Add sk_uid(const struct sock *sk) helper to factorize the needed
READ_ONCE() annotations, and add corresponding WRITE_ONCE()
where needed.
Fixes: 86741ec254 ("net: core: Add a UID field to struct sock.")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Lorenzo Colitti <lorenzo@google.com>
Reviewed-by: Maciej Żenczykowski <maze@google.com>
Link: https://patch.msgid.link/20250620133001.4090592-2-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Most UDP tunnels bind a socket to a local port, with ANY address, no
peer and no interface index specified.
Additionally it's quite common to have a single tunnel device per
namespace.
Track in each namespace the UDP tunnel socket respecting the above.
When only a single one is present, store a reference in the netns.
When such reference is not NULL, UDP tunnel GRO lookup just need to
match the incoming packet destination port vs the socket local port.
The tunnel socket never sets the reuse[port] flag[s]. When bound to no
address and interface, no other socket can exist in the same netns
matching the specified local port.
Matching packets with non-local destination addresses will be
aggregated, and eventually segmented as needed - no behavior changes
intended.
Restrict the optimization to kernel sockets only: it covers all the
relevant use-cases, and user-space owned sockets could be disconnected
and rebound after setup_udp_tunnel_sock(), breaking the uniqueness
assumption
Note that the UDP tunnel socket reference is stored into struct
netns_ipv4 for both IPv4 and IPv6 tunnels. That is intentional to keep
all the fastpath-related netns fields in the same struct and allow
cacheline-based optimization. Currently both the IPv4 and IPv6 socket
pointer share the same cacheline as the `udp_table` field.
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/41d16bc8d1257d567f9344c445b4ae0b4a91ede4.1744040675.git.pabeni@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Revert "udp_tunnel: use static call for GRO hooks when possible"
This reverts commit 311b36574c.
Revert "udp_tunnel: create a fastpath GRO lookup."
This reverts commit 8d4880db37.
There are multiple small issues with the series. In the interest
of unblocking the merge window let's opt for a revert.
Link: https://lore.kernel.org/cover.1742557254.git.pabeni@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Most UDP tunnels bind a socket to a local port, with ANY address, no
peer and no interface index specified.
Additionally it's quite common to have a single tunnel device per
namespace.
Track in each namespace the UDP tunnel socket respecting the above.
When only a single one is present, store a reference in the netns.
When such reference is not NULL, UDP tunnel GRO lookup just need to
match the incoming packet destination port vs the socket local port.
The tunnel socket never sets the reuse[port] flag[s]. When bound to no
address and interface, no other socket can exist in the same netns
matching the specified local port.
Matching packets with non-local destination addresses will be
aggregated, and eventually segmented as needed - no behavior changes
intended.
Note that the UDP tunnel socket reference is stored into struct
netns_ipv4 for both IPv4 and IPv6 tunnels. That is intentional to keep
all the fastpath-related netns fields in the same struct and allow
cacheline-based optimization. Currently both the IPv4 and IPv6 socket
pointer share the same cacheline as the `udp_table` field.
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/4d5c319c4471161829f50cb8436841de81a5edae.1741718157.git.pabeni@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
SKB_DROP_REASON_UDP_CSUM can be used in four locations
when dropping a packet because of a wrong UDP checksum.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20250307102002.2095238-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
This initializes tclass and dontfrag before cmsg parsing, removing the
need for explicit checks against -1 in each caller.
Leave hlimit set to -1, because its full initialization
(in ip6_sk_dst_hoplimit) requires more state (dst, flowi6, ..).
This also prepares for calling sockcm_init in a follow-on patch.
Signed-off-by: Willem de Bruijn <willemb@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Link: https://patch.msgid.link/20250214222720.3205500-7-willemdebruijn.kernel@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Commit 4094871db1 ("udp: only do GSO if # of segs > 1") avoided GSO
for small packets. But the kernel currently dismisses GSO requests only
after checking MTU/PMTU on gso_size. This means any packets, regardless
of their payload sizes, could be dropped when PMTU becomes smaller than
requested gso_size. We encountered this issue in production and it
caused a reliability problem that new QUIC connection cannot be
established before PMTU cache expired, while non GSO sockets still
worked fine at the same time.
Ideally, do not check any GSO related constraints when payload size is
smaller than requested gso_size, and return EMSGSIZE instead of EINVAL
on MTU/PMTU check failure to be more specific on the error cause.
Fixes: 4094871db1 ("udp: only do GSO if # of segs > 1")
Signed-off-by: Yan Zhai <yan@cloudflare.com>
Suggested-by: Willem de Bruijn <willemdebruijn.kernel@gmail.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If a UDP socket changes its local address while it's receiving
datagrams, as a result of connect(), there is a period during which
a lookup operation might fail to find it, after the address is changed
but before the secondary hash (port and address) and the four-tuple
hash (local and remote ports and addresses) are updated.
Secondary hash chains were introduced by commit 30fff9231f ("udp:
bind() optimisation") and, as a result, a rehash operation became
needed to make a bound socket reachable again after a connect().
This operation was introduced by commit 719f835853 ("udp: add
rehash on connect()") which isn't however a complete fix: the
socket will be found once the rehashing completes, but not while
it's pending.
This is noticeable with a socat(1) server in UDP4-LISTEN mode, and a
client sending datagrams to it. After the server receives the first
datagram (cf. _xioopen_ipdgram_listen()), it issues a connect() to
the address of the sender, in order to set up a directed flow.
Now, if the client, running on a different CPU thread, happens to
send a (subsequent) datagram while the server's socket changes its
address, but is not rehashed yet, this will result in a failed
lookup and a port unreachable error delivered to the client, as
apparent from the following reproducer:
LEN=$(($(cat /proc/sys/net/core/wmem_default) / 4))
dd if=/dev/urandom bs=1 count=${LEN} of=tmp.in
while :; do
taskset -c 1 socat UDP4-LISTEN:1337,null-eof OPEN:tmp.out,create,trunc &
sleep 0.1 || sleep 1
taskset -c 2 socat OPEN:tmp.in UDP4:localhost:1337,shut-null
wait
done
where the client will eventually get ECONNREFUSED on a write()
(typically the second or third one of a given iteration):
2024/11/13 21:28:23 socat[46901] E write(6, 0x556db2e3c000, 8192): Connection refused
This issue was first observed as a seldom failure in Podman's tests
checking UDP functionality while using pasta(1) to connect the
container's network namespace, which leads us to a reproducer with
the lookup error resulting in an ICMP packet on a tap device:
LOCAL_ADDR="$(ip -j -4 addr show|jq -rM '.[] | .addr_info[0] | select(.scope == "global").local')"
while :; do
./pasta --config-net -p pasta.pcap -u 1337 socat UDP4-LISTEN:1337,null-eof OPEN:tmp.out,create,trunc &
sleep 0.2 || sleep 1
socat OPEN:tmp.in UDP4:${LOCAL_ADDR}:1337,shut-null
wait
cmp tmp.in tmp.out
done
Once this fails:
tmp.in tmp.out differ: char 8193, line 29
we can finally have a look at what's going on:
$ tshark -r pasta.pcap
1 0.000000 :: ? ff02::16 ICMPv6 110 Multicast Listener Report Message v2
2 0.168690 88.198.0.161 ? 88.198.0.164 UDP 8234 60260 ? 1337 Len=8192
3 0.168767 88.198.0.161 ? 88.198.0.164 UDP 8234 60260 ? 1337 Len=8192
4 0.168806 88.198.0.161 ? 88.198.0.164 UDP 8234 60260 ? 1337 Len=8192
5 0.168827 c6:47:05:8d:dc:04 ? Broadcast ARP 42 Who has 88.198.0.161? Tell 88.198.0.164
6 0.168851 9a:55:9a:55:9a:55 ? c6:47:05:8d:dc:04 ARP 42 88.198.0.161 is at 9a:55:9a:55:9a:55
7 0.168875 88.198.0.161 ? 88.198.0.164 UDP 8234 60260 ? 1337 Len=8192
8 0.168896 88.198.0.164 ? 88.198.0.161 ICMP 590 Destination unreachable (Port unreachable)
9 0.168926 88.198.0.161 ? 88.198.0.164 UDP 8234 60260 ? 1337 Len=8192
10 0.168959 88.198.0.161 ? 88.198.0.164 UDP 8234 60260 ? 1337 Len=8192
11 0.168989 88.198.0.161 ? 88.198.0.164 UDP 4138 60260 ? 1337 Len=4096
12 0.169010 88.198.0.161 ? 88.198.0.164 UDP 42 60260 ? 1337 Len=0
On the third datagram received, the network namespace of the container
initiates an ARP lookup to deliver the ICMP message.
In another variant of this reproducer, starting the client with:
strace -f pasta --config-net -u 1337 socat UDP4-LISTEN:1337,null-eof OPEN:tmp.out,create,trunc 2>strace.log &
and connecting to the socat server using a loopback address:
socat OPEN:tmp.in UDP4:localhost:1337,shut-null
we can more clearly observe a sendmmsg() call failing after the
first datagram is delivered:
[pid 278012] connect(173, 0x7fff96c95fc0, 16) = 0
[...]
[pid 278012] recvmmsg(173, 0x7fff96c96020, 1024, MSG_DONTWAIT, NULL) = -1 EAGAIN (Resource temporarily unavailable)
[pid 278012] sendmmsg(173, 0x561c5ad0a720, 1, MSG_NOSIGNAL) = 1
[...]
[pid 278012] sendmmsg(173, 0x561c5ad0a720, 1, MSG_NOSIGNAL) = -1 ECONNREFUSED (Connection refused)
and, somewhat confusingly, after a connect() on the same socket
succeeded.
Until commit 4cdeeee925 ("net: udp: prefer listeners bound to an
address"), the race between receive address change and lookup didn't
actually cause visible issues, because, once the lookup based on the
secondary hash chain failed, we would still attempt a lookup based on
the primary hash (destination port only), and find the socket with the
outdated secondary hash.
That change, however, dropped port-only lookups altogether, as side
effect, making the race visible.
To fix this, while avoiding the need to make address changes and
rehash atomic against lookups, reintroduce primary hash lookups as
fallback, if lookups based on four-tuple and secondary hashes fail.
To this end, introduce a simplified lookup implementation, which
doesn't take care of SO_REUSEPORT groups: if we have one, there are
multiple sockets that would match the four-tuple or secondary hash,
meaning that we can't run into this race at all.
v2:
- instead of synchronising lookup operations against address change
plus rehash, reintroduce a simplified version of the original
primary hash lookup as fallback
v1:
- fix build with CONFIG_IPV6=n: add ifdef around sk_v6_rcv_saddr
usage (Kuniyuki Iwashima)
- directly use sk_rcv_saddr for IPv4 receive addresses instead of
fetching inet_rcv_saddr (Kuniyuki Iwashima)
- move inet_update_saddr() to inet_hashtables.h and use that
to set IPv4/IPv6 addresses as suitable (Kuniyuki Iwashima)
- rebase onto net-next, update commit message accordingly
Reported-by: Ed Santiago <santiago@redhat.com>
Link: https://github.com/containers/podman/issues/24147
Analysed-by: David Gibson <david@gibson.dropbear.id.au>
Fixes: 30fff9231f ("udp: bind() optimisation")
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
The Linux socket API currently allows setting SO_PRIORITY at the
socket level, applying a uniform priority to all packets sent through
that socket. The exception to this is IP_TOS, when the priority value
is calculated during the handling of
ancillary data, as implemented in commit f02db315b8 ("ipv4: IP_TOS
and IP_TTL can be specified as ancillary data").
However, this is a computed
value, and there is currently no mechanism to set a custom priority
via control messages prior to this patch.
According to this patch, if SO_PRIORITY is specified as ancillary data,
the packet is sent with the priority value set through
sockc->priority, overriding the socket-level values
set via the traditional setsockopt() method. This is analogous to
the existing support for SO_MARK, as implemented in
commit c6af0c227a ("ip: support SO_MARK cmsg").
If both cmsg SO_PRIORITY and IP_TOS are passed, then the one that
takes precedence is the last one in the cmsg list.
This patch has the side effect that raw_send_hdrinc now interprets cmsg
IP_TOS.
Reviewed-by: Willem de Bruijn <willemb@google.com>
Suggested-by: Ferenc Fejes <fejes@inf.elte.hu>
Signed-off-by: Anna Emese Nyiri <annaemesenyiri@gmail.com>
Link: https://patch.msgid.link/20241213084457.45120-3-annaemesenyiri@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Implement ipv6 udp hash4 like that in ipv4. The major difference is that
the hash value should be calculated with udp6_ehashfn(). Besides,
ipv4-mapped ipv6 address is handled before hash() and rehash(). Export
udp_ehashfn because now we use it in udpv6 rehash.
Core procedures of hash/unhash/rehash are same as ipv4, and udpv4 and
udpv6 share the same udptable, so some functions in ipv4 hash4 can also
be shared.
Co-developed-by: Cambda Zhu <cambda@linux.alibaba.com>
Signed-off-by: Cambda Zhu <cambda@linux.alibaba.com>
Co-developed-by: Fred Chen <fred.cc@alibaba-inc.com>
Signed-off-by: Fred Chen <fred.cc@alibaba-inc.com>
Co-developed-by: Yubing Qiu <yubing.qiuyubing@alibaba-inc.com>
Signed-off-by: Yubing Qiu <yubing.qiuyubing@alibaba-inc.com>
Signed-off-by: Philo Lu <lulie@linux.alibaba.com>
Acked-by: Willem de Bruijn <willemb@google.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Currently, the udp_table has two hash table, the port hash and portaddr
hash. Usually for UDP servers, all sockets have the same local port and
addr, so they are all on the same hash slot within a reuseport group.
In some applications, UDP servers use connect() to manage clients. In
particular, when firstly receiving from an unseen 4 tuple, a new socket
is created and connect()ed to the remote addr:port, and then the fd is
used exclusively by the client.
Once there are connected sks in a reuseport group, udp has to score all
sks in the same hash2 slot to find the best match. This could be
inefficient with a large number of connections, resulting in high
softirq overhead.
To solve the problem, this patch implement 4-tuple hash for connected
udp sockets. During connect(), hash4 slot is updated, as well as a
corresponding counter, hash4_cnt, in hslot2. In __udp4_lib_lookup(),
hslot4 will be searched firstly if the counter is non-zero. Otherwise,
hslot2 is used like before. Note that only connected sockets enter this
hash4 path, while un-connected ones are not affected.
hlist_nulls is used for hash4, because we probably move to another hslot
wrongly when lookup with concurrent rehash. Then we check nulls at the
list end to see if we should restart lookup. Because udp does not use
SLAB_TYPESAFE_BY_RCU, we don't need to touch sk_refcnt when lookup.
Stress test results (with 1 cpu fully used) are shown below, in pps:
(1) _un-connected_ socket as server
[a] w/o hash4: 1,825176
[b] w/ hash4: 1,831750 (+0.36%)
(2) 500 _connected_ sockets as server
[c] w/o hash4: 290860 (only 16% of [a])
[d] w/ hash4: 1,889658 (+3.1% compared with [b])
With hash4, compute_score is skipped when lookup, so [d] is slightly
better than [b].
Co-developed-by: Cambda Zhu <cambda@linux.alibaba.com>
Signed-off-by: Cambda Zhu <cambda@linux.alibaba.com>
Co-developed-by: Fred Chen <fred.cc@alibaba-inc.com>
Signed-off-by: Fred Chen <fred.cc@alibaba-inc.com>
Co-developed-by: Yubing Qiu <yubing.qiuyubing@alibaba-inc.com>
Signed-off-by: Yubing Qiu <yubing.qiuyubing@alibaba-inc.com>
Signed-off-by: Philo Lu <lulie@linux.alibaba.com>
Acked-by: Willem de Bruijn <willemb@google.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Preparing for udp 4-tuple hash (uhash4 for short).
To implement uhash4 without cache line missing when lookup, hslot2 is
used to record the number of hashed sockets in hslot4. Thus adding a new
struct udp_hslot_main with field hash4_cnt, which is used by hash2. The
new struct is used to avoid doubling the size of udp_hslot.
Before uhash4 lookup, firstly checking hash4_cnt to see if there are
hashed sks in hslot4. Because hslot2 is always used in lookup, there is
no cache line miss.
Related helpers are updated, and use the helpers as possible.
uhash4 is implemented in following patches.
Signed-off-by: Philo Lu <lulie@linux.alibaba.com>
Acked-by: Willem de Bruijn <willemb@google.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
If:
1) the user requested USO, but
2) there is not enough payload for GSO to kick in, and
3) the egress device doesn't offer checksum offload, then
we want to compute the L4 checksum in software early on.
In the case when we are not taking the GSO path, but it has been requested,
the software checksum fallback in skb_segment doesn't get a chance to
compute the full checksum, if the egress device can't do it. As a result we
end up sending UDP datagrams with only a partial checksum filled in, which
the peer will discard.
Fixes: 10154dbded ("udp: Allow GSO transmit from devices with no checksum offload")
Reported-by: Ivan Babrou <ivan@cloudflare.com>
Signed-off-by: Jakub Sitnicki <jakub@cloudflare.com>
Acked-by: Willem de Bruijn <willemdebruijn.kernel@gmail.com>
Cc: stable@vger.kernel.org
Link: https://patch.msgid.link/20241011-uso-swcsum-fixup-v2-1-6e1ddc199af9@cloudflare.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Following helpers do not touch their 'struct net' argument.
- udp6_lib_lookup()
- __udp6_lib_lookup()
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20240802134029.3748005-6-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Remove duplicate included header file trace/events/udp.h and the
following warning reported by make includecheck:
trace/events/udp.h is included more than once
Compile-tested only.
Signed-off-by: Thorsten Blum <thorsten.blum@toblux.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://patch.msgid.link/20240706071132.274352-2-thorsten.blum@toblux.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Today sending a UDP GSO packet from a TUN device results in an EIO error:
import fcntl, os, struct
from socket import *
TUNSETIFF = 0x400454CA
IFF_TUN = 0x0001
IFF_NO_PI = 0x1000
UDP_SEGMENT = 103
tun_fd = os.open("/dev/net/tun", os.O_RDWR)
ifr = struct.pack("16sH", b"tun0", IFF_TUN | IFF_NO_PI)
fcntl.ioctl(tun_fd, TUNSETIFF, ifr)
os.system("ip addr add 192.0.2.1/24 dev tun0")
os.system("ip link set dev tun0 up")
s = socket(AF_INET, SOCK_DGRAM)
s.setsockopt(SOL_UDP, UDP_SEGMENT, 1200)
s.sendto(b"x" * 3000, ("192.0.2.2", 9)) # EIO
This is due to a check in the udp stack if the egress device offers
checksum offload. While TUN/TAP devices, by default, don't advertise this
capability because it requires support from the TUN/TAP reader.
However, the GSO stack has a software fallback for checksum calculation,
which we can use. This way we don't force UDP_SEGMENT users to handle the
EIO error and implement a segmentation fallback.
Lift the restriction so that UDP_SEGMENT can be used with any egress
device. We also need to adjust the UDP GSO code to match the GSO stack
expectation about ip_summed field, as set in commit 8d63bee643 ("net:
avoid skb_warn_bad_offload false positives on UFO"). Otherwise we will hit
the bad offload check.
Users should, however, expect a potential performance impact when
batch-sending packets with UDP_SEGMENT without checksum offload on the
egress device. In such case the packet payload is read twice: first during
the sendmsg syscall when copying data from user memory, and then in the GSO
stack for checksum computation. This double memory read can be less
efficient than a regular sendmsg where the checksum is calculated during
the initial data copy from user memory.
Signed-off-by: Jakub Sitnicki <jakub@cloudflare.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20240626-linux-udpgso-v2-1-422dfcbd6b48@cloudflare.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Replace kfree_skb_reason with sk_skb_reason_drop and pass the receiving
socket to the tracepoint.
Reported-by: kernel test robot <lkp@intel.com>
Closes: https://lore.kernel.org/r/202406011751.NpVN0sSk-lkp@intel.com/
Signed-off-by: Yan Zhai <yan@cloudflare.com>
Acked-by: Jesper Dangaard Brouer <hawk@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
Cross-merge networking fixes after downstream PR.
Conflicts:
include/linux/filter.h
kernel/bpf/core.c
66e13b615a ("bpf: verifier: prevent userspace memory access")
d503a04f8b ("bpf: Add support for certain atomics in bpf_arena to x86 JIT")
https://lore.kernel.org/all/20240429114939.210328b0@canb.auug.org.au/
No adjacent changes.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Commits a602456 ("udp: Add GRO functions to UDP socket") and 57c67ff ("udp:
additional GRO support") introduce incorrect usage of {ip,ipv6}_hdr in the
complete phase of gro. The functions always return skb->network_header,
which in the case of encapsulated packets at the gro complete phase, is
always set to the innermost L3 of the packet. That means that calling
{ip,ipv6}_hdr for skbs which completed the GRO receive phase (both in
gro_list and *_gro_complete) when parsing an encapsulated packet's _outer_
L3/L4 may return an unexpected value.
This incorrect usage leads to a bug in GRO's UDP socket lookup.
udp{4,6}_lib_lookup_skb functions use ip_hdr/ipv6_hdr respectively. These
*_hdr functions return network_header which will point to the innermost L3,
resulting in the wrong offset being used in __udp{4,6}_lib_lookup with
encapsulated packets.
This patch adds network_offset and inner_network_offset to napi_gro_cb, and
makes sure both are set correctly.
To fix the issue, network_offsets union is used inside napi_gro_cb, in
which both the outer and the inner network offsets are saved.
Reproduction example:
Endpoint configuration example (fou + local address bind)
# ip fou add port 6666 ipproto 4
# ip link add name tun1 type ipip remote 2.2.2.1 local 2.2.2.2 encap fou encap-dport 5555 encap-sport 6666 mode ipip
# ip link set tun1 up
# ip a add 1.1.1.2/24 dev tun1
Netperf TCP_STREAM result on net-next before patch is applied:
net-next main, GRO enabled:
$ netperf -H 1.1.1.2 -t TCP_STREAM -l 5
Recv Send Send
Socket Socket Message Elapsed
Size Size Size Time Throughput
bytes bytes bytes secs. 10^6bits/sec
131072 16384 16384 5.28 2.37
net-next main, GRO disabled:
$ netperf -H 1.1.1.2 -t TCP_STREAM -l 5
Recv Send Send
Socket Socket Message Elapsed
Size Size Size Time Throughput
bytes bytes bytes secs. 10^6bits/sec
131072 16384 16384 5.01 2745.06
patch applied, GRO enabled:
$ netperf -H 1.1.1.2 -t TCP_STREAM -l 5
Recv Send Send
Socket Socket Message Elapsed
Size Size Size Time Throughput
bytes bytes bytes secs. 10^6bits/sec
131072 16384 16384 5.01 2877.38
Fixes: a6024562ff ("udp: Add GRO functions to UDP socket")
Signed-off-by: Richard Gobert <richardbgobert@gmail.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Instead of (struct rt6_info *)dst casts, we can use :
#define dst_rt6_info(_ptr) \
container_of_const(_ptr, struct rt6_info, dst)
Some places needed missing const qualifiers :
ip6_confirm_neigh(), ipv6_anycast_destination(),
ipv6_unicast_destination(), has_gateway()
v2: added missing parts (David Ahern)
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
If "udp_cmsg_send()" returned 0 (i.e. only UDP cmsg),
"connected" should not be set to 0. Otherwise it stops
the connected socket from using the cached route.
Fixes: 2e8de85763 ("udp: add gso segment cmsg")
Signed-off-by: Yick Xie <yick.xie@gmail.com>
Cc: stable@vger.kernel.org
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://lore.kernel.org/r/20240418170610.867084-1-yick.xie@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
We've observed a 7-12% performance regression in iperf3 UDP ipv4 and
ipv6 tests with multiple sockets on Zen3 cpus, which we traced back to
commit f0ea27e7bf ("udp: re-score reuseport groups when connected
sockets are present"). The failing tests were those that would spawn
UDP sockets per-cpu on systems that have a high number of cpus.
Unsurprisingly, it is not caused by the extra re-scoring of the reused
socket, but due to the compiler no longer inlining compute_score, once
it has the extra call site in udp4_lib_lookup2. This is augmented by
the "Safe RET" mitigation for SRSO, needed in our Zen3 cpus.
We could just explicitly inline it, but compute_score() is quite a large
function, around 300b. Inlining in two sites would almost double
udp4_lib_lookup2, which is a silly thing to do just to workaround a
mitigation. Instead, this patch shuffles the code a bit to avoid the
multiple calls to compute_score. Since it is a static function used in
one spot, the compiler can safely fold it in, as it did before, without
increasing the text size.
With this patch applied I ran my original iperf3 testcases. The failing
cases all looked like this (ipv4):
iperf3 -c 127.0.0.1 --udp -4 -f K -b $R -l 8920 -t 30 -i 5 -P 64 -O 2
where $R is either 1G/10G/0 (max, unlimited). I ran 3 times each.
baseline is v6.9-rc3. harmean == harmonic mean; CV == coefficient of
variation.
ipv4:
1G 10G MAX
HARMEAN (CV) HARMEAN (CV) HARMEAN (CV)
baseline 1743852.66(0.0208) 1725933.02(0.0167) 1705203.78(0.0386)
patched 1968727.61(0.0035) 1962283.22(0.0195) 1923853.50(0.0256)
ipv6:
1G 10G MAX
HARMEAN (CV) HARMEAN (CV) HARMEAN (CV)
baseline 1729020.03(0.0028) 1691704.49(0.0243) 1692251.34(0.0083)
patched 1900422.19(0.0067) 1900968.01(0.0067) 1568532.72(0.1519)
This restores the performance we had before the change above with this
benchmark. We obviously don't expect any real impact when mitigations
are disabled, but just to be sure it also doesn't regresses:
mitigations=off ipv4:
1G 10G MAX
HARMEAN (CV) HARMEAN (CV) HARMEAN (CV)
baseline 3230279.97(0.0066) 3229320.91(0.0060) 2605693.19(0.0697)
patched 3242802.36(0.0073) 3239310.71(0.0035) 2502427.19(0.0882)
Cc: Lorenz Bauer <lmb@isovalent.com>
Fixes: f0ea27e7bf ("udp: re-score reuseport groups when connected sockets are present")
Signed-off-by: Gabriel Krisman Bertazi <krisman@suse.de>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
