fix: resolve pause/snapshot failures and CoW exhaustion on large VMs

Remove hard 10s timeout from Firecracker HTTP client — callers already
pass context.Context with appropriate deadlines, and 20GB+ memfile
writes easily exceed 10s.

Ensure CoW file is at least as large as the origin rootfs. Previously,
WRENN_DEFAULT_ROOTFS_SIZE=30Gi expanded the base image to 30GB but the
default 5GB CoW could not hold all writes, causing dm-snapshot
invalidation and EIO on all guest I/O.

Destroy frozen VMs in resumeOnError instead of leaving zombies that
report "running" but can't execute. Use fresh context for the resume
attempt so a cancelled caller context doesn't falsely trigger destroy.

Increase CP→Agent ResponseHeaderTimeout from 45s to 5min and
PrepareSnapshot timeout from 3s to 30s for large-memory VMs.

After failed pause, ping agent to detect destroyed sandboxes and mark
DB status as "error" instead of reverting to "running".

internal/sandbox/manager.go

@@ -186,9 +186,12 @@ func (m *Manager) Create(ctx context.Context, sandboxID string, teamID, template
 	}
 
 	// Create dm-snapshot with per-sandbox CoW file.
+	// CoW must be at least as large as the origin — if every block is
+	// rewritten, the CoW stores a full copy. Undersized CoW causes
+	// dm-snapshot invalidation → EIO on all guest I/O.
 	dmName := "wrenn-" + sandboxID
 	cowPath := filepath.Join(layout.SandboxesDir(m.cfg.WrennDir), fmt.Sprintf("%s.cow", sandboxID))
-	cowSize := int64(diskSizeMB) * 1024 * 1024
+	cowSize := max(int64(diskSizeMB)*1024*1024, originSize)
 	dmDev, err := devicemapper.CreateSnapshot(dmName, originLoop, cowPath, originSize, cowSize)
 	if err != nil {
 		m.loops.Release(baseRootfs)
@@ -391,11 +394,13 @@ func (m *Manager) Pause(ctx context.Context, sandboxID string) error {
 	slog.Debug("pause: envd client idle connections closed", "id", sandboxID)
 
 	// Step 0c: Signal envd to quiesce (stop port scanner/forwarder, mark
-	// connections for post-restore cleanup). The 3s timeout also gives time
+	// connections for post-restore cleanup). Also drops page cache which
-	// for the FINs from Step 0b to be processed by the guest kernel.
+	// can take significant time on large-memory VMs (20GB+). The timeout
+	// also gives time for the FINs from Step 0b to be processed by the
+	// guest kernel.
 	// Best-effort: a failure is logged but does not abort the pause.
 	func() {
-		prepCtx, prepCancel := context.WithTimeout(ctx, 3*time.Second)
+		prepCtx, prepCancel := context.WithTimeout(ctx, 30*time.Second)
 		defer prepCancel()
 		if err := sb.client.PrepareSnapshot(prepCtx); err != nil {
 			slog.Warn("pause: pre-snapshot quiesce failed (best-effort)", "id", sandboxID, "error", err)
@@ -423,12 +428,24 @@ func (m *Manager) Pause(ctx context.Context, sandboxID string) error {
 
 	// resumeOnError unpauses the VM so the sandbox stays usable when a
 	// post-freeze step fails. If the resume itself fails, the sandbox is
-	// left frozen — the caller should destroy it. It also resets the
+	// frozen and unrecoverable — destroy it to avoid a zombie that reports
-	// connection tracker so the sandbox can accept proxy connections again.
+	// "running" but can't execute anything.
 	resumeOnError := func() {
 		sb.connTracker.Reset()
-		if err := m.vm.Resume(ctx, sandboxID); err != nil {
+		// Use a fresh context — the caller's ctx may already be cancelled
-			slog.Error("failed to resume VM after pause error — sandbox is frozen", "id", sandboxID, "error", err)
+		// (e.g. CP-side ResponseHeaderTimeout fired), which would make the
+		// resume fail immediately and destroy a perfectly resumable VM.
+		resumeCtx, resumeCancel := context.WithTimeout(context.Background(), 30*time.Second)
+		defer resumeCancel()
+		if err := m.vm.Resume(resumeCtx, sandboxID); err != nil {
+			slog.Error("failed to resume VM after pause error — destroying frozen sandbox", "id", sandboxID, "error", err)
+			m.cleanup(context.Background(), sb)
+			m.mu.Lock()
+			delete(m.boxes, sandboxID)
+			m.mu.Unlock()
+			if m.onDestroy != nil {
+				m.onDestroy(sandboxID)
+			}
 		}
 	}
 
@@ -444,6 +461,7 @@ func (m *Manager) Pause(ctx context.Context, sandboxID string) error {
 
 	snapshotStart := time.Now()
 	if err := m.vm.Snapshot(ctx, sandboxID, snapPath, rawMemPath, snapshotType); err != nil {
+		slog.Error("pause: snapshot failed", "id", sandboxID, "type", snapshotType, "elapsed", time.Since(snapshotStart), "error", err)
 		warnErr("snapshot dir cleanup error", sandboxID, os.RemoveAll(pauseDir))
 		resumeOnError()
 		return fmt.Errorf("create VM snapshot: %w", err)
@@ -1134,7 +1152,7 @@ func (m *Manager) createFromSnapshot(ctx context.Context, sandboxID string, team
 
 	dmName := "wrenn-" + sandboxID
 	cowPath := filepath.Join(layout.SandboxesDir(m.cfg.WrennDir), fmt.Sprintf("%s.cow", sandboxID))
-	cowSize := int64(diskSizeMB) * 1024 * 1024
+	cowSize := max(int64(diskSizeMB)*1024*1024, originSize)
 	dmDev, err := devicemapper.CreateSnapshot(dmName, originLoop, cowPath, originSize, cowSize)
 	if err != nil {
 		source.Close()

internal/vm/fc.go

@@ -8,7 +8,6 @@ import (
 	"io"
 	"net"
 	"net/http"
-	"time"
 )
 
 // fcClient talks to the Firecracker HTTP API over a Unix socket.
@@ -27,7 +26,9 @@ func newFCClient(socketPath string) *fcClient {
 					return d.DialContext(ctx, "unix", socketPath)
 				},
 			},
-			Timeout: 10 * time.Second,
+			// No global timeout — callers pass context.Context with appropriate
+			// deadlines. A fixed 10s timeout was too short for snapshot/resume
+			// operations on large-memory VMs (20GB+ memfiles).
 		},
 	}
 }

pkg/lifecycle/hostpool.go

@@ -47,7 +47,7 @@ func NewHostClientPoolTLS(tlsCfg *tls.Config) *HostClientPool {
 		TLSNextProto:          make(map[string]func(authority string, c *tls.Conn) http.RoundTripper),
 		MaxIdleConnsPerHost:   20,
 		IdleConnTimeout:       90 * time.Second,
-		ResponseHeaderTimeout: 45 * time.Second,
+		ResponseHeaderTimeout: 5 * time.Minute,
 		DialContext: (&net.Dialer{
 			Timeout:   10 * time.Second,
 			KeepAlive: 30 * time.Second,

pkg/service/sandbox.go

@@ -239,12 +239,26 @@ func (s *SandboxService) Pause(ctx context.Context, sandboxID, teamID pgtype.UUI
 	if _, err := agent.PauseSandbox(ctx, connect.NewRequest(&pb.PauseSandboxRequest{
 		SandboxId: sandboxIDStr,
 	})); err != nil {
-		// Revert status on failure.
+		// Check if the agent still has this sandbox. If it was destroyed
-		if _, dbErr := s.DB.UpdateSandboxStatus(ctx, db.UpdateSandboxStatusParams{
+		// (e.g. frozen VM couldn't be resumed), mark as "error" instead of
-			ID: sandboxID, Status: "running",
+		// reverting to "running" — which would create a ghost record.
+		// Use a fresh context since the original ctx may already be expired.
+		revertStatus := "running"
+		pingCtx, pingCancel := context.WithTimeout(context.Background(), 10*time.Second)
+		if _, pingErr := agent.PingSandbox(pingCtx, connect.NewRequest(&pb.PingSandboxRequest{
+			SandboxId: sandboxIDStr,
+		})); pingErr != nil {
+			revertStatus = "error"
+			slog.Warn("sandbox gone from agent after failed pause, marking as error", "sandbox_id", sandboxIDStr)
+		}
+		pingCancel()
+		dbCtx, dbCancel := context.WithTimeout(context.Background(), 5*time.Second)
+		if _, dbErr := s.DB.UpdateSandboxStatus(dbCtx, db.UpdateSandboxStatusParams{
+			ID: sandboxID, Status: revertStatus,
 		}); dbErr != nil {
 			slog.Warn("failed to revert sandbox status after pause error", "sandbox_id", sandboxIDStr, "error", dbErr)
 		}
+		dbCancel()
 		return db.Sandbox{}, fmt.Errorf("agent pause: %w", err)
 	}