wg_netns_pre_exit() manually acquires rtnl_lock() inside the
pernet .pre_exit callback. This causes a hung task when another
thread holds rtnl_mutex - the cleanup_net workqueue (or the
setup_net failure rollback path) blocks indefinitely in
wg_netns_pre_exit() waiting to acquire the lock.
Convert to .exit_rtnl, introduced in commit 7a60d91c69 ("net:
Add ->exit_rtnl() hook to struct pernet_operations."), where the
framework already holds RTNL and batches all callbacks under a
single rtnl_lock()/rtnl_unlock() pair, eliminating the contention
window.
The rcu_assign_pointer(wg->creating_net, NULL) is safe to move
from .pre_exit to .exit_rtnl (which runs after synchronize_rcu())
because all RCU readers of creating_net either use maybe_get_net()
- which returns NULL for a dying namespace with zero refcount - or
access net->user_ns which remains valid throughout the entire
ops_undo_list sequence.
Reported-by: syzbot+f2fbf7478a35a94c8b7c@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?id=cb64c22a492202ca929e18262fdb8cb89e635c70
Signed-off-by: Shardul Bankar <shardul.b@mpiricsoftware.com>
[ Jason: added __net_exit and __read_mostly annotations that were missing. ]
Fixes: 900575aa33 ("wireguard: device: avoid circular netns references")
Cc: stable@vger.kernel.org
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Link: https://patch.msgid.link/20260414153944.2742252-5-Jason@zx2c4.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Currently if a user enqueue a work item using schedule_delayed_work() the
used wq is "system_wq" (per-cpu wq) while queue_delayed_work() use
WORK_CPU_UNBOUND (used when a cpu is not specified). The same applies to
schedule_work() that is using system_wq and queue_work(), that makes use
again of WORK_CPU_UNBOUND.
This lack of consistentcy cannot be addressed without refactoring the API.
alloc_workqueue() treats all queues as per-CPU by default, while unbound
workqueues must opt-in via WQ_UNBOUND.
This default is suboptimal: most workloads benefit from unbound queues,
allowing the scheduler to place worker threads where they’re needed and
reducing noise when CPUs are isolated.
This change adds a new WQ_PERCPU flag at the network subsystem, to explicitly
request the use of the per-CPU behavior. Both flags coexist for one release
cycle to allow callers to transition their calls.
Once migration is complete, WQ_UNBOUND can be removed and unbound will
become the implicit default.
With the introduction of the WQ_PERCPU flag (equivalent to !WQ_UNBOUND),
any alloc_workqueue() caller that doesn’t explicitly specify WQ_UNBOUND
must now use WQ_PERCPU.
All existing users have been updated accordingly.
Suggested-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Marco Crivellari <marco.crivellari@suse.com>
Link: https://patch.msgid.link/20250918142427.309519-4-marco.crivellari@suse.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Prepare for adding an enum type for NAPI threaded states by adding
netif_threaded_enable API. De-export the existing netif_set_threaded API
and only use it internally. Update existing drivers to use
netif_threaded_enable instead of the de-exported netif_set_threaded.
Note that dev_set_threaded used by mt76 debugfs file is unchanged.
Signed-off-by: Samiullah Khawaja <skhawaja@google.com>
Link: https://patch.msgid.link/20250723013031.2911384-3-skhawaja@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Commit cc34acd577 ("docs: net: document new locking reality")
introduced netif_ vs dev_ function semantics: the former expects locked
netdev, the latter takes care of the locking. We don't strictly
follow this semantics on either side, but there are more dev_xxx handlers
now that don't fit. Rename them to netif_xxx where appropriate.
Note that one dev_set_threaded call still remains in mt76 for debugfs file.
Signed-off-by: Stanislav Fomichev <sdf@fomichev.me>
Link: https://patch.msgid.link/20250717172333.1288349-7-sdf@fomichev.me
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Enable threaded NAPI by default for WireGuard devices in response to low
performance behavior that we observed when multiple tunnels (and thus
multiple wg devices) are deployed on a single host. This affects any
kind of multi-tunnel deployment, regardless of whether the tunnels share
the same endpoints or not (i.e., a VPN concentrator type of gateway
would also be affected).
The problem is caused by the fact that, in case of a traffic surge that
involves multiple tunnels at the same time, the polling of the NAPI
instance of all these wg devices tends to converge onto the same core,
causing underutilization of the CPU and bottlenecking performance.
This happens because NAPI polling is hosted by default in softirq
context, but the WireGuard driver only raises this softirq after the rx
peer queue has been drained, which doesn't happen during high traffic.
In this case, the softirq already active on a core is reused instead of
raising a new one.
As a result, once two or more tunnel softirqs have been scheduled on
the same core, they remain pinned there until the surge ends.
In our experiments, this almost always leads to all tunnel NAPIs being
handled on a single core shortly after a surge begins, limiting
scalability to less than 3× the performance of a single tunnel, despite
plenty of unused CPU cores being available.
The proposed mitigation is to enable threaded NAPI for all WireGuard
devices. This moves the NAPI polling context to a dedicated per-device
kernel thread, allowing the scheduler to balance the load across all
available cores.
On our 32-core gateways, enabling threaded NAPI yields a ~4× performance
improvement with 16 tunnels, increasing throughput from ~13 Gbps to
~48 Gbps. Meanwhile, CPU usage on the receiver (which is the bottleneck)
jumps from 20% to 100%.
We have found no performance regressions in any scenario we tested.
Single-tunnel throughput remains unchanged.
More details are available in our Netdev paper.
Link: https://netdevconf.info/0x18/docs/netdev-0x18-paper23-talk-paper.pdf
Signed-off-by: Mirco Barone <mirco.barone@polito.it>
Fixes: e7096c131e ("net: WireGuard secure network tunnel")
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Link: https://patch.msgid.link/20250605120616.2808744-1-Jason@zx2c4.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
timer_delete[_sync]() replaces del_timer[_sync](). Convert the whole tree
over and remove the historical wrapper inlines.
Conversion was done with coccinelle plus manual fixups where necessary.
Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@kernel.org>
Add two helper functions - rtnl_newlink_link_net() and
rtnl_newlink_peer_net() for netns fallback logic. Peer netns falls back
to link netns, and link netns falls back to source netns.
Convert the use of params->net in netdevice drivers to one of the helper
functions for clarity.
Signed-off-by: Xiao Liang <shaw.leon@gmail.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250219125039.18024-4-shaw.leon@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
There are 4 net namespaces involved when creating links:
- source netns - where the netlink socket resides,
- target netns - where to put the device being created,
- link netns - netns associated with the device (backend),
- peer netns - netns of peer device.
Currently, two nets are passed to newlink() callback - "src_net"
parameter and "dev_net" (implicitly in net_device). They are set as
follows, depending on netlink attributes in the request.
+------------+-------------------+---------+---------+
| peer netns | IFLA_LINK_NETNSID | src_net | dev_net |
+------------+-------------------+---------+---------+
| | absent | source | target |
| absent +-------------------+---------+---------+
| | present | link | link |
+------------+-------------------+---------+---------+
| | absent | peer | target |
| present +-------------------+---------+---------+
| | present | peer | link |
+------------+-------------------+---------+---------+
When IFLA_LINK_NETNSID is present, the device is created in link netns
first and then moved to target netns. This has some side effects,
including extra ifindex allocation, ifname validation and link events.
These could be avoided if we create it in target netns from
the beginning.
On the other hand, the meaning of src_net parameter is ambiguous. It
varies depending on how parameters are passed. It is the effective
link (or peer netns) by design, but some drivers ignore it and use
dev_net instead.
To provide more netns context for drivers, this patch packs existing
newlink() parameters, along with the source netns, link netns and peer
netns, into a struct. The old "src_net" is renamed to "net" to avoid
confusion with real source netns, and will be deprecated later. The use
of src_net are converted to params->net trivially.
Signed-off-by: Xiao Liang <shaw.leon@gmail.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://patch.msgid.link/20250219125039.18024-3-shaw.leon@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Advertise GSO_MAX_SIZE as TSO max size in order support BIG TCP for wireguard.
This helps to improve wireguard performance a bit when enabled as it allows
wireguard to aggregate larger skbs in wg_packet_consume_data_done() via
napi_gro_receive(), but also allows the stack to build larger skbs on xmit
where the driver then segments them before encryption inside wg_xmit().
We've seen a 15% improvement in TCP stream performance.
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Link: https://patch.msgid.link/20241117212030.629159-5-Jason@zx2c4.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
NETIF_F_LLTX can't be changed via Ethtool and is not a feature,
rather an attribute, very similar to IFF_NO_QUEUE (and hot).
Free one netdev_features_t bit and make it a "hot" private flag.
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Commit 3e2f544dd8 ("net: get stats64 if device if driver is
configured") moved the callback to dev_get_tstats64() to net core, so,
unless the driver is doing some custom stats collection, it does not
need to set .ndo_get_stats64.
Since this driver is now relying in NETDEV_PCPU_STAT_TSTATS, then, it
doesn't need to set the dev_get_tstats64() generic .ndo_get_stats64
function pointer.
Signed-off-by: Breno Leitao <leitao@debian.org>
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Reviewed-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
With commit 34d21de99c ("net: Move {l,t,d}stats allocation to core
and convert veth & vrf"), stats allocation could be done on net core
instead of in this driver.
With this new approach, the driver doesn't have to bother with error
handling (allocation failure checking, making sure free happens in the
right spot, etc). This is core responsibility now.
Remove the allocation in this driver and leverage the network core
allocation instead.
Signed-off-by: Breno Leitao <leitao@debian.org>
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Reviewed-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
wg_xmit() can be called concurrently, KCSAN reported [1]
some device stats updates can be lost.
Use DEV_STATS_INC() for this unlikely case.
[1]
BUG: KCSAN: data-race in wg_xmit / wg_xmit
read-write to 0xffff888104239160 of 8 bytes by task 1375 on cpu 0:
wg_xmit+0x60f/0x680 drivers/net/wireguard/device.c:231
__netdev_start_xmit include/linux/netdevice.h:4918 [inline]
netdev_start_xmit include/linux/netdevice.h:4932 [inline]
xmit_one net/core/dev.c:3543 [inline]
dev_hard_start_xmit+0x11b/0x3f0 net/core/dev.c:3559
...
read-write to 0xffff888104239160 of 8 bytes by task 1378 on cpu 1:
wg_xmit+0x60f/0x680 drivers/net/wireguard/device.c:231
__netdev_start_xmit include/linux/netdevice.h:4918 [inline]
netdev_start_xmit include/linux/netdevice.h:4932 [inline]
xmit_one net/core/dev.c:3543 [inline]
dev_hard_start_xmit+0x11b/0x3f0 net/core/dev.c:3559
...
v2: also change wg_packet_consume_data_done() (Hangbin Liu)
and wg_packet_purge_staged_packets()
Fixes: e7096c131e ("net: WireGuard secure network tunnel")
Reported-by: syzbot <syzkaller@googlegroups.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Jason A. Donenfeld <Jason@zx2c4.com>
Cc: Hangbin Liu <liuhangbin@gmail.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Reviewed-by: Hangbin Liu <liuhangbin@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
When a virtual machine forks, it's important that WireGuard clear
existing sessions so that different plaintexts are not transmitted using
the same key+nonce, which can result in catastrophic cryptographic
failure. To accomplish this, we simply hook into the newly added vmfork
notifier.
As a bonus, it turns out that, like the vmfork registration function,
the PM registration function is stubbed out when CONFIG_PM_SLEEP is not
set, so we can actually just remove the maze of ifdefs, which makes it
really quite clean to support both notifiers at once.
Cc: Dominik Brodowski <linux@dominikbrodowski.net>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Theodore Ts'o <tytso@mit.edu>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Apparently the spinlock on incoming_handshake's skb_queue is highly
contended, and a torrent of handshake or cookie packets can bring the
data plane to its knees, simply by virtue of enqueueing the handshake
packets to be processed asynchronously. So, we try switching this to a
ring buffer to hopefully have less lock contention. This alleviates the
problem somewhat, though it still isn't perfect, so future patches will
have to improve this further. However, it at least doesn't completely
diminish the data plane.
Reported-by: Streun Fabio <fstreun@student.ethz.ch>
Reported-by: Joel Wanner <joel.wanner@inf.ethz.ch>
Fixes: e7096c131e ("net: WireGuard secure network tunnel")
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Each peer's endpoint contains a dst_cache entry that takes a reference
to another netdev. When the containing namespace exits, we take down the
socket and prevent future sockets from being created (by setting
creating_net to NULL), which removes that potential reference on the
netns. However, it doesn't release references to the netns that a netdev
cached in dst_cache might be taking, so the netns still might fail to
exit. Since the socket is gimped anyway, we can simply clear all the
dst_caches (by way of clearing the endpoint src), which will release all
references.
However, the current dst_cache_reset function only releases those
references lazily. But it turns out that all of our usages of
wg_socket_clear_peer_endpoint_src are called from contexts that are not
exactly high-speed or bottle-necked. For example, when there's
connection difficulty, or when userspace is reconfiguring the interface.
And in particular for this patch, when the netns is exiting. So for
those cases, it makes more sense to call dst_release immediately. For
that, we add a small helper function to dst_cache.
This patch also adds a test to netns.sh from Hangbin Liu to ensure this
doesn't regress.
Tested-by: Hangbin Liu <liuhangbin@gmail.com>
Reported-by: Xiumei Mu <xmu@redhat.com>
Cc: Toke Høiland-Jørgensen <toke@redhat.com>
Cc: Paolo Abeni <pabeni@redhat.com>
Fixes: 900575aa33 ("wireguard: device: avoid circular netns references")
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Having two ring buffers per-peer means that every peer results in two
massive ring allocations. On an 8-core x86_64 machine, this commit
reduces the per-peer allocation from 18,688 bytes to 1,856 bytes, which
is an 90% reduction. Ninety percent! With some single-machine
deployments approaching 500,000 peers, we're talking about a reduction
from 7 gigs of memory down to 700 megs of memory.
In order to get rid of these per-peer allocations, this commit switches
to using a list-based queueing approach. Currently GSO fragments are
chained together using the skb->next pointer (the skb_list_* singly
linked list approach), so we form the per-peer queue around the unused
skb->prev pointer (which sort of makes sense because the links are
pointing backwards). Use of skb_queue_* is not possible here, because
that is based on doubly linked lists and spinlocks. Multiple cores can
write into the queue at any given time, because its writes occur in the
start_xmit path or in the udp_recv path. But reads happen in a single
workqueue item per-peer, amounting to a multi-producer, single-consumer
paradigm.
The MPSC queue is implemented locklessly and never blocks. However, it
is not linearizable (though it is serializable), with a very tight and
unlikely race on writes, which, when hit (some tiny fraction of the
0.15% of partial adds on a fully loaded 16-core x86_64 system), causes
the queue reader to terminate early. However, because every packet sent
queues up the same workqueue item after it is fully added, the worker
resumes again, and stopping early isn't actually a problem, since at
that point the packet wouldn't have yet been added to the encryption
queue. These properties allow us to avoid disabling interrupts or
spinning. The design is based on Dmitry Vyukov's algorithm [1].
Performance-wise, ordinarily list-based queues aren't preferable to
ringbuffers, because of cache misses when following pointers around.
However, we *already* have to follow the adjacent pointers when working
through fragments, so there shouldn't actually be any change there. A
potential downside is that dequeueing is a bit more complicated, but the
ptr_ring structure used prior had a spinlock when dequeueing, so all and
all the difference appears to be a wash.
Actually, from profiling, the biggest performance hit, by far, of this
commit winds up being atomic_add_unless(count, 1, max) and atomic_
dec(count), which account for the majority of CPU time, according to
perf. In that sense, the previous ring buffer was superior in that it
could check if it was full by head==tail, which the list-based approach
cannot do.
But all and all, this enables us to get massive memory savings, allowing
WireGuard to scale for real world deployments, without taking much of a
performance hit.
[1] http://www.1024cores.net/home/lock-free-algorithms/queues/intrusive-mpsc-node-based-queue
Reviewed-by: Dmitry Vyukov <dvyukov@google.com>
Reviewed-by: Toke Høiland-Jørgensen <toke@redhat.com>
Fixes: e7096c131e ("net: WireGuard secure network tunnel")
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
If skb->protocol doesn't match the actual skb->data header, it's
probably not a good idea to pass it off to icmp{,v6}_ndo_send, which is
expecting to reply to a valid IP packet. So this commit has that early
mismatch case jump to a later error label.
Fixes: e7096c131e ("net: WireGuard secure network tunnel")
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
The definition of IS_ERR() already applies the unlikely() notation
when checking the error status of the passed pointer. For this
reason there is no need to have the same notation outside of
IS_ERR() itself.
Clean up code by removing redundant notation.
Signed-off-by: Antonio Quartulli <a@unstable.cc>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Replace ip_tunnel_get_stats64() with the new identical core function
dev_get_tstats64().
Reviewed-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Heiner Kallweit <hkallweit1@gmail.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
WireGuard uses skb->protocol to determine packet type, and bails out if
it's not set or set to something it's not expecting. For AF_PACKET
injection, we need to support its call chain of:
packet_sendmsg -> packet_snd -> packet_parse_headers ->
dev_parse_header_protocol -> parse_protocol
Without a valid parse_protocol, this returns zero, and wireguard then
rejects the skb. So, this wires up the ip_tunnel handler for layer 3
packets for that case.
Reported-by: Hans Wippel <ndev@hwipl.net>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
