diff --git a/bpf/doc.go b/bpf/doc.go
index 633dc90e..a715ab26 100644
--- a/bpf/doc.go
+++ b/bpf/doc.go
@@ -2,10 +2,80 @@
 // Use of this source code is governed by a BSD-style
 // license that can be found in the LICENSE file.
 
-// Package bpf implements marshalling and unmarshalling of programs
-// for the Berkeley Packet Filter virtual machine.
-//
-// TODO: brief overview of the BPF virtual machine (registers, scratch, packet access, execution constraints)
-//
-// TODO: simple BPF program examples
+/*
+
+Package bpf implements marshalling and unmarshalling of programs
+for the Berkeley Packet Filter virtual machine.
+
+BPF's main use is to specify a packet filter for network taps, so
+that the kernel doesn't have to expensively copy every packet it
+sees to userspace. However, it's been repurposed to other areas
+where running user code in-kernel is needed. For example, Linux's
+seccomp uses BPF to apply security policies to system calls. For
+simplicity, this documentation refers only to packets, but other
+uses of BPF have their own data payloads.
+
+BPF programs run in a restricted virtual machine. It has almost no
+access to kernel functions, and while conditional branches are
+allowed, they can only jump forwards, to guarantee that there are no
+infinite loops.
+
+The virtual machine
+
+The BPF VM is an accumulator machine. Its main register, called
+register A, is an implicit source and destination in all arithmetic
+and logic operations. The machine also has 16 scratch registers for
+temporary storage, and an indirection register (register X) for
+indirect memory access. All registers are 32 bits wide.
+
+Each run of a BPF program is given one packet, which is placed in the
+VM's read-only "main memory". LoadAbsolute and LoadIndirect
+instructions can fetch up to 32 bits at a time into register A for
+examination.
+
+The goal of a BPF program is to produce and return a verdict (uint32),
+which tells the kernel what to do with the packet. In the context of
+packet filtering, the returned value is the number of bytes of the
+packet to forward to userspace, or 0 to ignore the packet. Other
+contexts like seccomp define their own return values.
+
+In order to simplify programs, attempts to read past the end of the
+packet terminate the program execution with a verdict of 0 (ignore
+packet). This means that the vast majority of BPF programs don't need
+to do any explicit bounds checking.
+
+In addition to the bytes of the packet, some BPF programs have access
+to extensions, which are essentially calls to kernel utility
+functions. Currently, the only extensions supported by this package
+are the Linux packet filter extensions.
+
+Examples
+
+This packet filter selects all ARP packets.
+
+	bpf.Assemble([]bpf.Instruction{
+		// Load "EtherType" field from the ethernet header.
+		bpf.LoadAbsolute{Off: 12, Size: 2},
+		// Skip over the next instruction if EtherType is not ARP.
+		bpf.JumpIf{Cond: bpf.JumpNotEqual, Val: 0x0806, SkipTrue: 1},
+		// Verdict is "send up to 4k of the packet to userspace."
+		bpf.RetConstant{Val: 4096},
+		// Verdict is "ignore packet."
+		bpf.RetConstant{Val: 0},
+	})
+
+This packet filter captures a random 1% sample of traffic.
+
+	bpf.Assemble([]bpf.Instruction{
+		// Get a 32-bit random number from the Linux kernel.
+		bpf.LoadExtension{Num: bpf.ExtRand},
+		// 1% dice roll?
+		bpf.JumpIf{Cond: bpf.JumpLessThan, Val: 2^32/100, SkipFalse: 1},
+		// Capture.
+		bpf.RetConstant{Val: 4096},
+		// Ignore.
+		bpf.RetConstant{Val: 0},
+	})
+
+*/
 package bpf // import "golang.org/x/net/bpf"