hakurei/internal/pkg/exec.go

package pkg

import (
	"bytes"
	"context"
	"errors"
	"os"
	"runtime"
	"slices"
	"strconv"
	"syscall"
	"time"

	"hakurei.app/container"
	"hakurei.app/container/check"
	"hakurei.app/container/fhs"
	"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 [File].
	A []Artifact
	// Whether to make the mount point writable via an invisible tmpfs upperdir.
	W bool
}

// 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}
}

// 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 context.
	ctx context.Context
	// Caller-supplied inner mount points.
	paths []ExecPath
	// Caller-supplied logging facility, passed through to [container] and used
	// internally to produce verbose output.
	msg message.Msg

	// Number of [Artifact] to concurrently cure. A value of 0 or lower is
	// equivalent to the value returned by [runtime.NumCPU].
	cures int

	// 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
}

// 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 (a *execNetArtifact) Kind() Kind { return KindExecNet }

// Params is [Checksum] concatenated with [KindExec] params.
func (a *execNetArtifact) Params() []byte {
	return slices.Concat(a.checksum[:], a.execArtifact.Params())
}

// Cure cures the [Artifact] by curing all its dependencies then running the
// container described by the caller. The container retains host networking.
func (a *execNetArtifact) Cure(c *CureContext) error {
	return a.cure(c, true)
}

// NewExec returns a new [Artifact] bounded by ctx, it cures all [Artifact]
// in paths at the specified maximum concurrent cures limit. Specified paths are
// bind mounted read-only in the specified order in the resulting container.
// 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.
//
// A cures value of 0 or lower is equivalent to the value returned by
// [runtime.NumCPU].
func NewExec(
	ctx context.Context,
	msg message.Msg,
	cures int,
	checksum *Checksum,

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

	paths ...ExecPath,
) Artifact {
	a := execArtifact{ctx, paths, msg, cures, dir, env, path, args}
	if checksum == nil {
		return &a
	}
	return &execNetArtifact{*checksum, a}
}

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

// Params returns paths, executable pathname and args concatenated together.
func (a *execArtifact) Params() []byte {
	var buf bytes.Buffer
	for _, p := range a.paths {
		if p.W {
			buf.WriteByte(1)
		} else {
			buf.WriteByte(0)
		}
		if p.P != nil {
			buf.WriteString(p.P.String())
		} else {
			buf.WriteString("invalid P\x00")
		}
		buf.WriteByte(0)
		for _, d := range p.A {
			id := Ident(d)
			buf.Write(id[:])
		}
		buf.WriteByte(0)
	}
	buf.WriteByte(0)
	buf.WriteString(a.dir.String())
	buf.WriteByte(0)
	for _, e := range a.env {
		buf.WriteString(e)
	}
	buf.WriteByte(0)
	buf.WriteString(a.path.String())
	buf.WriteByte(0)
	for _, arg := range a.args {
		buf.WriteString(arg)
	}
	return buf.Bytes()
}

// 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...)
}

// Cure cures the [Artifact] by curing all its dependencies then running the
// container described by the caller.
func (a *execArtifact) Cure(c *CureContext) (err error) {
	return a.cure(c, false)
}

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

// 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(c *CureContext, hostNet bool) (err error) {
	cures := a.cures
	if cures < 1 {
		cures = runtime.NumCPU()
	}

	overlayWorkIndex := -1
	type curePath struct {
		// Copied from ExecPath.P.
		dst *check.Absolute
		// Cured from ExecPath.A.
		src []*check.Absolute
	}
	paths := make([]curePath, len(a.paths))
	for i, p := range a.paths {
		if p.P == nil || len(p.A) == 0 {
			return os.ErrInvalid
		}
		if p.P.Is(AbsWork) {
			overlayWorkIndex = i
		}
		paths[i].dst = p.P
		paths[i].src = make([]*check.Absolute, len(p.A))
	}

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

	if len(paths) > 0 {
		type cureArtifact struct {
			// Index of pending Artifact in paths.
			index [2]int
			// Pending artifact.
			a Artifact
		}
		ac := make(chan cureArtifact, artifactCount)
		for i, p := range a.paths {
			for j, d := range p.A {
				ac <- cureArtifact{[2]int{i, j}, d}
			}
		}

		type cureRes struct {
			// Index of result in paths.
			index [2]int
			// Cured pathname.
			pathname *check.Absolute
			// Error returned by c.
			err error
		}
		res := make(chan cureRes)

		for i := 0; i < cures; i++ {
			go func() {
				for d := range ac {
					// computing and encoding identifier is expensive
					if a.msg.IsVerbose() {
						a.msg.Verbosef("curing %s...", Encode(Ident(d.a)))
					}

					var cr cureRes
					cr.index = d.index
					cr.pathname, _, cr.err = c.Cure(d.a)
					res <- cr
				}
			}()
		}

		var count int
		errs := make([]error, 0, artifactCount)
		for cr := range res {
			count++

			if cr.err != nil {
				errs = append(errs, cr.err)
			} else {
				paths[cr.index[0]].src[cr.index[1]] = cr.pathname
			}

			if count == artifactCount {
				break
			}
		}
		close(ac)
		if err = errors.Join(errs...); err != nil {
			return
		}
	}

	ctx, cancel := context.WithCancel(a.ctx)
	defer cancel()

	z := container.New(ctx, a.msg)
	z.WaitDelay = execWaitDelay
	z.SeccompPresets |= std.PresetStrict
	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 a.msg.IsVerbose() {
		z.Stdout, z.Stderr = os.Stdout, os.Stderr
	}

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

	temp, work := c.GetTempDir(), c.GetWorkDir()
	for i, b := range 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.src...,
			)
			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.dst, tempUpper, tempWork, b.src...)
		} else if len(b.src) == 1 {
			z.Bind(b.src[0], b.dst, 0)
		} else {
			z.OverlayReadonly(b.dst, b.src...)
		}
	}
	if overlayWorkIndex < 0 {
		z.Bind(
			work,
			AbsWork,
			std.BindWritable|std.BindEnsure,
		)
	}
	z.Bind(
		c.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
}