backend/s3: use memory pool for buffer allocations

Currently each multipart upload allocated his own buffers, which after
file upload was garbaged. Next files couldn't leverage already allocated
memory which resulted in inefficent memory management. This change
introduces backend memory pool keeping memory chunks which can be
used during object operations.

Fixes #3967

backend/s3/s3.go

@@ -56,6 +56,7 @@ import (
 	"github.com/rclone/rclone/lib/bucket"
 	"github.com/rclone/rclone/lib/encoder"
 	"github.com/rclone/rclone/lib/pacer"
+	"github.com/rclone/rclone/lib/pool"
 	"github.com/rclone/rclone/lib/readers"
 	"github.com/rclone/rclone/lib/rest"
 	"golang.org/x/sync/errgroup"
@@ -841,11 +842,24 @@ In Ceph, this can be increased with the "rgw list buckets max chunk" option.
 			// - doubled / encoding
 			// - trailing / encoding
 			// so that AWS keys are always valid file names
-			Default: (encoder.EncodeInvalidUtf8 |
+			Default: encoder.EncodeInvalidUtf8 |
 				encoder.EncodeSlash |
-				encoder.EncodeDot),
-		}},
-	})
+				encoder.EncodeDot,
+		}, {
+			Name:     "memory_pool_flush_time",
+			Default:  memoryPoolFlushTime,
+			Advanced: true,
+			Help: `How often internal memory buffer pools will be flushed.
+
+Uploads which requires additional buffers (f.e multipart) will use memory pool for allocations.
+This option controls how often unused buffers will be removed from the pool.`,
+		}, {
+			Name:     "memory_pool_use_mmap",
+			Default:  memoryPoolUseMmap,
+			Advanced: true,
+			Help:     `Whether to use mmap buffers in internal memory pool.`,
+		},
+		}})
 }
 
 // Constants
@@ -859,6 +873,9 @@ const (
 	defaultUploadCutoff = fs.SizeSuffix(200 * 1024 * 1024)
 	maxUploadCutoff     = fs.SizeSuffix(5 * 1024 * 1024 * 1024)
 	minSleep            = 10 * time.Millisecond // In case of error, start at 10ms sleep.
+
+	memoryPoolFlushTime = fs.Duration(time.Minute) // flush the cached buffers after this long
+	memoryPoolUseMmap   = false
 )
 
 // Options defines the configuration for this backend
@@ -887,6 +904,8 @@ type Options struct {
 	LeavePartsOnError     bool                 `config:"leave_parts_on_error"`
 	ListChunk             int64                `config:"list_chunk"`
 	Enc                   encoder.MultiEncoder `config:"encoding"`
+	MemoryPoolFlushTime   fs.Duration          `config:"memory_pool_flush_time"`
+	MemoryPoolUseMmap     bool                 `config:"memory_pool_use_mmap"`
 }
 
 // Fs represents a remote s3 server
@@ -902,6 +921,8 @@ type Fs struct {
 	cache         *bucket.Cache        // cache for bucket creation status
 	pacer         *fs.Pacer            // To pace the API calls
 	srv           *http.Client         // a plain http client
+	poolMu        sync.Mutex           // mutex protecting memory pools map
+	pools         map[int64]*pool.Pool // memory pools
 }
 
 // Object describes a s3 object
@@ -1180,6 +1201,7 @@ func NewFs(name, root string, m configmap.Mapper) (fs.Fs, error) {
 	if err != nil {
 		return nil, err
 	}
+
 	f := &Fs{
 		name:  name,
 		opt:   *opt,
@@ -1188,7 +1210,9 @@ func NewFs(name, root string, m configmap.Mapper) (fs.Fs, error) {
 		pacer: fs.NewPacer(pacer.NewS3(pacer.MinSleep(minSleep))),
 		cache: bucket.NewCache(),
 		srv:   fshttp.NewClient(fs.Config),
+		pools: make(map[int64]*pool.Pool),
 	}
+
 	f.setRoot(root)
 	f.features = (&fs.Features{
 		ReadMimeType:      true,
@@ -1875,6 +1899,22 @@ func (f *Fs) Hashes() hash.Set {
 	return hash.Set(hash.MD5)
 }
 
+func (f *Fs) getMemoryPool(size int64) *pool.Pool {
+	f.poolMu.Lock()
+	defer f.poolMu.Unlock()
+
+	_, ok := f.pools[size]
+	if !ok {
+		f.pools[size] = pool.New(
+			time.Duration(f.opt.MemoryPoolFlushTime),
+			int(f.opt.ChunkSize),
+			f.opt.UploadConcurrency*fs.Config.Transfers,
+			f.opt.MemoryPoolUseMmap,
+		)
+	}
+	return f.pools[size]
+}
+
 // ------------------------------------------------------------
 
 // Fs returns the parent Fs
@@ -2078,16 +2118,7 @@ func (o *Object) uploadMultipart(ctx context.Context, req *s3.PutObjectInput, si
 	if concurrency < 1 {
 		concurrency = 1
 	}
-	bufs := make(chan []byte, concurrency)
-	defer func() {
-		// empty the channel on exit
-		close(bufs)
-		for range bufs {
-		}
-	}()
-	for i := 0; i < concurrency; i++ {
-		bufs <- nil
-	}
+	tokens := pacer.NewTokenDispenser(concurrency)
 
 	// calculate size of parts
 	partSize := int(f.opt.ChunkSize)
@@ -2108,6 +2139,8 @@ func (o *Object) uploadMultipart(ctx context.Context, req *s3.PutObjectInput, si
 		}
 	}
 
+	memPool := f.getMemoryPool(int64(partSize))
+
 	var cout *s3.CreateMultipartUploadOutput
 	err = f.pacer.Call(func() (bool, error) {
 		var err error
@@ -2159,11 +2192,9 @@ func (o *Object) uploadMultipart(ctx context.Context, req *s3.PutObjectInput, si
 	)
 
 	for partNum := int64(1); !finished; partNum++ {
-		// Get a block of memory from the channel (which limits concurrency)
-		buf := <-bufs
-		if buf == nil {
-			buf = make([]byte, partSize)
-		}
+		// Get a block of memory from the pool and token which limits concurrency.
+		tokens.Get()
+		buf := memPool.Get()
 
 		// Fail fast, in case an errgroup managed function returns an error
 		// gCtx is cancelled. There is no point in uploading all the other parts.
@@ -2226,8 +2257,9 @@ func (o *Object) uploadMultipart(ctx context.Context, req *s3.PutObjectInput, si
 				return false, nil
 			})
 
-			// return the memory
-			bufs <- buf[:partSize]
+			// return the memory and token
+			memPool.Put(buf[:partSize])
+			tokens.Put()
 
 			if err != nil {
 				return errors.Wrap(err, "multipart upload failed to upload part")