You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
tailscale/net/tstun/wrap.go

747 lines
23 KiB
Go

wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
// Copyright (c) 2020 Tailscale Inc & AUTHORS All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
// Package tstun provides a TUN struct implementing the tun.Device interface
// with additional features as required by wgengine.
package tstun
import (
"errors"
"fmt"
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
"io"
"os"
"strings"
"sync"
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
"sync/atomic"
"time"
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
"go4.org/mem"
"golang.zx2c4.com/wireguard/device"
"golang.zx2c4.com/wireguard/tun"
net/packet: remove the custom IP4/IP6 types in favor of netaddr.IP. Upstream netaddr has a change that makes it alloc-free, so it's safe to use in hot codepaths. This gets rid of one of the many IP types in our codebase. Performance is currently worse across the board. This is likely due in part to netaddr.IP being a larger value type (4b -> 24b for IPv4, 16b -> 24b for IPv6), and in other part due to missing low-hanging fruit optimizations in netaddr. However, the regression is less bad than it looks at first glance, because we'd micro-optimized packet.IP* in the past few weeks. This change drops us back to roughly where we were at the 1.2 release, but with the benefit of a significant code and architectural simplification. name old time/op new time/op delta pkg:tailscale.com/net/packet goos:linux goarch:amd64 Decode/tcp4-8 12.2ns ± 5% 29.7ns ± 2% +142.32% (p=0.008 n=5+5) Decode/tcp6-8 12.6ns ± 3% 65.1ns ± 2% +418.47% (p=0.008 n=5+5) Decode/udp4-8 11.8ns ± 3% 30.5ns ± 2% +157.94% (p=0.008 n=5+5) Decode/udp6-8 27.1ns ± 1% 65.7ns ± 2% +142.36% (p=0.016 n=4+5) Decode/icmp4-8 24.6ns ± 2% 30.5ns ± 2% +23.65% (p=0.016 n=4+5) Decode/icmp6-8 22.9ns ±51% 65.5ns ± 2% +186.19% (p=0.008 n=5+5) Decode/igmp-8 18.1ns ±44% 30.2ns ± 1% +66.89% (p=0.008 n=5+5) Decode/unknown-8 20.8ns ± 1% 10.6ns ± 9% -49.11% (p=0.016 n=4+5) pkg:tailscale.com/wgengine/filter goos:linux goarch:amd64 Filter/icmp4-8 30.5ns ± 1% 77.9ns ± 3% +155.01% (p=0.008 n=5+5) Filter/tcp4_syn_in-8 43.7ns ± 3% 123.0ns ± 3% +181.72% (p=0.008 n=5+5) Filter/tcp4_syn_out-8 24.5ns ± 2% 45.7ns ± 6% +86.22% (p=0.008 n=5+5) Filter/udp4_in-8 64.8ns ± 1% 210.0ns ± 2% +223.87% (p=0.008 n=5+5) Filter/udp4_out-8 119ns ± 0% 278ns ± 0% +133.78% (p=0.016 n=4+5) Filter/icmp6-8 40.3ns ± 2% 204.4ns ± 4% +407.70% (p=0.008 n=5+5) Filter/tcp6_syn_in-8 35.3ns ± 3% 199.2ns ± 2% +464.95% (p=0.008 n=5+5) Filter/tcp6_syn_out-8 32.8ns ± 2% 81.0ns ± 2% +147.10% (p=0.008 n=5+5) Filter/udp6_in-8 106ns ± 2% 290ns ± 2% +174.48% (p=0.008 n=5+5) Filter/udp6_out-8 184ns ± 2% 314ns ± 3% +70.43% (p=0.016 n=4+5) pkg:tailscale.com/wgengine/tstun goos:linux goarch:amd64 Write-8 9.02ns ± 3% 8.92ns ± 1% ~ (p=0.421 n=5+5) name old alloc/op new alloc/op delta pkg:tailscale.com/net/packet goos:linux goarch:amd64 Decode/tcp4-8 0.00B 0.00B ~ (all equal) Decode/tcp6-8 0.00B 0.00B ~ (all equal) Decode/udp4-8 0.00B 0.00B ~ (all equal) Decode/udp6-8 0.00B 0.00B ~ (all equal) Decode/icmp4-8 0.00B 0.00B ~ (all equal) Decode/icmp6-8 0.00B 0.00B ~ (all equal) Decode/igmp-8 0.00B 0.00B ~ (all equal) Decode/unknown-8 0.00B 0.00B ~ (all equal) pkg:tailscale.com/wgengine/filter goos:linux goarch:amd64 Filter/icmp4-8 0.00B 0.00B ~ (all equal) Filter/tcp4_syn_in-8 0.00B 0.00B ~ (all equal) Filter/tcp4_syn_out-8 0.00B 0.00B ~ (all equal) Filter/udp4_in-8 0.00B 0.00B ~ (all equal) Filter/udp4_out-8 16.0B ± 0% 64.0B ± 0% +300.00% (p=0.008 n=5+5) Filter/icmp6-8 0.00B 0.00B ~ (all equal) Filter/tcp6_syn_in-8 0.00B 0.00B ~ (all equal) Filter/tcp6_syn_out-8 0.00B 0.00B ~ (all equal) Filter/udp6_in-8 0.00B 0.00B ~ (all equal) Filter/udp6_out-8 48.0B ± 0% 64.0B ± 0% +33.33% (p=0.008 n=5+5) name old allocs/op new allocs/op delta pkg:tailscale.com/net/packet goos:linux goarch:amd64 Decode/tcp4-8 0.00 0.00 ~ (all equal) Decode/tcp6-8 0.00 0.00 ~ (all equal) Decode/udp4-8 0.00 0.00 ~ (all equal) Decode/udp6-8 0.00 0.00 ~ (all equal) Decode/icmp4-8 0.00 0.00 ~ (all equal) Decode/icmp6-8 0.00 0.00 ~ (all equal) Decode/igmp-8 0.00 0.00 ~ (all equal) Decode/unknown-8 0.00 0.00 ~ (all equal) pkg:tailscale.com/wgengine/filter goos:linux goarch:amd64 Filter/icmp4-8 0.00 0.00 ~ (all equal) Filter/tcp4_syn_in-8 0.00 0.00 ~ (all equal) Filter/tcp4_syn_out-8 0.00 0.00 ~ (all equal) Filter/udp4_in-8 0.00 0.00 ~ (all equal) Filter/udp4_out-8 1.00 ± 0% 1.00 ± 0% ~ (all equal) Filter/icmp6-8 0.00 0.00 ~ (all equal) Filter/tcp6_syn_in-8 0.00 0.00 ~ (all equal) Filter/tcp6_syn_out-8 0.00 0.00 ~ (all equal) Filter/udp6_in-8 0.00 0.00 ~ (all equal) Filter/udp6_out-8 1.00 ± 0% 1.00 ± 0% ~ (all equal) Signed-off-by: David Anderson <danderson@tailscale.com>
4 years ago
"inet.af/netaddr"
"tailscale.com/disco"
"tailscale.com/net/packet"
"tailscale.com/tstime/mono"
"tailscale.com/types/ipproto"
"tailscale.com/types/key"
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
"tailscale.com/types/logger"
"tailscale.com/types/pad32"
"tailscale.com/util/clientmetric"
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
"tailscale.com/wgengine/filter"
)
const maxBufferSize = device.MaxMessageSize
// PacketStartOffset is the minimal amount of leading space that must exist
// before &packet[offset] in a packet passed to Read, Write, or InjectInboundDirect.
// This is necessary to avoid reallocation in wireguard-go internals.
const PacketStartOffset = device.MessageTransportHeaderSize
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
// MaxPacketSize is the maximum size (in bytes)
// of a packet that can be injected into a tstun.Wrapper.
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
const MaxPacketSize = device.MaxContentSize
const tapDebug = false // for super verbose TAP debugging
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
var (
// ErrClosed is returned when attempting an operation on a closed Wrapper.
ErrClosed = errors.New("device closed")
// ErrFiltered is returned when the acted-on packet is rejected by a filter.
ErrFiltered = errors.New("packet dropped by filter")
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
)
var (
errPacketTooBig = errors.New("packet too big")
errOffsetTooBig = errors.New("offset larger than buffer length")
errOffsetTooSmall = errors.New("offset smaller than PacketStartOffset")
)
// parsedPacketPool holds a pool of Parsed structs for use in filtering.
// This is needed because escape analysis cannot see that parsed packets
// do not escape through {Pre,Post}Filter{In,Out}.
var parsedPacketPool = sync.Pool{New: func() interface{} { return new(packet.Parsed) }}
// FilterFunc is a packet-filtering function with access to the Wrapper device.
// It must not hold onto the packet struct, as its backing storage will be reused.
type FilterFunc func(*packet.Parsed, *Wrapper) filter.Response
// Wrapper augments a tun.Device with packet filtering and injection.
type Wrapper struct {
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
logf logger.Logf
// tdev is the underlying Wrapper device.
tdev tun.Device
isTAP bool // whether tdev is a TAP device
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
closeOnce sync.Once
_ pad32.Four
lastActivityAtomic mono.Time // time of last send or receive
net/packet: remove the custom IP4/IP6 types in favor of netaddr.IP. Upstream netaddr has a change that makes it alloc-free, so it's safe to use in hot codepaths. This gets rid of one of the many IP types in our codebase. Performance is currently worse across the board. This is likely due in part to netaddr.IP being a larger value type (4b -> 24b for IPv4, 16b -> 24b for IPv6), and in other part due to missing low-hanging fruit optimizations in netaddr. However, the regression is less bad than it looks at first glance, because we'd micro-optimized packet.IP* in the past few weeks. This change drops us back to roughly where we were at the 1.2 release, but with the benefit of a significant code and architectural simplification. name old time/op new time/op delta pkg:tailscale.com/net/packet goos:linux goarch:amd64 Decode/tcp4-8 12.2ns ± 5% 29.7ns ± 2% +142.32% (p=0.008 n=5+5) Decode/tcp6-8 12.6ns ± 3% 65.1ns ± 2% +418.47% (p=0.008 n=5+5) Decode/udp4-8 11.8ns ± 3% 30.5ns ± 2% +157.94% (p=0.008 n=5+5) Decode/udp6-8 27.1ns ± 1% 65.7ns ± 2% +142.36% (p=0.016 n=4+5) Decode/icmp4-8 24.6ns ± 2% 30.5ns ± 2% +23.65% (p=0.016 n=4+5) Decode/icmp6-8 22.9ns ±51% 65.5ns ± 2% +186.19% (p=0.008 n=5+5) Decode/igmp-8 18.1ns ±44% 30.2ns ± 1% +66.89% (p=0.008 n=5+5) Decode/unknown-8 20.8ns ± 1% 10.6ns ± 9% -49.11% (p=0.016 n=4+5) pkg:tailscale.com/wgengine/filter goos:linux goarch:amd64 Filter/icmp4-8 30.5ns ± 1% 77.9ns ± 3% +155.01% (p=0.008 n=5+5) Filter/tcp4_syn_in-8 43.7ns ± 3% 123.0ns ± 3% +181.72% (p=0.008 n=5+5) Filter/tcp4_syn_out-8 24.5ns ± 2% 45.7ns ± 6% +86.22% (p=0.008 n=5+5) Filter/udp4_in-8 64.8ns ± 1% 210.0ns ± 2% +223.87% (p=0.008 n=5+5) Filter/udp4_out-8 119ns ± 0% 278ns ± 0% +133.78% (p=0.016 n=4+5) Filter/icmp6-8 40.3ns ± 2% 204.4ns ± 4% +407.70% (p=0.008 n=5+5) Filter/tcp6_syn_in-8 35.3ns ± 3% 199.2ns ± 2% +464.95% (p=0.008 n=5+5) Filter/tcp6_syn_out-8 32.8ns ± 2% 81.0ns ± 2% +147.10% (p=0.008 n=5+5) Filter/udp6_in-8 106ns ± 2% 290ns ± 2% +174.48% (p=0.008 n=5+5) Filter/udp6_out-8 184ns ± 2% 314ns ± 3% +70.43% (p=0.016 n=4+5) pkg:tailscale.com/wgengine/tstun goos:linux goarch:amd64 Write-8 9.02ns ± 3% 8.92ns ± 1% ~ (p=0.421 n=5+5) name old alloc/op new alloc/op delta pkg:tailscale.com/net/packet goos:linux goarch:amd64 Decode/tcp4-8 0.00B 0.00B ~ (all equal) Decode/tcp6-8 0.00B 0.00B ~ (all equal) Decode/udp4-8 0.00B 0.00B ~ (all equal) Decode/udp6-8 0.00B 0.00B ~ (all equal) Decode/icmp4-8 0.00B 0.00B ~ (all equal) Decode/icmp6-8 0.00B 0.00B ~ (all equal) Decode/igmp-8 0.00B 0.00B ~ (all equal) Decode/unknown-8 0.00B 0.00B ~ (all equal) pkg:tailscale.com/wgengine/filter goos:linux goarch:amd64 Filter/icmp4-8 0.00B 0.00B ~ (all equal) Filter/tcp4_syn_in-8 0.00B 0.00B ~ (all equal) Filter/tcp4_syn_out-8 0.00B 0.00B ~ (all equal) Filter/udp4_in-8 0.00B 0.00B ~ (all equal) Filter/udp4_out-8 16.0B ± 0% 64.0B ± 0% +300.00% (p=0.008 n=5+5) Filter/icmp6-8 0.00B 0.00B ~ (all equal) Filter/tcp6_syn_in-8 0.00B 0.00B ~ (all equal) Filter/tcp6_syn_out-8 0.00B 0.00B ~ (all equal) Filter/udp6_in-8 0.00B 0.00B ~ (all equal) Filter/udp6_out-8 48.0B ± 0% 64.0B ± 0% +33.33% (p=0.008 n=5+5) name old allocs/op new allocs/op delta pkg:tailscale.com/net/packet goos:linux goarch:amd64 Decode/tcp4-8 0.00 0.00 ~ (all equal) Decode/tcp6-8 0.00 0.00 ~ (all equal) Decode/udp4-8 0.00 0.00 ~ (all equal) Decode/udp6-8 0.00 0.00 ~ (all equal) Decode/icmp4-8 0.00 0.00 ~ (all equal) Decode/icmp6-8 0.00 0.00 ~ (all equal) Decode/igmp-8 0.00 0.00 ~ (all equal) Decode/unknown-8 0.00 0.00 ~ (all equal) pkg:tailscale.com/wgengine/filter goos:linux goarch:amd64 Filter/icmp4-8 0.00 0.00 ~ (all equal) Filter/tcp4_syn_in-8 0.00 0.00 ~ (all equal) Filter/tcp4_syn_out-8 0.00 0.00 ~ (all equal) Filter/udp4_in-8 0.00 0.00 ~ (all equal) Filter/udp4_out-8 1.00 ± 0% 1.00 ± 0% ~ (all equal) Filter/icmp6-8 0.00 0.00 ~ (all equal) Filter/tcp6_syn_in-8 0.00 0.00 ~ (all equal) Filter/tcp6_syn_out-8 0.00 0.00 ~ (all equal) Filter/udp6_in-8 0.00 0.00 ~ (all equal) Filter/udp6_out-8 1.00 ± 0% 1.00 ± 0% ~ (all equal) Signed-off-by: David Anderson <danderson@tailscale.com>
4 years ago
destIPActivity atomic.Value // of map[netaddr.IP]func()
destMACAtomic atomic.Value // of [6]byte
discoKey atomic.Value // of key.DiscoPublic
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
// buffer stores the oldest unconsumed packet from tdev.
// It is made a static buffer in order to avoid allocations.
buffer [maxBufferSize]byte
// bufferConsumedMu protects bufferConsumed from concurrent sends and closes.
// It does not prevent send-after-close, only data races.
bufferConsumedMu sync.Mutex
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
// bufferConsumed synchronizes access to buffer (shared by Read and poll).
//
// Close closes bufferConsumed. There may be outstanding sends to bufferConsumed
// when that happens; we catch any resulting panics.
// This lets us avoid expensive multi-case selects.
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
bufferConsumed chan struct{}
// closed signals poll (by closing) when the device is closed.
closed chan struct{}
// outboundMu protects outbound from concurrent sends and closes.
// It does not prevent send-after-close, only data races.
outboundMu sync.Mutex
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
// outbound is the queue by which packets leave the TUN device.
//
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
// The directions are relative to the network, not the device:
// inbound packets arrive via UDP and are written into the TUN device;
// outbound packets are read from the TUN device and sent out via UDP.
// This queue is needed because although inbound writes are synchronous,
// the other direction must wait on a Wireguard goroutine to poll it.
//
// Empty reads are skipped by Wireguard, so it is always legal
// to discard an empty packet instead of sending it through t.outbound.
//
// Close closes outbound. There may be outstanding sends to outbound
// when that happens; we catch any resulting panics.
// This lets us avoid expensive multi-case selects.
outbound chan tunReadResult
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
// eventsUpDown yields up and down tun.Events that arrive on a Wrapper's events channel.
eventsUpDown chan tun.Event
// eventsOther yields non-up-and-down tun.Events that arrive on a Wrapper's events channel.
eventsOther chan tun.Event
wgengine/bench: speed test for channels, sockets, and wireguard-go. This tries to generate traffic at a rate that will saturate the receiver, without overdoing it, even in the event of packet loss. It's unrealistically more aggressive than TCP (which will back off quickly in case of packet loss) but less silly than a blind test that just generates packets as fast as it can (which can cause all the CPU to be absorbed by the transmitter, giving an incorrect impression of how much capacity the total system has). Initial indications are that a syscall about every 10 packets (TCP bulk delivery) is roughly the same speed as sending every packet through a channel. A syscall per packet is about 5x-10x slower than that. The whole tailscale wireguard-go + magicsock + packet filter combination is about 4x slower again, which is better than I thought we'd do, but probably has room for improvement. Note that in "full" tailscale, there is also a tundev read/write for every packet, effectively doubling the syscall overhead per packet. Given these numbers, it seems like read/write syscalls are only 25-40% of the total CPU time used in tailscale proper, so we do have significant non-syscall optimization work to do too. Sample output: $ GOMAXPROCS=2 go test -bench . -benchtime 5s ./cmd/tailbench goos: linux goarch: amd64 pkg: tailscale.com/cmd/tailbench cpu: Intel(R) Core(TM) i7-4785T CPU @ 2.20GHz BenchmarkTrivialNoAlloc/32-2 56340248 93.85 ns/op 340.98 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/124-2 57527490 99.27 ns/op 1249.10 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivialNoAlloc/1024-2 52537773 111.3 ns/op 9200.39 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/32-2 41878063 135.6 ns/op 236.04 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/124-2 41270439 138.4 ns/op 896.02 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTrivial/1024-2 36337252 154.3 ns/op 6635.30 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkBlockingChannel/32-2 12171654 494.3 ns/op 64.74 MB/s 0 %lost 1791 B/op 0 allocs/op BenchmarkBlockingChannel/124-2 12149956 507.8 ns/op 244.17 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkBlockingChannel/1024-2 11034754 528.8 ns/op 1936.42 MB/s 0 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/32-2 8960622 2195 ns/op 14.58 MB/s 8.825 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/124-2 3014614 2224 ns/op 55.75 MB/s 11.18 %lost 1792 B/op 1 allocs/op BenchmarkNonlockingChannel/1024-2 3234915 1688 ns/op 606.53 MB/s 3.765 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/32-2 8457559 764.1 ns/op 41.88 MB/s 5.945 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/124-2 5497726 1030 ns/op 120.38 MB/s 12.14 %lost 1792 B/op 1 allocs/op BenchmarkDoubleChannel/1024-2 7985656 1360 ns/op 752.86 MB/s 13.57 %lost 1792 B/op 1 allocs/op BenchmarkUDP/32-2 1652134 3695 ns/op 8.66 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/124-2 1621024 3765 ns/op 32.94 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkUDP/1024-2 1553750 3825 ns/op 267.72 MB/s 0 %lost 176 B/op 3 allocs/op BenchmarkTCP/32-2 11056336 503.2 ns/op 63.60 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/124-2 11074869 533.7 ns/op 232.32 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkTCP/1024-2 8934968 671.4 ns/op 1525.20 MB/s 0 %lost 0 B/op 0 allocs/op BenchmarkWireGuardTest/32-2 1403702 4547 ns/op 7.04 MB/s 14.37 %lost 467 B/op 3 allocs/op BenchmarkWireGuardTest/124-2 780645 7927 ns/op 15.64 MB/s 1.537 %lost 420 B/op 3 allocs/op BenchmarkWireGuardTest/1024-2 512671 11791 ns/op 86.85 MB/s 0.5206 %lost 411 B/op 3 allocs/op PASS ok tailscale.com/wgengine/bench 195.724s Updates #414. Signed-off-by: Avery Pennarun <apenwarr@tailscale.com>
4 years ago
// filter atomically stores the currently active packet filter
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
filter atomic.Value // of *filter.Filter
// filterFlags control the verbosity of logging packet drops/accepts.
filterFlags filter.RunFlags
// PreFilterIn is the inbound filter function that runs before the main filter
// and therefore sees the packets that may be later dropped by it.
PreFilterIn FilterFunc
// PostFilterIn is the inbound filter function that runs after the main filter.
PostFilterIn FilterFunc
// PreFilterOut is the outbound filter function that runs before the main filter
// and therefore sees the packets that may be later dropped by it.
PreFilterOut FilterFunc
// PostFilterOut is the outbound filter function that runs after the main filter.
PostFilterOut FilterFunc
// OnTSMPPongReceived, if non-nil, is called whenever a TSMP pong arrives.
OnTSMPPongReceived func(packet.TSMPPongReply)
// PeerAPIPort, if non-nil, returns the peerapi port that's
// running for the given IP address.
PeerAPIPort func(netaddr.IP) (port uint16, ok bool)
// disableFilter disables all filtering when set. This should only be used in tests.
disableFilter bool
// disableTSMPRejected disables TSMP rejected responses. For tests.
disableTSMPRejected bool
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
}
// tunReadResult is the result of a TUN read: Some data and an error.
// The byte slice is not interpreted in the usual way for a Read method.
// See the comment in the middle of Wrap.Read.
type tunReadResult struct {
data []byte
err error
}
func WrapTAP(logf logger.Logf, tdev tun.Device) *Wrapper {
return wrap(logf, tdev, true)
}
func Wrap(logf logger.Logf, tdev tun.Device) *Wrapper {
return wrap(logf, tdev, false)
}
func wrap(logf logger.Logf, tdev tun.Device, isTAP bool) *Wrapper {
tun := &Wrapper{
logf: logger.WithPrefix(logf, "tstun: "),
isTAP: isTAP,
tdev: tdev,
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
// bufferConsumed is conceptually a condition variable:
// a goroutine should not block when setting it, even with no listeners.
bufferConsumed: make(chan struct{}, 1),
closed: make(chan struct{}),
// outbound can be unbuffered; the buffer is an optimization.
outbound: make(chan tunReadResult, 1),
eventsUpDown: make(chan tun.Event),
eventsOther: make(chan tun.Event),
// TODO(dmytro): (highly rate-limited) hexdumps should happen on unknown packets.
filterFlags: filter.LogAccepts | filter.LogDrops,
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
}
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
go tun.poll()
go tun.pumpEvents()
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
// The buffer starts out consumed.
tun.bufferConsumed <- struct{}{}
tun.noteActivity()
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
return tun
}
// SetDestIPActivityFuncs sets a map of funcs to run per packet
// destination (the map keys).
//
// The map ownership passes to the Wrapper. It must be non-nil.
func (t *Wrapper) SetDestIPActivityFuncs(m map[netaddr.IP]func()) {
net/packet: remove the custom IP4/IP6 types in favor of netaddr.IP. Upstream netaddr has a change that makes it alloc-free, so it's safe to use in hot codepaths. This gets rid of one of the many IP types in our codebase. Performance is currently worse across the board. This is likely due in part to netaddr.IP being a larger value type (4b -> 24b for IPv4, 16b -> 24b for IPv6), and in other part due to missing low-hanging fruit optimizations in netaddr. However, the regression is less bad than it looks at first glance, because we'd micro-optimized packet.IP* in the past few weeks. This change drops us back to roughly where we were at the 1.2 release, but with the benefit of a significant code and architectural simplification. name old time/op new time/op delta pkg:tailscale.com/net/packet goos:linux goarch:amd64 Decode/tcp4-8 12.2ns ± 5% 29.7ns ± 2% +142.32% (p=0.008 n=5+5) Decode/tcp6-8 12.6ns ± 3% 65.1ns ± 2% +418.47% (p=0.008 n=5+5) Decode/udp4-8 11.8ns ± 3% 30.5ns ± 2% +157.94% (p=0.008 n=5+5) Decode/udp6-8 27.1ns ± 1% 65.7ns ± 2% +142.36% (p=0.016 n=4+5) Decode/icmp4-8 24.6ns ± 2% 30.5ns ± 2% +23.65% (p=0.016 n=4+5) Decode/icmp6-8 22.9ns ±51% 65.5ns ± 2% +186.19% (p=0.008 n=5+5) Decode/igmp-8 18.1ns ±44% 30.2ns ± 1% +66.89% (p=0.008 n=5+5) Decode/unknown-8 20.8ns ± 1% 10.6ns ± 9% -49.11% (p=0.016 n=4+5) pkg:tailscale.com/wgengine/filter goos:linux goarch:amd64 Filter/icmp4-8 30.5ns ± 1% 77.9ns ± 3% +155.01% (p=0.008 n=5+5) Filter/tcp4_syn_in-8 43.7ns ± 3% 123.0ns ± 3% +181.72% (p=0.008 n=5+5) Filter/tcp4_syn_out-8 24.5ns ± 2% 45.7ns ± 6% +86.22% (p=0.008 n=5+5) Filter/udp4_in-8 64.8ns ± 1% 210.0ns ± 2% +223.87% (p=0.008 n=5+5) Filter/udp4_out-8 119ns ± 0% 278ns ± 0% +133.78% (p=0.016 n=4+5) Filter/icmp6-8 40.3ns ± 2% 204.4ns ± 4% +407.70% (p=0.008 n=5+5) Filter/tcp6_syn_in-8 35.3ns ± 3% 199.2ns ± 2% +464.95% (p=0.008 n=5+5) Filter/tcp6_syn_out-8 32.8ns ± 2% 81.0ns ± 2% +147.10% (p=0.008 n=5+5) Filter/udp6_in-8 106ns ± 2% 290ns ± 2% +174.48% (p=0.008 n=5+5) Filter/udp6_out-8 184ns ± 2% 314ns ± 3% +70.43% (p=0.016 n=4+5) pkg:tailscale.com/wgengine/tstun goos:linux goarch:amd64 Write-8 9.02ns ± 3% 8.92ns ± 1% ~ (p=0.421 n=5+5) name old alloc/op new alloc/op delta pkg:tailscale.com/net/packet goos:linux goarch:amd64 Decode/tcp4-8 0.00B 0.00B ~ (all equal) Decode/tcp6-8 0.00B 0.00B ~ (all equal) Decode/udp4-8 0.00B 0.00B ~ (all equal) Decode/udp6-8 0.00B 0.00B ~ (all equal) Decode/icmp4-8 0.00B 0.00B ~ (all equal) Decode/icmp6-8 0.00B 0.00B ~ (all equal) Decode/igmp-8 0.00B 0.00B ~ (all equal) Decode/unknown-8 0.00B 0.00B ~ (all equal) pkg:tailscale.com/wgengine/filter goos:linux goarch:amd64 Filter/icmp4-8 0.00B 0.00B ~ (all equal) Filter/tcp4_syn_in-8 0.00B 0.00B ~ (all equal) Filter/tcp4_syn_out-8 0.00B 0.00B ~ (all equal) Filter/udp4_in-8 0.00B 0.00B ~ (all equal) Filter/udp4_out-8 16.0B ± 0% 64.0B ± 0% +300.00% (p=0.008 n=5+5) Filter/icmp6-8 0.00B 0.00B ~ (all equal) Filter/tcp6_syn_in-8 0.00B 0.00B ~ (all equal) Filter/tcp6_syn_out-8 0.00B 0.00B ~ (all equal) Filter/udp6_in-8 0.00B 0.00B ~ (all equal) Filter/udp6_out-8 48.0B ± 0% 64.0B ± 0% +33.33% (p=0.008 n=5+5) name old allocs/op new allocs/op delta pkg:tailscale.com/net/packet goos:linux goarch:amd64 Decode/tcp4-8 0.00 0.00 ~ (all equal) Decode/tcp6-8 0.00 0.00 ~ (all equal) Decode/udp4-8 0.00 0.00 ~ (all equal) Decode/udp6-8 0.00 0.00 ~ (all equal) Decode/icmp4-8 0.00 0.00 ~ (all equal) Decode/icmp6-8 0.00 0.00 ~ (all equal) Decode/igmp-8 0.00 0.00 ~ (all equal) Decode/unknown-8 0.00 0.00 ~ (all equal) pkg:tailscale.com/wgengine/filter goos:linux goarch:amd64 Filter/icmp4-8 0.00 0.00 ~ (all equal) Filter/tcp4_syn_in-8 0.00 0.00 ~ (all equal) Filter/tcp4_syn_out-8 0.00 0.00 ~ (all equal) Filter/udp4_in-8 0.00 0.00 ~ (all equal) Filter/udp4_out-8 1.00 ± 0% 1.00 ± 0% ~ (all equal) Filter/icmp6-8 0.00 0.00 ~ (all equal) Filter/tcp6_syn_in-8 0.00 0.00 ~ (all equal) Filter/tcp6_syn_out-8 0.00 0.00 ~ (all equal) Filter/udp6_in-8 0.00 0.00 ~ (all equal) Filter/udp6_out-8 1.00 ± 0% 1.00 ± 0% ~ (all equal) Signed-off-by: David Anderson <danderson@tailscale.com>
4 years ago
t.destIPActivity.Store(m)
}
// SetDiscoKey sets the current discovery key.
//
// It is only used for filtering out bogus traffic when network
// stack(s) get confused; see Issue 1526.
func (t *Wrapper) SetDiscoKey(k key.DiscoPublic) {
t.discoKey.Store(k)
}
// isSelfDisco reports whether packet p
// looks like a Disco packet from ourselves.
// See Issue 1526.
func (t *Wrapper) isSelfDisco(p *packet.Parsed) bool {
if p.IPProto != ipproto.UDP {
return false
}
pkt := p.Payload()
discobs, ok := disco.Source(pkt)
if !ok {
return false
}
discoSrc := key.DiscoPublicFromRaw32(mem.B(discobs))
selfDiscoPub, ok := t.discoKey.Load().(key.DiscoPublic)
return ok && selfDiscoPub == discoSrc
}
func (t *Wrapper) Close() error {
var err error
t.closeOnce.Do(func() {
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
close(t.closed)
t.bufferConsumedMu.Lock()
close(t.bufferConsumed)
t.bufferConsumedMu.Unlock()
t.outboundMu.Lock()
close(t.outbound)
t.outboundMu.Unlock()
err = t.tdev.Close()
})
return err
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
}
// isClosed reports whether t is closed.
func (t *Wrapper) isClosed() bool {
select {
case <-t.closed:
return true
default:
return false
}
}
// pumpEvents copies events from t.tdev to t.eventsUpDown and t.eventsOther.
// pumpEvents exits when t.tdev.events or t.closed is closed.
// pumpEvents closes t.eventsUpDown and t.eventsOther when it exits.
func (t *Wrapper) pumpEvents() {
defer close(t.eventsUpDown)
defer close(t.eventsOther)
src := t.tdev.Events()
for {
// Retrieve an event from the TUN device.
var event tun.Event
var ok bool
select {
case <-t.closed:
return
case event, ok = <-src:
if !ok {
return
}
}
// Pass along event to the correct recipient.
// Though event is a bitmask, in practice there is only ever one bit set at a time.
dst := t.eventsOther
if event&(tun.EventUp|tun.EventDown) != 0 {
dst = t.eventsUpDown
}
select {
case <-t.closed:
return
case dst <- event:
}
}
}
// EventsUpDown returns a TUN event channel that contains all Up and Down events.
func (t *Wrapper) EventsUpDown() chan tun.Event {
return t.eventsUpDown
}
// Events returns a TUN event channel that contains all non-Up, non-Down events.
// It is named Events because it is the set of events that we want to expose to wireguard-go,
// and Events is the name specified by the wireguard-go tun.Device interface.
func (t *Wrapper) Events() chan tun.Event {
return t.eventsOther
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
}
func (t *Wrapper) File() *os.File {
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
return t.tdev.File()
}
func (t *Wrapper) Flush() error {
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
return t.tdev.Flush()
}
func (t *Wrapper) MTU() (int, error) {
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
return t.tdev.MTU()
}
func (t *Wrapper) Name() (string, error) {
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
return t.tdev.Name()
}
// allowSendOnClosedChannel suppresses panics due to sending on a closed channel.
// This allows us to avoid synchronization between poll and Close.
// Such synchronization (particularly multi-case selects) is too expensive
// for code like poll or Read that is on the hot path of every packet.
// If this makes you sad or angry, you may want to join our
// weekly Go Performance Delinquents Anonymous meetings on Monday nights.
func allowSendOnClosedChannel() {
r := recover()
if r == nil {
return
}
e, _ := r.(error)
if e != nil && e.Error() == "send on closed channel" {
return
}
panic(r)
}
const ethernetFrameSize = 14 // 2 six byte MACs, 2 bytes ethertype
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
// poll polls t.tdev.Read, placing the oldest unconsumed packet into t.buffer.
// This is needed because t.tdev.Read in general may block (it does on Windows),
// so packets may be stuck in t.outbound if t.Read called t.tdev.Read directly.
func (t *Wrapper) poll() {
for range t.bufferConsumed {
DoRead:
var n int
var err error
// Read may use memory in t.buffer before PacketStartOffset for mandatory headers.
// This is the rationale behind the tun.Wrapper.{Read,Write} interfaces
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
// and the reason t.buffer has size MaxMessageSize and not MaxContentSize.
// In principle, read errors are not fatal (but wireguard-go disagrees).
// We loop here until we get a non-empty (or failed) read.
// We don't need this loop for correctness,
// but wireguard-go will skip an empty read,
// so we might as well avoid the send through t.outbound.
for n == 0 && err == nil {
if t.isClosed() {
return
}
if t.isTAP {
n, err = t.tdev.Read(t.buffer[:], PacketStartOffset-ethernetFrameSize)
if tapDebug {
s := fmt.Sprintf("% x", t.buffer[:])
for strings.HasSuffix(s, " 00") {
s = strings.TrimSuffix(s, " 00")
}
t.logf("TAP read %v, %v: %s", n, err, s)
}
} else {
n, err = t.tdev.Read(t.buffer[:], PacketStartOffset)
}
}
if t.isTAP {
if err == nil {
ethernetFrame := t.buffer[PacketStartOffset-ethernetFrameSize:][:n]
if t.handleTAPFrame(ethernetFrame) {
goto DoRead
}
}
// Fall through. We got an IP packet.
if n >= ethernetFrameSize {
n -= ethernetFrameSize
}
if tapDebug {
t.logf("tap regular frame: %x", t.buffer[PacketStartOffset:PacketStartOffset+n])
}
}
t.sendOutbound(tunReadResult{data: t.buffer[PacketStartOffset : PacketStartOffset+n], err: err})
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
}
}
// sendBufferConsumed does t.bufferConsumed <- struct{}{}.
// It protects against any panics or data races that that send could cause.
func (t *Wrapper) sendBufferConsumed() {
defer allowSendOnClosedChannel()
t.bufferConsumedMu.Lock()
defer t.bufferConsumedMu.Unlock()
t.bufferConsumed <- struct{}{}
}
// sendOutbound does t.outboundMu <- r.
// It protects against any panics or data races that that send could cause.
func (t *Wrapper) sendOutbound(r tunReadResult) {
defer allowSendOnClosedChannel()
t.outboundMu.Lock()
defer t.outboundMu.Unlock()
t.outbound <- r
}
var magicDNSIPPort = netaddr.MustParseIPPort("100.100.100.100:0")
func (t *Wrapper) filterOut(p *packet.Parsed) filter.Response {
// Fake ICMP echo responses to MagicDNS (100.100.100.100).
if p.IsEchoRequest() && p.Dst == magicDNSIPPort {
header := p.ICMP4Header()
header.ToResponse()
outp := packet.Generate(&header, p.Payload())
t.InjectInboundCopy(outp)
return filter.DropSilently // don't pass on to OS; already handled
}
// Issue 1526 workaround: if we sent disco packets over
// Tailscale from ourselves, then drop them, as that shouldn't
// happen unless a networking stack is confused, as it seems
// macOS in Network Extension mode might be.
if p.IPProto == ipproto.UDP && // disco is over UDP; avoid isSelfDisco call for TCP/etc
t.isSelfDisco(p) {
t.logf("[unexpected] received self disco out packet over tstun; dropping")
metricPacketOutDropSelfDisco.Add(1)
return filter.DropSilently
}
if t.PreFilterOut != nil {
if res := t.PreFilterOut(p, t); res.IsDrop() {
// Handled by userspaceEngine.handleLocalPackets (quad-100 DNS primarily).
return res
}
}
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
filt, _ := t.filter.Load().(*filter.Filter)
if filt == nil {
return filter.Drop
}
if filt.RunOut(p, t.filterFlags) != filter.Accept {
metricPacketOutDropFilter.Add(1)
return filter.Drop
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
}
if t.PostFilterOut != nil {
if res := t.PostFilterOut(p, t); res.IsDrop() {
return res
}
}
return filter.Accept
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
}
// noteActivity records that there was a read or write at the current time.
func (t *Wrapper) noteActivity() {
t.lastActivityAtomic.StoreAtomic(mono.Now())
}
// IdleDuration reports how long it's been since the last read or write to this device.
//
// Its value should only be presumed accurate to roughly 10ms granularity.
// If there's never been activity, the duration is since the wrapper was created.
func (t *Wrapper) IdleDuration() time.Duration {
return mono.Since(t.lastActivityAtomic.LoadAtomic())
}
func (t *Wrapper) Read(buf []byte, offset int) (int, error) {
res, ok := <-t.outbound
if !ok {
// Wrapper is closed.
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
return 0, io.EOF
}
if res.err != nil {
return 0, res.err
}
metricPacketOut.Add(1)
pkt := res.data
n := copy(buf[offset:], pkt)
// t.buffer has a fixed location in memory.
// If the packet is not from t.buffer, then it is an injected packet.
// &pkt[0] can be used because empty packets do not reach t.outbound.
isInjectedPacket := &pkt[0] != &t.buffer[PacketStartOffset]
if !isInjectedPacket {
// We are done with t.buffer. Let poll re-use it.
t.sendBufferConsumed()
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
}
p := parsedPacketPool.Get().(*packet.Parsed)
defer parsedPacketPool.Put(p)
p.Decode(buf[offset : offset+n])
net/packet: remove the custom IP4/IP6 types in favor of netaddr.IP. Upstream netaddr has a change that makes it alloc-free, so it's safe to use in hot codepaths. This gets rid of one of the many IP types in our codebase. Performance is currently worse across the board. This is likely due in part to netaddr.IP being a larger value type (4b -> 24b for IPv4, 16b -> 24b for IPv6), and in other part due to missing low-hanging fruit optimizations in netaddr. However, the regression is less bad than it looks at first glance, because we'd micro-optimized packet.IP* in the past few weeks. This change drops us back to roughly where we were at the 1.2 release, but with the benefit of a significant code and architectural simplification. name old time/op new time/op delta pkg:tailscale.com/net/packet goos:linux goarch:amd64 Decode/tcp4-8 12.2ns ± 5% 29.7ns ± 2% +142.32% (p=0.008 n=5+5) Decode/tcp6-8 12.6ns ± 3% 65.1ns ± 2% +418.47% (p=0.008 n=5+5) Decode/udp4-8 11.8ns ± 3% 30.5ns ± 2% +157.94% (p=0.008 n=5+5) Decode/udp6-8 27.1ns ± 1% 65.7ns ± 2% +142.36% (p=0.016 n=4+5) Decode/icmp4-8 24.6ns ± 2% 30.5ns ± 2% +23.65% (p=0.016 n=4+5) Decode/icmp6-8 22.9ns ±51% 65.5ns ± 2% +186.19% (p=0.008 n=5+5) Decode/igmp-8 18.1ns ±44% 30.2ns ± 1% +66.89% (p=0.008 n=5+5) Decode/unknown-8 20.8ns ± 1% 10.6ns ± 9% -49.11% (p=0.016 n=4+5) pkg:tailscale.com/wgengine/filter goos:linux goarch:amd64 Filter/icmp4-8 30.5ns ± 1% 77.9ns ± 3% +155.01% (p=0.008 n=5+5) Filter/tcp4_syn_in-8 43.7ns ± 3% 123.0ns ± 3% +181.72% (p=0.008 n=5+5) Filter/tcp4_syn_out-8 24.5ns ± 2% 45.7ns ± 6% +86.22% (p=0.008 n=5+5) Filter/udp4_in-8 64.8ns ± 1% 210.0ns ± 2% +223.87% (p=0.008 n=5+5) Filter/udp4_out-8 119ns ± 0% 278ns ± 0% +133.78% (p=0.016 n=4+5) Filter/icmp6-8 40.3ns ± 2% 204.4ns ± 4% +407.70% (p=0.008 n=5+5) Filter/tcp6_syn_in-8 35.3ns ± 3% 199.2ns ± 2% +464.95% (p=0.008 n=5+5) Filter/tcp6_syn_out-8 32.8ns ± 2% 81.0ns ± 2% +147.10% (p=0.008 n=5+5) Filter/udp6_in-8 106ns ± 2% 290ns ± 2% +174.48% (p=0.008 n=5+5) Filter/udp6_out-8 184ns ± 2% 314ns ± 3% +70.43% (p=0.016 n=4+5) pkg:tailscale.com/wgengine/tstun goos:linux goarch:amd64 Write-8 9.02ns ± 3% 8.92ns ± 1% ~ (p=0.421 n=5+5) name old alloc/op new alloc/op delta pkg:tailscale.com/net/packet goos:linux goarch:amd64 Decode/tcp4-8 0.00B 0.00B ~ (all equal) Decode/tcp6-8 0.00B 0.00B ~ (all equal) Decode/udp4-8 0.00B 0.00B ~ (all equal) Decode/udp6-8 0.00B 0.00B ~ (all equal) Decode/icmp4-8 0.00B 0.00B ~ (all equal) Decode/icmp6-8 0.00B 0.00B ~ (all equal) Decode/igmp-8 0.00B 0.00B ~ (all equal) Decode/unknown-8 0.00B 0.00B ~ (all equal) pkg:tailscale.com/wgengine/filter goos:linux goarch:amd64 Filter/icmp4-8 0.00B 0.00B ~ (all equal) Filter/tcp4_syn_in-8 0.00B 0.00B ~ (all equal) Filter/tcp4_syn_out-8 0.00B 0.00B ~ (all equal) Filter/udp4_in-8 0.00B 0.00B ~ (all equal) Filter/udp4_out-8 16.0B ± 0% 64.0B ± 0% +300.00% (p=0.008 n=5+5) Filter/icmp6-8 0.00B 0.00B ~ (all equal) Filter/tcp6_syn_in-8 0.00B 0.00B ~ (all equal) Filter/tcp6_syn_out-8 0.00B 0.00B ~ (all equal) Filter/udp6_in-8 0.00B 0.00B ~ (all equal) Filter/udp6_out-8 48.0B ± 0% 64.0B ± 0% +33.33% (p=0.008 n=5+5) name old allocs/op new allocs/op delta pkg:tailscale.com/net/packet goos:linux goarch:amd64 Decode/tcp4-8 0.00 0.00 ~ (all equal) Decode/tcp6-8 0.00 0.00 ~ (all equal) Decode/udp4-8 0.00 0.00 ~ (all equal) Decode/udp6-8 0.00 0.00 ~ (all equal) Decode/icmp4-8 0.00 0.00 ~ (all equal) Decode/icmp6-8 0.00 0.00 ~ (all equal) Decode/igmp-8 0.00 0.00 ~ (all equal) Decode/unknown-8 0.00 0.00 ~ (all equal) pkg:tailscale.com/wgengine/filter goos:linux goarch:amd64 Filter/icmp4-8 0.00 0.00 ~ (all equal) Filter/tcp4_syn_in-8 0.00 0.00 ~ (all equal) Filter/tcp4_syn_out-8 0.00 0.00 ~ (all equal) Filter/udp4_in-8 0.00 0.00 ~ (all equal) Filter/udp4_out-8 1.00 ± 0% 1.00 ± 0% ~ (all equal) Filter/icmp6-8 0.00 0.00 ~ (all equal) Filter/tcp6_syn_in-8 0.00 0.00 ~ (all equal) Filter/tcp6_syn_out-8 0.00 0.00 ~ (all equal) Filter/udp6_in-8 0.00 0.00 ~ (all equal) Filter/udp6_out-8 1.00 ± 0% 1.00 ± 0% ~ (all equal) Signed-off-by: David Anderson <danderson@tailscale.com>
4 years ago
if m, ok := t.destIPActivity.Load().(map[netaddr.IP]func()); ok {
if fn := m[p.Dst.IP()]; fn != nil {
net/packet: remove the custom IP4/IP6 types in favor of netaddr.IP. Upstream netaddr has a change that makes it alloc-free, so it's safe to use in hot codepaths. This gets rid of one of the many IP types in our codebase. Performance is currently worse across the board. This is likely due in part to netaddr.IP being a larger value type (4b -> 24b for IPv4, 16b -> 24b for IPv6), and in other part due to missing low-hanging fruit optimizations in netaddr. However, the regression is less bad than it looks at first glance, because we'd micro-optimized packet.IP* in the past few weeks. This change drops us back to roughly where we were at the 1.2 release, but with the benefit of a significant code and architectural simplification. name old time/op new time/op delta pkg:tailscale.com/net/packet goos:linux goarch:amd64 Decode/tcp4-8 12.2ns ± 5% 29.7ns ± 2% +142.32% (p=0.008 n=5+5) Decode/tcp6-8 12.6ns ± 3% 65.1ns ± 2% +418.47% (p=0.008 n=5+5) Decode/udp4-8 11.8ns ± 3% 30.5ns ± 2% +157.94% (p=0.008 n=5+5) Decode/udp6-8 27.1ns ± 1% 65.7ns ± 2% +142.36% (p=0.016 n=4+5) Decode/icmp4-8 24.6ns ± 2% 30.5ns ± 2% +23.65% (p=0.016 n=4+5) Decode/icmp6-8 22.9ns ±51% 65.5ns ± 2% +186.19% (p=0.008 n=5+5) Decode/igmp-8 18.1ns ±44% 30.2ns ± 1% +66.89% (p=0.008 n=5+5) Decode/unknown-8 20.8ns ± 1% 10.6ns ± 9% -49.11% (p=0.016 n=4+5) pkg:tailscale.com/wgengine/filter goos:linux goarch:amd64 Filter/icmp4-8 30.5ns ± 1% 77.9ns ± 3% +155.01% (p=0.008 n=5+5) Filter/tcp4_syn_in-8 43.7ns ± 3% 123.0ns ± 3% +181.72% (p=0.008 n=5+5) Filter/tcp4_syn_out-8 24.5ns ± 2% 45.7ns ± 6% +86.22% (p=0.008 n=5+5) Filter/udp4_in-8 64.8ns ± 1% 210.0ns ± 2% +223.87% (p=0.008 n=5+5) Filter/udp4_out-8 119ns ± 0% 278ns ± 0% +133.78% (p=0.016 n=4+5) Filter/icmp6-8 40.3ns ± 2% 204.4ns ± 4% +407.70% (p=0.008 n=5+5) Filter/tcp6_syn_in-8 35.3ns ± 3% 199.2ns ± 2% +464.95% (p=0.008 n=5+5) Filter/tcp6_syn_out-8 32.8ns ± 2% 81.0ns ± 2% +147.10% (p=0.008 n=5+5) Filter/udp6_in-8 106ns ± 2% 290ns ± 2% +174.48% (p=0.008 n=5+5) Filter/udp6_out-8 184ns ± 2% 314ns ± 3% +70.43% (p=0.016 n=4+5) pkg:tailscale.com/wgengine/tstun goos:linux goarch:amd64 Write-8 9.02ns ± 3% 8.92ns ± 1% ~ (p=0.421 n=5+5) name old alloc/op new alloc/op delta pkg:tailscale.com/net/packet goos:linux goarch:amd64 Decode/tcp4-8 0.00B 0.00B ~ (all equal) Decode/tcp6-8 0.00B 0.00B ~ (all equal) Decode/udp4-8 0.00B 0.00B ~ (all equal) Decode/udp6-8 0.00B 0.00B ~ (all equal) Decode/icmp4-8 0.00B 0.00B ~ (all equal) Decode/icmp6-8 0.00B 0.00B ~ (all equal) Decode/igmp-8 0.00B 0.00B ~ (all equal) Decode/unknown-8 0.00B 0.00B ~ (all equal) pkg:tailscale.com/wgengine/filter goos:linux goarch:amd64 Filter/icmp4-8 0.00B 0.00B ~ (all equal) Filter/tcp4_syn_in-8 0.00B 0.00B ~ (all equal) Filter/tcp4_syn_out-8 0.00B 0.00B ~ (all equal) Filter/udp4_in-8 0.00B 0.00B ~ (all equal) Filter/udp4_out-8 16.0B ± 0% 64.0B ± 0% +300.00% (p=0.008 n=5+5) Filter/icmp6-8 0.00B 0.00B ~ (all equal) Filter/tcp6_syn_in-8 0.00B 0.00B ~ (all equal) Filter/tcp6_syn_out-8 0.00B 0.00B ~ (all equal) Filter/udp6_in-8 0.00B 0.00B ~ (all equal) Filter/udp6_out-8 48.0B ± 0% 64.0B ± 0% +33.33% (p=0.008 n=5+5) name old allocs/op new allocs/op delta pkg:tailscale.com/net/packet goos:linux goarch:amd64 Decode/tcp4-8 0.00 0.00 ~ (all equal) Decode/tcp6-8 0.00 0.00 ~ (all equal) Decode/udp4-8 0.00 0.00 ~ (all equal) Decode/udp6-8 0.00 0.00 ~ (all equal) Decode/icmp4-8 0.00 0.00 ~ (all equal) Decode/icmp6-8 0.00 0.00 ~ (all equal) Decode/igmp-8 0.00 0.00 ~ (all equal) Decode/unknown-8 0.00 0.00 ~ (all equal) pkg:tailscale.com/wgengine/filter goos:linux goarch:amd64 Filter/icmp4-8 0.00 0.00 ~ (all equal) Filter/tcp4_syn_in-8 0.00 0.00 ~ (all equal) Filter/tcp4_syn_out-8 0.00 0.00 ~ (all equal) Filter/udp4_in-8 0.00 0.00 ~ (all equal) Filter/udp4_out-8 1.00 ± 0% 1.00 ± 0% ~ (all equal) Filter/icmp6-8 0.00 0.00 ~ (all equal) Filter/tcp6_syn_in-8 0.00 0.00 ~ (all equal) Filter/tcp6_syn_out-8 0.00 0.00 ~ (all equal) Filter/udp6_in-8 0.00 0.00 ~ (all equal) Filter/udp6_out-8 1.00 ± 0% 1.00 ± 0% ~ (all equal) Signed-off-by: David Anderson <danderson@tailscale.com>
4 years ago
fn()
}
}
// Do not filter injected packets.
if !isInjectedPacket && !t.disableFilter {
response := t.filterOut(p)
if response != filter.Accept {
metricPacketOutDrop.Add(1)
// Wireguard considers read errors fatal; pretend nothing was read
return 0, nil
}
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
}
t.noteActivity()
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
return n, nil
}
func (t *Wrapper) filterIn(buf []byte) filter.Response {
p := parsedPacketPool.Get().(*packet.Parsed)
defer parsedPacketPool.Put(p)
p.Decode(buf)
if p.IPProto == ipproto.TSMP {
if pingReq, ok := p.AsTSMPPing(); ok {
t.noteActivity()
t.injectOutboundPong(p, pingReq)
return filter.DropSilently
} else if data, ok := p.AsTSMPPong(); ok {
if f := t.OnTSMPPongReceived; f != nil {
f(data)
}
}
}
// Issue 1526 workaround: if we see disco packets over
// Tailscale from ourselves, then drop them, as that shouldn't
// happen unless a networking stack is confused, as it seems
// macOS in Network Extension mode might be.
if p.IPProto == ipproto.UDP && // disco is over UDP; avoid isSelfDisco call for TCP/etc
t.isSelfDisco(p) {
t.logf("[unexpected] received self disco in packet over tstun; dropping")
metricPacketInDropSelfDisco.Add(1)
return filter.DropSilently
}
if t.PreFilterIn != nil {
if res := t.PreFilterIn(p, t); res.IsDrop() {
return res
}
}
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
filt, _ := t.filter.Load().(*filter.Filter)
if filt == nil {
return filter.Drop
}
outcome := filt.RunIn(p, t.filterFlags)
// Let peerapi through the filter; its ACLs are handled at L7,
// not at the packet level.
if outcome != filter.Accept &&
p.IPProto == ipproto.TCP &&
p.TCPFlags&packet.TCPSyn != 0 &&
t.PeerAPIPort != nil {
if port, ok := t.PeerAPIPort(p.Dst.IP()); ok && port == p.Dst.Port() {
outcome = filter.Accept
}
}
if outcome != filter.Accept {
metricPacketInDropFilter.Add(1)
// Tell them, via TSMP, we're dropping them due to the ACL.
// Their host networking stack can translate this into ICMP
// or whatnot as required. But notably, their GUI or tailscale CLI
// can show them a rejection history with reasons.
if p.IPVersion == 4 && p.IPProto == ipproto.TCP && p.TCPFlags&packet.TCPSyn != 0 && !t.disableTSMPRejected {
rj := packet.TailscaleRejectedHeader{
IPSrc: p.Dst.IP(),
IPDst: p.Src.IP(),
Src: p.Src,
Dst: p.Dst,
Proto: p.IPProto,
Reason: packet.RejectedDueToACLs,
}
if filt.ShieldsUp() {
rj.Reason = packet.RejectedDueToShieldsUp
}
pkt := packet.Generate(rj, nil)
t.InjectOutbound(pkt)
// TODO(bradfitz): also send a TCP RST, after the TSMP message.
}
return filter.Drop
}
if t.PostFilterIn != nil {
if res := t.PostFilterIn(p, t); res.IsDrop() {
return res
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
}
}
return filter.Accept
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
}
// Write accepts an incoming packet. The packet begins at buf[offset:],
// like wireguard-go/tun.Device.Write.
func (t *Wrapper) Write(buf []byte, offset int) (int, error) {
metricPacketIn.Add(1)
if !t.disableFilter {
if t.filterIn(buf[offset:]) != filter.Accept {
metricPacketInDrop.Add(1)
// If we're not accepting the packet, lie to wireguard-go and pretend
// that everything is okay with a nil error, so wireguard-go
// doesn't log about this Write "failure".
//
// We return len(buf), but the ill-defined wireguard-go/tun.Device.Write
// method doesn't specify how the offset affects the return value.
// In fact, the Linux implementation does one of two different things depending
// on how the /dev/net/tun was created. But fortunately the wireguard-go
// code ignores the int return and only looks at the error:
//
// device/receive.go: _, err = device.tun.device.Write(....)
//
// TODO(bradfitz): fix upstream interface docs, implementation.
return len(buf), nil
}
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
}
t.noteActivity()
return t.tdevWrite(buf, offset)
}
func (t *Wrapper) tdevWrite(buf []byte, offset int) (int, error) {
if t.isTAP {
return t.tapWrite(buf, offset)
}
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
return t.tdev.Write(buf, offset)
}
func (t *Wrapper) GetFilter() *filter.Filter {
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
filt, _ := t.filter.Load().(*filter.Filter)
return filt
}
func (t *Wrapper) SetFilter(filt *filter.Filter) {
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
t.filter.Store(filt)
}
// InjectInboundDirect makes the Wrapper device behave as if a packet
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
// with the given contents was received from the network.
// It blocks and does not take ownership of the packet.
// The injected packet will not pass through inbound filters.
//
// The packet contents are to start at &buf[offset].
// offset must be greater or equal to PacketStartOffset.
// The space before &buf[offset] will be used by Wireguard.
func (t *Wrapper) InjectInboundDirect(buf []byte, offset int) error {
if len(buf) > MaxPacketSize {
return errPacketTooBig
}
if len(buf) < offset {
return errOffsetTooBig
}
if offset < PacketStartOffset {
return errOffsetTooSmall
}
// Write to the underlying device to skip filters.
_, err := t.tdevWrite(buf, offset)
return err
}
// InjectInboundCopy takes a packet without leading space,
// reallocates it to conform to the InjectInboundDirect interface
// and calls InjectInboundDirect on it. Injecting a nil packet is a no-op.
func (t *Wrapper) InjectInboundCopy(packet []byte) error {
// We duplicate this check from InjectInboundDirect here
// to avoid wasting an allocation on an oversized packet.
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
if len(packet) > MaxPacketSize {
return errPacketTooBig
}
if len(packet) == 0 {
return nil
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
}
buf := make([]byte, PacketStartOffset+len(packet))
copy(buf[PacketStartOffset:], packet)
return t.InjectInboundDirect(buf, PacketStartOffset)
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
}
func (t *Wrapper) injectOutboundPong(pp *packet.Parsed, req packet.TSMPPingRequest) {
pong := packet.TSMPPongReply{
Data: req.Data,
}
if t.PeerAPIPort != nil {
pong.PeerAPIPort, _ = t.PeerAPIPort(pp.Dst.IP())
}
switch pp.IPVersion {
case 4:
h4 := pp.IP4Header()
h4.ToResponse()
pong.IPHeader = h4
case 6:
h6 := pp.IP6Header()
h6.ToResponse()
pong.IPHeader = h6
default:
return
}
t.InjectOutbound(packet.Generate(pong, nil))
}
// InjectOutbound makes the Wrapper device behave as if a packet
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
// with the given contents was sent to the network.
// It does not block, but takes ownership of the packet.
// The injected packet will not pass through outbound filters.
// Injecting an empty packet is a no-op.
func (t *Wrapper) InjectOutbound(packet []byte) error {
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
if len(packet) > MaxPacketSize {
return errPacketTooBig
}
if len(packet) == 0 {
return nil
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
}
t.sendOutbound(tunReadResult{data: packet})
return nil
wgengine: wrap tun.Device to support filtering and packet injection (#358) Right now, filtering and packet injection in wgengine depend on a patch to wireguard-go that probably isn't suitable for upstreaming. This need not be the case: wireguard-go/tun.Device is an interface. For example, faketun.go implements it to mock a TUN device for testing. This patch implements the same interface to provide filtering and packet injection at the tunnel device level, at which point the wireguard-go patch should no longer be necessary. This patch has the following performance impact on i7-7500U @ 2.70GHz, tested in the following namespace configuration: ┌────────────────┐ ┌─────────────────────────────────┐ ┌────────────────┐ │ $ns1 │ │ $ns0 │ │ $ns2 │ │ client0 │ │ tailcontrol, logcatcher │ │ client1 │ │ ┌─────┐ │ │ ┌──────┐ ┌──────┐ │ │ ┌─────┐ │ │ │vethc│───────┼────┼──│vethrc│ │vethrs│──────┼─────┼──│veths│ │ │ ├─────┴─────┐ │ │ ├──────┴────┐ ├──────┴────┐ │ │ ├─────┴─────┐ │ │ │10.0.0.2/24│ │ │ │10.0.0.1/24│ │10.0.1.1/24│ │ │ │10.0.1.2/24│ │ │ └───────────┘ │ │ └───────────┘ └───────────┘ │ │ └───────────┘ │ └────────────────┘ └─────────────────────────────────┘ └────────────────┘ Before: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 557.0 (±8.5) Mbits/sec | 3.03 (±0.02) Gbits/sec | --------------------------------------------------- After: --------------------------------------------------- | TCP send | UDP send | |------------------------|------------------------| | 544.8 (±1.6) Mbits/sec | 3.13 (±0.02) Gbits/sec | --------------------------------------------------- The impact on receive performance is similar. Signed-off-by: Dmytro Shynkevych <dmytro@tailscale.com>
5 years ago
}
// Unwrap returns the underlying tun.Device.
func (t *Wrapper) Unwrap() tun.Device {
return t.tdev
}
var (
metricPacketIn = clientmetric.NewGauge("tstun_in_from_wg")
metricPacketInDrop = clientmetric.NewGauge("tstun_in_from_wg_drop")
metricPacketInDropFilter = clientmetric.NewGauge("tstun_in_from_wg_drop_filter")
metricPacketInDropSelfDisco = clientmetric.NewGauge("tstun_in_from_wg_drop_self_disco")
metricPacketOut = clientmetric.NewGauge("tstun_out_to_wg")
metricPacketOutDrop = clientmetric.NewGauge("tstun_out_to_wg_drop")
metricPacketOutDropFilter = clientmetric.NewGauge("tstun_out_to_wg_drop_filter")
metricPacketOutDropSelfDisco = clientmetric.NewGauge("tstun_out_to_wg_drop_self_disco")
)