ping: Fix stale flows kept alive by unrelated ICMP traffic

Unconnected raw ICMP sockets receive every ICMP packet arriving at the
host, so any ICMP traffic refreshed the read deadline of every flow and
stale flows (with their raw sockets and goroutines) were only reclaimed
by LRU eviction while processing all host ICMP traffic in the meantime.

Expire flows based on their own activity only, and on Linux attach a
classic BPF ident filter to each raw socket so other flows' packets are
dropped in the kernel instead of waking every flow.

ping/filter_linux.go

@@ -0,0 +1,99 @@
+package ping
+
+import (
+	"sync"
+	"syscall"
+
+	"github.com/sagernet/sing-tun/internal/gtcpip/header"
+	"github.com/sagernet/sing/common"
+	"github.com/sagernet/sing/common/control"
+
+	"golang.org/x/net/bpf"
+	"golang.org/x/sys/unix"
+)
+
+type identFilterState struct {
+	access     sync.Mutex
+	attached   bool
+	disabled   bool
+	identifier uint16
+}
+
+// The kernel clones every matching-protocol packet into every unconnected raw
+// ICMP socket, so without a socket filter each flow receives and discards all
+// other flows' traffic.
+func (c *Conn) updateIdentFilter(wireIdentifier uint16) {
+	if !c.privileged {
+		return
+	}
+	syscallConn, isSyscallConn := common.Cast[syscall.Conn](c.conn)
+	if !isSyscallConn {
+		return
+	}
+	state := &c.identFilter
+	state.access.Lock()
+	defer state.access.Unlock()
+	if state.disabled {
+		return
+	}
+	if state.attached {
+		if state.identifier == wireIdentifier {
+			return
+		}
+		state.attached = false
+		state.disabled = true
+		_ = control.Conn(syscallConn, func(fd uintptr) error {
+			return unix.SetsockoptInt(int(fd), unix.SOL_SOCKET, unix.SO_DETACH_FILTER, 0)
+		})
+		return
+	}
+	rawInstructions, err := identFilterProgram(c.destination.Is6(), wireIdentifier)
+	if err != nil {
+		state.disabled = true
+		return
+	}
+	filter := make([]unix.SockFilter, len(rawInstructions))
+	for i, instruction := range rawInstructions {
+		filter[i] = unix.SockFilter{Code: instruction.Op, Jt: instruction.Jt, Jf: instruction.Jf, K: instruction.K}
+	}
+	program := unix.SockFprog{Len: uint16(len(filter)), Filter: &filter[0]}
+	err = control.Conn(syscallConn, func(fd uintptr) error {
+		return unix.SetsockoptSockFprog(int(fd), unix.SOL_SOCKET, unix.SO_ATTACH_FILTER, &program)
+	})
+	if err != nil {
+		state.disabled = true
+		return
+	}
+	state.attached = true
+	state.identifier = wireIdentifier
+}
+
+func identFilterProgram(is6 bool, wireIdentifier uint16) ([]bpf.RawInstruction, error) {
+	if !is6 {
+		// Raw ICMPv4 sockets deliver the full packet including the IP header.
+		return bpf.Assemble([]bpf.Instruction{
+			bpf.LoadMemShift{Off: 0},
+			bpf.LoadIndirect{Off: 0, Size: 1},
+			bpf.JumpIf{Cond: bpf.JumpEqual, Val: uint32(header.ICMPv4EchoReply), SkipTrue: 3},
+			bpf.JumpIf{Cond: bpf.JumpEqual, Val: uint32(header.ICMPv4DstUnreachable), SkipTrue: 4},
+			bpf.JumpIf{Cond: bpf.JumpEqual, Val: uint32(header.ICMPv4TimeExceeded), SkipTrue: 3},
+			bpf.RetConstant{Val: 0},
+			bpf.LoadIndirect{Off: 4, Size: 2},
+			bpf.JumpIf{Cond: bpf.JumpEqual, Val: uint32(wireIdentifier), SkipFalse: 1},
+			bpf.RetConstant{Val: 0xffffffff},
+			bpf.RetConstant{Val: 0},
+		})
+	}
+	// Raw ICMPv6 sockets deliver the ICMPv6 message without the IP header.
+	return bpf.Assemble([]bpf.Instruction{
+		bpf.LoadAbsolute{Off: 0, Size: 1},
+		bpf.JumpIf{Cond: bpf.JumpEqual, Val: uint32(header.ICMPv6EchoReply), SkipTrue: 3},
+		bpf.JumpIf{Cond: bpf.JumpGreaterThan, Val: uint32(header.ICMPv6ParamProblem), SkipTrue: 1},
+		bpf.RetConstant{Val: 0xffffffff},
+		bpf.RetConstant{Val: 0},
+		bpf.LoadAbsolute{Off: 4, Size: 2},
+		bpf.JumpIf{Cond: bpf.JumpEqual, Val: uint32(wireIdentifier), SkipFalse: 1},
+		bpf.RetConstant{Val: 0xffffffff},
+		bpf.RetConstant{Val: 0},
+	})
+}