hakurei/internal/pkg/exec.go

package pkg

import (
	"bufio"
	"context"
	"errors"
	"fmt"
	"io"
	"os"
	"os/exec"
	"path"
	"slices"
	"strconv"
	"syscall"
	"time"
	"unique"

	"hakurei.app/container"
	"hakurei.app/container/check"
	"hakurei.app/container/fhs"
	"hakurei.app/container/seccomp"
	"hakurei.app/container/std"
	"hakurei.app/message"
)

// AbsWork is the container pathname [CureContext.GetWorkDir] is mounted on.
var AbsWork = fhs.AbsRoot.Append("work/")

// ExecPath is a slice of [Artifact] and the [check.Absolute] pathname to make
// it available at under in the container.
type ExecPath struct {
	// Pathname in the container mount namespace.
	P *check.Absolute
	// Artifacts to mount on the pathname, must contain at least one [Artifact].
	// If there are multiple entries or W is true, P is set up as an overlay
	// mount, and entries of A must not implement [FileArtifact].
	A []Artifact
	// Whether to make the mount point writable via the temp directory.
	W bool
}

// layers returns pathnames collected from A deduplicated by checksum.
func (p *ExecPath) layers(f *FContext) []*check.Absolute {
	msg := f.GetMessage()

	layers := make([]*check.Absolute, 0, len(p.A))
	checksums := make(map[unique.Handle[Checksum]]struct{}, len(p.A))
	for i := range p.A {
		d := p.A[len(p.A)-1-i]
		pathname, checksum := f.GetArtifact(d)
		if _, ok := checksums[checksum]; ok {
			if msg.IsVerbose() {
				msg.Verbosef(
					"promoted layer %d as %s",
					len(p.A)-1-i, reportName(d, f.cache.Ident(d)),
				)
			}
			continue
		}
		checksums[checksum] = struct{}{}
		layers = append(layers, pathname)
	}
	slices.Reverse(layers)
	return layers
}

// Path returns a populated [ExecPath].
func Path(pathname *check.Absolute, writable bool, a ...Artifact) ExecPath {
	return ExecPath{pathname, a, writable}
}

// MustPath is like [Path], but takes a string pathname via [check.MustAbs].
func MustPath(pathname string, writable bool, a ...Artifact) ExecPath {
	return ExecPath{check.MustAbs(pathname), a, writable}
}

const (
	// ExecTimeoutDefault replaces out of range [NewExec] timeout values.
	ExecTimeoutDefault = 15 * time.Minute
	// ExecTimeoutMax is the arbitrary upper bound of [NewExec] timeout.
	ExecTimeoutMax = 48 * time.Hour
)

// An execArtifact is an [Artifact] that produces output by running a program
// part of another [Artifact] in a [container] to produce its output.
//
// Methods of execArtifact does not modify any struct field or underlying arrays
// referred to by slices.
type execArtifact struct {
	// Caller-supplied user-facing reporting name, guaranteed to be nonzero
	// during initialisation.
	name string
	// Caller-supplied inner mount points.
	paths []ExecPath

	// Passed through to [container.Params].
	dir *check.Absolute
	// Passed through to [container.Params].
	env []string
	// Passed through to [container.Params].
	path *check.Absolute
	// Passed through to [container.Params].
	args []string

	// Duration the initial process is allowed to run. The zero value is
	// equivalent to [ExecTimeoutDefault].
	timeout time.Duration

	// Caller-supplied exclusivity value, returned as is by IsExclusive.
	exclusive bool
}

var _ fmt.Stringer = new(execArtifact)

// execNetArtifact is like execArtifact but implements [KnownChecksum] and has
// its resulting container keep the host net namespace.
type execNetArtifact struct {
	checksum Checksum

	execArtifact
}

var _ KnownChecksum = new(execNetArtifact)

// Checksum returns the caller-supplied checksum.
func (a *execNetArtifact) Checksum() Checksum { return a.checksum }

// Kind returns the hardcoded [Kind] constant.
func (*execNetArtifact) Kind() Kind { return KindExecNet }

// Cure cures the [Artifact] in the container described by the caller. The
// container retains host networking.
func (a *execNetArtifact) Cure(f *FContext) error {
	return a.cure(f, true)
}

// NewExec returns a new [Artifact] that executes the program path in a
// container with specified paths bind mounted read-only in order. A private
// instance of /proc and /dev is made available to the container.
//
// The working and temporary directories are both created and mounted writable
// on [AbsWork] and [fhs.AbsTmp] respectively. If one or more paths target
// [AbsWork], the final entry is set up as a writable overlay mount on /work for
// which the upperdir is the host side work directory. In this configuration,
// the W field is ignored, and the program must avoid causing whiteout files to
// be created. Cure fails if upperdir ends up with entries other than directory,
// regular or symlink.
//
// If checksum is non-nil, the resulting [Artifact] implements [KnownChecksum]
// and its container runs in the host net namespace.
//
// The container is allowed to run for the specified duration before the initial
// process and all processes originating from it is terminated. A zero or
// negative timeout value is equivalent tp [ExecTimeoutDefault], a timeout value
// greater than [ExecTimeoutMax] is equivalent to [ExecTimeoutMax].
//
// The user-facing name and exclusivity value are not accessible from the
// container and does not affect curing outcome. Because of this, it is omitted
// from parameter data for computing identifier.
func NewExec(
	name string,
	checksum *Checksum,
	timeout time.Duration,
	exclusive bool,

	dir *check.Absolute,
	env []string,
	pathname *check.Absolute,
	args []string,

	paths ...ExecPath,
) Artifact {
	if name == "" {
		name = "exec-" + path.Base(pathname.String())
	}
	if timeout <= 0 {
		timeout = ExecTimeoutDefault
	}
	if timeout > ExecTimeoutMax {
		timeout = ExecTimeoutMax
	}
	a := execArtifact{name, paths, dir, env, pathname, args, timeout, exclusive}
	if checksum == nil {
		return &a
	}
	return &execNetArtifact{*checksum, a}
}

// Kind returns the hardcoded [Kind] constant.
func (*execArtifact) Kind() Kind { return KindExec }

// Params writes paths, executable pathname and args.
func (a *execArtifact) Params(ctx *IContext) {
	ctx.WriteString(a.name)

	ctx.WriteUint32(uint32(len(a.paths)))
	for _, p := range a.paths {
		if p.P != nil {
			ctx.WriteString(p.P.String())
		} else {
			ctx.WriteString("invalid P\x00")
		}

		ctx.WriteUint32(uint32(len(p.A)))
		for _, d := range p.A {
			ctx.WriteIdent(d)
		}

		if p.W {
			ctx.WriteUint32(1)
		} else {
			ctx.WriteUint32(0)
		}
	}

	ctx.WriteString(a.dir.String())

	ctx.WriteUint32(uint32(len(a.env)))
	for _, e := range a.env {
		ctx.WriteString(e)
	}

	ctx.WriteString(a.path.String())

	ctx.WriteUint32(uint32(len(a.args)))
	for _, arg := range a.args {
		ctx.WriteString(arg)
	}

	ctx.WriteUint32(uint32(a.timeout & 0xffffffff))
	ctx.WriteUint32(uint32(a.timeout >> 32))

	if a.exclusive {
		ctx.WriteUint32(1)
	} else {
		ctx.WriteUint32(0)
	}
}

// readExecArtifact interprets IR values and returns the address of execArtifact
// or execNetArtifact.
func readExecArtifact(r *IRReader, net bool) Artifact {
	r.DiscardAll()

	name := r.ReadString()

	sz := r.ReadUint32()
	if sz > irMaxDeps {
		panic(ErrIRDepend)
	}
	paths := make([]ExecPath, sz)
	for i := range paths {
		paths[i].P = check.MustAbs(r.ReadString())

		sz = r.ReadUint32()
		if sz > irMaxDeps {
			panic(ErrIRDepend)
		}
		paths[i].A = make([]Artifact, sz)
		for j := range paths[i].A {
			paths[i].A[j] = r.ReadIdent()
		}

		paths[i].W = r.ReadUint32() != 0
	}

	dir := check.MustAbs(r.ReadString())

	sz = r.ReadUint32()
	if sz > irMaxValues {
		panic(ErrIRValues)
	}
	env := make([]string, sz)
	for i := range env {
		env[i] = r.ReadString()
	}

	pathname := check.MustAbs(r.ReadString())

	sz = r.ReadUint32()
	if sz > irMaxValues {
		panic(ErrIRValues)
	}
	args := make([]string, sz)
	for i := range args {
		args[i] = r.ReadString()
	}

	timeout := time.Duration(r.ReadUint32())
	timeout |= time.Duration(r.ReadUint32()) << 32

	exclusive := r.ReadUint32() != 0

	checksum, ok := r.Finalise()

	var checksumP *Checksum
	if net {
		if !ok {
			panic(ErrExpectedChecksum)
		}
		checksumVal := checksum.Value()
		checksumP = &checksumVal
	} else {
		if ok {
			panic(ErrUnexpectedChecksum)
		}
	}

	return NewExec(
		name, checksumP, timeout, exclusive, dir, env, pathname, args, paths...,
	)
}

func init() {
	register(KindExec,
		func(r *IRReader) Artifact { return readExecArtifact(r, false) })
	register(KindExecNet,
		func(r *IRReader) Artifact { return readExecArtifact(r, true) })
}

// Dependencies returns a slice of all artifacts collected from caller-supplied
// [ExecPath].
func (a *execArtifact) Dependencies() []Artifact {
	artifacts := make([][]Artifact, 0, len(a.paths))
	for _, p := range a.paths {
		artifacts = append(artifacts, p.A)
	}
	return slices.Concat(artifacts...)
}

// IsExclusive returns the caller-supplied exclusivity value.
func (a *execArtifact) IsExclusive() bool { return a.exclusive }

// String returns the caller-supplied reporting name.
func (a *execArtifact) String() string { return a.name }

// Cure cures the [Artifact] in the container described by the caller.
func (a *execArtifact) Cure(f *FContext) (err error) {
	return a.cure(f, false)
}

const (
	// execWaitDelay is passed through to [container.Params].
	execWaitDelay = time.Nanosecond
)

// scanVerbose prefixes program output for a verbose [message.Msg].
func scanVerbose(
	msg message.Msg,
	done chan<- struct{},
	prefix string,
	r io.Reader,
) {
	defer close(done)
	s := bufio.NewScanner(r)
	s.Buffer(
		make([]byte, bufio.MaxScanTokenSize),
		bufio.MaxScanTokenSize<<12,
	)
	for s.Scan() {
		msg.Verbose(prefix, s.Text())
	}
	if err := s.Err(); err != nil && !errors.Is(err, os.ErrClosed) {
		msg.Verbose("*"+prefix, err)
	}
}

// cure is like Cure but allows optional host net namespace. This is used for
// the [KnownChecksum] variant where networking is allowed.
func (a *execArtifact) cure(f *FContext, hostNet bool) (err error) {
	overlayWorkIndex := -1
	for i, p := range a.paths {
		if p.P == nil || len(p.A) == 0 {
			return os.ErrInvalid
		}
		if p.P.Is(AbsWork) {
			overlayWorkIndex = i
		}
	}

	var artifactCount int
	for _, p := range a.paths {
		artifactCount += len(p.A)
	}

	ctx, cancel := context.WithTimeout(f.Unwrap(), a.timeout)
	defer cancel()

	z := container.New(ctx, f.GetMessage())
	z.WaitDelay = execWaitDelay
	z.SeccompPresets |= std.PresetStrict & ^std.PresetDenyNS
	z.SeccompFlags |= seccomp.AllowMultiarch
	z.ParentPerm = 0700
	z.HostNet = hostNet
	z.Hostname = "cure"
	if z.HostNet {
		z.Hostname = "cure-net"
	}
	z.Uid, z.Gid = (1<<10)-1, (1<<10)-1
	if msg := f.GetMessage(); msg.IsVerbose() {
		var stdout, stderr io.ReadCloser
		if stdout, err = z.StdoutPipe(); err != nil {
			return
		}
		if stderr, err = z.StderrPipe(); err != nil {
			_ = stdout.Close()
			return
		}
		defer func() {
			if err != nil && !errors.As(err, new(*exec.ExitError)) {
				_ = stdout.Close()
				_ = stderr.Close()
			}
		}()

		stdoutDone, stderrDone := make(chan struct{}), make(chan struct{})
		go scanVerbose(msg, stdoutDone, "("+a.name+":1)", stdout)
		go scanVerbose(msg, stderrDone, "("+a.name+":2)", stderr)
		defer func() { <-stdoutDone; <-stderrDone }()
	}

	z.Dir, z.Env, z.Path, z.Args = a.dir, a.env, a.path, a.args
	z.Grow(len(a.paths) + 4)

	temp, work := f.GetTempDir(), f.GetWorkDir()
	for i, b := range a.paths {
		if i == overlayWorkIndex {
			if err = os.MkdirAll(work.String(), 0700); err != nil {
				return
			}
			tempWork := temp.Append(".work")
			if err = os.MkdirAll(tempWork.String(), 0700); err != nil {
				return
			}
			z.Overlay(
				AbsWork,
				work,
				tempWork,
				b.layers(f)...,
			)
			continue
		}

		if a.paths[i].W {
			tempUpper, tempWork := temp.Append(
				".upper", strconv.Itoa(i),
			), temp.Append(
				".work", strconv.Itoa(i),
			)
			if err = os.MkdirAll(tempUpper.String(), 0700); err != nil {
				return
			}
			if err = os.MkdirAll(tempWork.String(), 0700); err != nil {
				return
			}
			z.Overlay(b.P, tempUpper, tempWork, b.layers(f)...)
		} else if len(b.A) == 1 {
			pathname, _ := f.GetArtifact(b.A[0])
			z.Bind(pathname, b.P, 0)
		} else {
			z.OverlayReadonly(b.P, b.layers(f)...)
		}
	}
	if overlayWorkIndex < 0 {
		z.Bind(
			work,
			AbsWork,
			std.BindWritable|std.BindEnsure,
		)
	}
	z.Bind(
		f.GetTempDir(),
		fhs.AbsTmp,
		std.BindWritable|std.BindEnsure,
	)
	z.Proc(fhs.AbsProc).Dev(fhs.AbsDev, true)

	if err = z.Start(); err != nil {
		return
	}
	if err = z.Serve(); err != nil {
		return
	}
	if err = z.Wait(); err != nil {
		return
	}

	// do not allow empty directories to succeed
	for {
		err = syscall.Rmdir(work.String())
		if err != syscall.EINTR {
			break
		}
	}
	if err != nil && errors.Is(err, syscall.ENOTEMPTY) {
		err = nil
	}
	return
}