fix(cloudhypervisor): balloon settle counts progress only past 1% of guest memory per stall window

docs/snapshots.md

@@ -33,7 +33,7 @@ A snapshot contains the full VM state:
 
 `snapshot save` and `vm hibernate` are refused while a vhost-user-fs share, a VFIO device or a hot-attached disk is present; see [Devices](devices.md).
 
-On Cloud Hypervisor, `snapshot save` and `vm hibernate` first let a running guest finish inflating its balloon, waiting up to a minute and moving on once inflation stalls for 2 s (the wait reads the live target, so it covers a target raised by `vm balloon`), so a `vm restore` or hibernate wake comes back with a settled balloon instead of inflating it again at a high CPU cost. A clone moves to the create-time balloon target after restore: it deflates from a raised snapshot target, so a golden saved at the default target keeps the settled balloon.
+On Cloud Hypervisor, `snapshot save` and `vm hibernate` first let a running guest finish inflating its balloon, waiting up to a minute and moving on once inflation frees less than 1% of guest memory in 2 s (the wait reads the live target, so it covers a target raised by `vm balloon`), so a `vm restore` or hibernate wake comes back with a settled balloon instead of inflating it again at a high CPU cost. A clone moves to the create-time balloon target after restore: it deflates from a raised snapshot target, so a golden saved at the default target keeps the settled balloon.
 
 ### Clone Constraints
 

hypervisor/cloudhypervisor/snapshot.go

@@ -21,6 +21,7 @@ import (
 const (
 	balloonPollInterval  = 100 * time.Millisecond
 	balloonStallWindow   = 2 * time.Second
+	balloonStallPercent  = 1
 	balloonSettleTimeout = time.Minute
 
 	ejectRenotifyInterval = 5 * time.Second
@@ -182,8 +183,9 @@ func waitBalloonSettled(ctx context.Context, hc *http.Client, vmID string, info
 		return nil
 	}
 	target := info.Config.Memory.Size - balloon.Size
+	step := info.Config.Memory.Size * balloonStallPercent / 100
 	start := time.Now()
-	progressAt := start
+	progressAt, progressFrom := start, info.MemoryActualSize
 	for actual := info.MemoryActualSize; actual > target; {
 		if time.Since(progressAt) >= balloonStallWindow || time.Since(start) >= balloonSettleTimeout {
 			log.WithFunc("cloudhypervisor.waitBalloonSettled").Warnf(ctx, "snapshot VM %s with balloon at %d of %d bytes", vmID, info.Config.Memory.Size-actual, balloon.Size)
@@ -198,10 +200,10 @@ func waitBalloonSettled(ctx context.Context, hc *http.Client, vmID string, info
 		if err != nil {
 			return err
 		}
-		if cur.MemoryActualSize < actual {
-			progressAt = time.Now()
-		}
 		actual = cur.MemoryActualSize
+		if progressFrom-actual >= step {
+			progressAt, progressFrom = time.Now(), actual
+		}
 	}
 	return nil
 }

hypervisor/cloudhypervisor/snapshot_test.go

@@ -246,7 +246,8 @@ func TestSnapshotPauseWaitsForBalloonToSettle(t *testing.T) {
 		{"paused guest cannot inflate", chStatePaused, &chBalloon{Size: balloon}, func(int) int64 { return mem }, 0},
 		{"inflating reaches target", chStateRunning, &chBalloon{Size: balloon}, func(polls int) int64 { return mem - min(int64(polls)*(512<<20), balloon) }, 4 * balloonPollInterval},
 		{"stalled under guest memory pressure", chStateRunning, &chBalloon{Size: balloon}, func(int) int64 { return mem - (512 << 20) }, balloonStallWindow},
-		{"slow inflation stops at the cap", chStateRunning, &chBalloon{Size: balloon}, func(polls int) int64 { return mem - int64(polls)<<12 }, balloonSettleTimeout},
+		{"trickling inflation counts as stalled", chStateRunning, &chBalloon{Size: balloon}, func(polls int) int64 { return mem - int64(polls)<<12 }, balloonStallWindow},
+		{"slow inflation stops at the cap", chStateRunning, &chBalloon{Size: mem * 3 / 4}, func(polls int) int64 { return mem - int64(polls)*(mem/1000) }, balloonSettleTimeout},
 	}
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {