Commit b7c7fba5b4

Stephen Gregoratto <dev@sgregoratto.me>
2022-01-22 08:14:07
Add classic BPF library
This library contains: - The global constants as used by C code. - An Insn struct that implements can generate all the BPF instructions. - A simple BPF virtual machine implementation that can be used for testing programs. This has complete code-coverage and has been extensively fuzzed.
1 parent 63ee6e6
Changed files (2)
lib
std
lib/std/x/net/bpf.zig
@@ -0,0 +1,1008 @@
+//! This package provides instrumentation for creating Berkeley Packet Filter[1]
+//! (BPF) programs, along with a simulator for running them.
+//!
+//! BPF is a mechanism for cheap, in-kernel packet filtering. Programs are
+//! attached to a network device and executed for every packet that flows
+//! through it. The program must then return a verdict: the amount of packet
+//! bytes that the kernel should copy into userspace. Execution speed is
+//! achieved by having programs run in a limited virtual machine, which has the
+//! added benefit of graceful failure in the face of buggy programs.
+//!
+//! The BPF virtual machine has a 32-bit word length and a small number of
+//! word-sized registers:
+//!
+//! - The accumulator, `a`: The source/destination of arithmetic and logic
+//!   operations.
+//! - The index register, `x`: Used as an offset for indirect memory access and
+//!   as a comparison value for conditional jumps.
+//! - The scratch memory store, `M[0]..M[15]`: Used for saving the value of a/x
+//!   for later use.
+//!
+//! The packet being examined is an array of bytes, and is addressed using plain
+//! array subscript notation, e.g. [10] for the byte at offset 10. An implicit
+//! program counter, `pc`, is intialized to zero and incremented for each instruction.
+//!
+//! The machine has a fixed instruction set with the following form, where the
+//! numbers represent bit length:
+//!
+//! ```
+//! ┌───────────┬──────┬──────┐
+//! │ opcode:16 │ jt:8 │ jt:8 │
+//! ├───────────┴──────┴──────┤
+//! │           k:32          │
+//! └─────────────────────────┘
+//! ```
+//!
+//! The `opcode` indicates the instruction class and its addressing mode.
+//! Opcodes are generated by performing binary addition on the 8-bit class and
+//! mode constants. For example, the opcode for loading a byte from the packet
+//! at X + 2, (`ldb [x + 2]`), is:
+//!
+//! ```
+//! LD | IND | B = 0x00 | 0x40 | 0x20
+//!              = 0x60
+//! ```
+//!
+//! `jt` is an offset used for conditional jumps, and increments the program
+//! counter by its amount if the comparison was true. Conversely, `jf`
+//! increments the counter if it was false. These fields are ignored in all
+//! other cases. `k` is a generic variable used for various purposes, most
+//! commonly as some sort of constant.
+//!
+//! This package contains opcode extensions used by different implementations,
+//! where "extension" is anything outside of the original that was imported into
+//! 4.4BSD[2]. These are marked with "EXTENSION", along with a list of
+//! implementations that use them.
+//!
+//! Most of the doc-comments use the BPF assembly syntax as described in the
+//! original paper[1]. For the sake of completeness, here is the complete
+//! instruction set, along with the extensions:
+//!
+//!```
+//! opcode  addressing modes
+//! ld      #k  #len    M[k]    [k]     [x + k]
+//! ldh     [k] [x + k]
+//! ldb     [k] [x + k]
+//! ldx     #k  #len    M[k]    4 * ([k] & 0xf) arc4random()
+//! st      M[k]
+//! stx     M[k]
+//! jmp     L
+//! jeq     #k, Lt, Lf
+//! jgt     #k, Lt, Lf
+//! jge     #k, Lt, Lf
+//! jset    #k, Lt, Lf
+//! add     #k  x
+//! sub     #k  x
+//! mul     #k  x
+//! div     #k  x
+//! or      #k  x
+//! and     #k  x
+//! lsh     #k  x
+//! rsh     #k  x
+//! neg     #k  x
+//! mod     #k  x
+//! xor     #k  x
+//! ret     #k  a
+//! tax
+//! txa
+//! ```
+//!
+//! Finally, a note on program design. The lack of backwards jumps leads to a
+//! "return early, return often" control flow. Take for example the program
+//! generated from the tcpdump filter `ip`:
+//!
+//! ```
+//! (000) ldh   [12]            ; Ethernet Packet Type
+//! (001) jeq   #0x86dd, 2, 7   ; ETHERTYPE_IPV6
+//! (002) ldb   [20]            ; IPv6 Next Header
+//! (003) jeq   #0x6, 10, 4     ; TCP
+//! (004) jeq   #0x2c, 5, 11    ; IPv6 Fragment Header
+//! (005) ldb   [54]            ; TCP Source Port
+//! (006) jeq   #0x6, 10, 11    ; IPPROTO_TCP
+//! (007) jeq   #0x800, 8, 11   ; ETHERTYPE_IP
+//! (008) ldb   [23]            ; IPv4 Protocol
+//! (009) jeq   #0x6, 10, 11    ; IPPROTO_TCP
+//! (010) ret   #262144         ; copy 0x40000
+//! (011) ret   #0              ; skip packet
+//! ```
+//! 
+//! Here we can make a few observations:
+//!
+//! - The problem "filter only tcp packets" has essentially been transformed
+//!   into a series of layer checks.
+//! - There are two distinct branches in the code, one for validating IPv4
+//!   headers and one for IPv6 headers.
+//! - Most conditional jumps in these branches lead directly to the last two
+//!   instructions, a pass or fail. Thus the goal of a program is to find the
+//!   fastest route to a pass/fail comparison.
+//!
+//! [1]: S. McCanne and V. Jacobson, "The BSD Packet Filter: A New Architecture
+//!      for User-level Packet Capture", Proceedings of the 1993 Winter USENIX.
+//! [2]: https://minnie.tuhs.org/cgi-bin/utree.pl?file=4.4BSD/usr/src/sys/net/bpf.h
+const std = @import("std");
+const builtin = @import("builtin");
+const native_endian = builtin.target.cpu.arch.endian();
+const mem = std.mem;
+const math = std.math;
+const random = std.crypto.random;
+const assert = std.debug.assert;
+const expectEqual = std.testing.expectEqual;
+const expectError = std.testing.expectError;
+const expect = std.testing.expect;
+
+// instruction classes
+/// ld, ldh, ldb: Load data into a.
+pub const LD = 0x00;
+/// ldx: Load data into x.
+pub const LDX = 0x01;
+/// st:  Store into scratch memory the value of a.
+pub const ST = 0x02;
+/// st:  Store into scratch memory the value of x.
+pub const STX = 0x03;
+/// alu: Wrapping arithmetic/bitwise operations on a using the value of k/x.
+pub const ALU = 0x04;
+/// jmp, jeq, jgt, je, jset: Increment the program counter based on a comparison
+/// between k/x and the accumulator.
+pub const JMP = 0x05;
+/// ret: Return a verdict using the value of k/the accumulator.
+pub const RET = 0x06;
+/// tax, txa: Register value copying between X and a.
+pub const MISC = 0x07;
+
+// Size of data to be loaded from the packet.
+/// ld: 32-bit full word.
+pub const W = 0x00;
+/// ldh: 16-bit half word.
+pub const H = 0x08;
+/// ldb: Single byte.
+pub const B = 0x10;
+
+// Addressing modes used for loads to a/x.
+/// #k: The immediate value stored in k.
+pub const IMM = 0x00;
+/// [k]: The value at offset k in the packet.
+pub const ABS = 0x20;
+/// [x + k]: The value at offset x + k in the packet.
+pub const IND = 0x40;
+/// M[k]: The value of the k'th scratch memory register.
+pub const MEM = 0x60;
+/// #len: The size of the packet.
+pub const LEN = 0x80;
+/// 4 * ([k] & 0xf): Four times the low four bits of the byte at offset k in the
+/// packet. This is used for efficiently loading the header length of an IP
+/// packet.
+pub const MSH = 0xa0;
+/// arc4random: 32-bit integer generated from a CPRNG (see arc4random(3)) loaded into a.
+/// EXTENSION. Defined for:
+/// - OpenBSD.
+pub const RND = 0xc0;
+
+// Modifiers for different instruction classes.
+/// Use the value of k for alu operations (add #k).
+/// Compare against the value of k for jumps (jeq #k, Lt, Lf).
+/// Return the value of k for returns (ret #k).
+pub const K = 0x00;
+/// Use the value of x for alu operations (add x).
+/// Compare against the value of X for jumps (jeq x, Lt, Lf).
+pub const X = 0x08;
+/// Return the value of a for returns (ret a).
+pub const A = 0x10;
+
+// ALU Operations on a using the value of k/x.
+// All arithmetic operations are defined to overflow the value of a.
+/// add: a = a + k
+///      a = a + x.
+pub const ADD = 0x00;
+/// sub: a = a - k
+///      a = a - x.
+pub const SUB = 0x10;
+/// mul: a = a * k
+///      a = a * x.
+pub const MUL = 0x20;
+/// div: a = a / k
+///      a = a / x.
+/// Truncated division.
+pub const DIV = 0x30;
+/// or:  a = a | k
+///      a = a | x.
+pub const OR = 0x40;
+/// and: a = a & k
+///      a = a & x.
+pub const AND = 0x50;
+/// lsh: a = a << k
+///      a = a << x.
+/// a = a << k, a = a << x.
+pub const LSH = 0x60;
+/// rsh: a = a >> k
+///      a = a >> x.
+pub const RSH = 0x70;
+/// neg: a = -a.
+/// Note that this isn't a binary negation, rather the value of `~a + 1`.
+pub const NEG = 0x80;
+/// mod: a = a % k
+///      a = a % x.
+/// EXTENSION. Defined for:
+///  - Linux.
+///  - NetBSD + Minix 3.
+///  - FreeBSD and derivitives.
+pub const MOD = 0x90;
+/// xor: a = a ^ k
+///      a = a ^ x.
+/// EXTENSION. Defined for:
+///  - Linux.
+///  - NetBSD + Minix 3.
+///  - FreeBSD and derivitives.
+pub const XOR = 0xa0;
+
+// Jump operations using a comparison between a and x/k.
+/// jmp L: pc += k.
+/// No comparison done here.
+pub const JA = 0x00;
+/// jeq    #k, Lt, Lf: pc += (a == k)    ? jt : jf.
+/// jeq     x, Lt, Lf: pc += (a == x)    ? jt : jf.
+pub const JEQ = 0x10;
+/// jgt    #k, Lt, Lf: pc += (a >  k)    ? jt : jf.
+/// jgt     x, Lt, Lf: pc += (a >  x)    ? jt : jf.
+pub const JGT = 0x20;
+/// jge    #k, Lt, Lf: pc += (a >= k)    ? jt : jf.
+/// jge     x, Lt, Lf: pc += (a >= x)    ? jt : jf.
+pub const JGE = 0x30;
+/// jset   #k, Lt, Lf: pc += (a & k > 0) ? jt : jf.
+/// jset    x, Lt, Lf: pc += (a & x > 0) ? jt : jf. 
+pub const JSET = 0x40;
+
+// Miscellaneous operations/register copy.
+/// tax: x = a.
+pub const TAX = 0x00;
+/// txa: a = x.
+pub const TXA = 0x80;
+
+/// The 16 registers in the scratch memory store as named enums.
+pub const Scratch = enum(u4) { m0, m1, m2, m3, m4, m5, m6, m7, m8, m9, m10, m11, m12, m13, m14, m15 };
+pub const MEMWORDS = 16;
+pub const MAXINSNS = switch (builtin.os.tag) {
+    .linux => 4096,
+    else => 512,
+};
+pub const MINBUFSIZE = 32;
+pub const MAXBUFSIZE = 1 << 21;
+
+pub const Insn = extern struct {
+    opcode: u16,
+    jt: u8,
+    jf: u8,
+    k: u32,
+
+    /// Implements the `std.fmt.format` API.
+    /// The formatting is similar to the output of tcpdump -dd.
+    pub fn format(
+        self: Insn,
+        comptime layout: []const u8,
+        opts: std.fmt.FormatOptions,
+        writer: anytype,
+    ) !void {
+        _ = opts;
+        if (comptime layout.len != 0 and layout[0] != 's')
+            @compileError("Unsupported format specifier for BPF Insn type '" ++ layout ++ "'.");
+
+        try std.fmt.format(
+            writer,
+            "Insn{{ 0x{X:0<2}, {d}, {d}, 0x{X:0<8} }}",
+            .{ self.opcode, self.jt, self.jf, self.k },
+        );
+    }
+
+    const Size = enum(u8) {
+        word = W,
+        half_word = H,
+        byte = B,
+    };
+
+    fn stmt(opcode: u16, k: u32) Insn {
+        return .{
+            .opcode = opcode,
+            .jt = 0,
+            .jf = 0,
+            .k = k,
+        };
+    }
+
+    pub fn ld_imm(value: u32) Insn {
+        return stmt(LD | IMM, value);
+    }
+
+    pub fn ld_abs(size: Size, offset: u32) Insn {
+        return stmt(LD | ABS | @enumToInt(size), offset);
+    }
+
+    pub fn ld_ind(size: Size, offset: u32) Insn {
+        return stmt(LD | IND | @enumToInt(size), offset);
+    }
+
+    pub fn ld_mem(reg: Scratch) Insn {
+        return stmt(LD | MEM, @enumToInt(reg));
+    }
+
+    pub fn ld_len() Insn {
+        return stmt(LD | LEN | W, 0);
+    }
+
+    pub fn ld_rnd() Insn {
+        return stmt(LD | RND | W, 0);
+    }
+
+    pub fn ldx_imm(value: u32) Insn {
+        return stmt(LDX | IMM, value);
+    }
+
+    pub fn ldx_mem(reg: Scratch) Insn {
+        return stmt(LDX | MEM, @enumToInt(reg));
+    }
+
+    pub fn ldx_len() Insn {
+        return stmt(LDX | LEN | W, 0);
+    }
+
+    pub fn ldx_msh(offset: u32) Insn {
+        return stmt(LDX | MSH | B, offset);
+    }
+
+    pub fn st(reg: Scratch) Insn {
+        return stmt(ST, @enumToInt(reg));
+    }
+    pub fn stx(reg: Scratch) Insn {
+        return stmt(STX, @enumToInt(reg));
+    }
+
+    const AluOp = enum(u16) {
+        add = ADD,
+        sub = SUB,
+        mul = MUL,
+        div = DIV,
+        @"or" = OR,
+        @"and" = AND,
+        lsh = LSH,
+        rsh = RSH,
+        mod = MOD,
+        xor = XOR,
+    };
+
+    const Source = enum(u16) {
+        k = K,
+        x = X,
+    };
+    const KOrX = union(Source) {
+        k: u32,
+        x: void,
+    };
+
+    pub fn alu_neg() Insn {
+        return stmt(ALU | NEG, 0);
+    }
+
+    pub fn alu(op: AluOp, source: KOrX) Insn {
+        return stmt(
+            ALU | @enumToInt(op) | @enumToInt(source),
+            if (source == .k) source.k else 0,
+        );
+    }
+
+    const JmpOp = enum(u16) {
+        jeq = JEQ,
+        jgt = JGT,
+        jge = JGE,
+        jset = JSET,
+    };
+
+    pub fn jmp_ja(location: u32) Insn {
+        return stmt(JMP | JA, location);
+    }
+
+    pub fn jmp(op: JmpOp, source: KOrX, jt: u8, jf: u8) Insn {
+        return Insn{
+            .opcode = JMP | @enumToInt(op) | @enumToInt(source),
+            .jt = jt,
+            .jf = jf,
+            .k = if (source == .k) source.k else 0,
+        };
+    }
+
+    const Verdict = enum(u16) {
+        k = K,
+        a = A,
+    };
+    const KOrA = union(Verdict) {
+        k: u32,
+        a: void,
+    };
+
+    pub fn ret(verdict: KOrA) Insn {
+        return stmt(
+            RET | @enumToInt(verdict),
+            if (verdict == .k) verdict.k else 0,
+        );
+    }
+
+    pub fn tax() Insn {
+        return stmt(MISC | TAX, 0);
+    }
+
+    pub fn txa() Insn {
+        return stmt(MISC | TXA, 0);
+    }
+};
+
+fn opcodeEqual(opcode: u16, insn: Insn) !void {
+    try expectEqual(opcode, insn.opcode);
+}
+
+test "opcodes" {
+    try opcodeEqual(0x00, Insn.ld_imm(0));
+    try opcodeEqual(0x20, Insn.ld_abs(.word, 0));
+    try opcodeEqual(0x28, Insn.ld_abs(.half_word, 0));
+    try opcodeEqual(0x30, Insn.ld_abs(.byte, 0));
+    try opcodeEqual(0x40, Insn.ld_ind(.word, 0));
+    try opcodeEqual(0x48, Insn.ld_ind(.half_word, 0));
+    try opcodeEqual(0x50, Insn.ld_ind(.byte, 0));
+    try opcodeEqual(0x60, Insn.ld_mem(.m0));
+    try opcodeEqual(0x80, Insn.ld_len());
+    try opcodeEqual(0xc0, Insn.ld_rnd());
+
+    try opcodeEqual(0x01, Insn.ldx_imm(0));
+    try opcodeEqual(0x61, Insn.ldx_mem(.m0));
+    try opcodeEqual(0x81, Insn.ldx_len());
+    try opcodeEqual(0xb1, Insn.ldx_msh(0));
+
+    try opcodeEqual(0x02, Insn.st(.m0));
+    try opcodeEqual(0x03, Insn.stx(.m0));
+
+    try opcodeEqual(0x04, Insn.alu(.add, .{ .k = 0 }));
+    try opcodeEqual(0x14, Insn.alu(.sub, .{ .k = 0 }));
+    try opcodeEqual(0x24, Insn.alu(.mul, .{ .k = 0 }));
+    try opcodeEqual(0x34, Insn.alu(.div, .{ .k = 0 }));
+    try opcodeEqual(0x44, Insn.alu(.@"or", .{ .k = 0 }));
+    try opcodeEqual(0x54, Insn.alu(.@"and", .{ .k = 0 }));
+    try opcodeEqual(0x64, Insn.alu(.lsh, .{ .k = 0 }));
+    try opcodeEqual(0x74, Insn.alu(.rsh, .{ .k = 0 }));
+    try opcodeEqual(0x94, Insn.alu(.mod, .{ .k = 0 }));
+    try opcodeEqual(0xa4, Insn.alu(.xor, .{ .k = 0 }));
+    try opcodeEqual(0x84, Insn.alu_neg());
+    try opcodeEqual(0x0c, Insn.alu(.add, .x));
+    try opcodeEqual(0x1c, Insn.alu(.sub, .x));
+    try opcodeEqual(0x2c, Insn.alu(.mul, .x));
+    try opcodeEqual(0x3c, Insn.alu(.div, .x));
+    try opcodeEqual(0x4c, Insn.alu(.@"or", .x));
+    try opcodeEqual(0x5c, Insn.alu(.@"and", .x));
+    try opcodeEqual(0x6c, Insn.alu(.lsh, .x));
+    try opcodeEqual(0x7c, Insn.alu(.rsh, .x));
+    try opcodeEqual(0x9c, Insn.alu(.mod, .x));
+    try opcodeEqual(0xac, Insn.alu(.xor, .x));
+
+    try opcodeEqual(0x05, Insn.jmp_ja(0));
+    try opcodeEqual(0x15, Insn.jmp(.jeq, .{ .k = 0 }, 0, 0));
+    try opcodeEqual(0x25, Insn.jmp(.jgt, .{ .k = 0 }, 0, 0));
+    try opcodeEqual(0x35, Insn.jmp(.jge, .{ .k = 0 }, 0, 0));
+    try opcodeEqual(0x45, Insn.jmp(.jset, .{ .k = 0 }, 0, 0));
+    try opcodeEqual(0x1d, Insn.jmp(.jeq, .x, 0, 0));
+    try opcodeEqual(0x2d, Insn.jmp(.jgt, .x, 0, 0));
+    try opcodeEqual(0x3d, Insn.jmp(.jge, .x, 0, 0));
+    try opcodeEqual(0x4d, Insn.jmp(.jset, .x, 0, 0));
+
+    try opcodeEqual(0x06, Insn.ret(.{ .k = 0 }));
+    try opcodeEqual(0x16, Insn.ret(.a));
+
+    try opcodeEqual(0x07, Insn.tax());
+    try opcodeEqual(0x87, Insn.txa());
+}
+
+pub const Error = error{
+    InvalidOpcode,
+    InvalidOffset,
+    InvalidLocation,
+    DivisionByZero,
+    NoReturn,
+};
+
+/// A simple implementation of the BPF virtual-machine.
+/// Use this to run/debug programs.
+pub fn simulate(
+    packet: []const u8,
+    filter: []const Insn,
+    byte_order: std.builtin.Endian,
+) Error!u32 {
+    assert(filter.len > 0 and filter.len < MAXINSNS);
+    assert(packet.len < MAXBUFSIZE);
+    const len = @intCast(u32, packet.len);
+
+    var a: u32 = 0;
+    var x: u32 = 0;
+    var m = mem.zeroes([MEMWORDS]u32);
+    var pc: usize = 0;
+
+    while (pc < filter.len) : (pc += 1) {
+        const i = filter[pc];
+        // Cast to a wider type to protect against overflow.
+        const k = @as(u64, i.k);
+        const remaining = filter.len - (pc + 1);
+
+        // Do validation/error checking here to compress the second switch.
+        switch (i.opcode) {
+            LD | ABS | W => if (k + @sizeOf(u32) - 1 >= packet.len) return error.InvalidOffset,
+            LD | ABS | H => if (k + @sizeOf(u16) - 1 >= packet.len) return error.InvalidOffset,
+            LD | ABS | B => if (k >= packet.len) return error.InvalidOffset,
+            LD | IND | W => if (k + x + @sizeOf(u32) - 1 >= packet.len) return error.InvalidOffset,
+            LD | IND | H => if (k + x + @sizeOf(u16) - 1 >= packet.len) return error.InvalidOffset,
+            LD | IND | B => if (k + x >= packet.len) return error.InvalidOffset,
+
+            LDX | MSH | B => if (k >= packet.len) return error.InvalidOffset,
+            ST, STX, LD | MEM, LDX | MEM => if (i.k >= MEMWORDS) return error.InvalidOffset,
+
+            JMP | JA => if (remaining <= i.k) return error.InvalidOffset,
+            JMP | JEQ | K,
+            JMP | JGT | K,
+            JMP | JGE | K,
+            JMP | JSET | K,
+            JMP | JEQ | X,
+            JMP | JGT | X,
+            JMP | JGE | X,
+            JMP | JSET | X,
+            => if (remaining <= i.jt or remaining <= i.jf) return error.InvalidLocation,
+            else => {},
+        }
+        switch (i.opcode) {
+            LD | IMM => a = i.k,
+            LD | MEM => a = m[i.k],
+            LD | LEN | W => a = len,
+            LD | RND | W => a = random.int(u32),
+            LD | ABS | W => a = mem.readInt(u32, packet[i.k..][0..@sizeOf(u32)], byte_order),
+            LD | ABS | H => a = mem.readInt(u16, packet[i.k..][0..@sizeOf(u16)], byte_order),
+            LD | ABS | B => a = packet[i.k],
+            LD | IND | W => a = mem.readInt(u32, packet[i.k + x ..][0..@sizeOf(u32)], byte_order),
+            LD | IND | H => a = mem.readInt(u16, packet[i.k + x ..][0..@sizeOf(u16)], byte_order),
+            LD | IND | B => a = packet[i.k + x],
+
+            LDX | IMM => x = i.k,
+            LDX | MEM => x = m[i.k],
+            LDX | LEN | W => x = len,
+            LDX | MSH | B => x = @as(u32, @truncate(u4, packet[i.k])) << 2,
+
+            ST => m[i.k] = a,
+            STX => m[i.k] = x,
+
+            ALU | ADD | K => a +%= i.k,
+            ALU | SUB | K => a -%= i.k,
+            ALU | MUL | K => a *%= i.k,
+            ALU | DIV | K => a = try math.divTrunc(u32, a, i.k),
+            ALU | OR | K => a |= i.k,
+            ALU | AND | K => a &= i.k,
+            ALU | LSH | K => a = math.shl(u32, a, i.k),
+            ALU | RSH | K => a = math.shr(u32, a, i.k),
+            ALU | MOD | K => a = try math.mod(u32, a, i.k),
+            ALU | XOR | K => a ^= i.k,
+            ALU | ADD | X => a +%= x,
+            ALU | SUB | X => a -%= x,
+            ALU | MUL | X => a *%= x,
+            ALU | DIV | X => a = try math.divTrunc(u32, a, x),
+            ALU | OR | X => a |= x,
+            ALU | AND | X => a &= x,
+            ALU | LSH | X => a = math.shl(u32, a, x),
+            ALU | RSH | X => a = math.shr(u32, a, x),
+            ALU | MOD | X => a = try math.mod(u32, a, x),
+            ALU | XOR | X => a ^= x,
+            ALU | NEG => a = @bitCast(u32, -%@bitCast(i32, a)),
+
+            JMP | JA => pc += i.k,
+            JMP | JEQ | K => pc += if (a == i.k) i.jt else i.jf,
+            JMP | JGT | K => pc += if (a > i.k) i.jt else i.jf,
+            JMP | JGE | K => pc += if (a >= i.k) i.jt else i.jf,
+            JMP | JSET | K => pc += if (a & i.k > 0) i.jt else i.jf,
+            JMP | JEQ | X => pc += if (a == x) i.jt else i.jf,
+            JMP | JGT | X => pc += if (a > x) i.jt else i.jf,
+            JMP | JGE | X => pc += if (a >= x) i.jt else i.jf,
+            JMP | JSET | X => pc += if (a & x > 0) i.jt else i.jf,
+
+            RET | K => return i.k,
+            RET | A => return a,
+
+            MISC | TAX => x = a,
+            MISC | TXA => a = x,
+            else => return error.InvalidOpcode,
+        }
+    }
+
+    return error.NoReturn;
+}
+
+// This program is the BPF form of the tcpdump filter:
+//
+//     tcpdump -dd 'ip host mirror.internode.on.net and tcp port ftp-data'
+//
+// As of January 2022, mirror.internode.on.net resolves to 150.101.135.3
+//
+// For reference, here's what it looks like in BPF assembler.
+// Note that the jumps are used for TCP/IP layer checks.
+//
+// ```
+//       ldh [12] (#proto)
+//       jeq #0x0800 (ETHERTYPE_IP), L1, fail
+// L1:   ld [26]
+//       jeq #150.101.135.3, L2, dest
+// dest: ld [30]
+//       jeq #150.101.135.3, L2, fail
+// L2:   ldb [23]
+//       jeq #0x6 (IPPROTO_TCP), L3, fail
+// L3:   ldh [20]
+//       jset #0x1fff, fail, plen
+// plen: ldx 4 * ([14] & 0xf)
+//       ldh [x + 14]
+//       jeq  #0x14 (FTP), pass, dstp
+// dstp: ldh [x + 16]
+//       jeq  #0x14 (FTP), pass, fail
+// pass: ret #0x40000
+// fail: ret #0
+// ```
+const tcpdump_filter = [_]Insn{
+    Insn.ld_abs(.half_word, 12),
+    Insn.jmp(.jeq, .{ .k = 0x800 }, 0, 14),
+    Insn.ld_abs(.word, 26),
+    Insn.jmp(.jeq, .{ .k = 0x96658703 }, 2, 0),
+    Insn.ld_abs(.word, 30),
+    Insn.jmp(.jeq, .{ .k = 0x96658703 }, 0, 10),
+    Insn.ld_abs(.byte, 23),
+    Insn.jmp(.jeq, .{ .k = 0x6 }, 0, 8),
+    Insn.ld_abs(.half_word, 20),
+    Insn.jmp(.jset, .{ .k = 0x1fff }, 6, 0),
+    Insn.ldx_msh(14),
+    Insn.ld_ind(.half_word, 14),
+    Insn.jmp(.jeq, .{ .k = 0x14 }, 2, 0),
+    Insn.ld_ind(.half_word, 16),
+    Insn.jmp(.jeq, .{ .k = 0x14 }, 0, 1),
+    Insn.ret(.{ .k = 0x40000 }),
+    Insn.ret(.{ .k = 0 }),
+};
+
+// This packet is the output of `ls` on mirror.internode.on.net:/, captured
+// using the filter above.
+//
+// zig fmt: off
+const ftp_data = [_]u8{
+    // ethernet - 14 bytes: IPv4(0x0800) from a4:71:74:ad:4b:f0 -> de:ad:be:ef:f0:0f
+    0xde, 0xad, 0xbe, 0xef, 0xf0, 0x0f, 0xa4, 0x71, 0x74, 0xad, 0x4b, 0xf0, 0x08, 0x00,
+    // IPv4 - 20 bytes: TCP data from 150.101.135.3 -> 192.168.1.3
+    0x45, 0x00, 0x01, 0xf2, 0x70, 0x3b, 0x40, 0x00, 0x37, 0x06, 0xf2, 0xb6,
+    0x96, 0x65, 0x87, 0x03, 0xc0, 0xa8, 0x01, 0x03,
+    // TCP - 32 bytes: Source port: 20 (FTP). Payload = 446 bytes
+    0x00, 0x14, 0x80, 0x6d, 0x35, 0x81, 0x2d, 0x40, 0x4f, 0x8a, 0x29, 0x9e, 0x80, 0x18, 0x00, 0x2e,
+    0x88, 0x8d, 0x00, 0x00, 0x01, 0x01, 0x08, 0x0a, 0x0b, 0x59, 0x5d, 0x09, 0x32, 0x8b, 0x51, 0xa0
+} ++
+    // Raw line-based FTP data - 446 bytes
+    "lrwxrwxrwx   1 root     root           12 Feb 14  2012 debian -> .pub2/debian\r\n" ++
+    "lrwxrwxrwx   1 root     root           15 Feb 14  2012 debian-cd -> .pub2/debian-cd\r\n" ++
+    "lrwxrwxrwx   1 root     root            9 Mar  9  2018 linux -> pub/linux\r\n" ++
+    "drwxr-xr-X   3 mirror   mirror       4096 Sep 20 08:10 pub\r\n" ++
+    "lrwxrwxrwx   1 root     root           12 Feb 14  2012 ubuntu -> .pub2/ubuntu\r\n" ++
+    "-rw-r--r--   1 root     root         1044 Jan 20  2015 welcome.msg\r\n";
+// zig fmt: on
+
+test "tcpdump filter" {
+    try expectEqual(
+        @as(u32, 0x40000),
+        try simulate(ftp_data, &tcpdump_filter, .Big),
+    );
+}
+
+fn expectPass(data: anytype, filter: []Insn) !void {
+    try expectEqual(
+        @as(u32, 0),
+        try simulate(mem.asBytes(data), filter, .Big),
+    );
+}
+
+fn expectFail(expected_error: anyerror, data: anytype, filter: []Insn) !void {
+    try expectError(
+        expected_error,
+        simulate(mem.asBytes(data), filter, native_endian),
+    );
+}
+
+test "simulator coverage" {
+    const some_data: packed struct {
+        foo: u32,
+        bar: u8,
+    } = .{
+        .foo = mem.nativeToBig(u32, 0xaabbccdd),
+        .bar = 0x7f,
+    };
+
+    try expectPass(&some_data, &.{
+        // ld  #10
+        // ldx #1
+        // st M[0]
+        // stx M[1]
+        // fail if A != 10
+        Insn.ld_imm(10),
+        Insn.ldx_imm(1),
+        Insn.st(.m0),
+        Insn.stx(.m1),
+        Insn.jmp(.jeq, .{ .k = 10 }, 1, 0),
+        Insn.ret(.{ .k = 1 }),
+        // ld [0]
+        // fail if A != 0xaabbccdd
+        Insn.ld_abs(.word, 0),
+        Insn.jmp(.jeq, .{ .k = 0xaabbccdd }, 1, 0),
+        Insn.ret(.{ .k = 2 }),
+        // ldh [0]
+        // fail if A != 0xaabb
+        Insn.ld_abs(.half_word, 0),
+        Insn.jmp(.jeq, .{ .k = 0xaabb }, 1, 0),
+        Insn.ret(.{ .k = 3 }),
+        // ldb [0]
+        // fail if A != 0xaa
+        Insn.ld_abs(.byte, 0),
+        Insn.jmp(.jeq, .{ .k = 0xaa }, 1, 0),
+        Insn.ret(.{ .k = 4 }),
+        // ld [x + 0]
+        // fail if A != 0xbbccdd7f
+        Insn.ld_ind(.word, 0),
+        Insn.jmp(.jeq, .{ .k = 0xbbccdd7f }, 1, 0),
+        Insn.ret(.{ .k = 5 }),
+        // ldh [x + 0]
+        // fail if A != 0xbbcc
+        Insn.ld_ind(.half_word, 0),
+        Insn.jmp(.jeq, .{ .k = 0xbbcc }, 1, 0),
+        Insn.ret(.{ .k = 6 }),
+        // ldb [x + 0]
+        // fail if A != 0xbb
+        Insn.ld_ind(.byte, 0),
+        Insn.jmp(.jeq, .{ .k = 0xbb }, 1, 0),
+        Insn.ret(.{ .k = 7 }),
+        // ld M[0]
+        // fail if A != 10
+        Insn.ld_mem(.m0),
+        Insn.jmp(.jeq, .{ .k = 10 }, 1, 0),
+        Insn.ret(.{ .k = 8 }),
+        // ld #len
+        // fail if A != 5
+        Insn.ld_len(),
+        Insn.jmp(.jeq, .{ .k = @sizeOf(@TypeOf(some_data)) }, 1, 0),
+        Insn.ret(.{ .k = 9 }),
+        // ld #0
+        // ld arc4random()
+        // fail if A == 0
+        Insn.ld_imm(0),
+        Insn.ld_rnd(),
+        Insn.jmp(.jgt, .{ .k = 0 }, 1, 0),
+        Insn.ret(.{ .k = 10 }),
+        // ld  #3
+        // ldx #10
+        // st M[2]
+        // txa
+        // fail if a != x
+        Insn.ld_imm(3),
+        Insn.ldx_imm(10),
+        Insn.st(.m2),
+        Insn.txa(),
+        Insn.jmp(.jeq, .x, 1, 0),
+        Insn.ret(.{ .k = 11 }),
+        // ldx M[2]
+        // fail if A <= X
+        Insn.ldx_mem(.m2),
+        Insn.jmp(.jgt, .x, 1, 0),
+        Insn.ret(.{ .k = 12 }),
+        // ldx #len
+        // fail if a <= x
+        Insn.ldx_len(),
+        Insn.jmp(.jgt, .x, 1, 0),
+        Insn.ret(.{ .k = 13 }),
+        // a = 4 * (0x7f & 0xf)
+        // x = 4 * ([4]  & 0xf)
+        // fail if a != x
+        Insn.ld_imm(4 * (0x7f & 0xf)),
+        Insn.ldx_msh(4),
+        Insn.jmp(.jeq, .x, 1, 0),
+        Insn.ret(.{ .k = 14 }),
+        // ld  #(u32)-1
+        // ldx #2
+        // add #1
+        // fail if a != 0
+        Insn.ld_imm(0xffffffff),
+        Insn.ldx_imm(2),
+        Insn.alu(.add, .{ .k = 1 }),
+        Insn.jmp(.jeq, .{ .k = 0 }, 1, 0),
+        Insn.ret(.{ .k = 15 }),
+        // sub #1
+        // fail if a != (u32)-1
+        Insn.alu(.sub, .{ .k = 1 }),
+        Insn.jmp(.jeq, .{ .k = 0xffffffff }, 1, 0),
+        Insn.ret(.{ .k = 16 }),
+        // add x
+        // fail if a != 1
+        Insn.alu(.add, .x),
+        Insn.jmp(.jeq, .{ .k = 1 }, 1, 0),
+        Insn.ret(.{ .k = 17 }),
+        // sub x
+        // fail if a != (u32)-1
+        Insn.alu(.sub, .x),
+        Insn.jmp(.jeq, .{ .k = 0xffffffff }, 1, 0),
+        Insn.ret(.{ .k = 18 }),
+        // ld #16
+        // mul #2
+        // fail if a != 32
+        Insn.ld_imm(16),
+        Insn.alu(.mul, .{ .k = 2 }),
+        Insn.jmp(.jeq, .{ .k = 32 }, 1, 0),
+        Insn.ret(.{ .k = 19 }),
+        // mul x
+        // fail if a != 64
+        Insn.alu(.mul, .x),
+        Insn.jmp(.jeq, .{ .k = 64 }, 1, 0),
+        Insn.ret(.{ .k = 20 }),
+        // div #2
+        // fail if a != 32
+        Insn.alu(.div, .{ .k = 2 }),
+        Insn.jmp(.jeq, .{ .k = 32 }, 1, 0),
+        Insn.ret(.{ .k = 21 }),
+        // div x
+        // fail if a != 16
+        Insn.alu(.div, .x),
+        Insn.jmp(.jeq, .{ .k = 16 }, 1, 0),
+        Insn.ret(.{ .k = 22 }),
+        // or #4
+        // fail if a != 20
+        Insn.alu(.@"or", .{ .k = 4 }),
+        Insn.jmp(.jeq, .{ .k = 20 }, 1, 0),
+        Insn.ret(.{ .k = 23 }),
+        // or x
+        // fail if a != 22
+        Insn.alu(.@"or", .x),
+        Insn.jmp(.jeq, .{ .k = 22 }, 1, 0),
+        Insn.ret(.{ .k = 24 }),
+        // and #6
+        // fail if a != 6
+        Insn.alu(.@"and", .{ .k = 0b110 }),
+        Insn.jmp(.jeq, .{ .k = 6 }, 1, 0),
+        Insn.ret(.{ .k = 25 }),
+        // and x
+        // fail if a != 2
+        Insn.alu(.@"and", .x),
+        Insn.jmp(.jeq, .x, 1, 0),
+        Insn.ret(.{ .k = 26 }),
+        // xor #15
+        // fail if a != 13
+        Insn.alu(.xor, .{ .k = 0b1111 }),
+        Insn.jmp(.jeq, .{ .k = 0b1101 }, 1, 0),
+        Insn.ret(.{ .k = 27 }),
+        // xor x
+        // fail if a != 15
+        Insn.alu(.xor, .x),
+        Insn.jmp(.jeq, .{ .k = 0b1111 }, 1, 0),
+        Insn.ret(.{ .k = 28 }),
+        // rsh #1
+        // fail if a != 7
+        Insn.alu(.rsh, .{ .k = 1 }),
+        Insn.jmp(.jeq, .{ .k = 0b0111 }, 1, 0),
+        Insn.ret(.{ .k = 29 }),
+        // rsh x
+        // fail if a != 1
+        Insn.alu(.rsh, .x),
+        Insn.jmp(.jeq, .{ .k = 0b0001 }, 1, 0),
+        Insn.ret(.{ .k = 30 }),
+        // lsh #1
+        // fail if a != 2
+        Insn.alu(.lsh, .{ .k = 1 }),
+        Insn.jmp(.jeq, .{ .k = 0b0010 }, 1, 0),
+        Insn.ret(.{ .k = 31 }),
+        // lsh x
+        // fail if a != 8
+        Insn.alu(.lsh, .x),
+        Insn.jmp(.jeq, .{ .k = 0b1000 }, 1, 0),
+        Insn.ret(.{ .k = 32 }),
+        // mod 6
+        // fail if a != 2
+        Insn.alu(.mod, .{ .k = 6 }),
+        Insn.jmp(.jeq, .{ .k = 2 }, 1, 0),
+        Insn.ret(.{ .k = 33 }),
+        // mod x
+        // fail if a != 0
+        Insn.alu(.mod, .x),
+        Insn.jmp(.jeq, .{ .k = 0 }, 1, 0),
+        Insn.ret(.{ .k = 34 }),
+        // tax
+        // neg
+        // fail if a != (u32)-2
+        Insn.txa(),
+        Insn.alu_neg(),
+        Insn.jmp(.jeq, .{ .k = ~@as(u32, 2) + 1 }, 1, 0),
+        Insn.ret(.{ .k = 35 }),
+        // ja #1 (skip the next instruction)
+        Insn.jmp_ja(1),
+        Insn.ret(.{ .k = 36 }),
+        // ld #20
+        // tax
+        // fail if a != 20
+        // fail if a != x
+        Insn.ld_imm(20),
+        Insn.tax(),
+        Insn.jmp(.jeq, .{ .k = 20 }, 1, 0),
+        Insn.ret(.{ .k = 37 }),
+        Insn.jmp(.jeq, .x, 1, 0),
+        Insn.ret(.{ .k = 38 }),
+        // ld #19
+        // fail if a == 20
+        // fail if a == x
+        // fail if a >= 20
+        // fail if a >= X
+        Insn.ld_imm(19),
+        Insn.jmp(.jeq, .{ .k = 20 }, 0, 1),
+        Insn.ret(.{ .k = 39 }),
+        Insn.jmp(.jeq, .x, 0, 1),
+        Insn.ret(.{ .k = 40 }),
+        Insn.jmp(.jgt, .{ .k = 20 }, 0, 1),
+        Insn.ret(.{ .k = 41 }),
+        Insn.jmp(.jgt, .x, 0, 1),
+        Insn.ret(.{ .k = 42 }),
+        // ld #21
+        // fail if a < 20
+        // fail if a < x
+        Insn.ld_imm(21),
+        Insn.jmp(.jgt, .{ .k = 20 }, 1, 0),
+        Insn.ret(.{ .k = 43 }),
+        Insn.jmp(.jgt, .x, 1, 0),
+        Insn.ret(.{ .k = 44 }),
+        // ldx #22
+        // fail if a < 22
+        // fail if a < x
+        Insn.ldx_imm(22),
+        Insn.jmp(.jge, .{ .k = 22 }, 0, 1),
+        Insn.ret(.{ .k = 45 }),
+        Insn.jmp(.jge, .x, 0, 1),
+        Insn.ret(.{ .k = 46 }),
+        // ld #23
+        // fail if a >= 22
+        // fail if a >= x
+        Insn.ld_imm(23),
+        Insn.jmp(.jge, .{ .k = 22 }, 1, 0),
+        Insn.ret(.{ .k = 47 }),
+        Insn.jmp(.jge, .x, 1, 0),
+        Insn.ret(.{ .k = 48 }),
+        // ldx #0b10100
+        // fail if a & 0b10100 == 0
+        // fail if a & x       == 0
+        Insn.ldx_imm(0b10100),
+        Insn.jmp(.jset, .{ .k = 0b10100 }, 1, 0),
+        Insn.ret(.{ .k = 47 }),
+        Insn.jmp(.jset, .x, 1, 0),
+        Insn.ret(.{ .k = 48 }),
+        // ldx #0
+        // fail if a & 0 > 0
+        // fail if a & x > 0
+        Insn.ldx_imm(0),
+        Insn.jmp(.jset, .{ .k = 0 }, 0, 1),
+        Insn.ret(.{ .k = 49 }),
+        Insn.jmp(.jset, .x, 0, 1),
+        Insn.ret(.{ .k = 50 }),
+        Insn.ret(.{ .k = 0 }),
+    });
+    try expectPass(&some_data, &.{
+        Insn.ld_imm(35),
+        Insn.ld_imm(0),
+        Insn.ret(.a),
+    });
+
+    // Errors
+    try expectFail(error.NoReturn, &some_data, &.{
+        Insn.ld_imm(10),
+    });
+    try expectFail(error.InvalidOpcode, &some_data, &.{
+        Insn.stmt(0x7f, 0xdeadbeef),
+    });
+    try expectFail(error.InvalidOffset, &some_data, &.{
+        Insn.stmt(LD | ABS | W, 10),
+    });
+    try expectFail(error.InvalidLocation, &some_data, &.{
+        Insn.jmp(.jeq, .{ .k = 0 }, 10, 0),
+    });
+    try expectFail(error.InvalidLocation, &some_data, &.{
+        Insn.jmp(.jeq, .{ .k = 0 }, 0, 10),
+    });
+}
lib/std/x.zig
@@ -9,6 +9,7 @@ pub const os = struct {
 pub const net = struct {
     pub const ip = @import("x/net/ip.zig");
     pub const tcp = @import("x/net/tcp.zig");
+    pub const bpf = @import("x/net/bpf.zig");
 };
 
 test {