prog.go

package ebpf

import (
	"bytes"
	"encoding/binary"
	"errors"
	"fmt"
	"math"
	"path/filepath"
	"runtime"
	"time"

	"github.com/cilium/ebpf/asm"
	"github.com/cilium/ebpf/btf"
	"github.com/cilium/ebpf/internal"
	"github.com/cilium/ebpf/internal/kallsyms"
	"github.com/cilium/ebpf/internal/linux"
	"github.com/cilium/ebpf/internal/platform"
	"github.com/cilium/ebpf/internal/sys"
	"github.com/cilium/ebpf/internal/sysenc"
	"github.com/cilium/ebpf/internal/unix"
)

// ErrNotSupported is returned whenever the kernel doesn't support a feature.
var ErrNotSupported = internal.ErrNotSupported

// errBadRelocation is returned when the verifier rejects a program due to a
// bad CO-RE relocation.
//
// This error is detected based on heuristics and therefore may not be reliable.
var errBadRelocation = errors.New("bad CO-RE relocation")

// errUnknownKfunc is returned when the verifier rejects a program due to an
// unknown kfunc.
//
// This error is detected based on heuristics and therefore may not be reliable.
var errUnknownKfunc = errors.New("unknown kfunc")

// ProgramID represents the unique ID of an eBPF program.
type ProgramID = sys.ProgramID

const (
	// Number of bytes to pad the output buffer for BPF_PROG_TEST_RUN.
	// This is currently the maximum of spare space allocated for SKB
	// and XDP programs, and equal to XDP_PACKET_HEADROOM + NET_IP_ALIGN.
	outputPad = 256 + 2
)

// minVerifierLogSize is the default number of bytes allocated for the
// verifier log.
const minVerifierLogSize = 64 * 1024

// maxVerifierLogSize is the maximum size of verifier log buffer the kernel
// will accept before returning EINVAL. May be increased to MaxUint32 in the
// future, but avoid the unnecessary EINVAL for now.
const maxVerifierLogSize = math.MaxUint32 >> 2

// maxVerifierAttempts is the maximum number of times the verifier will retry
// loading a program with a growing log buffer before giving up. Since we double
// the log size on every attempt, this is the absolute maximum number of
// attempts before the buffer reaches [maxVerifierLogSize].
const maxVerifierAttempts = 30

// ProgramOptions control loading a program into the kernel.
type ProgramOptions struct {
	// Bitmap controlling the detail emitted by the kernel's eBPF verifier log.
	// LogLevel-type values can be ORed together to request specific kinds of
	// verifier output. See the documentation on [ebpf.LogLevel] for details.
	//
	//  opts.LogLevel = (ebpf.LogLevelBranch | ebpf.LogLevelStats)
	//
	// If left to its default value, the program will first be loaded without
	// verifier output enabled. Upon error, the program load will be repeated
	// with LogLevelBranch and the given (or default) LogSize value.
	//
	// Unless LogDisabled is set, setting this to a non-zero value will enable the verifier
	// log, populating the [ebpf.Program.VerifierLog] field on successful loads
	// and including detailed verifier errors if the program is rejected. This
	// will always allocate an output buffer, but will result in only a single
	// attempt at loading the program.
	LogLevel LogLevel

	// Starting size of the verifier log buffer. If the verifier log is larger
	// than this size, the buffer will be grown to fit the entire log. Leave at
	// its default value unless troubleshooting.
	LogSizeStart uint32

	// Disables the verifier log completely, regardless of other options.
	LogDisabled bool

	// Type information used for CO-RE relocations.
	//
	// This is useful in environments where the kernel BTF is not available
	// (containers) or where it is in a non-standard location. Defaults to
	// use the kernel BTF from a well-known location if nil.
	KernelTypes *btf.Spec

	// Type information used for CO-RE relocations of kernel modules,
	// indexed by module name.
	//
	// This is useful in environments where the kernel BTF is not available
	// (containers) or where it is in a non-standard location. Defaults to
	// use the kernel module BTF from a well-known location if nil.
	KernelModuleTypes map[string]*btf.Spec
}

// ProgramSpec defines a Program.
type ProgramSpec struct {
	// Name is passed to the kernel as a debug aid.
	//
	// Unsupported characters will be stripped.
	Name string

	// Type determines at which hook in the kernel a program will run.
	Type ProgramType

	// AttachType of the program, needed to differentiate allowed context
	// accesses in some newer program types like CGroupSockAddr.
	//
	// Available on kernels 4.17 and later.
	AttachType AttachType

	// Name of a kernel data structure or function to attach to. Its
	// interpretation depends on Type and AttachType.
	AttachTo string

	// The program to attach to. Must be provided manually.
	AttachTarget *Program

	// The name of the ELF section this program originated from.
	SectionName string

	Instructions asm.Instructions

	// Flags is passed to the kernel and specifies additional program
	// load attributes.
	Flags uint32

	// License of the program. Some helpers are only available if
	// the license is deemed compatible with the GPL.
	//
	// See https://www.kernel.org/doc/html/latest/process/license-rules.html#id1
	License string

	// Version used by Kprobe programs.
	//
	// Deprecated on kernels 5.0 and later. Leave empty to let the library
	// detect this value automatically.
	KernelVersion uint32

	// The byte order this program was compiled for, may be nil.
	ByteOrder binary.ByteOrder
}

// Copy returns a copy of the spec.
func (ps *ProgramSpec) Copy() *ProgramSpec {
	if ps == nil {
		return nil
	}

	cpy := *ps
	cpy.Instructions = make(asm.Instructions, len(ps.Instructions))
	copy(cpy.Instructions, ps.Instructions)
	return &cpy
}

// Tag calculates the kernel tag for a series of instructions.
//
// Use asm.Instructions.Tag if you need to calculate for non-native endianness.
func (ps *ProgramSpec) Tag() (string, error) {
	return ps.Instructions.Tag(internal.NativeEndian)
}

// kernelModule returns the kernel module providing the symbol in
// ProgramSpec.AttachTo, if any. Returns an empty string if the symbol is not
// present or not part of a kernel module.
func (ps *ProgramSpec) kernelModule() (string, error) {
	if ps.targetsKernelModule() {
		return kallsyms.Module(ps.AttachTo)
	}

	return "", nil
}

// targetsKernelModule returns true if the program supports being attached to a
// symbol provided by a kernel module.
func (ps *ProgramSpec) targetsKernelModule() bool {
	if ps.AttachTo == "" {
		return false
	}

	switch ps.Type {
	case Tracing:
		switch ps.AttachType {
		case AttachTraceFEntry, AttachTraceFExit:
			return true
		}
	case Kprobe:
		return true
	}

	return false
}

// VerifierError is returned by [NewProgram] and [NewProgramWithOptions] if a
// program is rejected by the verifier.
//
// Use [errors.As] to access the error.
type VerifierError = internal.VerifierError

// Program represents BPF program loaded into the kernel.
//
// It is not safe to close a Program which is used by other goroutines.
type Program struct {
	// Contains the output of the kernel verifier if enabled,
	// otherwise it is empty.
	VerifierLog string

	fd         *sys.FD
	name       string
	pinnedPath string
	typ        ProgramType
}

// NewProgram creates a new Program.
//
// See [NewProgramWithOptions] for details.
//
// Returns a [VerifierError] containing the full verifier log if the program is
// rejected by the kernel.
func NewProgram(spec *ProgramSpec) (*Program, error) {
	return NewProgramWithOptions(spec, ProgramOptions{})
}

// NewProgramWithOptions creates a new Program.
//
// Loading a program for the first time will perform
// feature detection by loading small, temporary programs.
//
// Returns a [VerifierError] containing the full verifier log if the program is
// rejected by the kernel.
func NewProgramWithOptions(spec *ProgramSpec, opts ProgramOptions) (*Program, error) {
	if spec == nil {
		return nil, errors.New("can't load a program from a nil spec")
	}

	prog, err := newProgramWithOptions(spec, opts, newBTFCache(&opts))
	if errors.Is(err, asm.ErrUnsatisfiedMapReference) {
		return nil, fmt.Errorf("cannot load program without loading its whole collection: %w", err)
	}
	return prog, err
}

var (
	coreBadLoad = []byte(fmt.Sprintf("(18) r10 = 0x%x\n", btf.COREBadRelocationSentinel))
	// This log message was introduced by ebb676daa1a3 ("bpf: Print function name in
	// addition to function id") which first appeared in v4.10 and has remained
	// unchanged since.
	coreBadCall  = []byte(fmt.Sprintf("invalid func unknown#%d\n", btf.COREBadRelocationSentinel))
	kfuncBadCall = []byte(fmt.Sprintf("invalid func unknown#%d\n", kfuncCallPoisonBase))
)

func newProgramWithOptions(spec *ProgramSpec, opts ProgramOptions, c *btf.Cache) (*Program, error) {
	if len(spec.Instructions) == 0 {
		return nil, errors.New("instructions cannot be empty")
	}

	if spec.Type == UnspecifiedProgram {
		return nil, errors.New("can't load program of unspecified type")
	}

	if spec.ByteOrder != nil && spec.ByteOrder != internal.NativeEndian {
		return nil, fmt.Errorf("can't load %s program on %s", spec.ByteOrder, internal.NativeEndian)
	}

	// Kernels before 5.0 (6c4fc209fcf9 "bpf: remove useless version check for prog load")
	// require the version field to be set to the value of the KERNEL_VERSION
	// macro for kprobe-type programs.
	// Overwrite Kprobe program version if set to zero or the magic version constant.
	kv := spec.KernelVersion
	if spec.Type == Kprobe && (kv == 0 || kv == internal.MagicKernelVersion) {
		v, err := linux.KernelVersion()
		if err != nil {
			return nil, fmt.Errorf("detecting kernel version: %w", err)
		}
		kv = v.Kernel()
	}

	p, progType := platform.DecodeConstant(spec.Type)
	if p != platform.Native {
		return nil, fmt.Errorf("program type %s (%s): %w", spec.Type, p, internal.ErrNotSupportedOnOS)
	}

	attr := &sys.ProgLoadAttr{
		ProgName:           maybeFillObjName(spec.Name),
		ProgType:           sys.ProgType(progType),
		ProgFlags:          spec.Flags,
		ExpectedAttachType: sys.AttachType(spec.AttachType),
		License:            sys.NewStringPointer(spec.License),
		KernVersion:        kv,
	}

	insns := make(asm.Instructions, len(spec.Instructions))
	copy(insns, spec.Instructions)

	kmodName, err := spec.kernelModule()
	if err != nil {
		return nil, fmt.Errorf("kernel module search: %w", err)
	}

	var b btf.Builder
	if err := applyRelocations(insns, kmodName, spec.ByteOrder, &b, c); err != nil {
		return nil, fmt.Errorf("apply CO-RE relocations: %w", err)
	}

	errExtInfos := haveProgramExtInfos()
	if !b.Empty() && errors.Is(errExtInfos, ErrNotSupported) {
		// There is at least one CO-RE relocation which relies on a stable local
		// type ID.
		// Return ErrNotSupported instead of E2BIG if there is no BTF support.
		return nil, errExtInfos
	}

	if errExtInfos == nil {
		// Only add func and line info if the kernel supports it. This allows
		// BPF compiled with modern toolchains to work on old kernels.
		fib, lib, err := btf.MarshalExtInfos(insns, &b)
		if err != nil {
			return nil, fmt.Errorf("marshal ext_infos: %w", err)
		}

		attr.FuncInfoRecSize = btf.FuncInfoSize
		attr.FuncInfoCnt = uint32(len(fib)) / btf.FuncInfoSize
		attr.FuncInfo = sys.SlicePointer(fib)

		attr.LineInfoRecSize = btf.LineInfoSize
		attr.LineInfoCnt = uint32(len(lib)) / btf.LineInfoSize
		attr.LineInfo = sys.SlicePointer(lib)
	}

	if !b.Empty() {
		handle, err := btf.NewHandle(&b)
		if err != nil {
			return nil, fmt.Errorf("load BTF: %w", err)
		}
		defer handle.Close()

		attr.ProgBtfFd = uint32(handle.FD())
	}

	kconfig, err := resolveKconfigReferences(insns)
	if err != nil {
		return nil, fmt.Errorf("resolve .kconfig: %w", err)
	}
	defer kconfig.Close()

	if err := resolveKsymReferences(insns); err != nil {
		return nil, fmt.Errorf("resolve .ksyms: %w", err)
	}

	if err := fixupAndValidate(insns); err != nil {
		return nil, err
	}

	handles, err := fixupKfuncs(insns)
	if err != nil {
		return nil, fmt.Errorf("fixing up kfuncs: %w", err)
	}
	defer handles.Close()

	if len(handles) > 0 {
		fdArray := handles.fdArray()
		attr.FdArray = sys.SlicePointer(fdArray)
	}

	buf := bytes.NewBuffer(make([]byte, 0, insns.Size()))
	err = insns.Marshal(buf, internal.NativeEndian)
	if err != nil {
		return nil, err
	}

	bytecode := buf.Bytes()
	attr.Insns = sys.SlicePointer(bytecode)
	attr.InsnCnt = uint32(len(bytecode) / asm.InstructionSize)

	if spec.AttachTarget != nil {
		targetID, err := findTargetInProgram(spec.AttachTarget, spec.AttachTo, spec.Type, spec.AttachType)
		if err != nil {
			return nil, fmt.Errorf("attach %s/%s: %w", spec.Type, spec.AttachType, err)
		}

		attr.AttachBtfId = targetID
		attr.AttachBtfObjFd = uint32(spec.AttachTarget.FD())
		defer runtime.KeepAlive(spec.AttachTarget)
	} else if spec.AttachTo != "" {
		module, targetID, err := findProgramTargetInKernel(spec.AttachTo, spec.Type, spec.AttachType)
		if err != nil && !errors.Is(err, errUnrecognizedAttachType) {
			// We ignore errUnrecognizedAttachType since AttachTo may be non-empty
			// for programs that don't attach anywhere.
			return nil, fmt.Errorf("attach %s/%s: %w", spec.Type, spec.AttachType, err)
		}

		attr.AttachBtfId = targetID
		if module != nil {
			attr.AttachBtfObjFd = uint32(module.FD())
			defer module.Close()
		}
	}

	if platform.IsWindows && opts.LogLevel != 0 {
		return nil, fmt.Errorf("log level: %w", internal.ErrNotSupportedOnOS)
	}

	var logBuf []byte
	var fd *sys.FD
	if opts.LogDisabled {
		// Loading with logging disabled should never retry.
		fd, err = sys.ProgLoad(attr)
		if err == nil {
			return &Program{"", fd, spec.Name, "", spec.Type}, nil
		}
	} else {
		// Only specify log size if log level is also specified. Setting size
		// without level results in EINVAL. Level will be bumped to LogLevelBranch
		// if the first load fails.
		if opts.LogLevel != 0 {
			attr.LogLevel = opts.LogLevel
			attr.LogSize = internal.Between(opts.LogSizeStart, minVerifierLogSize, maxVerifierLogSize)
		}

		attempts := 1
		for {
			if attr.LogLevel != 0 {
				logBuf = make([]byte, attr.LogSize)
				attr.LogBuf = sys.SlicePointer(logBuf)
			}

			fd, err = sys.ProgLoad(attr)
			if err == nil {
				return &Program{unix.ByteSliceToString(logBuf), fd, spec.Name, "", spec.Type}, nil
			}

			if !retryLogAttrs(attr, opts.LogSizeStart, err) {
				break
			}

			if attempts >= maxVerifierAttempts {
				return nil, fmt.Errorf("load program: %w (bug: hit %d verifier attempts)", err, maxVerifierAttempts)
			}
			attempts++
		}
	}

	end := bytes.IndexByte(logBuf, 0)
	if end < 0 {
		end = len(logBuf)
	}

	tail := logBuf[max(end-256, 0):end]
	switch {
	case errors.Is(err, unix.EPERM):
		if len(logBuf) > 0 && logBuf[0] == 0 {
			// EPERM due to RLIMIT_MEMLOCK happens before the verifier, so we can
			// check that the log is empty to reduce false positives.
			return nil, fmt.Errorf("load program: %w (MEMLOCK may be too low, consider rlimit.RemoveMemlock)", err)
		}

	case errors.Is(err, unix.EFAULT):
		// EFAULT is returned when the kernel hits a verifier bug, and always
		// overrides ENOSPC, defeating the buffer growth strategy. Warn the user
		// that they may need to increase the buffer size manually.
		return nil, fmt.Errorf("load program: %w (hit verifier bug, increase LogSizeStart to fit the log and check dmesg)", err)

	case errors.Is(err, unix.EINVAL):
		if bytes.Contains(tail, coreBadCall) {
			err = errBadRelocation
			break
		} else if bytes.Contains(tail, kfuncBadCall) {
			err = errUnknownKfunc
			break
		}

	case errors.Is(err, unix.EACCES):
		if bytes.Contains(tail, coreBadLoad) {
			err = errBadRelocation
			break
		}
	}

	// hasFunctionReferences may be expensive, so check it last.
	if (errors.Is(err, unix.EINVAL) || errors.Is(err, unix.EPERM)) &&
		hasFunctionReferences(spec.Instructions) {
		if err := haveBPFToBPFCalls(); err != nil {
			return nil, fmt.Errorf("load program: %w", err)
		}
	}

	return nil, internal.ErrorWithLog("load program", err, logBuf)
}

func retryLogAttrs(attr *sys.ProgLoadAttr, startSize uint32, err error) bool {
	if attr.LogSize == maxVerifierLogSize {
		// Maximum buffer size reached, don't grow or retry.
		return false
	}

	// ENOSPC means the log was enabled on the previous iteration, so we only
	// need to grow the buffer.
	if errors.Is(err, unix.ENOSPC) {
		if attr.LogTrueSize != 0 {
			// Kernel supports LogTrueSize and previous iteration undershot the buffer
			// size. Try again with the given true size.
			attr.LogSize = attr.LogTrueSize
			return true
		}

		// Ensure the size doesn't overflow.
		const factor = 2
		if attr.LogSize >= maxVerifierLogSize/factor {
			attr.LogSize = maxVerifierLogSize
			return true
		}

		// Make an educated guess how large the buffer should be by multiplying. Due
		// to int division, this rounds down odd sizes.
		attr.LogSize = internal.Between(attr.LogSize, minVerifierLogSize, maxVerifierLogSize/factor)
		attr.LogSize *= factor

		return true
	}

	if attr.LogLevel == 0 {
		// Loading the program failed, it wasn't a buffer-related error, and the log
		// was disabled the previous iteration. Enable basic logging and retry.
		attr.LogLevel = LogLevelBranch
		attr.LogSize = internal.Between(startSize, minVerifierLogSize, maxVerifierLogSize)
		return true
	}

	// Loading the program failed for a reason other than buffer size and the log
	// was already enabled the previous iteration. Don't retry.
	return false
}

// NewProgramFromFD creates a program from a raw fd.
//
// You should not use fd after calling this function.
//
// Requires at least Linux 4.10. Returns an error on Windows.
func NewProgramFromFD(fd int) (*Program, error) {
	f, err := sys.NewFD(fd)
	if err != nil {
		return nil, err
	}

	return newProgramFromFD(f)
}

// NewProgramFromID returns the program for a given id.
//
// Returns ErrNotExist, if there is no eBPF program with the given id.
func NewProgramFromID(id ProgramID) (*Program, error) {
	fd, err := sys.ProgGetFdById(&sys.ProgGetFdByIdAttr{
		Id: uint32(id),
	})
	if err != nil {
		return nil, fmt.Errorf("get program by id: %w", err)
	}

	return newProgramFromFD(fd)
}

func newProgramFromFD(fd *sys.FD) (*Program, error) {
	info, err := newProgramInfoFromFd(fd)
	if err != nil {
		fd.Close()
		return nil, fmt.Errorf("discover program type: %w", err)
	}

	return &Program{"", fd, info.Name, "", info.Type}, nil
}

func (p *Program) String() string {
	if p.name != "" {
		return fmt.Sprintf("%s(%s)#%v", p.typ, p.name, p.fd)
	}
	return fmt.Sprintf("%s(%v)", p.typ, p.fd)
}

// Type returns the underlying type of the program.
func (p *Program) Type() ProgramType {
	return p.typ
}

// Info returns metadata about the program.
//
// Requires at least 4.10.
func (p *Program) Info() (*ProgramInfo, error) {
	return newProgramInfoFromFd(p.fd)
}

// Handle returns a reference to the program's type information in the kernel.
//
// Returns ErrNotSupported if the kernel has no BTF support, or if there is no
// BTF associated with the program.
func (p *Program) Handle() (*btf.Handle, error) {
	info, err := p.Info()
	if err != nil {
		return nil, err
	}

	id, ok := info.BTFID()
	if !ok {
		return nil, fmt.Errorf("program %s: retrieve BTF ID: %w", p, ErrNotSupported)
	}

	return btf.NewHandleFromID(id)
}

// FD gets the file descriptor of the Program.
//
// It is invalid to call this function after Close has been called.
func (p *Program) FD() int {
	return p.fd.Int()
}

// Clone creates a duplicate of the Program.
//
// Closing the duplicate does not affect the original, and vice versa.
//
// Cloning a nil Program returns nil.
func (p *Program) Clone() (*Program, error) {
	if p == nil {
		return nil, nil
	}

	dup, err := p.fd.Dup()
	if err != nil {
		return nil, fmt.Errorf("can't clone program: %w", err)
	}

	return &Program{p.VerifierLog, dup, p.name, "", p.typ}, nil
}

// Pin persists the Program on the BPF virtual file system past the lifetime of
// the process that created it
//
// Calling Pin on a previously pinned program will overwrite the path, except when
// the new path already exists. Re-pinning across filesystems is not supported.
//
// This requires bpffs to be mounted above fileName.
// See https://docs.cilium.io/en/stable/network/kubernetes/configuration/#mounting-bpffs-with-systemd
func (p *Program) Pin(fileName string) error {
	if err := sys.Pin(p.pinnedPath, fileName, p.fd); err != nil {
		return err
	}
	p.pinnedPath = fileName
	return nil
}

// Unpin removes the persisted state for the Program from the BPF virtual filesystem.
//
// Failed calls to Unpin will not alter the state returned by IsPinned.
//
// Unpinning an unpinned Program returns nil.
func (p *Program) Unpin() error {
	if err := sys.Unpin(p.pinnedPath); err != nil {
		return err
	}
	p.pinnedPath = ""
	return nil
}

// IsPinned returns true if the Program has a non-empty pinned path.
func (p *Program) IsPinned() bool {
	return p.pinnedPath != ""
}

// Close the Program's underlying file descriptor, which could unload
// the program from the kernel if it is not pinned or attached to a
// kernel hook.
func (p *Program) Close() error {
	if p == nil {
		return nil
	}

	return p.fd.Close()
}

// Various options for Run'ing a Program
type RunOptions struct {
	// Program's data input. Required field.
	//
	// The kernel expects at least 14 bytes input for an ethernet header for
	// XDP and SKB programs.
	Data []byte
	// Program's data after Program has run. Caller must allocate. Optional field.
	DataOut []byte
	// Program's context input. Optional field.
	Context interface{}
	// Program's context after Program has run. Must be a pointer or slice. Optional field.
	ContextOut interface{}
	// Minimum number of times to run Program. Optional field. Defaults to 1.
	//
	// The program may be executed more often than this due to interruptions, e.g.
	// when runtime.AllThreadsSyscall is invoked.
	Repeat uint32
	// Optional flags.
	Flags uint32
	// CPU to run Program on. Optional field.
	// Note not all program types support this field.
	CPU uint32
	// Called whenever the syscall is interrupted, and should be set to testing.B.ResetTimer
	// or similar. Typically used during benchmarking. Optional field.
	//
	// Deprecated: use [testing.B.ReportMetric] with unit "ns/op" instead.
	Reset func()
}

// Test runs the Program in the kernel with the given input and returns the
// value returned by the eBPF program.
//
// Note: the kernel expects at least 14 bytes input for an ethernet header for
// XDP and SKB programs.
//
// This function requires at least Linux 4.12.
func (p *Program) Test(in []byte) (uint32, []byte, error) {
	// Older kernels ignore the dataSizeOut argument when copying to user space.
	// Combined with things like bpf_xdp_adjust_head() we don't really know what the final
	// size will be. Hence we allocate an output buffer which we hope will always be large
	// enough, and panic if the kernel wrote past the end of the allocation.
	// See https://patchwork.ozlabs.org/cover/1006822/
	var out []byte
	if len(in) > 0 {
		out = make([]byte, len(in)+outputPad)
	}

	opts := RunOptions{
		Data:    in,
		DataOut: out,
		Repeat:  1,
	}

	ret, _, err := p.run(&opts)
	if err != nil {
		return ret, nil, fmt.Errorf("test program: %w", err)
	}
	return ret, opts.DataOut, nil
}

// Run runs the Program in kernel with given RunOptions.
//
// Note: the same restrictions from Test apply.
func (p *Program) Run(opts *RunOptions) (uint32, error) {
	if opts == nil {
		opts = &RunOptions{}
	}

	ret, _, err := p.run(opts)
	if err != nil {
		return ret, fmt.Errorf("run program: %w", err)
	}
	return ret, nil
}

// Benchmark runs the Program with the given input for a number of times
// and returns the time taken per iteration.
//
// Returns the result of the last execution of the program and the time per
// run or an error. reset is called whenever the benchmark syscall is
// interrupted, and should be set to testing.B.ResetTimer or similar.
//
// This function requires at least Linux 4.12.
func (p *Program) Benchmark(in []byte, repeat int, reset func()) (uint32, time.Duration, error) {
	if uint(repeat) > math.MaxUint32 {
		return 0, 0, fmt.Errorf("repeat is too high")
	}

	opts := RunOptions{
		Data:   in,
		Repeat: uint32(repeat),
		Reset:  reset,
	}

	ret, total, err := p.run(&opts)
	if err != nil {
		return ret, total, fmt.Errorf("benchmark program: %w", err)
	}
	return ret, total, nil
}

var haveProgRun = internal.NewFeatureTest("BPF_PROG_RUN", func() error {
	if platform.IsWindows {
		return nil
	}

	prog, err := NewProgram(&ProgramSpec{
		// SocketFilter does not require privileges on newer kernels.
		Type: SocketFilter,
		Instructions: asm.Instructions{
			asm.LoadImm(asm.R0, 0, asm.DWord),
			asm.Return(),
		},
		License: "MIT",
	})
	if err != nil {
		// This may be because we lack sufficient permissions, etc.
		return err
	}
	defer prog.Close()

	in := internal.EmptyBPFContext
	attr := sys.ProgRunAttr{
		ProgFd:     uint32(prog.FD()),
		DataSizeIn: uint32(len(in)),
		DataIn:     sys.SlicePointer(in),
	}

	err = sys.ProgRun(&attr)
	switch {
	case errors.Is(err, unix.EINVAL):
		// Check for EINVAL specifically, rather than err != nil since we
		// otherwise misdetect due to insufficient permissions.
		return internal.ErrNotSupported

	case errors.Is(err, unix.EINTR):
		// We know that PROG_TEST_RUN is supported if we get EINTR.
		return nil

	case errors.Is(err, sys.ENOTSUPP):
		// The first PROG_TEST_RUN patches shipped in 4.12 didn't include
		// a test runner for SocketFilter. ENOTSUPP means PROG_TEST_RUN is
		// supported, but not for the program type used in the probe.
		return nil
	}

	return err
}, "4.12", "windows:0.20")

func (p *Program) run(opts *RunOptions) (uint32, time.Duration, error) {
	if uint(len(opts.Data)) > math.MaxUint32 {
		return 0, 0, fmt.Errorf("input is too long")
	}

	if err := haveProgRun(); err != nil {
		return 0, 0, err
	}

	var ctxIn []byte
	if opts.Context != nil {
		var err error
		ctxIn, err = binary.Append(nil, internal.NativeEndian, opts.Context)
		if err != nil {
			return 0, 0, fmt.Errorf("cannot serialize context: %v", err)
		}
	}

	var ctxOut []byte
	if opts.ContextOut != nil {
		ctxOut = make([]byte, binary.Size(opts.ContextOut))
	}

	attr := sys.ProgRunAttr{
		ProgFd:      p.fd.Uint(),
		DataSizeIn:  uint32(len(opts.Data)),
		DataSizeOut: uint32(len(opts.DataOut)),
		DataIn:      sys.SlicePointer(opts.Data),
		DataOut:     sys.SlicePointer(opts.DataOut),
		Repeat:      uint32(opts.Repeat),
		CtxSizeIn:   uint32(len(ctxIn)),
		CtxSizeOut:  uint32(len(ctxOut)),
		CtxIn:       sys.SlicePointer(ctxIn),
		CtxOut:      sys.SlicePointer(ctxOut),
		Flags:       opts.Flags,
		Cpu:         opts.CPU,
	}

retry:
	for {
		err := sys.ProgRun(&attr)
		if err == nil {
			break retry
		}

		if errors.Is(err, unix.EINTR) {
			if attr.Repeat <= 1 {
				// Older kernels check whether enough repetitions have been
				// executed only after checking for pending signals.
				//
				//     run signal? done? run ...
				//
				// As a result we can get EINTR for repeat==1 even though
				// the program was run exactly once. Treat this as a
				// successful run instead.
				//
				// Since commit 607b9cc92bd7 ("bpf: Consolidate shared test timing code")
				// the conditions are reversed:
				//     run done? signal? ...
				break retry
			}

			if opts.Reset != nil {
				opts.Reset()
			}
			continue retry
		}

		if errors.Is(err, sys.ENOTSUPP) {
			return 0, 0, fmt.Errorf("kernel doesn't support running %s: %w", p.Type(), ErrNotSupported)
		}

		return 0, 0, err
	}

	if opts.DataOut != nil {
		if int(attr.DataSizeOut) > cap(opts.DataOut) {
			// Houston, we have a problem. The program created more data than we allocated,
			// and the kernel wrote past the end of our buffer.
			panic("kernel wrote past end of output buffer")
		}
		opts.DataOut = opts.DataOut[:int(attr.DataSizeOut)]
	}

	if len(ctxOut) != 0 {
		b := bytes.NewReader(ctxOut)
		if err := binary.Read(b, internal.NativeEndian, opts.ContextOut); err != nil {
			return 0, 0, fmt.Errorf("failed to decode ContextOut: %v", err)
		}
	}

	total := time.Duration(attr.Duration) * time.Nanosecond
	return attr.Retval, total, nil
}

func unmarshalProgram(buf sysenc.Buffer) (*Program, error) {
	var id uint32
	if err := buf.Unmarshal(&id); err != nil {
		return nil, err
	}

	// Looking up an entry in a nested map or prog array returns an id,
	// not an fd.
	return NewProgramFromID(ProgramID(id))
}

func marshalProgram(p *Program, length int) ([]byte, error) {
	if p == nil {
		return nil, errors.New("can't marshal a nil Program")
	}

	if length != 4 {
		return nil, fmt.Errorf("can't marshal program to %d bytes", length)
	}

	buf := make([]byte, 4)
	internal.NativeEndian.PutUint32(buf, p.fd.Uint())
	return buf, nil
}

// LoadPinnedProgram loads a Program from a pin (file) on the BPF virtual
// filesystem.
//
// Requires at least Linux 4.11.
func LoadPinnedProgram(fileName string, opts *LoadPinOptions) (*Program, error) {
	fd, typ, err := sys.ObjGetTyped(&sys.ObjGetAttr{
		Pathname:  sys.NewStringPointer(fileName),
		FileFlags: opts.Marshal(),
	})
	if err != nil {
		return nil, err
	}

	if typ != sys.BPF_TYPE_PROG {
		_ = fd.Close()
		return nil, fmt.Errorf("%s is not a Program", fileName)
	}

	info, err := newProgramInfoFromFd(fd)
	if err != nil {
		_ = fd.Close()
		return nil, fmt.Errorf("info for %s: %w", fileName, err)
	}

	var progName string
	if haveObjName() == nil {
		progName = info.Name
	} else {
		progName = filepath.Base(fileName)
	}

	return &Program{"", fd, progName, fileName, info.Type}, nil
}

// ProgramGetNextID returns the ID of the next eBPF program.
//
// Returns ErrNotExist, if there is no next eBPF program.
func ProgramGetNextID(startID ProgramID) (ProgramID, error) {
	attr := &sys.ProgGetNextIdAttr{Id: uint32(startID)}
	return ProgramID(attr.NextId), sys.ProgGetNextId(attr)
}

// BindMap binds map to the program and is only released once program is released.
//
// This may be used in cases where metadata should be associated with the program
// which otherwise does not contain any references to the map.
func (p *Program) BindMap(m *Map) error {
	attr := &sys.ProgBindMapAttr{
		ProgFd: uint32(p.FD()),
		MapFd:  uint32(m.FD()),
	}

	return sys.ProgBindMap(attr)
}

var errUnrecognizedAttachType = errors.New("unrecognized attach type")

// find an attach target type in the kernel.
//
// name, progType and attachType determine which type we need to attach to.
//
// The attach target may be in a loaded kernel module.
// In that case the returned handle will be non-nil.
// The caller is responsible for closing the handle.
//
// Returns errUnrecognizedAttachType if the combination of progType and attachType
// is not recognised.
func findProgramTargetInKernel(name string, progType ProgramType, attachType AttachType) (*btf.Handle, btf.TypeID, error) {
	type match struct {
		p ProgramType
		a AttachType
	}

	var (
		typeName, featureName string
		target                btf.Type
	)

	switch (match{progType, attachType}) {
	case match{LSM, AttachLSMMac}:
		typeName = "bpf_lsm_" + name
		featureName = name + " LSM hook"
		target = (*btf.Func)(nil)
	case match{Tracing, AttachTraceIter}:
		typeName = "bpf_iter_" + name
		featureName = name + " iterator"
		target = (*btf.Func)(nil)
	case match{Tracing, AttachTraceFEntry}:
		typeName = name
		featureName = fmt.Sprintf("fentry %s", name)
		target = (*btf.Func)(nil)
	case match{Tracing, AttachTraceFExit}:
		typeName = name
		featureName = fmt.Sprintf("fexit %s", name)
		target = (*btf.Func)(nil)
	case match{Tracing, AttachModifyReturn}:
		typeName = name
		featureName = fmt.Sprintf("fmod_ret %s", name)
		target = (*btf.Func)(nil)
	case match{Tracing, AttachTraceRawTp}:
		typeName = fmt.Sprintf("btf_trace_%s", name)
		featureName = fmt.Sprintf("raw_tp %s", name)
		target = (*btf.Typedef)(nil)
	default:
		return nil, 0, errUnrecognizedAttachType
	}

	spec, err := btf.LoadKernelSpec()
	if err != nil {
		return nil, 0, fmt.Errorf("load kernel spec: %w", err)
	}

	spec, module, err := findTargetInKernel(spec, typeName, &target)
	if errors.Is(err, btf.ErrNotFound) {
		return nil, 0, &internal.UnsupportedFeatureError{Name: featureName}
	}
	// See cilium/ebpf#894. Until we can disambiguate between equally-named kernel
	// symbols, we should explicitly refuse program loads. They will not reliably
	// do what the caller intended.
	if errors.Is(err, btf.ErrMultipleMatches) {
		return nil, 0, fmt.Errorf("attaching to ambiguous kernel symbol is not supported: %w", err)
	}
	if err != nil {
		return nil, 0, fmt.Errorf("find target for %s: %w", featureName, err)
	}

	id, err := spec.TypeID(target)
	if err != nil {
		module.Close()
		return nil, 0, err
	}

	return module, id, nil
}

// findTargetInKernel attempts to find a named type in the current kernel.
//
// target will point at the found type after a successful call. Searches both
// vmlinux and any loaded modules.
//
// Returns a non-nil handle if the type was found in a module, or btf.ErrNotFound
// if the type wasn't found at all.
func findTargetInKernel(kernelSpec *btf.Spec, typeName string, target *btf.Type) (*btf.Spec, *btf.Handle, error) {
	if kernelSpec == nil {
		return nil, nil, fmt.Errorf("nil kernelSpec: %w", btf.ErrNotFound)
	}

	err := kernelSpec.TypeByName(typeName, target)
	if errors.Is(err, btf.ErrNotFound) {
		spec, module, err := findTargetInModule(kernelSpec, typeName, target)
		if err != nil {
			return nil, nil, fmt.Errorf("find target in modules: %w", err)
		}
		return spec, module, nil
	}
	if err != nil {
		return nil, nil, fmt.Errorf("find target in vmlinux: %w", err)
	}
	return kernelSpec, nil, err
}

// findTargetInModule attempts to find a named type in any loaded module.
//
// base must contain the kernel's types and is used to parse kmod BTF. Modules
// are searched in the order they were loaded.
//
// Returns btf.ErrNotFound if the target can't be found in any module.
func findTargetInModule(base *btf.Spec, typeName string, target *btf.Type) (*btf.Spec, *btf.Handle, error) {
	it := new(btf.HandleIterator)
	defer it.Handle.Close()

	for it.Next() {
		info, err := it.Handle.Info()
		if err != nil {
			return nil, nil, fmt.Errorf("get info for BTF ID %d: %w", it.ID, err)
		}

		if !info.IsModule() {
			continue
		}

		spec, err := it.Handle.Spec(base)
		if err != nil {
			return nil, nil, fmt.Errorf("parse types for module %s: %w", info.Name, err)
		}

		err = spec.TypeByName(typeName, target)
		if errors.Is(err, btf.ErrNotFound) {
			continue
		}
		if err != nil {
			return nil, nil, fmt.Errorf("lookup type in module %s: %w", info.Name, err)
		}

		return spec, it.Take(), nil
	}
	if err := it.Err(); err != nil {
		return nil, nil, fmt.Errorf("iterate modules: %w", err)
	}

	return nil, nil, btf.ErrNotFound
}

// find an attach target type in a program.
//
// Returns errUnrecognizedAttachType.
func findTargetInProgram(prog *Program, name string, progType ProgramType, attachType AttachType) (btf.TypeID, error) {
	type match struct {
		p ProgramType
		a AttachType
	}

	var typeName string
	switch (match{progType, attachType}) {
	case match{Extension, AttachNone},
		match{Tracing, AttachTraceFEntry},
		match{Tracing, AttachTraceFExit}:
		typeName = name
	default:
		return 0, errUnrecognizedAttachType
	}

	btfHandle, err := prog.Handle()
	if err != nil {
		return 0, fmt.Errorf("load target BTF: %w", err)
	}
	defer btfHandle.Close()

	spec, err := btfHandle.Spec(nil)
	if err != nil {
		return 0, err
	}

	var targetFunc *btf.Func
	err = spec.TypeByName(typeName, &targetFunc)
	if err != nil {
		return 0, fmt.Errorf("find target %s: %w", typeName, err)
	}

	return spec.TypeID(targetFunc)
}

func newBTFCache(opts *ProgramOptions) *btf.Cache {
	c := btf.NewCache()
	if opts.KernelTypes != nil {
		c.KernelTypes = opts.KernelTypes
		c.KernelModules = opts.KernelModuleTypes
	}
	return c
}