eth/downloader: prioritize block fetch based on chain position, cap memory use

1 files changed, 282 insertions, 134 deletions

diff --git a/eth/downloader/queue.go b/eth/downloader/queue.go
index 1b63a5ffb..eae567052 100644
--- a/eth/downloader/queue.go
+++ b/eth/downloader/queue.go
@@ -1,201 +1,349 @@
 package downloader
 
 import (
+	"errors"
 	"fmt"
-	"math"
 	"sync"
 	"time"
 
 	"github.com/ethereum/go-ethereum/common"
 	"github.com/ethereum/go-ethereum/core/types"
-	"gopkg.in/fatih/set.v0"
+	"gopkg.in/karalabe/cookiejar.v2/collections/prque"
 )
 
+const (
+	blockCacheLimit = 4096 // Maximum number of blocks to cache before throttling the download
+)
+
+// fetchRequest is a currently running block retrieval operation.
+type fetchRequest struct {
+	Peer   *peer               // Peer to which the request was sent
+	Hashes map[common.Hash]int // Requested hashes with their insertion index (priority)
+	Time   time.Time           // Time when the request was made
+}
+
 // queue represents hashes that are either need fetching or are being fetched
 type queue struct {
-	hashPool    *set.Set
-	fetchPool   *set.Set
-	blockHashes *set.Set
+	hashPool    map[common.Hash]int // Pending hashes, mapping to their insertion index (priority)
+	hashQueue   *prque.Prque        // Priority queue of the block hashes to fetch
+	hashCounter int                 // Counter indexing the added hashes to ensure retrieval order
+
+	pendPool  map[string]*fetchRequest // Currently pending block retrieval operations
+	pendCount int                      // Number of pending block fetches (to throttle the download)
 
-	mu       sync.Mutex
-	fetching map[string]*chunk
+	blockPool   map[common.Hash]int // Hash-set of the downloaded data blocks, mapping to cache indexes
+	blockCache  []*types.Block      // Downloaded but not yet delivered blocks
+	blockOffset int                 // Offset of the first cached block in the block-chain
 
-	blockOffset int
-	blocks      []*types.Block
+	lock sync.RWMutex
 }
 
-func newqueue() *queue {
+// newQueue creates a new download queue for scheduling block retrieval.
+func newQueue() *queue {
 	return &queue{
-		hashPool:    set.New(),
-		fetchPool:   set.New(),
-		blockHashes: set.New(),
-		fetching:    make(map[string]*chunk),
+		hashPool:  make(map[common.Hash]int),
+		hashQueue: prque.New(),
+		pendPool:  make(map[string]*fetchRequest),
+		blockPool: make(map[common.Hash]int),
 	}
 }
 
-func (c *queue) reset() {
-	c.mu.Lock()
-	defer c.mu.Unlock()
+// Reset clears out the queue contents.
+func (q *queue) Reset() {
+	q.lock.Lock()
+	defer q.lock.Unlock()
 
-	c.resetNoTS()
+	q.hashPool = make(map[common.Hash]int)
+	q.hashQueue.Reset()
+	q.hashCounter = 0
+
+	q.pendPool = make(map[string]*fetchRequest)
+	q.pendCount = 0
+
+	q.blockPool = make(map[common.Hash]int)
+	q.blockOffset = 0
+	q.blockCache = nil
 }
-func (c *queue) resetNoTS() {
-	c.blockOffset = 0
-	c.hashPool.Clear()
-	c.fetchPool.Clear()
-	c.blockHashes.Clear()
-	c.blocks = nil
-	c.fetching = make(map[string]*chunk)
+
+// Done checks if all the downloads have been retrieved, wiping the queue.
+func (q *queue) Done() {
+	q.lock.Lock()
+	defer q.lock.Unlock()
+
+	if len(q.blockCache) == 0 {
+		q.Reset()
+	}
 }
 
-func (c *queue) size() int {
-	return c.hashPool.Size() + c.blockHashes.Size() + c.fetchPool.Size()
+// Size retrieves the number of hashes in the queue, returning separately for
+// pending and already downloaded.
+func (q *queue) Size() (int, int) {
+	q.lock.RLock()
+	defer q.lock.RUnlock()
+
+	return len(q.hashPool), len(q.blockPool)
 }
 
-// reserve a `max` set of hashes for `p` peer.
-func (c *queue) get(p *peer, max int) *chunk {
-	c.mu.Lock()
-	defer c.mu.Unlock()
+// Pending retrieves the number of hashes pending for retrieval.
+func (q *queue) Pending() int {
+	q.lock.RLock()
+	defer q.lock.RUnlock()
 
-	// return nothing if the pool has been depleted
-	if c.hashPool.Size() == 0 {
-		return nil
-	}
+	return q.hashQueue.Size()
+}
 
-	limit := int(math.Min(float64(max), float64(c.hashPool.Size())))
-	// Create a new set of hashes
-	hashes, i := set.New(), 0
-	c.hashPool.Each(func(v interface{}) bool {
-		// break on limit
-		if i == limit {
-			return false
-		}
-		// skip any hashes that have previously been requested from the peer
-		if p.ignored.Has(v) {
-			return true
-		}
+// InFlight retrieves the number of fetch requests currently in flight.
+func (q *queue) InFlight() int {
+	q.lock.RLock()
+	defer q.lock.RUnlock()
 
-		hashes.Add(v)
-		i++
+	return len(q.pendPool)
+}
 
+// Throttle checks if the download should be throttled (active block fetches
+// exceed block cache).
+func (q *queue) Throttle() bool {
+	q.lock.RLock()
+	defer q.lock.RUnlock()
+
+	return q.pendCount >= len(q.blockCache)-len(q.blockPool)
+}
+
+// Has checks if a hash is within the download queue or not.
+func (q *queue) Has(hash common.Hash) bool {
+	q.lock.RLock()
+	defer q.lock.RUnlock()
+
+	if _, ok := q.hashPool[hash]; ok {
+		return true
+	}
+	if _, ok := q.blockPool[hash]; ok {
 		return true
-	})
-	// if no hashes can be requested return a nil chunk
-	if hashes.Size() == 0 {
-		return nil
 	}
+	return false
+}
 
-	// remove the fetchable hashes from hash pool
-	c.hashPool.Separate(hashes)
-	c.fetchPool.Merge(hashes)
+// Insert adds a set of hashes for the download queue for scheduling.
+func (q *queue) Insert(hashes []common.Hash) {
+	q.lock.Lock()
+	defer q.lock.Unlock()
 
-	// Create a new chunk for the seperated hashes. The time is being used
-	// to reset the chunk (timeout)
-	chunk := &chunk{p, hashes, time.Now()}
-	// register as 'fetching' state
-	c.fetching[p.id] = chunk
+	// Insert all the hashes prioritized in the arrival order
+	for i, hash := range hashes {
+		index := q.hashCounter + i
 
-	// create new chunk for peer
-	return chunk
+		q.hashPool[hash] = index
+		q.hashQueue.Push(hash, float32(index)) // Highest gets schedules first
+	}
+	// Update the hash counter for the next batch of inserts
+	q.hashCounter += len(hashes)
 }
 
-func (c *queue) has(hash common.Hash) bool {
-	return c.hashPool.Has(hash) || c.fetchPool.Has(hash) || c.blockHashes.Has(hash)
+// GetHeadBlock retrieves the first block from the cache, or nil if it hasn't
+// been downloaded yet (or simply non existent).
+func (q *queue) GetHeadBlock() *types.Block {
+	q.lock.RLock()
+	defer q.lock.RUnlock()
+
+	if len(q.blockCache) == 0 {
+		return nil
+	}
+	return q.blockCache[0]
 }
 
-func (c *queue) getBlock(hash common.Hash) *types.Block {
-	c.mu.Lock()
-	defer c.mu.Unlock()
+// GetBlock retrieves a downloaded block, or nil if non-existent.
+func (q *queue) GetBlock(hash common.Hash) *types.Block {
+	q.lock.RLock()
+	defer q.lock.RUnlock()
 
-	if !c.blockHashes.Has(hash) {
+	// Short circuit if the block hasn't been downloaded yet
+	index, ok := q.blockPool[hash]
+	if !ok {
 		return nil
 	}
-
-	for _, block := range c.blocks {
-		if block.Hash() == hash {
-			return block
-		}
+	// Return the block if it's still available in the cache
+	if q.blockOffset <= index && index < q.blockOffset+len(q.blockCache) {
+		return q.blockCache[index-q.blockOffset]
 	}
 	return nil
 }
 
-// deliver delivers a chunk to the queue that was requested of the peer
-func (c *queue) deliver(id string, blocks []*types.Block) (err error) {
-	c.mu.Lock()
-	defer c.mu.Unlock()
-
-	chunk := c.fetching[id]
-	// If the chunk was never requested simply ignore it
-	if chunk != nil {
-		delete(c.fetching, id)
-		// check the length of the returned blocks. If the length of blocks is 0
-		// we'll assume the peer doesn't know about the chain.
-		if len(blocks) == 0 {
-			// So we can ignore the blocks we didn't know about
-			chunk.peer.ignored.Merge(chunk.hashes)
-		}
+// TakeBlocks retrieves and permanently removes a batch of blocks from the cache.
+// The head parameter is required to prevent a race condition where concurrent
+// takes may fail parent verifications.
+func (q *queue) TakeBlocks(head *types.Block) types.Blocks {
+	q.lock.Lock()
+	defer q.lock.Unlock()
 
-		// Add the blocks
-		for i, block := range blocks {
-			// See (1) for future limitation
-			n := int(block.NumberU64()) - c.blockOffset
-			if n > len(c.blocks) || n < 0 {
-				// set the error and set the blocks which could be processed
-				// abort the rest of the blocks (FIXME this could be improved)
-				err = fmt.Errorf("received block which overflow (N=%v O=%v)", block.Number(), c.blockOffset)
-				blocks = blocks[:i]
-				break
-			}
-			c.blocks[n] = block
+	// Short circuit if the head block's different
+	if len(q.blockCache) == 0 || q.blockCache[0] != head {
+		return nil
+	}
+	// Otherwise accumulate all available blocks
+	var blocks types.Blocks
+	for _, block := range q.blockCache {
+		if block == nil {
+			break
 		}
-		// seperate the blocks and the hashes
-		blockHashes := chunk.fetchedHashes(blocks)
-		// merge block hashes
-		c.blockHashes.Merge(blockHashes)
-		// Add back whatever couldn't be delivered
-		c.hashPool.Merge(chunk.hashes)
-		// Remove the hashes from the fetch pool
-		c.fetchPool.Separate(chunk.hashes)
+		blocks = append(blocks, block)
+		delete(q.blockPool, block.Hash())
+	}
+	// Delete the blocks from the slice and let them be garbage collected
+	// without this slice trick the blocks would stay in memory until nil
+	// would be assigned to q.blocks
+	copy(q.blockCache, q.blockCache[len(blocks):])
+	for k, n := len(q.blockCache)-len(blocks), len(q.blockCache); k < n; k++ {
+		q.blockCache[k] = nil
 	}
+	q.blockOffset += len(blocks)
 
-	return
+	return blocks
 }
 
-func (c *queue) alloc(offset, size int) {
-	c.mu.Lock()
-	defer c.mu.Unlock()
+// Reserve reserves a set of hashes for the given peer, skipping any previously
+// failed download.
+func (q *queue) Reserve(p *peer, max int) *fetchRequest {
+	q.lock.Lock()
+	defer q.lock.Unlock()
 
-	if c.blockOffset < offset {
-		c.blockOffset = offset
+	// Short circuit if the pool has been depleted
+	if q.hashQueue.Empty() {
+		return nil
 	}
-
-	// (1) XXX at some point we could limit allocation to memory and use the disk
-	// to store future blocks.
-	if len(c.blocks) < size {
-		c.blocks = append(c.blocks, make([]*types.Block, size)...)
+	// Retrieve a batch of hashes, skipping previously failed ones
+	send := make(map[common.Hash]int)
+	skip := make(map[common.Hash]int)
+
+	for len(send) < max && !q.hashQueue.Empty() {
+		hash, priority := q.hashQueue.Pop()
+		if p.ignored.Has(hash) {
+			skip[hash.(common.Hash)] = int(priority)
+		} else {
+			send[hash.(common.Hash)] = int(priority)
+		}
+	}
+	// Merge all the skipped hashes back
+	for hash, index := range skip {
+		q.hashQueue.Push(hash, float32(index))
+	}
+	// Assemble and return the block download request
+	if len(send) == 0 {
+		return nil
 	}
+	request := &fetchRequest{
+		Peer:   p,
+		Hashes: send,
+		Time:   time.Now(),
+	}
+	q.pendPool[p.id] = request
+	q.pendCount += len(request.Hashes)
+
+	return request
 }
 
-// puts puts sets of hashes on to the queue for fetching
-func (c *queue) put(hashes *set.Set) {
-	c.mu.Lock()
-	defer c.mu.Unlock()
+// Cancel aborts a fetch request, returning all pending hashes to the queue.
+func (q *queue) Cancel(request *fetchRequest) {
+	q.lock.Lock()
+	defer q.lock.Unlock()
 
-	c.hashPool.Merge(hashes)
+	for hash, index := range request.Hashes {
+		q.hashQueue.Push(hash, float32(index))
+	}
+	delete(q.pendPool, request.Peer.id)
+	q.pendCount -= len(request.Hashes)
 }
 
-type chunk struct {
-	peer   *peer
-	hashes *set.Set
-	itime  time.Time
+// Expire checks for in flight requests that exceeded a timeout allowance,
+// canceling them and returning the responsible peers for penalization.
+func (q *queue) Expire(timeout time.Duration) []string {
+	q.lock.Lock()
+	defer q.lock.Unlock()
+
+	// Iterate over the expired requests and return each to the queue
+	peers := []string{}
+	for id, request := range q.pendPool {
+		if time.Since(request.Time) > timeout {
+			for hash, index := range request.Hashes {
+				q.hashQueue.Push(hash, float32(index))
+			}
+			q.pendCount -= len(request.Hashes)
+			peers = append(peers, id)
+		}
+	}
+	// Remove the expired requests from the pending pool
+	for _, id := range peers {
+		delete(q.pendPool, id)
+	}
+	return peers
 }
 
-func (ch *chunk) fetchedHashes(blocks []*types.Block) *set.Set {
-	fhashes := set.New()
+// Deliver injects a block retrieval response into the download queue.
+func (q *queue) Deliver(id string, blocks []*types.Block) (err error) {
+	q.lock.Lock()
+	defer q.lock.Unlock()
+
+	// Short circuit if the blocks were never requested
+	request := q.pendPool[id]
+	if request == nil {
+		return errors.New("no fetches pending")
+	}
+	delete(q.pendPool, id)
+
+	// Mark all the hashes in the request as non-pending
+	q.pendCount -= len(request.Hashes)
+
+	// If no blocks were retrieved, mark them as unavailable for the origin peer
+	if len(blocks) == 0 {
+		for hash, _ := range request.Hashes {
+			request.Peer.ignored.Add(hash)
+		}
+	}
+	// Iterate over the downloaded blocks and add each of them
+	errs := make([]error, 0)
 	for _, block := range blocks {
-		fhashes.Add(block.Hash())
+		// Skip any blocks that fall outside the cache range
+		index := int(block.NumberU64()) - q.blockOffset
+		if index >= len(q.blockCache) || index < 0 {
+			//fmt.Printf("block cache overflown (N=%v O=%v, C=%v)", block.Number(), q.blockOffset, len(q.blockCache))
+			continue
+		}
+		// Skip any blocks that were not requested
+		hash := block.Hash()
+		if _, ok := request.Hashes[hash]; !ok {
+			errs = append(errs, fmt.Errorf("non-requested block %v", hash))
+			continue
+		}
+		// Otherwise merge the block and mark the hash block
+		q.blockCache[index] = block
+
+		delete(request.Hashes, hash)
+		delete(q.hashPool, hash)
+		q.blockPool[hash] = int(block.NumberU64())
 	}
-	ch.hashes.Separate(fhashes)
+	// Return all failed fetches to the queue
+	for hash, index := range request.Hashes {
+		q.hashQueue.Push(hash, float32(index))
+	}
+	if len(errs) != 0 {
+		return fmt.Errorf("multiple failures: %v", errs)
+	}
+	return nil
+}
 
-	return fhashes
+// Alloc ensures that the block cache is the correct size, given a starting
+// offset, and a memory cap.
+func (q *queue) Alloc(offset int) {
+	q.lock.Lock()
+	defer q.lock.Unlock()
+
+	if q.blockOffset < offset {
+		q.blockOffset = offset
+	}
+	size := len(q.hashPool)
+	if size > blockCacheLimit {
+		size = blockCacheLimit
+	}
+	if len(q.blockCache) < size {
+		q.blockCache = append(q.blockCache, make([]*types.Block, size-len(q.blockCache))...)
+	}
 }