cache: Only start at high watermark (#10403)

Currently, cache purges are triggered as soon as the low watermark is exceeded.
To reduce IO this should only be done when reaching the high watermark.
This simplifies checks and reduces all calls for a GC to go through
`dcache.diskSpaceAvailable(size)`. While a comment claims that 
`dcache.triggerGC <- struct{}{}` was non-blocking I don't see how 
that was possible. Instead, we add a 1 size to the queue channel 
and use channel  semantics to avoid blocking when a GC has 
already been requested.

`bytesToClear` now takes the high watermark into account to it will 
not request any bytes to be cleared until that is reached.

cmd/disk-cache-backend.go

@@ -157,7 +157,7 @@ func newDiskCache(ctx context.Context, dir string, config cache.Config) (*diskCa
 	}
 	cache := diskCache{
 		dir:           dir,
-		triggerGC:     make(chan struct{}),
+		triggerGC:     make(chan struct{}, 1),
 		stats:         CacheDiskStats{Dir: dir},
 		quotaPct:      quotaPct,
 		after:         config.After,
@@ -174,7 +174,7 @@ func newDiskCache(ctx context.Context, dir string, config cache.Config) (*diskCa
 		nsMutex: newNSLock(false),
 	}
 	go cache.purgeWait(ctx)
-	cache.diskUsageHigh() // update if cache usage is already high.
+	cache.diskSpaceAvailable(0) // update if cache usage is already high.
 	cache.NewNSLockFn = func(ctx context.Context, cachePath string) RWLocker {
 		return cache.nsMutex.NewNSLock(ctx, nil, cachePath, "")
 	}
@@ -203,9 +203,9 @@ func (c *diskCache) diskUsageLow() bool {
 	return low
 }
 
-// Returns if the disk usage reaches high water mark w.r.t the configured cache quota.
-// gc starts if high water mark reached.
-func (c *diskCache) diskUsageHigh() bool {
+// Returns if the disk usage reaches  or exceeds configured cache quota when size is added.
+// If current usage without size exceeds high watermark a GC is automatically queued.
+func (c *diskCache) diskSpaceAvailable(size int64) bool {
 	gcTriggerPct := c.quotaPct * c.highWatermark / 100
 	di, err := disk.GetInfo(c.dir)
 	if err != nil {
@@ -214,27 +214,30 @@ func (c *diskCache) diskUsageHigh() bool {
 		logger.LogIf(ctx, err)
 		return false
 	}
-	usedPercent := (di.Total - di.Free) * 100 / di.Total
-	high := int(usedPercent) >= gcTriggerPct
-	atomic.StoreUint64(&c.stats.UsagePercent, usedPercent)
-	if high {
+	if di.Total == 0 {
+		logger.Info("diskCache: Received 0 total disk size")
+		return false
+	}
+	usedPercent := float64(di.Total-di.Free) * 100 / float64(di.Total)
+	if usedPercent >= float64(gcTriggerPct) {
 		atomic.StoreInt32(&c.stats.UsageState, 1)
+		c.queueGC()
 	}
-	return high
+	atomic.StoreUint64(&c.stats.UsagePercent, uint64(usedPercent))
+
+	// Recalculate percentage with provided size added.
+	usedPercent = float64(di.Total-di.Free+uint64(size)) * 100 / float64(di.Total)
+
+	return usedPercent < float64(c.quotaPct)
 }
 
-// Returns if size space can be allocated without exceeding
-// max disk usable for caching
-func (c *diskCache) diskAvailable(size int64) bool {
-	di, err := disk.GetInfo(c.dir)
-	if err != nil {
-		reqInfo := (&logger.ReqInfo{}).AppendTags("cachePath", c.dir)
-		ctx := logger.SetReqInfo(GlobalContext, reqInfo)
-		logger.LogIf(ctx, err)
-		return false
+// queueGC will queue a GC.
+// Calling this function is always non-blocking.
+func (c *diskCache) queueGC() {
+	select {
+	case c.triggerGC <- struct{}{}:
+	default:
 	}
-	usedPercent := (di.Total - (di.Free - uint64(size))) * 100 / di.Total
-	return int(usedPercent) < c.quotaPct
 }
 
 // toClear returns how many bytes should be cleared to reach the low watermark quota.
@@ -247,7 +250,7 @@ func (c *diskCache) toClear() uint64 {
 		logger.LogIf(ctx, err)
 		return 0
 	}
-	return bytesToClear(int64(di.Total), int64(di.Free), uint64(c.quotaPct), uint64(c.lowWatermark))
+	return bytesToClear(int64(di.Total), int64(di.Free), uint64(c.quotaPct), uint64(c.lowWatermark), uint64(c.highWatermark))
 }
 
 var (
@@ -658,8 +661,7 @@ func newCacheEncryptMetadata(bucket, object string, metadata map[string]string)
 
 // Caches the object to disk
 func (c *diskCache) Put(ctx context.Context, bucket, object string, data io.Reader, size int64, rs *HTTPRangeSpec, opts ObjectOptions, incHitsOnly bool) error {
-	if c.diskUsageHigh() {
-		c.triggerGC <- struct{}{}
+	if !c.diskSpaceAvailable(size) {
 		io.Copy(ioutil.Discard, data)
 		return errDiskFull
 	}
@@ -688,7 +690,7 @@ func (c *diskCache) Put(ctx context.Context, bucket, object string, data io.Read
 	if rs != nil {
 		return c.putRange(ctx, bucket, object, data, size, rs, opts)
 	}
-	if !c.diskAvailable(size) {
+	if !c.diskSpaceAvailable(size) {
 		return errDiskFull
 	}
 	if err := os.MkdirAll(cachePath, 0777); err != nil {
@@ -730,7 +732,7 @@ func (c *diskCache) putRange(ctx context.Context, bucket, object string, data io
 	if err != nil {
 		return err
 	}
-	if !c.diskAvailable(rlen) {
+	if !c.diskSpaceAvailable(rlen) {
 		return errDiskFull
 	}
 	cachePath := getCacheSHADir(c.dir, bucket, object)

cmd/disk-cache-utils.go

@@ -489,9 +489,15 @@ func (f *fileScorer) queueString() string {
 // bytesToClear() returns the number of bytes to clear to reach low watermark
 // w.r.t quota given disk total and free space, quota in % allocated to cache
 // and low watermark % w.r.t allowed quota.
-func bytesToClear(total, free int64, quotaPct, lowWatermark uint64) uint64 {
-	used := (total - free)
+// If the high watermark hasn't been reached 0 will be returned.
+func bytesToClear(total, free int64, quotaPct, lowWatermark, highWatermark uint64) uint64 {
+	used := total - free
 	quotaAllowed := total * (int64)(quotaPct) / 100
-	lowWMUsage := (total * (int64)(lowWatermark*quotaPct) / (100 * 100))
+	highWMUsage := total * (int64)(highWatermark*quotaPct) / (100 * 100)
+	if used < highWMUsage {
+		return 0
+	}
+	// Return bytes needed to reach low watermark.
+	lowWMUsage := total * (int64)(lowWatermark*quotaPct) / (100 * 100)
 	return (uint64)(math.Min(float64(quotaAllowed), math.Max(0.0, float64(used-lowWMUsage))))
 }

cmd/disk-cache-utils_test.go

@@ -149,22 +149,26 @@ func TestNewFileScorer(t *testing.T) {
 }
 func TestBytesToClear(t *testing.T) {
 	testCases := []struct {
-		total        int64
-		free         int64
-		quotaPct     uint64
-		watermarkLow uint64
-		expected     uint64
+		total         int64
+		free          int64
+		quotaPct      uint64
+		watermarkLow  uint64
+		watermarkHigh uint64
+		expected      uint64
 	}{
-		{1000, 800, 40, 90, 0},
-		{1000, 200, 40, 90, 400},
-		{1000, 400, 40, 90, 240},
-		{1000, 600, 40, 90, 40},
-		{1000, 600, 40, 70, 120},
-		{1000, 1000, 90, 70, 0},
-		{1000, 0, 90, 70, 370},
+		{total: 1000, free: 800, quotaPct: 40, watermarkLow: 90, watermarkHigh: 90, expected: 0},
+		{total: 1000, free: 200, quotaPct: 40, watermarkLow: 90, watermarkHigh: 90, expected: 400},
+		{total: 1000, free: 400, quotaPct: 40, watermarkLow: 90, watermarkHigh: 90, expected: 240},
+		{total: 1000, free: 600, quotaPct: 40, watermarkLow: 90, watermarkHigh: 90, expected: 40},
+		{total: 1000, free: 600, quotaPct: 40, watermarkLow: 70, watermarkHigh: 70, expected: 120},
+		{total: 1000, free: 1000, quotaPct: 90, watermarkLow: 70, watermarkHigh: 70, expected: 0},
+
+		// High not yet reached..
+		{total: 1000, free: 250, quotaPct: 100, watermarkLow: 50, watermarkHigh: 90, expected: 0},
+		{total: 1000, free: 250, quotaPct: 100, watermarkLow: 50, watermarkHigh: 90, expected: 0},
 	}
 	for i, tc := range testCases {
-		toClear := bytesToClear(tc.total, tc.free, tc.quotaPct, tc.watermarkLow)
+		toClear := bytesToClear(tc.total, tc.free, tc.quotaPct, tc.watermarkLow, tc.watermarkHigh)
 		if tc.expected != toClear {
 			t.Errorf("test %d expected %v, got %v", i, tc.expected, toClear)
 		}

cmd/disk-cache.go

@@ -284,12 +284,6 @@ func (c *cacheObjects) GetObjectNInfo(ctx context.Context, bucket, object string
 	// Reaching here implies cache miss
 	c.cacheStats.incMiss()
 
-	// Since we got here, we are serving the request from backend,
-	// and also adding the object to the cache.
-	if dcache.diskUsageHigh() {
-		dcache.triggerGC <- struct{}{} // this is non-blocking
-	}
-
 	bkReader, bkErr := c.GetObjectNInfoFn(ctx, bucket, object, rs, h, lockType, opts)
 
 	if bkErr != nil {
@@ -306,7 +300,9 @@ func (c *cacheObjects) GetObjectNInfo(ctx context.Context, bucket, object string
 	if cacheErr == nil {
 		bkReader.ObjInfo.CacheLookupStatus = CacheHit
 	}
-	if !dcache.diskAvailable(objInfo.Size) {
+
+	// Check if we can add it without exceeding total cache size.
+	if !dcache.diskSpaceAvailable(objInfo.Size) {
 		return bkReader, bkErr
 	}
 
@@ -612,9 +608,10 @@ func (c *cacheObjects) PutObject(ctx context.Context, bucket, object string, r *
 	}
 
 	// fetch from backend if there is no space on cache drive
-	if !dcache.diskAvailable(size) {
+	if !dcache.diskSpaceAvailable(size) {
 		return putObjectFn(ctx, bucket, object, r, opts)
 	}
+
 	if opts.ServerSideEncryption != nil {
 		dcache.Delete(ctx, bucket, object)
 		return putObjectFn(ctx, bucket, object, r, opts)
@@ -721,7 +718,9 @@ func (c *cacheObjects) gc(ctx context.Context) {
 			}
 			for _, dcache := range c.cache {
 				if dcache != nil {
-					dcache.triggerGC <- struct{}{}
+					// Check if there is disk.
+					// Will queue a GC scan if at high watermark.
+					dcache.diskSpaceAvailable(0)
 				}
 			}
 		}

docs/disk-caching/README.md

@@ -29,6 +29,10 @@ export MINIO_CACHE_WATERMARK_HIGH=90
 minio gateway s3
 ```
 
+The `CACHE_WATERMARK` numbers are percentages of `CACHE_QUOTA`. 
+In the example above this means that  `MINIO_CACHE_WATERMARK_LOW` is effectively `0.8 * 0.7 * 100 = 56%` and the `MINIO_CACHE_WATERMARK_LOW` is effectively `0.8 * 0.9 * 100 = 72%` of total disk space.     
+
+
 ### 3. Test your setup
 
 To test this setup, access the MinIO gateway via browser or [`mc`](https://docs.min.io/docs/minio-client-quickstart-guide). You’ll see the uploaded files are accessible from all the MinIO endpoints.