This version has the various slew of bug fixes and compat fixes and
such, but the most interesting thing from an OpenWRT perspective is that
WireGuard now plays nicely with cake and fq_codel. I'll be very
interested to hear from OpenWRT users whether this makes a measurable
difference. Usual set of full changes follows.
This release aligns with the changes I sent to DaveM for 5.7-rc7 and were
pushed to net.git about 45 minutes ago.
* qemu: use newer iproute2 for gcc-10
* qemu: add -fcommon for compiling ping with gcc-10
These enable the test suite to compile with gcc-10.
* noise: read preshared key while taking lock
Matt noticed a benign data race when porting the Linux code to OpenBSD.
* queueing: preserve flow hash across packet scrubbing
* noise: separate receive counter from send counter
WireGuard now works with fq_codel, cake, and other qdiscs that make use of
skb->hash. This should significantly improve latency spikes related to
buffer bloat. Here's a before and after graph from some data Toke measured:
https://data.zx2c4.com/removal-of-buffer-bloat-in-wireguard.png
* compat: support RHEL 8 as 8.2, drop 8.1 support
* compat: support CentOS 8 explicitly
* compat: RHEL7 backported the skb hash renamings
The usual RHEL churn.
* compat: backport renamed/missing skb hash members
The new support for fq_codel and friends meant more backporting work.
* compat: ip6_dst_lookup_flow was backported to 4.14, 4.9, and 4.4
The main motivation for releasing this now: three stable kernels were released
at the same time, with a patch that necessitated updating in our compat layer.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* compat: timeconst.h is a generated artifact
Before we were trying to check for timeconst.h by looking in the kernel
source directory. This isn't quite correct on configurations in which
the object directory is separate from the kernel source directory, for
example when using O="elsewhere" as a make option when building the
kernel. The correct fix is to use $(CURDIR), which should point to
where we want.
* compat: use bash instead of bc for HZ-->USEC calculation
This should make packaging somewhat easier, as bash is generally already
available (at least for dkms), whereas bc isn't provided by distros by
default in their build meta packages.
* socket: remove errant restriction on looping to self
It's already possible to create two different interfaces and loop
packets between them. This has always been possible with tunnels in the
kernel, and isn't specific to wireguard. Therefore, the networking stack
already needs to deal with that. At the very least, the packet winds up
exceeding the MTU and is discarded at that point. So, since this is
already something that happens, there's no need to forbid the not very
exceptional case of routing a packet back to the same interface; this
loop is no different than others, and we shouldn't special case it, but
rather rely on generic handling of loops in general. This also makes it
easier to do interesting things with wireguard such as onion routing.
At the same time, we add a selftest for this, ensuring that both onion
routing works and infinite routing loops do not crash the kernel. We
also add a test case for wireguard interfaces nesting packets and
sending traffic between each other, as well as the loop in this case
too. We make sure to send some throughput-heavy traffic for this use
case, to stress out any possible recursion issues with the locks around
workqueues.
* send: cond_resched() when processing tx ringbuffers
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. This commit adds a cond_resched() to
the bottom of each loop iteration, so that these workers don't hog the
core. We don't do this on encryption/decryption because the compat
module here uses simd_relax, which already includes a call to schedule
in preempt_enable.
* selftests: initalize ipv6 members to NULL to squelch clang warning
This fixes a worthless warning from clang.
* send/receive: use explicit unlikely branch instead of implicit coalescing
Some code readibility cleanups.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* compat: support latest suse 15.1 and 15.2
* compat: support RHEL 7.8's faulty siphash backport
* compat: error out if bc is missing
* compat: backport hsiphash_1u32 for tests
We now have improved support for RHEL 7.8, SUSE 15.[12], and Ubuntu 16.04.
* compat: include sch_generic.h header for skb_reset_tc
A fix for a compiler error on kernels with weird configs.
* compat: import latest fixes for ptr_ring
* compat: don't assume READ_ONCE barriers on old kernels
* compat: kvmalloc_array is not required anyway
ptr_ring.h from upstream was imported, with compat modifications, to our
compat layer, to receive the latest fixes.
* compat: prefix icmp[v6]_ndo_send with __compat
Some distros that backported icmp[v6]_ndo_send still try to build the compat
module in some corner case circumstances, resulting in errors. Work around
this with the usual __compat games.
* compat: ip6_dst_lookup_flow was backported to 3.16.83
* compat: ip6_dst_lookup_flow was backported to 4.19.119
Greg and Ben backported the ip6_dst_lookup_flow patches to stable kernels,
causing breaking in our compat module, which these changes fix.
* git: add gitattributes so tarball doesn't have gitignore files
Distros won't need to clean this up manually now.
* crypto: do not export symbols
These don't do anything and only increased file size.
* queueing: cleanup ptr_ring in error path of packet_queue_init
Sultan Alsawaf reported a memory leak on an error path.
* main: mark as in-tree
Now that we're upstream, there's no need to set the taint flag.
* receive: use tunnel helpers for decapsulating ECN markings
ECN markings are now decapsulated using RFC6040 instead of the old RFC3168.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Recent backports to 5.5 and 5.4 broke our compat layer. This release is
to keep things running with the latest upstream stable kernels.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* queueing: backport skb_reset_redirect change from 5.6
* version: bump
This release has only one slight change, to put it closer to the 5.6
codebase, but its main purpose is to bump us to a 1.0.y version number.
Now that WireGuard 1.0.0 has been released for Linux 5.6 [1], we can put
the same number on the backport compat codebase.
When OpenWRT bumps to Linux 5.6, we'll be able to drop this package
entirely, which I look forward to seeing.
[1] https://lists.zx2c4.com/pipermail/wireguard/2020-March/005206.html
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
WireGuard had a brief professional security audit. The auditors didn't find
any vulnerabilities, but they did suggest one defense-in-depth suggestion to
protect against potential API misuse down the road, mentioned below. This
compat snapshot corresponds with the patches I just pushed to Dave for
5.6-rc7.
* curve25519-x86_64: avoid use of r12
This buys us 100 extra cycles, which isn't much, but it winds up being even
faster on PaX kernels, which use r12 as a RAP register.
* wireguard: queueing: account for skb->protocol==0
This is the defense-in-depth change. We deal with skb->protocol==0 just fine,
but the advice to deal explicitly with it seems like a good idea.
* receive: remove dead code from default packet type case
A default case of a particular switch statement should never be hit, so
instead of printing a pretty debug message there, we full-on WARN(), so that
we get bug reports.
* noise: error out precomputed DH during handshake rather than config
All peer keys will now be addable, even if they're low order. However, no
handshake messages will be produced successfully. This is a more consistent
behavior with other low order keys, where the handshake just won't complete if
they're being used anywhere.
* send: use normaler alignment formula from upstream
We're trying to keep a minimal delta with upstream for the compat backport.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* send: cleanup skb padding calculation
* socket: remove useless synchronize_net
Sorry for the back-to-back releases. This fixes a regression spotted by Eric
Dumazet.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* compat: support building for RHEL-8.2
* compat: remove RHEL-7.6 workaround
Bleeding edge RHEL users should be content now (which includes the actual
RedHat employees I've been talking to about getting this into the RHEL kernel
itself). Also, we remove old hacks for versions we no longer support anyway.
* allowedips: remove previously added list item when OOM fail
* noise: reject peers with low order public keys
With this now being upstream, we benefit from increased fuzzing coverage of
the code, uncovering these two bugs.
* netns: ensure non-addition of peers with failed precomputation
* netns: tie socket waiting to target pid
An added test to our test suite for the above and a small fix for high-load CI
scenarios.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* Makefile: strip prefixed v from version.h
This fixes a mistake in dmesg output and when parsing the sysfs entry in the
filesystem.
* device: skb_list_walk_safe moved upstream
This is a 5.6 change, which we won't support here, but it does make the code
cleaner, so we make this change to keep things in sync.
* curve25519: x86_64: replace with formally verified implementation
This comes from INRIA's HACL*/Vale. It implements the same algorithm and
implementation strategy as the code it replaces, only this code has been
formally verified, sans the base point multiplication, which uses code
similar to prior, only it uses the formally verified field arithmetic
alongside reproducable ladder generation steps. This doesn't have a
pure-bmi2 version, which means haswell no longer benefits, but the
increased (doubled) code complexity is not worth it for a single
generation of chips that's already old.
Performance-wise, this is around 1% slower on older microarchitectures,
and slightly faster on newer microarchitectures, mainly 10nm ones or
backports of 10nm to 14nm. This implementation is "everest" below:
Xeon E5-2680 v4 (Broadwell)
armfazh: 133340 cycles per call
everest: 133436 cycles per call
Xeon Gold 5120 (Sky Lake Server)
armfazh: 112636 cycles per call
everest: 113906 cycles per call
Core i5-6300U (Sky Lake Client)
armfazh: 116810 cycles per call
everest: 117916 cycles per call
Core i7-7600U (Kaby Lake)
armfazh: 119523 cycles per call
everest: 119040 cycles per call
Core i7-8750H (Coffee Lake)
armfazh: 113914 cycles per call
everest: 113650 cycles per call
Core i9-9880H (Coffee Lake Refresh)
armfazh: 112616 cycles per call
everest: 114082 cycles per call
Core i3-8121U (Cannon Lake)
armfazh: 113202 cycles per call
everest: 111382 cycles per call
Core i7-8265U (Whiskey Lake)
armfazh: 127307 cycles per call
everest: 127697 cycles per call
Core i7-8550U (Kaby Lake Refresh)
armfazh: 127522 cycles per call
everest: 127083 cycles per call
Xeon Platinum 8275CL (Cascade Lake)
armfazh: 114380 cycles per call
everest: 114656 cycles per call
Achieving these kind of results with formally verified code is quite
remarkable, especialy considering that performance is favorable for
newer chips.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* socket: mark skbs as not on list when receiving via gro
Certain drivers will pass gro skbs to udp, at which point the udp driver
simply iterates through them and passes them off to encap_rcv, which is
where we pick up. At the moment, we're not attempting to coalesce these
into bundles, but we also don't want to wind up having cascaded lists of
skbs treated separately. The right behavior here, then, is to just mark
each incoming one as not on a list. This can be seen in practice, for
example, with Qualcomm's rmnet_perf driver. This lead to crashes on
OnePlus devices and possibly other Qualcomm 4.14 devices. But I fear
that it could lead to issues on other drivers on weird OpenWRT routers.
This commit is upstream in net-next as:
https://git.kernel.org/pub/scm/linux/kernel/git/netdev/net-next.git/commit/?id=736775d06bac60d7a353e405398b48b2bd8b1e54
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
As announced on the mailing list, WireGuard will be in Linux 5.6. As a
result, the wg(8) tool, used by OpenWRT in the same manner as ip(8), is
moving to its own wireguard-tools repo. Meanwhile, the out-of-tree
kernel module for kernels 3.10 - 5.5 moved to its own wireguard-linux-
compat repo. Yesterday, releases were cut out of these repos, so this
commit bumps packages to match. Since wg(8) and the compat kernel module
are versioned and released separately, we create a wireguard-tools
Makefile to contain the source for the new tools repo. Later, when
OpenWRT moves permanently to Linux 5.6, we'll drop the original module
package, leaving only the tools. So this commit shuffles the build
definition around a bit but is basically the same idea as before.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
edad0d6 version: bump snapshot
0e38a3c compat: ipv6_dst_lookup_flow was backported to 5.3 and 5.4
2e52c41 wg-quick: linux: use already configured addresses instead of in-memory
3721521 tools: adjust wg.8 syntax for consistency in COMMANDS section
21a1498 wg-quick: linux: try both iptables(8) and nft(8) on teardown
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
1ec6ece version: bump snapshot
e13de91 main: remove unused include <linux/version.h>
72eb17c wg-quick: linux: support older nft(8)
1d8e978 global: fix up spelling
e02713e wg-quick: linux: add support for nft and prefer it
b4e3a83 compat: support building for RHEL-8.1 instead of RHEL-8.0
f29e3ac socket: convert to ipv6_dst_lookup_flow for 5.5
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
* wg-quick: linux: suppress error when finding unused table
This fixes a spurious warning messages seen with recent versions of iproute2
and kernels.
* wg-quick: linux: ensure postdown hooks execute
* wg-quick: linux: have remove_iptables return true
* wg-quick: linux: iptables-* -w is not widely supported
Adding in iptables had some hiccups. For the record, I'm very unhappy about
having to put any firewalling code into wg-quick(8). We'll of course need to
support nftables too at some point if this continues. I'm investigating with
upstream the possibility of adding a sysctl to patch the issue that iptables
is handling now, so hopefully at somepoint down the line we'll be able to shed
this dependency once again.
* send: use kfree_skb_list
* device: prepare skb_list_walk_safe for upstreaming
* send: avoid touching skb->{next,prev} directly
Suggestions from LKML.
* ipc: make sure userspace communication frees wgdevice
Free things properly on error paths.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* messages: recalculate rekey max based on a one minute flood
* allowedips: safely dereference rcu roots
* socket: remove redundant check of new4
* allowedips: avoid double lock in selftest error case
* tools: add syncconf command
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
8eb8443 version: bump snapshot
be09cf5 wg-quick: android: use Binder for setting DNS on Android 10
4716f85 noise: recompare stamps after taking write lock
54db197 netlink: allow preventing creation of new peers when updating
f1b87d1 netns: add test for failing 5.3 FIB changes
a3539c4 qemu: bump default version
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
Changes since 0.0.20190702:
define conversion constants for ancient kernels
android: refactor and add incoming allow rules
enforce that unused bits of flags are zero
immediately rekey all peers after changing device private key
support running in OpenVZ environments
do not run bc on clean target
skip peers with invalid keys
account for upstream configuration maze changes
openbsd: fix alternate routing table syntax
account for android-4.9 backport of addr_gen_mode
don't fail down when using systemd-resolved
allow specifying kernel release
enforce named pipe ownership and use protected prefix
work around ubuntu breakage
support newer PaX
don't rewrite siphash when it's from compat
squelch warnings for stack limit on broken kernel configs
support rhel/centos 7.7
Signed-off-by: Brandy Krueger <krueger.brandy24@gmail.com>
* curve25519: not all linkers support bmi2 and adx
This should allow WireGuard to build on older toolchains.
* global: switch to coarse ktime
Our prior use of fast ktime before meant that sometimes, depending on how
broken the motherboard was, we'd wind up calling into the HPET slow path. Here
we move to coarse ktime which is always super speedy. In the process we had to
fix the resolution of the clock, as well as introduce a new interface for it,
landing in 5.3. Older kernels fall back to a fast-enough mechanism based on
jiffies.
https://lore.kernel.org/lkml/tip-e3ff9c3678b4d80e22d2557b68726174578eaf52@git.kernel.org/https://lore.kernel.org/lkml/20190621203249.3909-3-Jason@zx2c4.com/
* netlink: cast struct over cb->args for type safety
This follow recent upstream changes such as:
https://lore.kernel.org/lkml/20190628144022.31376-1-Jason@zx2c4.com/
* peer: use LIST_HEAD macro
Style nit.
* receive: queue dead packets to napi queue instead of empty rx_queue
This mitigates a WARN_ON being triggered by the workqueue code. It was quite
hard to trigger, except sporadically, or reliably with a PC Engines ALIX, an
extremely slow board with an AMD LX800 that Ryan Whelan of Axatrax was kind
enough to mail me.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
There was an issue with the backport compat layer in yesterday's snapshot,
causing issues on certain (mostly Atom) Intel chips on kernels older than
4.2, due to the use of xgetbv without checking cpu flags for xsave support.
This manifested itself simply at module load time. Indeed it's somewhat tricky
to support 33 different kernel versions (3.10+), plus weird distro
frankenkernels.
If OpenWRT doesn't support < 4.2, you probably don't need to apply this.
But it also can't hurt, and probably best to stay updated.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* tools: add wincompat layer to wg(8)
Consistent with a lot of the Windows work we've been doing this last cycle,
wg(8) now supports the WireGuard for Windows app by talking through a named
pipe. You can compile this as `PLATFORM=windows make -C src/tools` with mingw.
Because programming things for Windows is pretty ugly, we've done this via a
separate standalone wincompat layer, so that we don't pollute our pretty *nix
utility.
* compat: udp_tunnel: force cast sk_data_ready
This is a hack to work around broken Android kernel wrapper scripts.
* wg-quick: freebsd: workaround SIOCGIFSTATUS race in FreeBSD kernel
FreeBSD had a number of kernel race conditions, some of which we can vaguely
work around. These are in the process of being fixed upstream, but probably
people won't update for a while.
* wg-quick: make darwin and freebsd path search strict like linux
Correctness.
* socket: set ignore_df=1 on xmit
This was intended from early on but didn't work on IPv6 without the ignore_df
flag. It allows sending fragments over IPv6.
* qemu: use newer iproute2 and kernel
* qemu: build iproute2 with libmnl support
* qemu: do not check for alignment with ubsan
The QEMU build system has been improved to compile newer versions. Linking
against libmnl gives us better error messages. As well, enabling the alignment
check on x86 UBSAN isn't realistic.
* wg-quick: look up existing routes properly
* wg-quick: specify protocol to ip(8), because of inconsistencies
The route inclusion check was wrong prior, and Linux 5.1 made it break
entirely. This makes a better invocation of `ip route show match`.
* netlink: use new strict length types in policy for 5.2
* kbuild: account for recent upstream changes
* zinc: arm64: use cpu_get_elf_hwcap accessor for 5.2
The usual churn of changes required for the upcoming 5.2.
* timers: add jitter on ack failure reinitiation
Correctness tweak in the timer system.
* blake2s,chacha: latency tweak
* blake2s: shorten ssse3 loop
In every odd-numbered round, instead of operating over the state
x00 x01 x02 x03
x05 x06 x07 x04
x10 x11 x08 x09
x15 x12 x13 x14
we operate over the rotated state
x03 x00 x01 x02
x04 x05 x06 x07
x09 x10 x11 x08
x14 x15 x12 x13
The advantage here is that this requires no changes to the 'x04 x05 x06 x07'
row, which is in the critical path. This results in a noticeable latency
improvement of roughly R cycles, for R diagonal rounds in the primitive. As
well, the blake2s AVX implementation is now SSSE3 and considerably shorter.
* tools: allow setting WG_ENDPOINT_RESOLUTION_RETRIES
System integrators can now specify things like
WG_ENDPOINT_RESOLUTION_RETRIES=infinity when building wg(8)-based init
scripts and services, or 0, or any other integer.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
With this change, the file is reduced from 5186 bytes to 4649 bytes that
its approximately 10.5 percent less memory consumption. For small
devices, sometimes every byte counts.
Also, all other protocol handler use tabs instead of spaces.
Signed-off-by: Florian Eckert <fe@dev.tdt.de>
* allowedips: initialize list head when removing intermediate nodes
Fix for an important regression in removing allowed IPs from the last
snapshot. We have new test cases to catch these in the future as well.
* tools: warn if an AllowedIP has a nonzero host part
If you try to run `wg set wg0 peer ... allowed-ips 192.168.1.82/24`, wg(8)
will now print a warning. Even though we mask this automatically down to
192.168.1.0/24, usually when people specify it like this, it's a mistake.
* wg-quick: add 'strip' subcommand
The new strip subcommand prints the config file to stdout after stripping
it of all wg-quick-specific options. This enables tricks such as:
`wg addconf $DEV <(wg-quick strip $DEV)`.
* tools: avoid unneccessary next_peer assignments in sort_peers()
Small C optimization the compiler was probably already doing.
* peerlookup: rename from hashtables
* allowedips: do not use __always_inline
* device: use skb accessor functions where possible
Suggested tweaks from Dave Miller.
* blake2s: simplify
* blake2s: remove outlen parameter from final
The blake2s implementation has been simplified, since we don't use any of the
fancy tree hashing parameters or the like. We also no longer separate the
output length at initialization time from the output length at finalization
time.
* global: the _bh variety of rcu helpers have been unified
* compat: nf_nat_core.h was removed upstream
* compat: backport skb_mark_not_on_list
The usual assortment of compat fixes for Linux 5.1.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Instead of creating host-routes depending on fwmark as (accidentally)
pushed by commit
1e8bb50b93 ("wireguard: do not add host-dependencies if fwmark is set")
use a new config option 'nohostroute' to explicitely prevent creation
of the route to the endpoint.
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
The 'fwmark' option is used to define routing traffic to
wireguard endpoints to go through specific routing tables.
In that case it doesn't make sense to setup routes for
host-dependencies in the 'main' table, so skip setting host
dependencies if 'fwmark' is set.
Signed-off-by: Daniel Golle <daniel@makrotopia.org>
* wg-quick: freebsd: allow loopback to work
FreeBSD adds a route for point-to-point destination addresses. We don't
really want to specify any destination address, but unfortunately we
have to. Before we tried to cheat by giving our own address as the
destination, but this had the unfortunate effect of preventing
loopback from working on our local ip address. We work around this with
yet another kludge: we set the destination address to 127.0.0.1. Since
127.0.0.1 is already assigned to an interface, this has the same effect
of not specifying a destination address, and therefore we accomplish the
intended behavior. Note that the bad behavior is still present in Darwin,
where such workaround does not exist.
* tools: remove unused check phony declaration
* highlighter: when subtracting char, cast to unsigned
* chacha20: name enums
* tools: fight compiler slightly harder
* tools: c_acc doesn't need to be initialized
* queueing: more reasonable allocator function convention
Usual nits.
* systemd: wg-quick should depend on nss-lookup.target
Since wg-quick(8) calls wg(8) which does hostname lookups, we should
probably only run this after we're allowed to look up hostnames.
* compat: backport ALIGN_DOWN
* noise: whiten the nanoseconds portion of the timestamp
This mitigates unrelated sidechannel attacks that think they can turn
WireGuard into a useful time oracle.
* hashtables: decouple hashtable allocations from the main device allocation
The hashtable allocations are quite large, and cause the device allocation in
the net framework to stall sometimes while it tries to find a contiguous
region that can fit the device struct. To fix the allocation stalls, decouple
the hashtable allocations from the device allocation and allocate the
hashtables with kvmalloc's implicit __GFP_NORETRY so that the allocations fall
back to vmalloc with little resistance.
* chacha20poly1305: permit unaligned strides on certain platforms
The map allocations required to fix this are mostly slower than unaligned
paths.
* noise: store clamped key instead of raw key
This causes `wg show` to now show the right thing. Useful for doing
comparisons.
* compat: ipv6_stub is sometimes null
On ancient kernels, ipv6_stub is sometimes null in cases where IPv6 has
been disabled with a command line flag or other failures.
* Makefile: don't duplicate code in install and modules-install
* Makefile: make the depmod path configurable
* queueing: net-next has changed signature of skb_probe_transport_header
A 5.1 change. This could change again, but for now it allows us to keep this
snapshot aligned with our upstream submissions.
* netlink: don't remove allowed ips for new peers
* peer: only synchronize_rcu_bh and traverse trie once when removing all peers
* allowedips: maintain per-peer list of allowedips
This is a rather big and important change that makes it much much faster to do
operations involving thousands of peers. Batch peer/allowedip addition and
clearing is several orders of magnitude faster now.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* tools: curve25519: handle unaligned loads/stores safely
This should fix sporadic crashes with `wg pubkey` on certain architectures.
* netlink: auth socket changes against namespace of socket
In WireGuard, the underlying UDP socket lives in the namespace where the
interface was created and doesn't move if the interface is moved. This
allows one to create the interface in some privileged place that has
Internet access, and then move it into a container namespace that only
has the WireGuard interface for egress. Consider the following
situation:
1. Interface created in namespace A. Socket therefore lives in namespace A.
2. Interface moved to namespace B. Socket remains in namespace A.
3. Namespace B now has access to the interface and changes the listen
port and/or fwmark of socket. Change is reflected in namespace A.
This behavior is arguably _fine_ and perhaps even expected or
acceptable. But there's also an argument to be made that B should have
A's cred to do so. So, this patch adds a simple ns_capable check.
* ratelimiter: build tests with !IPV6
Should reenable building in debug mode for systems without IPv6.
* noise: replace getnstimeofday64 with ktime_get_real_ts64
* ratelimiter: totalram_pages is now a function
* qemu: enable FP on MIPS
Linux 5.0 support.
* keygen-html: bring back pure javascript implementation
Benoît Viguier has proofs that values will stay well within 2^53. We
also have an improved carry function that's much simpler. Probably more
constant time than emscripten's 64-bit integers.
* contrib: introduce simple highlighter library
This is the highlighter library being used in:
- https://twitter.com/EdgeSecurity/status/1085294681003454465
- https://twitter.com/EdgeSecurity/status/1081953278248796165
It's included here as a contrib example, so that others can paste it into
their own GUI clients for having the same strictly validating highlighting.
* netlink: use __kernel_timespec for handshake time
This readies us for Y2038. See https://lwn.net/Articles/776435/ for more info.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* chacha20,poly1305: fix up for win64
* poly1305: only export neon symbols when in use
* poly1305: cleanup leftover debugging changes
* crypto: resolve target prefix on buggy kernels
* chacha20,poly1305: don't do compiler testing in generator and remove xor helper
* crypto: better path resolution and more specific generated .S
* poly1305: make frame pointers for auxiliary calls
* chacha20,poly1305: do not use xlate
This should fix up the various build errors, warnings, and insertion errors
introduced by the previous snapshot, where we added some significant
refactoring. In short, we're trying to port to using Andy Polyakov's original
perlasm files, and this means quite a lot of work to re-do that had stableized
in our old .S.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* Zinc no longer ships generated assembly code. Rather, we now
bundle in the original perlasm generator for it. The primary purpose
of this snapshot is to get testing of this.
* Clarify the peer removal logic and make lifetimes more precise.
* Use READ_ONCE for is_valid and is_dead.
* No need to use atomic when the recounter is mutex protected.
* Fix up macros and annotations in allowedips.
* Increment drop counter when staged packets are dropped.
* Use static constants instead of enums for 64-bit values in selftest.
* Mark large constants as ULL in poly1305-donna64.
* Fix sparse warnings in allowedips debugging code.
* Do not use wg_peer_get_maybe_zero in timer callbacks, since we now can
carefully control the lifetime of these functions and ensure they never
execute after dropping the last reference.
* Cleanup hashing in ratelimiter.
* Do not guard timer removals, since del_timer is always okay.
* We now check for PM_AUTOSLEEP, which makes the clear*on-suspend decision a
bit more general.
* Set csum_level to ~0, since the poly1305 authenticator certainly means
that no data was modified in transit.
* Use CHECKSUM_PARTIAL check for skb_checksum_help instead of
skb_checksum_setup check.
* wg.8: specify that wg(8) shows runtime info too
* wg.8: AllowedIPs isn't actually required
* keygen-html: add missing glue macro
* wg-quick: android: do not choke on empty allowed-ips
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
ba2ab5d version: bump snapshot
5f59c76 tools: wg-quick: wait for interface to disappear on freebsd
ac7e7a3 tools: don't fail if a netlink interface dump is inconsistent
8432585 main: get rid of unloaded debug message
139e57c tools: compile on gnu99
d65817c tools: use libc's endianness macro if no compiler macro
f985de2 global: give if statements brackets and other cleanups
b3a5d8a main: change module description
296d505 device: use textual error labels always
8bde328 allowedips: swap endianness early on
a650d49 timers: avoid using control statements in macro
db4dd93 allowedips: remove control statement from macro by rewriting
780a597 global: more nits
06b1236 global: rename struct wireguard_ to struct wg_
205dd46 netlink: do not stuff index into nla type
2c6b57b qemu: kill after 20 minutes
6f2953d compat: look in Kbuild and Makefile since they differ based on arch
a93d7e4 create-patch: blacklist instead of whitelist
8d53657 global: prefix functions used in callbacks with wg_
123f85c compat: don't output for grep errors
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
64750c1 version: bump snapshot
f11a2b8 global: style nits
4b34b6a crypto: clean up remaining .h->.c
06d9fc8 allowedips: document additional nobs
c32b5f9 makefile: do more generic wildcard so as to avoid rename issues
20f48d8 crypto: use BIT(i) & bitmap instead of (bitmap >> i) & 1
b6e09f6 crypto: disable broken implementations in selftests
fd50f77 compat: clang cannot handle __builtin_constant_p
bddaca7 compat: make asm/simd.h conditional on its existence
b4ba33e compat: account for ancient ARM assembler
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
* Account for big-endian 2^26 conversion in Poly1305.
* Account for big-endian NEON in Curve25519.
* Fix macros in big-endian AArch64 code so that this will actually run there
at all.
* Prefer if (IS_ENABLED(...)) over ifdef mazes when possible.
* Call simd_relax() within any preempt-disabling glue code every once in a
while so as not to increase latency if folks pass in super long buffers.
* Prefer compiler-defined architecture macros in assembly code, which puts us
in closer alignment with upstream CRYPTOGAMS code, and is cleaner.
* Non-static symbols are prefixed with wg_ to avoid polluting the global
namespace.
* Return a bool from simd_relax() indicating whether or not we were
rescheduled.
* Reflect the proper simd conditions on arm.
* Do not reorder lines in Kbuild files for the simd asm-generic addition,
since we don't want to cause merge conflicts.
* WARN() if the selftests fail in Zinc, since if this is an initcall, it won't
block module loading, so we want to be loud.
* Document some interdependencies beside include statements.
* Add missing static statement to fpu init functions.
* Use union in chacha to access state words as a flat matrix, instead of
casting a struct to a u8 and hoping all goes well. Then, by passing around
that array as a struct for as long as possible, we can update counter[0]
instead of state[12] in the generic blocks, which makes it clearer what's
happening.
* Remove __aligned(32) for chacha20_ctx since we no longer use vmovdqa on x86,
and the other implementations do not require that kind of alignment either.
* Submit patch to ARM tree for adjusting RiscPC's cflags to be -march=armv3 so
that we can build code that uses umull.
* Allow CONFIG_ARM[64] to imply [!]CONFIG_64BIT, and use zinc arch config
variables consistently throughout.
* Document rationale for the 2^26->2^64/32 conversion in code comments.
* Convert all of remaining BUG_ON to WARN_ON.
* Replace `bxeq lr` with `reteq lr` in ARM assembler to be compatible with old
ISAs via the macro in <asm/assembler.h>.
* Do not allow WireGuard to be a built-in if IPv6 is a module.
* Writeback the base register and reorder multiplications in the NEON x25519
implementation.
* Try all combinations of different implementations in selftests, so that
potential bugs are more immediately unearthed.
* Self tests and SIMD glue code work with #include, which lets the compiler
optimize these. Previously these files were .h, because they were included,
but a simple grep of the kernel tree shows 259 other files that carry out
this same pattern. Only they prefer to instead name the files with a .c
instead of a .h, so we now follow the convention.
* Support many more platforms in QEMU, especially big endian ones.
* Kernels < 3.17 don't have read_cpuid_part, so fix building there.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
33523a5 version: bump snapshot
0759480 curve25519-hacl64: reduce stack usage under KASAN
b9ab0fc chacha20: add bounds checking to selftests
2e99d19 chacha20-mips32r2: reduce stack and branches in loop, refactor jumptable handling
d6ac367 qemu: bump musl
28d8b7e crypto: make constant naming scheme consistent
56c4ea9 hchacha20: keep in native endian in words
0c3c0bc chacha20-arm: remove unused preambles
3dcd246 chacha20-arm: updated scalar code from Andy
6b9d5ca poly1305-mips64: remove useless preprocessor error
3ff3990 crypto-arm: rework KERNEL_MODE_NEON handling again
dd2f91e crypto: flatten out makefile
67a3cfb curve25519-fiat32: work around m68k compiler stack frame bug
9aa2943 allowedips: work around kasan stack frame bug in selftest
317b318 chacha20-arm: use new scalar implementation
b715e3b crypto-arm: rework KERNEL_MODE_NEON handling
77b07d9 global: reduce stack frame size
ddc2bd6 chacha20: add chunked selftest and test sliding alignments and hchacha20
2eead02 chacha20-mips32r2: reduce jumptable entry size and stack usage
a0ac620 chacha20-mips32r2: use simpler calling convention
09247c0 chacha20-arm: go with Ard's version to optimize for Cortex-A7
a329e0a chacha20-mips32r2: remove reorder directives
3b22533 chacha20-mips32r2: fix typo to allow reorder again
d4ac6bb poly1305-mips32r2: remove all reorder directives
197a30c global: put SPDX identifier on its own line
305806d ratelimiter: disable selftest with KASAN
4e06236 crypto: do not waste space on selftest items
5e0fd08 netlink: reverse my christmas trees
a61ea8b crypto: explicitly dual license
b161aff poly1305: account for simd being toggled off midway
470a0c5 allowedips: change from BUG_ON to WARN_ON
aa9e090 chacha20: prefer crypto_xor_cpy to avoid memmove
1b0adf5 poly1305: no need to trick gcc 8.1
a849803 blake2s: simplify final function
073f3d1 poly1305: better module description
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
* blake2s-x86_64: fix whitespace errors
* crypto: do not use compound literals in selftests
* crypto: make sure UML is properly disabled
* kconfig: make NEON depend on CPU_V7
* poly1305: rename finish to final
* chacha20: add constant for words in block
* curve25519-x86_64: remove useless define
* poly1305: precompute 5*r in init instead of blocks
* chacha20-arm: swap scalar and neon functions
* simd: add __must_check annotation
* poly1305: do not require simd context for arch
* chacha20-x86_64: cascade down implementations
* crypto: pass simd by reference
* chacha20-x86_64: don't activate simd for small blocks
* poly1305-x86_64: don't activate simd for small blocks
* crypto: do not use -include trick
* crypto: turn Zinc into individual modules
* chacha20poly1305: relax simd between sg chunks
* chacha20-x86_64: more limited cascade
* crypto: allow for disabling simd in zinc modules
* poly1305-x86_64: show full struct for state
* chacha20-x86_64: use correct cut off for avx512-vl
* curve25519-arm: only compile if symbols will be used
* chacha20poly1305: add __init to selftest helper functions
* chacha20: add independent self test
Tons of improvements all around the board to our cryptography library,
including some performance boosts with how we handle SIMD for small packets.
* send/receive: reduce number of sg entries
This quells a powerpc stack usage warning.
* global: remove non-essential inline annotations
We now allow the compiler to determine whether or not to inline certain
functions, while still manually choosing so for a few performance-critical
sections.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* curve25519: arm: do not modify sp directly
* compat: support neon.h on old kernels
* compat: arch-namespace certain includes
* compat: move simd.h from crypto to compat since it's going upstream
This fixes a decent amount of compat breakage and thumb2-mode breakage
introduced by our move to Zinc.
* crypto: use CRYPTOGAMS license
Rather than using code from OpenSSL, use code directly from AndyP.
* poly1305: rewrite self tests from scratch
* poly1305: switch to donna
This makes our C Poly1305 implementation a bit more intensely tested and also
faster, especially on 64-bit systems. It also sets the stage for moving to a
HACL* implementation when that's ready.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* Kconfig: use new-style help marker
* global: run through clang-format
* uapi: reformat
* global: satisfy check_patch.pl errors
* global: prefer sizeof(*pointer) when possible
* global: always find OOM unlikely
Tons of style cleanups.
* crypto: use unaligned helpers
We now avoid unaligned accesses for generic users of the crypto API.
* crypto: import zinc
More style cleanups and a rearrangement of the crypto routines to fit how this
is going to work upstream. This required some fairly big changes to our build
system, so there may be some build errors we'll have to address in subsequent
snapshots.
* compat: rng_is_initialized made it into 4.19
We therefore don't need it in the compat layer anymore.
* curve25519-hacl64: use formally verified C for comparisons
The previous code had been proved in Z3, but this new code from upstream
KreMLin is directly generated from the F*, which is preferable. The
assembly generated is identical.
* curve25519-x86_64: let the compiler decide when/how to load constants
Small performance boost.
* curve25519-arm: reformat
* curve25519-arm: cleanups from lkml
* curve25519-arm: add spaces after commas
* curve25519-arm: use ordinary prolog and epilogue
* curve25519-arm: do not waste 32 bytes of stack
* curve25519-arm: prefix immediates with #
This incorporates ASM nits from upstream review.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* send: switch handshake stamp to an atomic
Rather than abusing the handshake lock, we're much better off just using
a boring atomic64 for this. It's simpler and performs better. Also, while
we're at it, we set the handshake stamp both before and after the
calculations, in case the calculations block for a really long time waiting
for the RNG to initialize.
* compat: better atomic acquire/release backport
This should fix compilation and correctness on several platforms.
* crypto: move simd context to specific type
This was a suggestion from Andy Lutomirski on LKML.
* chacha20poly1305: selftest: use arrays for test vectors
We no longer have lines so long that they're rejected by SMTP servers.
* qemu: add easy git harness
This makes it a bit easier to use our qemu harness for testing our mainline
integration tree.
* curve25519-x86_64: avoid use of r12
This causes problems with RAP and KERNEXEC for PaX, as r12 is a
reserved register.
* chacha20: use memmove in case buffers overlap
A small correctness fix that we never actually hit in WireGuard but is
important especially for moving this into a general purpose library.
* curve25519-hacl64: simplify u64_eq_mask
* curve25519-hacl64: correct u64_gte_mask
Two bitmath fixes from Samuel, which come complete with a z3 script proving
their correctness.
* timers: include header in right file
This fixes compilation in some environments.
* netlink: don't start over iteration on multipart non-first allowedips
Matt Layher found a bug where a netlink dump of peers would never terminate in
some circumstances, causing wg(8) to keep trying forever. We now have a fix as
well as a unit test to mitigate this, and we'll be looking to create a fuzzer
out of Matt's nice library.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Changelog taken from the version announcement
> == Changes ==
>
> * chacha20poly1305: selftest: split up test vector constants
>
> The test vectors are encoded as long strings -- really long strings -- and
> apparently RFC821 doesn't like lines longer than 998.
> https://cr.yp.to/smtp/message.html
>
> * queueing: keep reference to peer after setting atomic state bit
>
> This fixes a regression introduced when preparing the LKML submission.
>
> * allowedips: prevent double read in kref
> * allowedips: avoid window of disappeared peer
> * hashtables: document immediate zeroing semantics
> * peer: ensure resources are freed when creation fails
> * queueing: document double-adding and reference conditions
> * queueing: ensure strictly ordered loads and stores
> * cookie: returned keypair might disappear if rcu lock not held
> * noise: free peer references on failure
> * peer: ensure destruction doesn't race
>
> Various fixes, as well as lots of code comment documentation, for a
> small variety of the less obvious aspects of object lifecycles,
> focused on correctness.
>
> * allowedips: free root inside of RCU callback
> * allowedips: use different macro names so as to avoid confusion
>
> These incorporate two suggestions from LKML.
>
> This snapshot contains commits from: Jason A. Donenfeld and Jann Horn.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Signed-off-by: Yousong Zhou <yszhou4tech@gmail.com>
This watchdog script tries to re-resolve hostnames for inactive WireGuard peers.
Use it for peers with a frequently changing dynamic IP.
persistent_keepalive must be set, recommended value is 25 seconds.
Run this script from cron every minute:
echo '* * * * * /usr/bin/wireguard_watchdog' >> /etc/crontabs/root
Signed-off-by: Aleksandr V. Piskunov <aleksandr.v.piskunov@gmail.com>
[bump the package release]
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
80b41cd version: bump snapshot
fe5f0f6 recieve: disable NAPI busy polling
e863f40 device: destroy workqueue before freeing queue
81a2e7e wg-quick: allow link local default gateway
95951af receive: use gro call instead of plain call
d9501f1 receive: account for zero or negative budget
e80799b tools: only error on wg show if all interfaces failk
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
[Added commit log to commit description]
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
* device: print daddr not saddr in missing peer error
* receive: style
Debug messages now make sense again.
* wg-quick: android: support excluding applications
Android now supports excluding certain apps (uids) from the tunnel.
* selftest: ratelimiter: improve chance of success via retry
* qemu: bump default kernel version
* qemu: decide debug kernel based on KERNEL_VERSION
Some improvements to our testing infrastructure.
* receive: use NAPI on the receive path
This is a big change that should both improve preemption latency (by not
disabling it unconditionally) and vastly improve rx performance on most
systems by using NAPI. The main purpose of this snapshot is to test out this
technique.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
dfd9827 version: bump snapshot
88729f0 wg-quick: android: prevent outgoing handshake packets from being dropped
1bb9daf compat: more robust ktime backport
68441fb global: use fast boottime instead of normal boottime
d0bd6dc global: use ktime boottime instead of jiffies
18822b8 tools: fix misspelling of strchrnul in comment
0f8718b manpages: eliminate whitespace at the end of the line
590c410 global: fix a few typos
bb76804 simd: add missing header
7e88174 poly1305: give linker the correct constant data section size
fd8dfd3 main: test poly1305 before chacha20poly1305
c754c59 receive: don't toggle bh
Compile-tested-for: ath79 Archer C7 v2
Run-tested-on: ath79 Archer C7 v2
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
0bc4230 version: bump snapshot
ed04799 poly1305: add missing string.h header
cbd4e34 compat: use stabler lkml links
caa718c ratelimiter: do not allow concurrent init and uninit
894ddae ratelimiter: mitigate reference underflow
0a8a62c receive: drop handshake packets if rng is not initialized
cad9e52 noise: wait for crng before taking locks
83c0690 netlink: maintain static_identity lock over entire private key update
0913f1c noise: take locks for ss precomputation
073f31a qemu: bump default kernel
bec4c48 wg-quick: android: don't forget to free compiled regexes
7ce2ef3 wg-quick: android: disable roaming to v6 networks when v4 is specified
9132be4 dns-hatchet: apply resolv.conf's selinux context to new resolv.conf
41a5747 simd: no need to restore fpu state when no preemption
6d7f0b0 simd: encapsulate fpu amortization into nice functions
f8b57d5 queueing: re-enable preemption periodically to lower latency
b7b193f queueing: remove useless spinlocks on sc
5bb62fe tools: getentropy requires 10.12
4e9f120 chacha20poly1305: use slow crypto on -rt kernels on arm too
Compiled-for: ar71xx, lantiq
Run-tested-on: ar71xx Archer C7 v2 & lantiq HH5a
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
This version bump was made upstream mostly for OpenWRT, and should fix
an issue with a null dst when on the flow offloading path.
While we're at it, Kevin and I are the only people actually taking care
of this package, so trim the maintainer list a bit.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
* chacha20poly1305: add mips32 implementation
"The OpenWRT Commit" - this significantly speeds up performance on cheap
plastic MIPS routers, and presumably the remaining MIPS32r2 super computers
out there.
* timers: reinitialize state on init
* timers: round up instead of down in slack_time
* timers: remove slack_time
* timers: clear send_keepalive timer on sending handshake response
* timers: no need to clear keepalive in persistent keepalive
Andrew He and I have helped simplify the timers and remove some old warts,
making the whole system a bit easier to analyze.
* tools: fix errno propagation and messages
Error messages are now more coherent.
* device: remove allowedips before individual peers
This avoids an O(n^2) traversal in favor of an O(n) one. Before systems with
many peers would grind when deleting the interface.
Signed-off-by: Jason A. Donenfeld <Jason@zx2c4.com>
Drop package/network/services/wireguard/patches/100-portability.patch
Instead pass 'PLATFORM=linux' to make since we are always building FOR
linux.
Signed-off-by: Kevin Darbyshire-Bryant <ldir@darbyshire-bryant.me.uk>
This makes it easier to distribute prefixes over a wireguard tunnel
interface, by simply setting the ip6prefix option in uci (just like with
other protocols).
Obviously, routing etc needs to be setup properly for things to work; this
just adds the config option so the prefix can be assigned to other
interfaces.
Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk>
Jason A. Donenfeld
Kevin Darbyshire-Bryant
Brandy Krueger
Florian Eckert
Daniel Golle
Daniel Engberg
Aleksandr V. Piskunov
Toke Høiland-Jørgensen
