Documentation
¶
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func LimitListener ¶
LimitListener returns a listener that accepts at most max connections at a time: a connection past it waits in the kernel's backlog until one closes (the header timeout frees the slots a client that says nothing holds). 0 = no limit.
func NewHTTPServer ¶
NewHTTPServer is the L7 proxy's HTTPS server with the limits that apply before any routing or authentication: a client that dribbles its headers or its body, or opens connections and says nothing, is cut instead of holding a connection and a goroutine forever. There is no write timeout (responses stream; an upgraded connection is bounded by the live check), and a hijacked connection has its deadlines cleared by the handler.
The client's address is the TCP peer: the load balancer keeps it (externalTrafficPolicy Local), so there is no header in front of the TLS to read and none is believed.
The server offers HTTP/2 and HTTP/1.1 (ALPN h2, http/1.1): a browser takes h2; the WebSocket of the console stays on HTTP/1.1. A tlsConfig that names its own protocols keeps them.
Types ¶
type L7Config ¶
type L7Config struct {
// The certificate and key of the wildcard Secret proxy-tls, where the chart mounts it.
TLSCertPath string `env:"TLS_CERT_PATH" envDefault:"/etc/proxy/tls/tls.crt"`
TLSKeyPath string `env:"TLS_KEY_PATH" envDefault:"/etc/proxy/tls/tls.key"`
// The access grants are verified with the access public keys of the tenants, kept in the Secrets
// tenant-<name>-access-keys of the access keys namespace (see l7.SecretKeys); the proxy has no key of its own and signs nothing.
BaseDomain string `env:"BASE_DOMAIN,notEmpty"`
Listen string `env:"LISTEN_HTTPS" envDefault:":8443"`
// CookieName is the host-only cookie a web host keeps its grant in (chart proxy.l7.grantCookieName). The default carries the __Host- prefix:
// a browser then accepts it only when it is Secure, has Path=/ and no Domain, so a sibling site under the parent domain cannot set or shadow
// it. The name is used everywhere (the cookie, the strip before the device, the filter of the device's own Set-Cookie).
CookieName string `env:"GRANT_COOKIE_NAME" envDefault:"__Host-cib_grant"`
// GrantMaxTTL is the longest exp - iat of a grant the proxy accepts (chart proxy.l7.grantMaxTTL); the agent reports it to the backend.
GrantMaxTTL time.Duration `env:"GRANT_MAX_TTL" envDefault:"8h"`
// PlatformOrigins are the pages that may post a grant (https://host[:port] or https://*.base[:port], as the embedding lists): the form
// comes from the platform page, so the Origin of the post must match one of them. Empty: every post is refused.
PlatformOrigins []string `env:"PLATFORM_ORIGINS" envSeparator:","`
// HealthAddr is where the readiness and liveness probes are served (the l7 container; the wg-demux container uses another port).
HealthAddr string `env:"HEALTH_ADDR" envDefault:":8081"`
// Instance names this replica in its traffic reports (the pod name).
Instance string `env:"POD_NAME"`
ReportInterval time.Duration `env:"REPORT_INTERVAL" envDefault:"1m"`
// LiveMaxLifetime caps one request or upgraded (WebSocket) connection; LiveCheckInterval is how often the
// open ones are checked against the access policy again, so a lock cuts them within about that time.
LiveMaxLifetime time.Duration `env:"LIVE_MAX_LIFETIME" envDefault:"12h"`
LiveCheckInterval time.Duration `env:"LIVE_CHECK_INTERVAL" envDefault:"10s"`
// SlowRequest is the upstream latency above which the access log flags a request as slow.
SlowRequest time.Duration `env:"SLOW_REQUEST" envDefault:"2s"`
// HTTP server limits (before any routing or authentication, so they stop slow-header and slow-body clients that
// would otherwise hold a connection and a goroutine forever): ReadHeaderTimeout bounds the request headers, ReadTimeout
// the whole request including the body, IdleTimeout a kept-alive connection with no request, MaxHeaderBytes the headers.
// There is no write timeout: responses stream, and an upgraded (WebSocket) connection lives at most LiveMaxLifetime.
ReadHeaderTimeout time.Duration `env:"READ_HEADER_TIMEOUT" envDefault:"10s"`
ReadTimeout time.Duration `env:"READ_TIMEOUT" envDefault:"5m"`
IdleTimeout time.Duration `env:"IDLE_TIMEOUT" envDefault:"2m"`
MaxHeaderBytes int `env:"MAX_HEADER_BYTES" envDefault:"65536"`
// What a client can hold open (R-17): MaxConnections is the connections of the server in all; LivePerClient, LivePerGroup and
// LiveTotal the requests in flight (upgraded connections included) per client of a group, per group and in all; AuthRate and
// AuthBurst limit the handoff path per peer address (all peers together may do twenty times that). 0 = unlimited.
// MaxConnections follows the memory limit of the chart (512Mi): a connection with a request in flight costs about 105 KiB, so
// 2500 are about 260 MiB, and with the informer caches and headroom they stay under the 80% soft limit. LiveTotal is not above
// it: a request in flight holds a connection. Above these numbers add replicas, do not raise the cap without the memory.
MaxConnections int `env:"MAX_CONNECTIONS" envDefault:"2500"`
LivePerClient int `env:"LIVE_PER_CLIENT" envDefault:"200"`
LivePerGroup int `env:"LIVE_PER_GROUP" envDefault:"1000"`
LiveTotal int `env:"LIVE_TOTAL" envDefault:"2500"`
AuthRate float64 `env:"AUTH_RATE" envDefault:"5"`
AuthBurst int `env:"AUTH_BURST" envDefault:"20"`
// The console (docs/specs/2026-11-12-device-access-and-checks.md): TerminalEnabled serves it on a host of its own per device (names.AccessHostLabel, the kind of access in the last character of the first label) under
// BaseDomain, relayed to the device access service pods of TerminalNamespace on TerminalPort. The limits mirror the chart values of the
// service (terminal.limits.*) and are enforced at the proxy as well.
TerminalEnabled bool `env:"TERMINAL_ENABLED" envDefault:"false"`
TerminalNamespace string `env:"TERMINAL_SERVICE_NAMESPACE"`
TerminalPort int `env:"TERMINAL_SERVICE_PORT" envDefault:"8080"`
TerminalPerClient int `env:"TERMINAL_PER_CLIENT" envDefault:"3"`
TerminalIdle time.Duration `env:"TERMINAL_IDLE_TIMEOUT" envDefault:"30m"`
TerminalLifetime time.Duration `env:"TERMINAL_MAX_LIFETIME" envDefault:"8h"`
// TerminalEmbedAllow and TerminalEmbedDeny are the lists of sites that may (and may not) show the console page in a frame; see
// l7.EmbedMatcher. Empty allow: no embedding.
TerminalEmbedAllow []string `env:"TERMINAL_EMBED_ALLOW" envSeparator:","`
TerminalEmbedDeny []string `env:"TERMINAL_EMBED_DENY" envSeparator:","`
}
func LoadL7Config ¶
type WGConfig ¶
type WGConfig struct {
// ListenAddr is the UDP address the demux binds; empty = ":" + names.WireGuardPort. The VPN pods listen on names.WireGuardPort too.
ListenAddr string `env:"UDP_LISTEN_ADDR"`
// HealthAddr is where the readiness and liveness probes are served.
HealthAddr string `env:"HEALTH_ADDR" envDefault:":8082"`
// The demux reads a public UDP port shared by every team, so what a stranger can make it hold or spend is capped
// (see demux.Limits). There is no limit per source address: one NAT hides a whole event behind one address.
// MaxEntries: conntrack entries in total (two per session). PartialTTL: how long a handshake in progress keeps its entry.
MaxEntries int `env:"DEMUX_MAX_ENTRIES" envDefault:"100000"`
PartialTTL time.Duration `env:"DEMUX_PARTIAL_TTL" envDefault:"15s"`
// All sources together may start this many handshakes per second (each costs a scan of the groups' keys) and send this many
// packets per second that match no session or come from an unexpected address (index guessing, spoofing, roaming).
GlobalHandshakeRate float64 `env:"DEMUX_GLOBAL_HANDSHAKE_RATE" envDefault:"2000"`
GlobalHandshakeBurst int `env:"DEMUX_GLOBAL_HANDSHAKE_BURST" envDefault:"4000"`
MissRate float64 `env:"DEMUX_MISS_RATE" envDefault:"2000"`
MissBurst int `env:"DEMUX_MISS_BURST" envDefault:"4000"`
RoamInterval time.Duration `env:"DEMUX_ROAM_INTERVAL" envDefault:"5s"`
// One session may send SessionRate packets per second, so that one participant cannot use up the shared demux.
// Readers is how many goroutines read the socket.
SessionRate float64 `env:"DEMUX_SESSION_RATE" envDefault:"15000"`
SessionBurst int `env:"DEMUX_SESSION_BURST" envDefault:"30000"`
Readers int `env:"DEMUX_READERS" envDefault:"4"`
// Mode is how transport packets are moved: userspace, or in the kernel by a TC or XDP program (needs NodeName).
Mode demux.Mode `env:"DEMUX_MODE" envDefault:"userspace"`
// PinPath is where the BPF maps are pinned, so a restart keeps the sessions.
PinPath string `env:"DEMUX_BPF_PIN_PATH" envDefault:"/sys/fs/bpf/cybericebox/wg-demux"`
NodeName string `env:"NODE_NAME"`
HostIP string `env:"HOST_IP"`
// ServiceName and ServiceNamespace name the Service whose address the clients' handshakes arrive at.
ServiceName string `env:"DEMUX_SERVICE_NAME" envDefault:"laboratory-proxy-wg"`
ServiceNamespace string `env:"DEMUX_SERVICE_NAMESPACE" envDefault:"laboratory-wg"`
MetricsAddr string `env:"DEMUX_METRICS_ADDR" envDefault:":8083"`
Resync time.Duration `env:"DEMUX_BPF_RESYNC" envDefault:"10s"`
}
func LoadWGConfig ¶
Directories
¶
| Path | Synopsis |
|---|---|
|
fastpath
Package fastpath is the in-kernel fast path of the WireGuard demux (modes tc and xdp): the eBPF programs of wg_fastpath.bpf.c, which forward transport packets of known sessions, and the pinned maps the demux keeps them in.
|
Package fastpath is the in-kernel fast path of the WireGuard demux (modes tc and xdp): the eBPF programs of wg_fastpath.bpf.c, which forward transport packets of known sessions, and the pinned maps the demux keeps them in. |