draft(uffd,fc): balloon free-page-hinting + envd reclaim on pause

Adds a pre-pause guest reclaim step (sync + drop_caches + compact_memory +
fstrim, run via the existing envd Process service with Connect-Timeout-Ms)
and a virtio-balloon free-page-hinting drain to MADV_DONTNEED freed pages
out of the memfile before the snapshot.

The balloon is installed with FPH=true whenever FPR is on; both behaviors
are off by default and gated by separate LD flags (free-page-hinting,
reclaim-on-pause), so they can be flipped at runtime without rebuilding
templates.

Author: ValentaTomas

packages/orchestrator/cmd/resume-build/main.go

@@ -72,8 +72,19 @@ func main() {
 	cmdSignalPause := flag.String("cmd-signal-pause", "", "execute command in sandbox, then wait for SIGUSR1 before pausing")
 	optimize := flag.Bool("optimize", false, "collect fresh prefetch mapping after pause (resumes snapshot to record page faults)")
 
+	// Pause-time reclaim/FPH overrides — both off by default. Pass >0 to enable.
+	fphTimeoutMs := flag.Int("fph-timeout-ms", 0, "override free-page-hinting-timeout-ms LD flag (0 = use LD default)")
+	reclaimTimeoutMs := flag.Int("reclaim-timeout-ms", 0, "override reclaim-on-pause-timeout-ms LD flag (0 = use LD default)")
+
 	flag.Parse()
 
+	if *fphTimeoutMs > 0 {
+		featureflags.NewIntFlag("free-page-hinting-timeout-ms", *fphTimeoutMs)
+	}
+	if *reclaimTimeoutMs > 0 {
+		featureflags.NewIntFlag("reclaim-on-pause-timeout-ms", *reclaimTimeoutMs)
+	}
+
 	if *fromBuild == "" {
 		log.Fatal("-from-build required")
 	}

packages/orchestrator/pkg/sandbox/fc/client.go

@@ -427,8 +427,11 @@ func (c *apiClient) startVM(ctx context.Context) error {
 	return nil
 }
 
-func (c *apiClient) enableFreePageReporting(ctx context.Context) error {
-	ctx, span := tracer.Start(ctx, "enable-free-page-reporting")
+// installBalloon installs the virtio-balloon pre-boot with target size 0.
+// FreePageHinting is always armed (pure runtime toggle, used by DrainBalloon);
+// FreePageReporting is set per template-build gate.
+func (c *apiClient) installBalloon(ctx context.Context, freePageReporting bool) error {
+	ctx, span := tracer.Start(ctx, "install-balloon")
 	defer span.End()
 
 	amountMib := int64(0)
@@ -439,7 +442,8 @@ func (c *apiClient) enableFreePageReporting(ctx context.Context) error {
 		Body: &models.Balloon{
 			AmountMib:         &amountMib,
 			DeflateOnOom:      &deflateOnOom,
-			FreePageReporting: true,
+			FreePageReporting: freePageReporting,
+			FreePageHinting:   true,
 		},
 	}
 
@@ -451,6 +455,33 @@ func (c *apiClient) enableFreePageReporting(ctx context.Context) error {
 	return nil
 }
 
+func (c *apiClient) startBalloonHinting(ctx context.Context, acknowledgeOnStop bool) error {
+	params := operations.StartBalloonHintingParams{
+		Context: ctx,
+		Body:    &models.BalloonStartCmd{AcknowledgeOnStop: acknowledgeOnStop},
+	}
+	_, err := c.client.Operations.StartBalloonHinting(&params)
+	if err != nil {
+		return fmt.Errorf("error starting balloon hinting: %w", err)
+	}
+
+	return nil
+}
+
+func (c *apiClient) describeBalloonHinting(ctx context.Context) (hostCmd, guestCmd int64, err error) {
+	params := operations.DescribeBalloonHintingParams{Context: ctx}
+	res, err := c.client.Operations.DescribeBalloonHinting(&params)
+	if err != nil {
+		return 0, 0, err
+	}
+	if res.Payload.HostCmd != nil {
+		hostCmd = *res.Payload.HostCmd
+	}
+	guestCmd = res.Payload.GuestCmd
+
+	return hostCmd, guestCmd, nil
+}
+
 func (c *apiClient) memoryMapping(ctx context.Context) (*memory.Mapping, error) {
 	params := operations.GetMemoryMappingsParams{
 		Context: ctx,

packages/orchestrator/pkg/sandbox/fc/process.go

@@ -133,6 +133,8 @@ type Process struct {
 	Exit *utils.ErrorOnce
 
 	client *apiClient
+
+	balloonInstalled bool
 }
 
 func NewProcess(
@@ -440,13 +442,13 @@ func (p *Process) Create(
 	telemetry.ReportEvent(ctx, "set fc entropy config")
 
 	if freePageReporting {
-		err = p.client.enableFreePageReporting(ctx)
-		if err != nil {
+		if err := p.client.installBalloon(ctx, freePageReporting); err != nil {
 			fcStopErr := p.Stop(ctx)
 
-			return errors.Join(fmt.Errorf("error enabling free page reporting: %w", err), fcStopErr)
+			return errors.Join(fmt.Errorf("error installing balloon device: %w", err), fcStopErr)
 		}
-		telemetry.ReportEvent(ctx, "enabled free page reporting")
+		p.balloonInstalled = true
+		telemetry.ReportEvent(ctx, "installed balloon device")
 	}
 
 	err = p.client.startVM(ctx)
@@ -710,6 +712,41 @@ func (p *Process) Pause(ctx context.Context) error {
 	return p.client.pauseVM(ctx)
 }
 
+// DrainBalloon triggers a free-page-hinting run and blocks until the guest
+// acknowledges or ctx fires. No-op when the balloon wasn't installed.
+func (p *Process) DrainBalloon(ctx context.Context) error {
+	if !p.balloonInstalled {
+		return nil
+	}
+
+	ctx, span := tracer.Start(ctx, "drain-balloon")
+	defer span.End()
+
+	if err := p.client.startBalloonHinting(ctx, true /* ackOnStop */); err != nil {
+		return fmt.Errorf("start balloon hinting: %w", err)
+	}
+
+	backoff := 5 * time.Millisecond
+	for {
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		case <-time.After(backoff):
+		}
+
+		host, guest, err := p.client.describeBalloonHinting(ctx)
+		if err != nil {
+			return fmt.Errorf("balloon hinting status: %w", err)
+		}
+		if guest >= host {
+			return nil
+		}
+		if backoff < 50*time.Millisecond {
+			backoff *= 2
+		}
+	}
+}
+
 // CreateSnapshot VM needs to be paused before creating a snapshot.
 func (p *Process) CreateSnapshot(ctx context.Context, snapfilePath string) error {
 	ctx, childSpan := tracer.Start(ctx, "create-snapshot-fc")

packages/orchestrator/pkg/sandbox/reclaim.go

@@ -0,0 +1,53 @@
+package sandbox
+
+import (
+	"context"
+	"fmt"
+	"net/http"
+	"strconv"
+	"time"
+
+	"connectrpc.com/connect"
+	"go.uber.org/zap"
+
+	"github.com/e2b-dev/infra/packages/shared/pkg/consts"
+	"github.com/e2b-dev/infra/packages/shared/pkg/grpc"
+	"github.com/e2b-dev/infra/packages/shared/pkg/grpc/envd/process"
+	"github.com/e2b-dev/infra/packages/shared/pkg/grpc/envd/process/processconnect"
+	"github.com/e2b-dev/infra/packages/shared/pkg/logger"
+)
+
+// Steps separated by ';' so each runs even if a previous one fails. On
+// timeout envd kills bash; the in-flight syscall finishes and remaining
+// steps are skipped.
+const reclaimScript = `sync; echo 3 > /proc/sys/vm/drop_caches 2>/dev/null; echo 1 > /proc/sys/vm/compact_memory 2>/dev/null; fstrim -av 2>/dev/null`
+
+// bestEffortReclaim asks envd to reclaim guest memory + disk before pause.
+// All failures are swallowed.
+func (s *Sandbox) bestEffortReclaim(ctx context.Context, timeout time.Duration) {
+	ctx, span := tracer.Start(ctx, "envd-reclaim")
+	defer span.End()
+
+	rcCtx, cancel := context.WithTimeout(ctx, timeout)
+	defer cancel()
+
+	addr := fmt.Sprintf("http://%s:%d", s.Slot.HostIPString(), consts.DefaultEnvdServerPort)
+	pc := processconnect.NewProcessClient(&http.Client{Transport: sandboxHttpClient.Transport}, addr)
+
+	req := connect.NewRequest(&process.StartRequest{
+		Process: &process.ProcessConfig{Cmd: "/bin/bash", Args: []string{"-c", reclaimScript}},
+	})
+	req.Header().Set("Connect-Timeout-Ms", strconv.FormatInt(int64(timeout/time.Millisecond), 10))
+	grpc.SetUserHeader(req.Header(), "root")
+
+	stream, err := pc.Start(rcCtx, req)
+	if err != nil {
+		logger.L().Warn(ctx, "envd reclaim failed", logger.WithSandboxID(s.Runtime.SandboxID), zap.Error(err))
+
+		return
+	}
+	defer stream.Close()
+
+	for stream.Receive() {
+	}
+}

packages/orchestrator/pkg/sandbox/sandbox.go

@@ -218,6 +218,8 @@ type Sandbox struct {
 	files   *storage.SandboxFiles
 	cleanup *Cleanup
 
+	featureFlags *featureflags.Client
+
 	process      *fc.Process
 	cgroupHandle *cgroup.CgroupHandle
 
@@ -458,7 +460,8 @@ func (f *Factory) CreateSandbox(
 		files:    sandboxFiles,
 		process:  fcHandle,
 
-		cleanup: cleanup,
+		cleanup:      cleanup,
+		featureFlags: f.featureFlags,
 
 		APIStoredConfig: apiConfigToStore,
 
@@ -799,7 +802,8 @@ func (f *Factory) ResumeSandbox(
 		files:    sandboxFiles,
 		process:  fcHandle,
 
-		cleanup: cleanup,
+		cleanup:      cleanup,
+		featureFlags: f.featureFlags,
 
 		APIStoredConfig: apiConfigToStore,
 		CABundle:        f.egressProxy.CABundle(),
@@ -1053,6 +1057,19 @@ func (s *Sandbox) Pause(
 	// Stop the health check before pausing the VM
 	s.Checks.Stop()
 
+	// Best-effort pre-pause guest reclaim, then FPH drain. Both run on the
+	// live VM and are non-fatal. Timeout=0 disables the step.
+	if t := time.Duration(s.featureFlags.IntFlag(ctx, featureflags.ReclaimOnPauseTimeoutMs)) * time.Millisecond; t > 0 {
+		s.bestEffortReclaim(ctx, t)
+	}
+	if t := time.Duration(s.featureFlags.IntFlag(ctx, featureflags.FreePageHintingTimeoutMs)) * time.Millisecond; t > 0 {
+		drainCtx, cancel := context.WithTimeout(ctx, t)
+		if err := s.process.DrainBalloon(drainCtx); err != nil {
+			telemetry.ReportError(ctx, "balloon hinting drain failed (continuing pause)", err)
+		}
+		cancel()
+	}
+
 	if err := s.process.Pause(ctx); err != nil {
 		return nil, fmt.Errorf("failed to pause VM: %w", err)
 	}

packages/shared/pkg/featureflags/flags.go

@@ -158,7 +158,11 @@ var (
 	BestOfKMaxOvercommit          = NewIntFlag("best-of-k-max-overcommit", 400)              // Default R=4 (stored as percentage, max over-commit ratio)
 	BestOfKAlpha                  = NewIntFlag("best-of-k-alpha", 50)                        // Default Alpha=0.5 (stored as percentage for int flag, current usage weight)
 	EnvdInitTimeoutMilliseconds   = NewIntFlag("envd-init-request-timeout-milliseconds", 50) // Timeout for envd init request in milliseconds
-	HostStatsSamplingInterval     = NewIntFlag("host-stats-sampling-interval", 5000)         // Host stats sampling interval in milliseconds (default 5s)
+	// 0 disables the step. Both default off; operator opts in once the host
+	// kernel has the FPH race fix and the fleet is ready.
+	FreePageHintingTimeoutMs      = NewIntFlag("free-page-hinting-timeout-ms", 0)
+	ReclaimOnPauseTimeoutMs       = NewIntFlag("reclaim-on-pause-timeout-ms", 0)
+	HostStatsSamplingInterval     = NewIntFlag("host-stats-sampling-interval", 5000) // Host stats sampling interval in milliseconds (default 5s)
 	MaxCacheWriterConcurrencyFlag = NewIntFlag("max-cache-writer-concurrency", 10)
 
 	// BuildCacheMaxUsagePercentage the maximum percentage of the cache disk storage