internal/outcome/process: use new store interface

This change also spawns shim before committing system state, leaving it blocking on the setup pipe. The internal/outcome/process structure is also entirely reworked to be much more readable and less error-prone, while enabling basic performance measurements. A long-standing bug where segment lock is not held during Commit is also resolved.

Closes #19.

Signed-off-by: Ophestra <cat@gensokyo.uk>

internal/outcome/process.go

@@ -4,6 +4,8 @@ import (
 	"context"
 	"encoding/gob"
 	"errors"
+	"iter"
+	"math"
 	"os"
 	"os/exec"
 	"strconv"
@@ -28,179 +30,6 @@ const (
 	shimSetupTimeout = 5 * time.Second
 )
 
-// mainState holds persistent state bound to outcome.main.
-type mainState struct {
-	// done is whether beforeExit has been called already.
-	done bool
-
-	// Populated on successful hsu startup.
-	cmd *exec.Cmd
-	// Cancels cmd, must be populated before cmd is populated.
-	cancel context.CancelFunc
-
-	store store.Compat
-
-	k *outcome
-	message.Msg
-	uintptr
-}
-
-const (
-	// mainNeedsRevert indicates the call to Commit has succeeded.
-	mainNeedsRevert uintptr = 1 << iota
-	// mainNeedsDestroy indicates the instance state entry is present in the store.
-	mainNeedsDestroy
-)
-
-// beforeExit must be called immediately before a call to [os.Exit].
-func (ms mainState) beforeExit(isFault bool) {
-	if ms.done {
-		panic("attempting to call beforeExit twice")
-	}
-	ms.done = true
-	defer ms.BeforeExit()
-
-	if isFault && ms.cancel != nil {
-		ms.cancel()
-	}
-
-	var hasErr bool
-	// updates hasErr but does not terminate
-	perror := func(err error, message string) {
-		hasErr = true
-		printMessageError(ms.GetLogger().Println, "cannot "+message+":", err)
-	}
-	exitCode := 1
-	defer func() {
-		if hasErr {
-			os.Exit(exitCode)
-		}
-	}()
-
-	// this also handles wait for a non-fault termination
-	if ms.cmd != nil {
-		select {
-		case err := <-func() chan error { w := make(chan error, 1); go func() { w <- ms.cmd.Wait(); ms.cancel() }(); return w }():
-			wstatus, ok := ms.cmd.ProcessState.Sys().(syscall.WaitStatus)
-			if ok {
-				if v := wstatus.ExitStatus(); v != 0 {
-					hasErr = true
-					exitCode = v
-				}
-			}
-
-			if ms.IsVerbose() {
-				if !ok {
-					if err != nil {
-						ms.Verbosef("wait: %v", err)
-					}
-				} else {
-					switch {
-					case wstatus.Exited():
-						ms.Verbosef("process %d exited with code %d", ms.cmd.Process.Pid, wstatus.ExitStatus())
-
-					case wstatus.CoreDump():
-						ms.Verbosef("process %d dumped core", ms.cmd.Process.Pid)
-
-					case wstatus.Signaled():
-						ms.Verbosef("process %d got %s", ms.cmd.Process.Pid, wstatus.Signal())
-
-					default:
-						ms.Verbosef("process %d exited with status %#x", ms.cmd.Process.Pid, wstatus)
-					}
-				}
-			}
-
-		case <-func() chan struct{} {
-			w := make(chan struct{})
-			// this ties waitDone to ctx with the additional compensated timeout duration
-			go func() { <-ms.k.ctx.Done(); time.Sleep(ms.k.state.Shim.WaitDelay + shimWaitTimeout); close(w) }()
-			return w
-		}():
-			ms.Resume()
-			// this is only reachable when shim did not exit within shimWaitTimeout, after its WaitDelay has elapsed.
-			// This is different from the container failing to terminate within its timeout period, as that is enforced
-			// by the shim. This path is instead reached when there is a lockup in shim preventing it from completing.
-			ms.GetLogger().Printf("process %d did not terminate", ms.cmd.Process.Pid)
-		}
-
-		ms.Resume()
-	}
-
-	if ms.uintptr&mainNeedsRevert != 0 {
-		if ok, err := ms.store.Do(ms.k.state.identity.unwrap(), func(c store.Cursor) {
-			if ms.uintptr&mainNeedsDestroy != 0 {
-				if err := c.Destroy(ms.k.state.id.unwrap()); err != nil {
-					perror(err, "destroy state entry")
-				}
-			}
-
-			var rt hst.Enablement
-			if states, err := c.Load(); err != nil {
-				// it is impossible to continue from this point;
-				// revert per-process state here to limit damage
-				ec := system.Process
-				if revertErr := ms.k.sys.Revert((*system.Criteria)(&ec)); revertErr != nil {
-					var joinError interface {
-						Unwrap() []error
-						error
-					}
-					if !errors.As(revertErr, &joinError) || joinError == nil {
-						perror(revertErr, "revert system setup")
-					} else {
-						for _, v := range joinError.Unwrap() {
-							perror(v, "revert system setup step")
-						}
-					}
-				}
-				perror(err, "load instance states")
-			} else {
-				ec := system.Process
-				if l := len(states); l == 0 {
-					ec |= system.User
-				} else {
-					ms.Verbosef("found %d instances, cleaning up without user-scoped operations", l)
-				}
-
-				// accumulate enablements of remaining launchers
-				for i, s := range states {
-					if s.Config != nil {
-						rt |= s.Config.Enablements.Unwrap()
-					} else {
-						ms.GetLogger().Printf("state entry %d does not contain config", i)
-					}
-				}
-
-				ec |= rt ^ (hst.EWayland | hst.EX11 | hst.EDBus | hst.EPulse)
-				if ms.IsVerbose() {
-					if ec > 0 {
-						ms.Verbose("reverting operations scope", system.TypeString(ec))
-					}
-				}
-
-				if err = ms.k.sys.Revert((*system.Criteria)(&ec)); err != nil {
-					perror(err, "revert system setup")
-				}
-			}
-		}); err != nil {
-			if ok {
-				perror(err, "unlock state store")
-			} else {
-				perror(err, "open state store")
-			}
-		}
-	} else if ms.uintptr&mainNeedsDestroy != 0 {
-		panic("unreachable")
-	}
-}
-
-// fatal calls printMessageError, performs necessary cleanup, followed by a call to [os.Exit](1).
-func (ms mainState) fatal(fallback string, ferr error) {
-	printMessageError(ms.GetLogger().Println, fallback, ferr)
-	ms.beforeExit(true)
-	os.Exit(1)
-}
-
 // main carries out outcome and terminates. main does not return.
 func (k *outcome) main(msg message.Msg) {
 	if k.ctx == nil || k.sys == nil || k.state == nil {
@@ -210,61 +39,281 @@ func (k *outcome) main(msg message.Msg) {
 	// read comp value early for early failure
 	hsuPath := internal.MustHsuPath()
 
-	// ms.beforeExit required beyond this point
+	const (
-	ms := mainState{Msg: msg, k: k}
+		// transitions to processCommit, or processFinal on failure
-
+		processStart = iota
-	if err := k.sys.Commit(); err != nil {
+		// transitions to processServe, or processLifecycle on failure
-		ms.fatal("cannot commit system setup:", err)
+		processCommit
-	}
+		// transitions to processLifecycle only
-	ms.uintptr |= mainNeedsRevert
+		processServe
-	ms.store = store.NewMulti(msg, k.state.sc.RunDirPath)
+		// transitions to processCleanup only
+		processLifecycle
+		// transitions to processFinal only
+		processCleanup
+		// execution terminates, must be the final state
+		processFinal
+	)
 
+	// for the shim process
 	ctx, cancel := context.WithCancel(k.ctx)
 	defer cancel()
-	ms.cancel = cancel
 
-	// shim starts and blocks on setup payload before container is started
 	var (
+		// state for next iteration
+		processState uintptr = processStart
+		// current state, must not be mutated directly
+		processStateCur uintptr = math.MaxUint
+		// point in time the current iteration began
+		processTime time.Time
+
+		// whether sys is currently in between a call to Commit and Revert
+		isBeforeRevert bool
+
+		// initialised during processStart if successful
+		handle *store.Handle
+		// initialised during processServe if state is saved
+		entryHandle *store.EntryHandle
+
+		// can be set in any state, used in processFinal
+		exitCode int
+
+		// shim process startup time,
+		// populated in processStart, accessed by processServe
 		startTime time.Time
+		// shim process as target uid,
+		// populated in processStart, accessed by processServe
+		shimCmd *exec.Cmd
+		// write end of shim setup pipe,
+		// populated in processStart, accessed by processServe
 		shimPipe *os.File
+
+		// perror cancels ctx and prints an error message
+		perror = func(err error, message string) {
+			cancel()
+			if shimPipe != nil {
+				if closeErr := shimPipe.Close(); closeErr != nil {
+					msg.Verbose(closeErr.Error())
+				}
+				shimPipe = nil
+			}
+			if exitCode == 0 {
+				exitCode = 1
+			}
+			printMessageError(msg.GetLogger().Println, "cannot "+message+":", err)
+		}
+
+		// perrorFatal cancels ctx, prints an error message, and sets the next state
+		perrorFatal = func(err error, message string, newState uintptr) {
+			perror(err, message)
+			processState = newState
+		}
 	)
-	if cmd, f, err := k.start(ctx, msg, hsuPath, &startTime); err != nil {
+
-		ms.fatal("cannot start shim:", err)
+	for {
-		panic("unreachable")
+		var processTimePrev time.Time
+		processTimePrev, processTime = processTime, time.Now()
+		var processStatePrev uintptr
+		processStatePrev, processStateCur = processStateCur, processState
+
+		if !processTimePrev.IsZero() && processStatePrev != processLifecycle {
+			msg.Verbosef("state %d took %d ms", processStatePrev, processTime.Sub(processTimePrev).Milliseconds())
+		}
+
+		switch processState {
+		case processStart:
+			if h, err := store.New(k.state.sc.RunDirPath.Append("state")).Handle(k.state.identity.unwrap()); err != nil {
+				perrorFatal(err, "obtain store segment handle", processFinal)
+				continue
 			} else {
-		ms.cmd, shimPipe = cmd, f
+				handle = h
 			}
 
-	// this starts the container, system setup must complete before this point
+			cmd, f, err := k.start(ctx, msg, hsuPath, &startTime)
-	if err := serveShim(msg, shimPipe, k.state); err != nil {
+			if err != nil {
-		ms.fatal("cannot serve shim payload:", err)
+				perrorFatal(err, "start shim", processFinal)
+				continue
+			} else {
+				shimCmd, shimPipe = cmd, f
 			}
 
-	// shim accepted setup payload, create process state
+			processState = processCommit
-	if ok, err := ms.store.Do(k.state.identity.unwrap(), func(c store.Cursor) {
+
-		if err := c.Save(&hst.State{
+		case processCommit:
+			if isBeforeRevert {
+				perrorFatal(newWithMessage("invalid transition to commit state"), "commit", processLifecycle)
+				continue
+			}
+
+			unlock, err := handle.Lock()
+			if err != nil {
+				perrorFatal(err, "acquire lock on store segment", processLifecycle)
+				continue
+			}
+			if entryHandle, err = handle.Save(&hst.State{
 				ID:      k.state.id.unwrap(),
 				PID:     os.Getpid(),
-			ShimPID: ms.cmd.Process.Pid,
+				ShimPID: shimCmd.Process.Pid,
 				Config:  k.config,
 				Time:    startTime,
 			}); err != nil {
-			ms.fatal("cannot save state entry:", err)
+				unlock()
+				// transition here to avoid the commit/revert cycle on the doomed instance
+				perrorFatal(err, "save instance state", processLifecycle)
+				continue
 			}
-	}); err != nil {
-		if ok {
-			ms.uintptr |= mainNeedsDestroy
-			ms.fatal("cannot unlock state store:", err)
-		} else {
-			ms.fatal("cannot open state store:", err)
-		}
-	}
-	// state in store at this point, destroy defunct state entry on termination
-	ms.uintptr |= mainNeedsDestroy
 
-	// beforeExit ties shim process to context
+			err = k.sys.Commit()
-	ms.beforeExit(false)
+			unlock()
-	os.Exit(0)
+			if err != nil {
+				perrorFatal(err, "commit system setup", processLifecycle)
+				continue
+			}
+			isBeforeRevert = true
+
+			processState = processServe
+
+		case processServe:
+			// this state transition to processLifecycle only
+			processState = processLifecycle
+
+			// this starts the container, system setup must complete before this point
+			if err := serveShim(msg, shimPipe, k.state); err != nil {
+				perror(err, "serve shim payload")
+				continue
+			} else {
+				shimPipe = nil // this is already closed by serveShim
+			}
+
+		case processLifecycle:
+			// this state transition to processCleanup only
+			processState = processCleanup
+
+			msg.Suspend()
+			select {
+			case err := <-func() chan error { w := make(chan error, 1); go func() { w <- shimCmd.Wait(); cancel() }(); return w }():
+				wstatus, ok := shimCmd.ProcessState.Sys().(syscall.WaitStatus)
+				if ok {
+					if v := wstatus.ExitStatus(); v != 0 {
+						exitCode = v
+					}
+				}
+
+				if msg.IsVerbose() {
+					if !ok {
+						if err != nil {
+							msg.Verbosef("wait: %v", err)
+						}
+					} else {
+						switch {
+						case wstatus.Exited():
+							msg.Verbosef("process %d exited with code %d", shimCmd.Process.Pid, wstatus.ExitStatus())
+
+						case wstatus.CoreDump():
+							msg.Verbosef("process %d dumped core", shimCmd.Process.Pid)
+
+						case wstatus.Signaled():
+							msg.Verbosef("process %d got %s", shimCmd.Process.Pid, wstatus.Signal())
+
+						default:
+							msg.Verbosef("process %d exited with status %#x", shimCmd.Process.Pid, wstatus)
+						}
+					}
+				}
+
+			case <-func() chan struct{} {
+				w := make(chan struct{})
+				// this ties processLifecycle to ctx with the additional compensated timeout duration
+				// to allow transition to the next state on a locked up shim
+				go func() { <-ctx.Done(); time.Sleep(k.state.Shim.WaitDelay + shimWaitTimeout); close(w) }()
+				return w
+			}():
+				// this is only reachable when wait did not return within shimWaitTimeout, after its WaitDelay has elapsed.
+				// This is different from the container failing to terminate within its timeout period, as that is enforced
+				// by the shim. This path is instead reached when there is a lockup in shim preventing it from completing.
+				msg.GetLogger().Printf("process %d did not terminate", shimCmd.Process.Pid)
+			}
+			msg.Resume()
+
+		case processCleanup:
+			// this state transition to processFinal only
+			processState = processFinal
+
+			unlock, err := handle.Lock()
+			if err != nil {
+				perror(err, "acquire lock on store segment")
+			}
+
+			if entryHandle != nil {
+				if err = entryHandle.Destroy(); err != nil {
+					perror(err, "destroy state entry")
+				}
+			}
+
+			if isBeforeRevert {
+				ec := system.Process
+
+				var entries iter.Seq[*store.EntryHandle]
+				if entries, _, err = handle.Entries(); err != nil {
+					// it is impossible to continue from this point,
+					// per-process state will be reverted to limit damage
+					perror(err, "read store segment entries")
+				} else {
+					// accumulate enablements of remaining instances
+					var (
+						// alive enablement bits
+						rt hst.Enablement
+						// alive instance count
+						n int
+					)
+					for eh := range entries {
+						var et hst.Enablement
+						if et, err = eh.Load(nil); err != nil {
+							perror(err, "read state header of instance "+eh.ID.String())
+						} else {
+							rt |= et
+							n++
+						}
+					}
+
+					if n == 0 {
+						ec |= system.User
+					} else {
+						msg.Verbosef("found %d instances, cleaning up without user-scoped operations", n)
+					}
+					ec |= rt ^ (hst.EWayland | hst.EX11 | hst.EDBus | hst.EPulse)
+					if msg.IsVerbose() {
+						if ec > 0 {
+							msg.Verbose("reverting operations scope", system.TypeString(ec))
+						}
+					}
+				}
+
+				if err = k.sys.Revert((*system.Criteria)(&ec)); err != nil {
+					var joinError interface {
+						Unwrap() []error
+						error
+					}
+					if !errors.As(err, &joinError) || joinError == nil {
+						perror(err, "revert system setup")
+					} else {
+						for _, v := range joinError.Unwrap() {
+							perror(v, "revert system setup step")
+						}
+					}
+				}
+				isBeforeRevert = false
+			}
+			unlock()
+
+		case processFinal:
+			msg.BeforeExit()
+			os.Exit(exitCode)
+
+		default: // not reached
+			k.fatalf("invalid transition from state %d to %d", processStatePrev, processState)
+			panic("unreachable")
+		}
+	}
 }
 
 // start starts the shim via cmd/hsu.
@@ -301,7 +350,6 @@ func (k *outcome) start(ctx context.Context, msg message.Msg,
 	}
 
 	msg.Verbosef("setuid helper at %s", hsuPath)
-	msg.Suspend()
 	if err := cmd.Start(); err != nil {
 		msg.Resume()
 		return cmd, shimPipe, &hst.AppError{Step: "start setuid wrapper", Err: err}
@@ -313,6 +361,10 @@ func (k *outcome) start(ctx context.Context, msg message.Msg,
 
 // serveShim serves outcomeState through the shim setup pipe.
 func serveShim(msg message.Msg, shimPipe *os.File, state *outcomeState) error {
+	if shimPipe == nil {
+		return newWithMessage("shim pipe not available")
+	}
+
 	if err := shimPipe.SetDeadline(time.Now().Add(shimSetupTimeout)); err != nil {
 		msg.Verbose(err.Error())
 	}