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>
In "wireguard: queueing: preserve flow hash across packet scrubbing", we
were required to slightly increase the size of the receive replay
counter to something still fairly small, but an increase nonetheless.
It turns out that we can recoup some of the additional memory overhead
by splitting up the prior union type into two distinct types. Before, we
used the same "noise_counter" union for both sending and receiving, with
sending just using a simple atomic64_t, while receiving used the full
replay counter checker. This meant that most of the memory being
allocated for the sending counter was being wasted. Since the old
"noise_counter" type increased in size in the prior commit, now is a
good time to split up that union type into a distinct "noise_replay_
counter" for receiving and a boring atomic64_t for sending, each using
neither more nor less memory than required.
Also, since sometimes the replay counter is accessed without
necessitating additional accesses to the bitmap, we can reduce cache
misses by hoisting the always-necessary lock above the bitmap in the
struct layout. We also change a "noise_replay_counter" stack allocation
to kmalloc in a -DDEBUG selftest so that KASAN doesn't trigger a stack
frame warning.
All and all, removing a bit of abstraction in this commit makes the code
simpler and smaller, in addition to the motivating memory usage
recuperation. For example, passing around raw "noise_symmetric_key"
structs is something that really only makes sense within noise.c, in the
one place where the sending and receiving keys can safely be thought of
as the same type of object; subsequent to that, it's important that we
uniformly access these through keypair->{sending,receiving}, where their
distinct roles are always made explicit. So this patch allows us to draw
that distinction clearly as well.
Fixes: e7096c131e ("net: WireGuard secure network tunnel")
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
It's important that we clear most header fields during encapsulation and
decapsulation, because the packet is substantially changed, and we don't
want any info leak or logic bug due to an accidental correlation. But,
for encapsulation, it's wrong to clear skb->hash, since it's used by
fq_codel and flow dissection in general. Without it, classification does
not proceed as usual. This change might make it easier to estimate the
number of innerflows by examining clustering of out of order packets,
but this shouldn't open up anything that can't already be inferred
otherwise (e.g. syn packet size inference), and fq_codel can be disabled
anyway.
Furthermore, it might be the case that the hash isn't used or queried at
all until after wireguard transmits the encrypted UDP packet, which
means skb->hash might still be zero at this point, and thus no hash
taken over the inner packet data. In order to address this situation, we
force a calculation of skb->hash before encrypting packet data.
Of course this means that fq_codel might transmit packets slightly more
out of order than usual. Toke did some testing on beefy machines with
high quantities of parallel flows and found that increasing the
reply-attack counter to 8192 takes care of the most pathological cases
pretty well.
Reported-by: Dave Taht <dave.taht@gmail.com>
Reviewed-and-tested-by: Toke Høiland-Jørgensen <toke@toke.dk>
Fixes: e7096c131e ("net: WireGuard secure network tunnel")
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
It's very unlikely that send will become true. It's nearly always false
between 0 and 120 seconds of a session, and in most cases becomes true
only between 120 and 121 seconds before becoming false again. So,
unlikely(send) is clearly the right option here.
What happened before was that we had this complex boolean expression
with multiple likely and unlikely clauses nested. Since this is
evaluated left-to-right anyway, the whole thing got converted to
unlikely. So, we can clean this up to better represent what's going on.
The generated code is the same.
Suggested-by: Sultan Alsawaf <sultan@kerneltoast.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Users with pathological hardware reported CPU stalls on CONFIG_
PREEMPT_VOLUNTARY=y, because the ringbuffers would stay full, meaning
these workers would never terminate. That turned out not to be okay on
systems without forced preemption, which Sultan observed. This commit
adds a cond_resched() to the bottom of each loop iteration, so that
these workers don't hog the core. Note that we don't need this on the
napi poll worker, since that terminates after its budget is expended.
Suggested-by: Sultan Alsawaf <sultan@kerneltoast.com>
Reported-by: Wang Jian <larkwang@gmail.com>
Fixes: e7096c131e ("net: WireGuard secure network tunnel")
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
This commit removes a useless newline at the end of a scope, which
doesn't add anything in the way of organization or readability.
Signed-off-by: Sultan Alsawaf <sultan@kerneltoast.com>
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
It turns out there's an easy way to get packets queued up while still
having an MTU of zero, and that's via persistent keep alive. This commit
makes sure that in whatever condition, we don't wind up dividing by
zero. Note that an MTU of zero for a wireguard interface is something
quasi-valid, so I don't think the correct fix is to limit it via
min_mtu. This can be reproduced easily with:
ip link add wg0 type wireguard
ip link add wg1 type wireguard
ip link set wg0 up mtu 0
ip link set wg1 up
wg set wg0 private-key <(wg genkey)
wg set wg1 listen-port 1 private-key <(wg genkey) peer $(wg show wg0 public-key)
wg set wg0 peer $(wg show wg1 public-key) persistent-keepalive 1 endpoint 127.0.0.1:1
However, while min_mtu=0 seems fine, it makes sense to restrict the
max_mtu. This commit also restricts the maximum MTU to the greatest
number for which rounding up to the padding multiple won't overflow a
signed integer. Packets this large were always rejected anyway
eventually, due to checks deeper in, but it seems more sound not to even
let the administrator configure something that won't work anyway.
We use this opportunity to clean up this function a bit so that it's
clear which paths we're expecting.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Cc: Eric Dumazet <eric.dumazet@gmail.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
WireGuard is a layer 3 secure networking tunnel made specifically for
the kernel, that aims to be much simpler and easier to audit than IPsec.
Extensive documentation and description of the protocol and
considerations, along with formal proofs of the cryptography, are
available at:
* https://www.wireguard.com/
* https://www.wireguard.com/papers/wireguard.pdf
This commit implements WireGuard as a simple network device driver,
accessible in the usual RTNL way used by virtual network drivers. It
makes use of the udp_tunnel APIs, GRO, GSO, NAPI, and the usual set of
networking subsystem APIs. It has a somewhat novel multicore queueing
system designed for maximum throughput and minimal latency of encryption
operations, but it is implemented modestly using workqueues and NAPI.
Configuration is done via generic Netlink, and following a review from
the Netlink maintainer a year ago, several high profile userspace tools
have already implemented the API.
This commit also comes with several different tests, both in-kernel
tests and out-of-kernel tests based on network namespaces, taking profit
of the fact that sockets used by WireGuard intentionally stay in the
namespace the WireGuard interface was originally created, exactly like
the semantics of userspace tun devices. See wireguard.com/netns/ for
pictures and examples.
The source code is fairly short, but rather than combining everything
into a single file, WireGuard is developed as cleanly separable files,
making auditing and comprehension easier. Things are laid out as
follows:
* noise.[ch], cookie.[ch], messages.h: These implement the bulk of the
cryptographic aspects of the protocol, and are mostly data-only in
nature, taking in buffers of bytes and spitting out buffers of
bytes. They also handle reference counting for their various shared
pieces of data, like keys and key lists.
* ratelimiter.[ch]: Used as an integral part of cookie.[ch] for
ratelimiting certain types of cryptographic operations in accordance
with particular WireGuard semantics.
* allowedips.[ch], peerlookup.[ch]: The main lookup structures of
WireGuard, the former being trie-like with particular semantics, an
integral part of the design of the protocol, and the latter just
being nice helper functions around the various hashtables we use.
* device.[ch]: Implementation of functions for the netdevice and for
rtnl, responsible for maintaining the life of a given interface and
wiring it up to the rest of WireGuard.
* peer.[ch]: Each interface has a list of peers, with helper functions
available here for creation, destruction, and reference counting.
* socket.[ch]: Implementation of functions related to udp_socket and
the general set of kernel socket APIs, for sending and receiving
ciphertext UDP packets, and taking care of WireGuard-specific sticky
socket routing semantics for the automatic roaming.
* netlink.[ch]: Userspace API entry point for configuring WireGuard
peers and devices. The API has been implemented by several userspace
tools and network management utility, and the WireGuard project
distributes the basic wg(8) tool.
* queueing.[ch]: Shared function on the rx and tx path for handling
the various queues used in the multicore algorithms.
* send.c: Handles encrypting outgoing packets in parallel on
multiple cores, before sending them in order on a single core, via
workqueues and ring buffers. Also handles sending handshake and cookie
messages as part of the protocol, in parallel.
* receive.c: Handles decrypting incoming packets in parallel on
multiple cores, before passing them off in order to be ingested via
the rest of the networking subsystem with GRO via the typical NAPI
poll function. Also handles receiving handshake and cookie messages
as part of the protocol, in parallel.
* timers.[ch]: Uses the timer wheel to implement protocol particular
event timeouts, and gives a set of very simple event-driven entry
point functions for callers.
* main.c, version.h: Initialization and deinitialization of the module.
* selftest/*.h: Runtime unit tests for some of the most security
sensitive functions.
* tools/testing/selftests/wireguard/netns.sh: Aforementioned testing
script using network namespaces.
This commit aims to be as self-contained as possible, implementing
WireGuard as a standalone module not needing much special handling or
coordination from the network subsystem. I expect for future
optimizations to the network stack to positively improve WireGuard, and
vice-versa, but for the time being, this exists as intentionally
standalone.
We introduce a menu option for CONFIG_WIREGUARD, as well as providing a
verbose debug log and self-tests via CONFIG_WIREGUARD_DEBUG.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Cc: David Miller <davem@davemloft.net>
Cc: Greg KH <gregkh@linuxfoundation.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: linux-crypto@vger.kernel.org
Cc: linux-kernel@vger.kernel.org
Cc: netdev@vger.kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>
Jason A. Donenfeld