| Age | Commit message (Collapse) | Author |
|---|
|
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>
|
|
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@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/20250916160951.541279-10-edumazet@google.com
Reviewed-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
|
|
Hosts under DOS attack can suffer from false sharing
in enqueue_to_backlog() : atomic_inc(&sd->dropped).
This is because sd->dropped can be touched from many cpus,
possibly residing on different NUMA nodes.
Generalize the sk_drop_counters infrastucture
added in commit c51613fa276f ("net: add sk->sk_drop_counters")
and use it to replace softnet_data.dropped
with NUMA friendly softnet_data.drop_counters.
This adds 64 bytes per cpu, maybe more in the future
if we increase the number of counters (currently 2)
per 'struct numa_drop_counters'.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20250909121942.1202585-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
When a packet flood hits one or more UDP sockets, many cpus
have to update sk->sk_drops.
This slows down other cpus, because currently
sk_drops is in sock_write_rx group.
Add a socket_drop_counters structure to udp sockets.
Using dedicated cache lines to hold drop counters
makes sure that consumers no longer suffer from
false sharing if/when producers only change sk->sk_drops.
This adds 128 bytes per UDP socket.
Tested with the following stress test, sending about 11 Mpps
to a dual socket AMD EPYC 7B13 64-Core.
super_netperf 20 -t UDP_STREAM -H DUT -l10 -- -n -P,1000 -m 120
Note: due to socket lookup, only one UDP socket is receiving
packets on DUT.
Then measure receiver (DUT) behavior. We can see both
consumer and BH handlers can process more packets per second.
Before:
nstat -n ; sleep 1 ; nstat | grep Udp
Udp6InDatagrams 615091 0.0
Udp6InErrors 3904277 0.0
Udp6RcvbufErrors 3904277 0.0
After:
nstat -n ; sleep 1 ; nstat | grep Udp
Udp6InDatagrams 816281 0.0
Udp6InErrors 7497093 0.0
Udp6RcvbufErrors 7497093 0.0
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20250826125031.1578842-5-edumazet@google.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
|
|
Require that iter->batch always contains a full bucket snapshot. This
invariant is important to avoid skipping or repeating sockets during
iteration when combined with the next few patches. Before, there were
two cases where a call to bpf_iter_udp_batch may only capture part of a
bucket:
1. When bpf_iter_udp_realloc_batch() returns -ENOMEM [1].
2. When more sockets are added to the bucket while calling
bpf_iter_udp_realloc_batch(), making the updated batch size
insufficient [2].
In cases where the batch size only covers part of a bucket, it is
possible to forget which sockets were already visited, especially if we
have to process a bucket in more than two batches. This forces us to
choose between repeating or skipping sockets, so don't allow this:
1. Stop iteration and propagate -ENOMEM up to userspace if reallocation
fails instead of continuing with a partial batch.
2. Try bpf_iter_udp_realloc_batch() with GFP_USER just as before, but if
we still aren't able to capture the full bucket, call
bpf_iter_udp_realloc_batch() again while holding the bucket lock to
guarantee the bucket does not change. On the second attempt use
GFP_NOWAIT since we hold onto the spin lock.
Introduce the udp_portaddr_for_each_entry_from macro and use it instead
of udp_portaddr_for_each_entry to make it possible to continue iteration
from an arbitrary socket. This is required for this patch in the
GFP_NOWAIT case to allow us to fill the rest of a batch starting from
the middle of a bucket and the later patch which skips sockets that were
already seen.
Testing all scenarios directly is a bit difficult, but I did some manual
testing to exercise the code paths where GFP_NOWAIT is used and where
ERR_PTR(err) is returned. I used the realloc test case included later
in this series to trigger a scenario where a realloc happens inside
bpf_iter_udp_batch and made a small code tweak to force the first
realloc attempt to allocate a too-small batch, thus requiring
another attempt with GFP_NOWAIT. Some printks showed both reallocs with
the tests passing:
Apr 25 23:16:24 crow kernel: go again GFP_USER
Apr 25 23:16:24 crow kernel: go again GFP_NOWAIT
With this setup, I also forced each of the bpf_iter_udp_realloc_batch
calls to return -ENOMEM to ensure that iteration ends and that the
read() in userspace fails.
[1]: https://lore.kernel.org/bpf/CABi4-ogUtMrH8-NVB6W8Xg_F_KDLq=yy-yu-tKr2udXE2Mu1Lg@mail.gmail.com/
[2]: https://lore.kernel.org/bpf/7ed28273-a716-4638-912d-f86f965e54bb@linux.dev/
Signed-off-by: Jordan Rife <jordan@jrife.io>
Signed-off-by: Martin KaFai Lau <martin.lau@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>
|
|
Add a new hash list, hash4, in udp table. It will be used to implement
4-tuple hash for connected udp sockets. This patch adds the hlist to
table, and implements helpers and the initialization. 4-tuple hash is
implemented in the following patch.
hash4 uses hlist_nulls to avoid moving wrongly onto another hlist due to
concurrent rehash, because rehash() can happen with lookup().
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>
|
|
git://git.kernel.org/pub/scm/linux/kernel/git/printk/linux
Pull printk updates from Petr Mladek:
- Use no_printk() instead of "if (0) printk()" constructs to avoid
generating printk index for messages disabled at compile time
- Remove deprecated strncpy/strcpy from printk.c
- Remove redundant CONFIG_BASE_FULL in favor of CONFIG_BASE_SMALL
* tag 'printk-for-6.10' of git://git.kernel.org/pub/scm/linux/kernel/git/printk/linux:
printk: cleanup deprecated uses of strncpy/strcpy
printk: Remove redundant CONFIG_BASE_FULL
printk: Change type of CONFIG_BASE_SMALL to bool
printk: Fix LOG_CPU_MAX_BUF_SHIFT when BASE_SMALL is enabled
ceph: Use no_printk() helper
dyndbg: Use *no_printk() helpers
dev_printk: Add and use dev_no_printk()
printk: Let no_printk() use _printk()
|
|
CONFIG_BASE_SMALL is currently a type int but is only used as a boolean.
So, change its type to bool and adapt all usages:
CONFIG_BASE_SMALL == 0 becomes !IS_ENABLED(CONFIG_BASE_SMALL) and
CONFIG_BASE_SMALL != 0 becomes IS_ENABLED(CONFIG_BASE_SMALL).
Reviewed-by: Petr Mladek <pmladek@suse.com>
Reviewed-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Reviewed-by: Masahiro Yamada <masahiroy@kernel.org>
Signed-off-by: Yoann Congal <yoann.congal@smile.fr>
Link: https://lore.kernel.org/r/20240505080343.1471198-3-yoann.congal@smile.fr
Signed-off-by: Petr Mladek <pmladek@suse.com>
|
|
The commit fc8b2a619469
("net: more strict VIRTIO_NET_HDR_GSO_UDP_L4 validation")
adds check of potential number of UDP segments vs
UDP_MAX_SEGMENTS in linux/virtio_net.h.
After this change certification test of USO guest-to-guest
transmit on Windows driver for virtio-net device fails,
for example with packet size of ~64K and mss of 536 bytes.
In general the USO should not be more restrictive than TSO.
Indeed, in case of unreasonably small mss a lot of segments
can cause queue overflow and packet loss on the destination.
Limit of 128 segments is good for any practical purpose,
with minimal meaningful mss of 536 the maximal UDP packet will
be divided to ~120 segments.
The number of segments for UDP packets is validated vs
UDP_MAX_SEGMENTS also in udp.c (v4,v6), this does not affect
quest-to-guest path but does affect packets sent to host, for
example.
It is important to mention that UDP_MAX_SEGMENTS is kernel-only
define and not available to user mode socket applications.
In order to request MSS smaller than MTU the applications
just uses setsockopt with SOL_UDP and UDP_SEGMENT and there is
no limitations on socket API level.
Fixes: fc8b2a619469 ("net: more strict VIRTIO_NET_HDR_GSO_UDP_L4 validation")
Signed-off-by: Yuri Benditovich <yuri.benditovich@daynix.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
When rx-udp-gro-forwarding is enabled UDP packets might be GROed when
being forwarded. If such packets might land in a tunnel this can cause
various issues and udp_gro_receive makes sure this isn't the case by
looking for a matching socket. This is performed in
udp4/6_gro_lookup_skb but only in the current netns. This is an issue
with tunneled packets when the endpoint is in another netns. In such
cases the packets will be GROed at the UDP level, which leads to various
issues later on. The same thing can happen with rx-gro-list.
We saw this with geneve packets being GROed at the UDP level. In such
case gso_size is set; later the packet goes through the geneve rx path,
the geneve header is pulled, the offset are adjusted and frag_list skbs
are not adjusted with regard to geneve. When those skbs hit
skb_fragment, it will misbehave. Different outcomes are possible
depending on what the GROed skbs look like; from corrupted packets to
kernel crashes.
One example is a BUG_ON[1] triggered in skb_segment while processing the
frag_list. Because gso_size is wrong (geneve header was pulled)
skb_segment thinks there is "geneve header size" of data in frag_list,
although it's in fact the next packet. The BUG_ON itself has nothing to
do with the issue. This is only one of the potential issues.
Looking up for a matching socket in udp_gro_receive is fragile: the
lookup could be extended to all netns (not speaking about performances)
but nothing prevents those packets from being modified in between and we
could still not find a matching socket. It's OK to keep the current
logic there as it should cover most cases but we also need to make sure
we handle tunnel packets being GROed too early.
This is done by extending the checks in udp_unexpected_gso: GSO packets
lacking the SKB_GSO_UDP_TUNNEL/_CSUM bits and landing in a tunnel must
be segmented.
[1] kernel BUG at net/core/skbuff.c:4408!
RIP: 0010:skb_segment+0xd2a/0xf70
__udp_gso_segment+0xaa/0x560
Fixes: 9fd1ff5d2ac7 ("udp: Support UDP fraglist GRO/GSO.")
Fixes: 36707061d6ba ("udp: allow forwarding of plain (non-fraglisted) UDP GRO packets")
Signed-off-by: Antoine Tenart <atenart@kernel.org>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
We want to re-organize the struct sock layout. The sk_peek_off
field location is problematic, as most protocols want it in the
RX read area, while UDP wants it on a cacheline different from
sk_receive_queue.
Create a local (inside udp_sock) copy of the 'peek offset is enabled'
flag and place it inside the same cacheline of reader_queue.
Check such flag before reading sk_peek_off. This will save potential
false sharing and cache misses in the fast-path.
Tested under UDP flood with small packets. The struct sock layout
update causes a 4% performance drop, and this patch restores completely
the original tput.
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://lore.kernel.org/r/67ab679c15fbf49fa05b3ffe05d91c47ab84f147.1708426665.git.pabeni@redhat.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
|
|
udp->pcflag, udp->pcslen and udp->pcrlen reads/writes are racy.
Move udp->pcflag to udp->udp_flags for atomicity,
and add READ_ONCE()/WRITE_ONCE() annotations for pcslen and pcrlen.
Fixes: ba4e58eca8aa ("[NET]: Supporting UDP-Lite (RFC 3828) in Linux")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
|
|
This flag is set but never read, we can remove it.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
|
|
Move udp->encap_enabled to udp->udp_flags.
Add udp_test_and_set_bit() helper to allow lockless
udp_tunnel_encap_enable() implementation.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
|
|
These are read locklessly, move them to udp_flags to fix data-races.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
|
|
syzbot reported that udp->gro_enabled can be read locklessly.
Use one atomic bit from udp->udp_flags.
Fixes: e20cf8d3f1f7 ("udp: implement GRO for plain UDP sockets.")
Reported-by: syzbot <syzkaller@googlegroups.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
|
|
syzbot reported that udp->no_check6_rx can be read locklessly.
Use one atomic bit from udp->udp_flags.
Fixes: 1c19448c9ba6 ("net: Make enabling of zero UDP6 csums more restrictive")
Reported-by: syzbot <syzkaller@googlegroups.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
|
|
syzbot reported that udp->no_check6_tx can be read locklessly.
Use one atomic bit from udp->udp_flags
Fixes: 1c19448c9ba6 ("net: Make enabling of zero UDP6 csums more restrictive")
Reported-by: syzbot <syzkaller@googlegroups.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
|
|
According to syzbot, it is time to use proper atomic flags
for various UDP flags.
Add udp_flags field, and convert udp->corkflag to first
bit in it.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
|
|
We can change udp_sk() to propagate const qualifier of its argument,
thanks to container_of_const()
This should avoid some potential errors caused by accidental
(const -> not_const) promotion.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Reviewed-by: Simon Horman <simon.horman@corigine.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The maximum hash table size is 64K due to the nature of the protocol. [0]
It's smaller than TCP, and fewer sockets can cause a performance drop.
On an EC2 c5.24xlarge instance (192 GiB memory), after running iperf3 in
different netns, creating 32Mi sockets without data transfer in the root
netns causes regression for the iperf3's connection.
uhash_entries sockets length Gbps
64K 1 1 5.69
1Mi 16 5.27
2Mi 32 4.90
4Mi 64 4.09
8Mi 128 2.96
16Mi 256 2.06
32Mi 512 1.12
The per-netns hash table breaks the lengthy lists into shorter ones. It is
useful on a multi-tenant system with thousands of netns. With smaller hash
tables, we can look up sockets faster, isolate noisy neighbours, and reduce
lock contention.
The max size of the per-netns table is 64K as well. This is because the
possible hash range by udp_hashfn() always fits in 64K within the same
netns and we cannot make full use of the whole buckets larger than 64K.
/* 0 < num < 64K -> X < hash < X + 64K */
(num + net_hash_mix(net)) & mask;
Also, the min size is 128. We use a bitmap to search for an available
port in udp_lib_get_port(). To keep the bitmap on the stack and not
fire the CONFIG_FRAME_WARN error at build time, we round up the table
size to 128.
The sysctl usage is the same with TCP:
$ dmesg | cut -d ' ' -f 6- | grep "UDP hash"
UDP hash table entries: 65536 (order: 9, 2097152 bytes, vmalloc)
# sysctl net.ipv4.udp_hash_entries
net.ipv4.udp_hash_entries = 65536 # can be changed by uhash_entries
# sysctl net.ipv4.udp_child_hash_entries
net.ipv4.udp_child_hash_entries = 0 # disabled by default
# ip netns add test1
# ip netns exec test1 sysctl net.ipv4.udp_hash_entries
net.ipv4.udp_hash_entries = -65536 # share the global table
# sysctl -w net.ipv4.udp_child_hash_entries=100
net.ipv4.udp_child_hash_entries = 100
# ip netns add test2
# ip netns exec test2 sysctl net.ipv4.udp_hash_entries
net.ipv4.udp_hash_entries = 128 # own a per-netns table with 2^n buckets
We could optimise the hash table lookup/iteration further by removing
the netns comparison for the per-netns one in the future. Also, we
could optimise the sparse udp_hslot layout by putting it in udp_table.
[0]: https://lore.kernel.org/netdev/4ACC2815.7010101@gmail.com/
Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Change the udp encap_err_rcv signature to match ip_icmp_error() and
ipv6_icmp_error() so that those can be used from the called function and
export them.
Signed-off-by: David Howells <dhowells@redhat.com>
cc: Marc Dionne <marc.dionne@auristor.com>
cc: linux-afs@lists.infradead.org
cc: netdev@vger.kernel.org
|
|
When the receiver process and the BH runs on different cores,
udp_rmem_release() experience a cache miss while accessing sk_rcvbuf,
as the latter shares the same cacheline with sk_forward_alloc, written
by the BH.
With this patch, UDP tracks the rcvbuf value and its update via custom
SOL_SOCKET socket options, and copies the forward memory threshold value
used by udp_rmem_release() in a different cacheline, already accessed by
the above function and uncontended.
Since the UDP socket init operation grown a bit, factor out the common
code between v4 and v6 in a shared helper.
Overall the above give a 10% peek throughput increase under UDP flood.
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Acked-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Because rxrpc pretends to be a tunnel on top of a UDP/UDP6 socket, allowing
it to siphon off UDP packets early in the handling of received UDP packets
thereby avoiding the packet going through the UDP receive queue, it doesn't
get ICMP packets through the UDP ->sk_error_report() callback. In fact, it
doesn't appear that there's any usable option for getting hold of ICMP
packets.
Fix this by adding a new UDP encap hook to distribute error messages for
UDP tunnels. If the hook is set, then the tunnel driver will be able to
see ICMP packets. The hook provides the offset into the packet of the UDP
header of the original packet that caused the notification.
An alternative would be to call the ->error_handler() hook - but that
requires that the skbuff be cloned (as ip_icmp_error() or ipv6_cmp_error()
do, though isn't really necessary or desirable in rxrpc's case is we want
to parse them there and then, not queue them).
Changes
=======
ver #3)
- Fixed an uninitialised variable.
ver #2)
- Fixed some missing CONFIG_AF_RXRPC_IPV6 conditionals.
Fixes: 5271953cad31 ("rxrpc: Use the UDP encap_rcv hook")
Signed-off-by: David Howells <dhowells@redhat.com>
|
|
Not used since added in v3.8.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
When passing up an UDP GSO packet with L4 aggregation, there is
no need to segment it at the vxlan level. We can propagate the
packet untouched and let it be segmented later, if needed.
Introduce an helper to allow let the UDP socket to accept any
L4 aggregation and use it in the vxlan driver.
v1 -> v2:
- updated to use the newly introduced UDP socket 'accept*' fields
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Currently the UDP protocol delivers GSO_FRAGLIST packets to
the sockets without the expected segmentation.
This change addresses the issue introducing and maintaining
a couple of new fields to explicitly accept SKB_GSO_UDP_L4
or GSO_FRAGLIST packets. Additionally updates udp_unexpected_gso()
accordingly.
UDP sockets enabling UDP_GRO stil keep accept_udp_fraglist
zeroed.
v1 -> v2:
- use 2 bits instead of a whole GSO bitmask (Willem)
Fixes: 9fd1ff5d2ac7 ("udp: Support UDP fraglist GRO/GSO.")
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Based on 1 normalized pattern(s):
this program is free software you can redistribute it and or modify
it under the terms of the gnu general public license as published by
the free software foundation either version 2 of the license or at
your option any later version
extracted by the scancode license scanner the SPDX license identifier
GPL-2.0-or-later
has been chosen to replace the boilerplate/reference in 3029 file(s).
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Reviewed-by: Allison Randal <allison@lohutok.net>
Cc: linux-spdx@vger.kernel.org
Link: https://lkml.kernel.org/r/20190527070032.746973796@linutronix.de
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
|
|
For both IPv4 and IPv6, if we can't match errors to a socket, try
tunnels before ignoring them. Look up a socket with the original source
and destination ports as found in the UDP packet inside the ICMP payload,
this will work for tunnels that force the same destination port for both
endpoints, i.e. VXLAN and GENEVE.
Actually, lwtunnels could break this assumption if they are configured by
an external control plane to have different destination ports on the
endpoints: in this case, we won't be able to trace ICMP messages back to
them.
For IPv6 redirect messages, call ip6_redirect() directly with the output
interface argument set to the interface we received the packet from (as
it's the very interface we should build the exception on), otherwise the
new nexthop will be rejected. There's no such need for IPv4.
Tunnels can now export an encap_err_lookup() operation that indicates a
match. Pass the packet to the lookup function, and if the tunnel driver
reports a matching association, continue with regular ICMP error handling.
v2:
- Added newline between network and transport header sets in
__udp{4,6}_lib_err_encap() (David Miller)
- Removed redundant skb_reset_network_header(skb); in
__udp4_lib_err_encap()
- Removed redundant reassignment of iph in __udp4_lib_err_encap()
(Sabrina Dubroca)
- Edited comment to __udp{4,6}_lib_err_encap() to reflect the fact this
won't work with lwtunnels configured to use asymmetric ports. By the way,
it's VXLAN, not VxLAN (Jiri Benc)
Signed-off-by: Stefano Brivio <sbrivio@redhat.com>
Reviewed-by: Sabrina Dubroca <sd@queasysnail.net>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
In some scenarios, the GRO engine can assemble an UDP GRO packet
that ultimately lands on a non GRO-enabled socket.
This patch tries to address the issue explicitly checking for the UDP
socket features before enqueuing the packet, and eventually segmenting
the unexpected GRO packet, as needed.
We must also cope with re-insertion requests: after segmentation the
UDP code calls the helper introduced by the previous patches, as needed.
Segmentation is performed by a common helper, which takes care of
updating socket and protocol stats is case of failure.
rfc v3 -> v1
- fix compile issues with rxrpc
- when gso_segment returns NULL, treat is as an error
- added 'ipv4' argument to udp_rcv_segment()
rfc v2 -> rfc v3
- moved udp_rcv_segment() into net/udp.h, account errors to socket
and ns, always return NULL or segs list
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
When UDP GRO is enabled, the UDP_GRO cmsg will carry the ingress
datagram size. User-space can use such info to compute the original
packets layout.
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This is the RX counterpart of commit bec1f6f69736 ("udp: generate gso
with UDP_SEGMENT"). When UDP_GRO is enabled, such socket is also
eligible for GRO in the rx path: UDP segments directed to such socket
are assembled into a larger GSO_UDP_L4 packet.
The core UDP GRO support is enabled with setsockopt(UDP_GRO).
Initial benchmark numbers:
Before:
udp rx: 1079 MB/s 769065 calls/s
After:
udp rx: 1466 MB/s 24877 calls/s
This change introduces a side effect in respect to UDP tunnels:
after a UDP tunnel creation, now the kernel performs a lookup per ingress
UDP packet, while before such lookup happened only if the ingress packet
carried a valid internal header csum.
rfc v2 -> rfc v3:
- fixed typos in macro name and comments
- really enforce UDP_GRO_CNT_MAX, instead of UDP_GRO_CNT_MAX + 1
- acquire socket lock in UDP_GRO setsockopt
rfc v1 -> rfc v2:
- use a new option to enable UDP GRO
- use static keys to protect the UDP GRO socket lookup
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
The *encap_needed static keys are enabled by UDP tunnels
and several UDP encapsulations type, but they are never
turned off. This can cause unneeded overall performance
degradation for systems where such features are used
transiently.
This patch introduces complete book-keeping for such keys,
decreasing the usage at socket destruction time, if needed,
and avoiding that the same socket could increase the key
usage multiple times.
rfc v3 -> v1:
- add socket lock around udp_tunnel_encap_enable()
rfc v2 -> rfc v3:
- use udp_tunnel_encap_enable() in setsockopt()
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Manage pending per-NAPI GRO packets via list_head.
Return an SKB pointer from the GRO receive handlers. When GRO receive
handlers return non-NULL, it means that this SKB needs to be completed
at this time and removed from the NAPI queue.
Several operations are greatly simplified by this transformation,
especially timing out the oldest SKB in the list when gro_count
exceeds MAX_GRO_SKBS, and napi_gro_flush() which walks the queue
in reverse order.
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Support generic segmentation offload for udp datagrams. Callers can
concatenate and send at once the payload of multiple datagrams with
the same destination.
To set segment size, the caller sets socket option UDP_SEGMENT to the
length of each discrete payload. This value must be smaller than or
equal to the relevant MTU.
A follow-up patch adds cmsg UDP_SEGMENT to specify segment size on a
per send call basis.
Total byte length may then exceed MTU. If not an exact multiple of
segment size, the last segment will be shorter.
The implementation adds a gso_size field to the udp socket, ip(v6)
cmsg cookie and inet_cork structure to be able to set the value at
setsockopt or cmsg time and to work with both lockless and corked
paths.
Initial benchmark numbers show UDP GSO about as expensive as TCP GSO.
tcp tso
3197 MB/s 54232 msg/s 54232 calls/s
6,457,754,262 cycles
tcp gso
1765 MB/s 29939 msg/s 29939 calls/s
11,203,021,806 cycles
tcp without tso/gso *
739 MB/s 12548 msg/s 12548 calls/s
11,205,483,630 cycles
udp
876 MB/s 14873 msg/s 624666 calls/s
11,205,777,429 cycles
udp gso
2139 MB/s 36282 msg/s 36282 calls/s
11,204,374,561 cycles
[*] after reverting commit 0a6b2a1dc2a2
("tcp: switch to GSO being always on")
Measured total system cycles ('-a') for one core while pinning both
the network receive path and benchmark process to that core:
perf stat -a -C 12 -e cycles \
./udpgso_bench_tx -C 12 -4 -D "$DST" -l 4
Note the reduction in calls/s with GSO. Bytes per syscall drops
increases from 1470 to 61818.
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
under udp flood the sk_receive_queue spinlock is heavily contended.
This patch try to reduce the contention on such lock adding a
second receive queue to the udp sockets; recvmsg() looks first
in such queue and, only if empty, tries to fetch the data from
sk_receive_queue. The latter is spliced into the newly added
queue every time the receive path has to acquire the
sk_receive_queue lock.
The accounting of forward allocated memory is still protected with
the sk_receive_queue lock, so udp_rmem_release() needs to acquire
both locks when the forward deficit is flushed.
On specific scenarios we can end up acquiring and releasing the
sk_receive_queue lock multiple times; that will be covered by
the next patch
Suggested-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Acked-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
In the udp_sock struct, the 'forward_deficit' and 'pcflag' fields
share the same cacheline. While the first is dirtied by
udp_recvmsg, the latter is read, possibly several times, by the
bottom half processing to discriminate between udp and udplite
sockets.
With this patch, sk->sk_protocol is used to check is the socket is
really an udplite one, avoiding some cache misses per
packet and improving the performance under udp_flood with
small packet up to 10%.
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
If udp_recvmsg() constantly releases sk_rmem_alloc
for every read packet, it gives opportunity for
producers to immediately grab spinlocks and desperatly
try adding another packet, causing false sharing.
We can add a simple heuristic to give the signal
by batches of ~25 % of the queue capacity.
This patch considerably increases performance under
flood by about 50 %, since the thread draining the queue
is no longer slowed by false sharing.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This patch adds GRO functions (gro_receive and gro_complete) to UDP
sockets. udp_gro_receive is changed to perform socket lookup on a
packet. If a socket is found the related GRO functions are called.
This features obsoletes using UDP offload infrastructure for GRO
(udp_offload). This has the advantage of not being limited to provide
offload on a per port basis, GRO is now applied to whatever individual
UDP sockets are bound to. This also allows the possbility of
"application defined GRO"-- that is we can attach something like
a BPF program to a UDP socket to perfrom GRO on an application
layer protocol.
Signed-off-by: Tom Herbert <tom@herbertland.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Tom Herbert would like not touching UDP socket refcnt for encapsulated
traffic. For this to happen, we need to use normal RCU rules, with a grace
period before freeing a socket. UDP sockets are not short lived in the
high usage case, so the added cost of call_rcu() should not be a concern.
This actually removes a lot of complexity in UDP stack.
Multicast receives no longer need to hold a bucket spinlock.
Note that ip early demux still needs to take a reference on the socket.
Same remark for functions used by xt_socket and xt_PROXY netfilter modules,
but this might be changed later.
Performance for a single UDP socket receiving flood traffic from
many RX queues/cpus.
Simple udp_rx using simple recvfrom() loop :
438 kpps instead of 374 kpps : 17 % increase of the peak rate.
v2: Addressed Willem de Bruijn feedback in multicast handling
- keep early demux break in __udp4_lib_demux_lookup()
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Tom Herbert <tom@herbertland.com>
Cc: Willem de Bruijn <willemb@google.com>
Tested-by: Tom Herbert <tom@herbertland.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
const qualifiers ease code review by making clear
which objects are not written in a function.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Move convert_csum from udp_sock to inet_sock. This allows the
possibility that we can use convert checksum for different types
of sockets and also allows convert checksum to be enabled from
inet layer (what we'll want to do when enabling IP_CHECKSUM cmsg).
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Add support for doing CHECKSUM_UNNECESSARY to CHECKSUM_COMPLETE
conversion in UDP tunneling path.
In the normal UDP path, we call skb_checksum_try_convert after locating
the UDP socket. The check is that checksum conversion is enabled for
the socket (new flag in UDP socket) and that checksum field is
non-zero.
In the UDP GRO path, we call skb_gro_checksum_try_convert after
checksum is validated and checksum field is non-zero. Since this is
already in GRO we assume that checksum conversion is always wanted.
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
RFC 6935 permits zero checksums to be used in IPv6 however this is
recommended only for certain tunnel protocols, it does not make
checksums completely optional like they are in IPv4.
This patch restricts the use of IPv6 zero checksums that was previously
intoduced. no_check6_tx and no_check6_rx have been added to control
the use of checksums in UDP6 RX and TX path. The normal
sk_no_check_{rx,tx} settings are not used (this avoids ambiguity when
dealing with a dual stack socket).
A helper function has been added (udp_set_no_check6) which can be
called by tunnel impelmentations to all zero checksums (send on the
socket, and accept them as valid).
Signed-off-by: Tom Herbert <therbert@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Users of udp encapsulation currently have an encap_rcv callback which they can
use to hook into the udp receive path.
In situations where a encapsulation user allocates resources associated with a
udp encap socket, it may be convenient to be able to also hook the proto
.destroy operation. For example, if an encap user holds a reference to the
udp socket, the destroy hook might be used to relinquish this reference.
This patch adds a socket destroy hook into udp, which is set and enabled
in the same way as the existing encap_rcv hook.
Signed-off-by: Tom Parkin <tparkin@katalix.com>
Signed-off-by: James Chapman <jchapman@katalix.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
This patch adds support in the kernel for offloading in the NIC Tx and Rx
checksumming for encapsulated packets (such as VXLAN and IP GRE).
For Tx encapsulation offload, the driver will need to set the right bits
in netdev->hw_enc_features. The protocol driver will have to set the
skb->encapsulation bit and populate the inner headers, so the NIC driver will
use those inner headers to calculate the csum in hardware.
For Rx encapsulation offload, the driver will need to set again the
skb->encapsulation flag and the skb->ip_csum to CHECKSUM_UNNECESSARY.
In that case the protocol driver should push the decapsulated packet up
to the stack, again with CHECKSUM_UNNECESSARY. In ether case, the protocol
driver should set the skb->encapsulation flag back to zero. Finally the
protocol driver should have NETIF_F_RXCSUM flag set in its features.
Signed-off-by: Joseph Gasparakis <joseph.gasparakis@intel.com>
Signed-off-by: Peter P Waskiewicz Jr <peter.p.waskiewicz.jr@intel.com>
Signed-off-by: Alexander Duyck <alexander.h.duyck@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Signed-off-by: David Howells <dhowells@redhat.com>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Michael Kerrisk <mtk.manpages@gmail.com>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Acked-by: Dave Jones <davej@redhat.com>
|
|
UDP bind() can be O(N^2) in some pathological cases.
Thanks to secondary hash tables, we can make it O(N)
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
|
Extends udp_table to contain a secondary hash table.
socket anchor for this second hash is free, because UDP
doesnt use skc_bind_node : We define an union to hold
both skc_bind_node & a new hlist_nulls_node udp_portaddr_node
udp_lib_get_port() inserts sockets into second hash chain
(additional cost of one atomic op)
udp_lib_unhash() deletes socket from second hash chain
(additional cost of one atomic op)
Note : No spinlock lockdep annotation is needed, because
lock for the secondary hash chain is always get after
lock for primary hash chain.
Signed-off-by: Eric Dumazet <eric.dumazet@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
|
