github.com/anacrolix/torrent@v1.61.0/requesting.go

package torrent

import (
	"cmp"
	"context"
	"encoding/gob"
	"fmt"
	"reflect"
	"runtime/pprof"
	"time"
	"unsafe"

	"github.com/RoaringBitmap/roaring"
	g "github.com/anacrolix/generics"
	"github.com/anacrolix/generics/heap"
	"github.com/anacrolix/log"
	"github.com/anacrolix/multiless"

	requestStrategy "github.com/anacrolix/torrent/internal/request-strategy"
	"github.com/anacrolix/torrent/metainfo"
	typedRoaring "github.com/anacrolix/torrent/typed-roaring"
)

type (
	// Since we have to store all the requests in memory, we can't reasonably exceed what could be
	// indexed with the memory space available.
	maxRequests = int
)

func (t *Torrent) requestStrategyPieceOrderState(i int) requestStrategy.PieceRequestOrderState {
	t.slogger().Debug("requestStrategyPieceOrderState", "pieceIndex", i)
	return requestStrategy.PieceRequestOrderState{
		Priority:     t.piece(i).purePriority(),
		Partial:      t.piecePartiallyDownloaded(i),
		Availability: t.piece(i).availability(),
	}
}

func init() {
	gob.Register(peerId{})
}

type peerId struct {
	*Peer
	ptr uintptr
}

func (p peerId) Uintptr() uintptr {
	return p.ptr
}

func (p peerId) GobEncode() (b []byte, _ error) {
	*(*reflect.SliceHeader)(unsafe.Pointer(&b)) = reflect.SliceHeader{
		Data: uintptr(unsafe.Pointer(&p.ptr)),
		Len:  int(unsafe.Sizeof(p.ptr)),
		Cap:  int(unsafe.Sizeof(p.ptr)),
	}
	return
}

func (p *peerId) GobDecode(b []byte) error {
	if uintptr(len(b)) != unsafe.Sizeof(p.ptr) {
		panic(len(b))
	}
	ptr := unsafe.Pointer(&b[0])
	p.ptr = *(*uintptr)(ptr)
	log.Printf("%p", ptr)
	dst := reflect.SliceHeader{
		Data: uintptr(unsafe.Pointer(&p.Peer)),
		Len:  int(unsafe.Sizeof(p.Peer)),
		Cap:  int(unsafe.Sizeof(p.Peer)),
	}
	copy(*(*[]byte)(unsafe.Pointer(&dst)), b)
	return nil
}

type (
	// A request index is a chunk indexed across the entire torrent. It's a single integer and can
	// be converted to a protocol request. TODO: This should be private.
	RequestIndex = requestStrategy.RequestIndex
	// This is request index but per-piece.
	chunkIndexType    = requestStrategy.ChunkIndex
	webseedSliceIndex RequestIndex
)

type desiredPeerRequests struct {
	requestIndexes []RequestIndex
	peer           *PeerConn
	pieceStates    []g.Option[requestStrategy.PieceRequestOrderState]
}

func (p *desiredPeerRequests) lessByValue(leftRequest, rightRequest RequestIndex) bool {
	t := p.peer.t
	leftPieceIndex := t.pieceIndexOfRequestIndex(leftRequest)
	rightPieceIndex := t.pieceIndexOfRequestIndex(rightRequest)
	ml := multiless.New()
	// Push requests that can't be served right now to the end. But we don't throw them away unless
	// there's a better alternative. This is for when we're using the fast extension and get choked
	// but our requests could still be good when we get unchoked.
	if p.peer.peerChoking {
		ml = ml.Bool(
			!p.peer.peerAllowedFast.Contains(leftPieceIndex),
			!p.peer.peerAllowedFast.Contains(rightPieceIndex),
		)
	}
	leftPiece := p.pieceStates[leftPieceIndex].UnwrapPtr()
	rightPiece := p.pieceStates[rightPieceIndex].UnwrapPtr()
	// Putting this first means we can steal requests from lesser-performing peers for our first few
	// new requests.
	priority := func() PiecePriority {
		// Technically we would be happy with the cached priority here, except we don't actually
		// cache it anymore, and Torrent.PiecePriority just does another lookup of *Piece to resolve
		// the priority through Piece.purePriority, which is probably slower.
		leftPriority := leftPiece.Priority
		rightPriority := rightPiece.Priority
		ml = ml.Int(
			-int(leftPriority),
			-int(rightPriority),
		)
		if !ml.Ok() {
			if leftPriority != rightPriority {
				panic("expected equal")
			}
		}
		return leftPriority
	}()
	if ml.Ok() {
		return ml.MustLess()
	}
	leftRequestState := t.requestState[leftRequest]
	rightRequestState := t.requestState[rightRequest]
	leftPeer := leftRequestState.peer.Value()
	rightPeer := rightRequestState.peer.Value()
	// Prefer chunks already requested from this peer.
	ml = ml.Bool(rightPeer == p.peer, leftPeer == p.peer)
	// Prefer unrequested chunks.
	ml = ml.Bool(rightPeer == nil, leftPeer == nil)
	if ml.Ok() {
		return ml.MustLess()
	}
	if leftPeer != nil {
		// The right peer should also be set, or we'd have resolved the computation by now.
		ml = ml.Uint64(
			rightPeer.requestState.Requests.GetCardinality(),
			leftPeer.requestState.Requests.GetCardinality(),
		)
		// Could either of the lastRequested be Zero? That's what checking an existing peer is for.
		leftLast := leftRequestState.when
		rightLast := rightRequestState.when
		if leftLast.IsZero() || rightLast.IsZero() {
			panic("expected non-zero last requested times")
		}
		// We want the most-recently requested on the left. Clients like Transmission serve requests
		// in received order, so the most recently-requested is the one that has the longest until
		// it will be served and therefore is the best candidate to cancel.
		ml = ml.CmpInt64(rightLast.Sub(leftLast).Nanoseconds())
	}
	// Just trigger on any webseed requests present on the Torrent. That suggests that the Torrent
	// or files are prioritized enough to compete with PeerConn requests. Later we could filter on
	// webseeds actually requesting or supporting requests for the pieces we're comparing.
	if t.hasActiveWebseedRequests() {
		// Prefer the highest possible request index, since webseeds prefer the lowest. Additionally,
		// this should mean remote clients serve in reverse order so we meet webseeds responses in
		// the middle.
		ml = ml.Cmp(-cmp.Compare(leftRequest, rightRequest))
	} else {
		ml = ml.Int(
			leftPiece.Availability,
			rightPiece.Availability)
		if priority == PiecePriorityReadahead {
			// TODO: For readahead in particular, it would be even better to consider distance from the
			// reader position so that reads earlier in a torrent don't starve reads later in the
			// torrent. This would probably require reconsideration of how readahead priority works.
			ml = ml.Int(leftPieceIndex, rightPieceIndex)
		} else {
			ml = ml.Int(t.pieceRequestOrder[leftPieceIndex], t.pieceRequestOrder[rightPieceIndex])
		}
		ml = multiless.EagerOrdered(ml, leftRequest, rightRequest)
	}
	// Prefer request indexes in order for storage write performance. Since the heap request heap
	// does not contain duplicates, if we order at the request index level we should never have any
	// ambiguity.
	return ml.MustLess()
}

type desiredRequestState struct {
	Requests   desiredPeerRequests
	Interested bool
}

func (cl *Client) getRequestablePieces(key clientPieceRequestOrderKeySumType, f requestStrategy.RequestPieceFunc) {
	input := key.getRequestStrategyInput(cl)
	order := cl.pieceRequestOrder[key].pieces
	requestStrategy.GetRequestablePieces(input, order, f)
}

func (t *Torrent) getRequestablePieces(f requestStrategy.RequestPieceFunc) {
	t.cl.getRequestablePieces(t.clientPieceRequestOrderKey(), f)
}

// This gets the best-case request state. That means handling pieces limited by capacity, preferring
// earlier pieces, low availability etc. It pays no attention to existing requests on the peer or
// other peers. Those are handled later.
func (p *PeerConn) getDesiredRequestState() (desired desiredRequestState) {
	t := p.t
	if !t.haveInfo() {
		return
	}
	if t.closed.IsSet() {
		return
	}
	if t.dataDownloadDisallowed.Bool() {
		return
	}
	requestHeap := desiredPeerRequests{
		peer:           p,
		pieceStates:    t.requestPieceStates,
		requestIndexes: t.requestIndexes,
	}
	clear(requestHeap.pieceStates)
	t.logPieceRequestOrder()
	// Caller-provided allocation for roaring bitmap iteration.
	var it typedRoaring.Iterator[RequestIndex]
	t.getRequestablePieces(
		func(ih metainfo.Hash, pieceIndex int, pieceExtra requestStrategy.PieceRequestOrderState) bool {
			if ih != *t.canonicalShortInfohash() {
				return true
			}
			if !p.peerHasPiece(pieceIndex) {
				return true
			}
			requestHeap.pieceStates[pieceIndex].Set(pieceExtra)
			allowedFast := p.peerAllowedFast.Contains(pieceIndex)
			t.iterUndirtiedRequestIndexesInPiece(&it, pieceIndex, func(r requestStrategy.RequestIndex) {
				if !allowedFast {
					// We must signal interest to request this. TODO: We could set interested if the
					// peers pieces (minus the allowed fast set) overlap with our missing pieces if
					// there are any readers, or any pending pieces.
					desired.Interested = true
					// We can make or will allow sustaining a request here if we're not choked, or
					// have made the request previously (presumably while unchoked), and haven't had
					// the peer respond yet (and the request was retained because we are using the
					// fast extension).
					if p.peerChoking && !p.requestState.Requests.Contains(r) {
						// We can't request this right now.
						return
					}
				}
				cancelled := &p.requestState.Cancelled
				if !cancelled.IsEmpty() && cancelled.Contains(r) {
					// Can't re-request while awaiting acknowledgement.
					return
				}
				requestHeap.requestIndexes = append(requestHeap.requestIndexes, r)
			})
			return true
		},
	)
	t.assertPendingRequests()
	desired.Requests = requestHeap
	return
}

// Update requests if there's a reason assigned.
func (p *PeerConn) maybeUpdateActualRequestState() {
	if p.needRequestUpdate == "" {
		return
	}
	p.updateRequestsWithReason(p.needRequestUpdate)
}

// Updates requests right now with the given reason. Clobbers any deferred reason if there was one.
// Does all the necessary checks and includes profiler tags to assign the overhead.
func (p *PeerConn) updateRequestsWithReason(reason updateRequestReason) {
	if p.closed.IsSet() {
		return
	}
	if reason == peerUpdateRequestsTimerReason {
		since := time.Since(p.lastRequestUpdate)
		if since < updateRequestsTimerDuration {
			panic(since)
		}
	}
	if p.t.cl.config.Debug {
		p.logger.Slogger().Debug("updating requests", "reason", p.needRequestUpdate)
	}
	pprof.Do(
		context.Background(),
		pprof.Labels("update request", string(reason)),
		func(_ context.Context) {
			p.updateRequests()
		},
	)
	// Is there any chance we should leave this to run again, and have the caller clear it if they
	// called with this reason?
	p.needRequestUpdate = ""
	p.lastRequestUpdate = time.Now()
	if enableUpdateRequestsTimer {
		p.updateRequestsTimer.Reset(updateRequestsTimerDuration)
	}
}

func (p *PeerConn) updateRequests() {
	next := p.getDesiredRequestState()
	p.applyRequestState(next)
	p.t.cacheNextRequestIndexesForReuse(next.Requests.requestIndexes)
}

func (t *Torrent) cacheNextRequestIndexesForReuse(slice []RequestIndex) {
	// The incoming slice can be smaller when getDesiredRequestState short circuits on some
	// conditions.
	if cap(slice) > cap(t.requestIndexes) {
		t.requestIndexes = slice[:0]
	}
}

// Whether we should allow sending not interested ("losing interest") to the peer. I noticed
// qBitTorrent seems to punish us for sending not interested when we're streaming and don't
// currently need anything.
func (p *Peer) allowSendNotInterested() bool {
	// Except for caching, we're not likely to lose pieces very soon.
	if p.t.haveAllPieces() {
		return true
	}
	all, known := p.peerHasAllPieces()
	if all || !known {
		return false
	}
	// Allow losing interest if we have all the pieces the peer has.
	return roaring.AndNot(p.peerPieces(), &p.t._completedPieces).IsEmpty()
}

// Transmit/action the request state to the peer. This includes work-stealing from other peers and
// some piece order randomization within the preferred state calculated earlier in next. Cancels are
// not done here, those are handled synchronously. We only track pending cancel acknowledgements.
func (p *PeerConn) applyRequestState(next desiredRequestState) {
	current := &p.requestState
	// Make interest sticky
	if !next.Interested && p.requestState.Interested {
		if !p.allowSendNotInterested() {
			next.Interested = true
		}
	}
	if !p.setInterested(next.Interested) {
		return
	}
	more := true
	orig := next.Requests.requestIndexes
	requestHeap := heap.InterfaceForSlice(
		&next.Requests.requestIndexes,
		next.Requests.lessByValue,
	)
	heap.Init(requestHeap)

	t := p.t
	originalRequestCount := current.Requests.GetCardinality()
	for {
		if requestHeap.Len() == 0 {
			break
		}
		numPending := maxRequests(current.Requests.GetCardinality() + current.Cancelled.GetCardinality())
		if numPending >= p.nominalMaxRequests() {
			break
		}
		req := heap.Pop(requestHeap)
		if cap(next.Requests.requestIndexes) != cap(orig) {
			panic("changed")
		}

		// don't add requests on receipt of a reject - because this causes request back
		// to potentially permanently unresponsive peers - which just adds network noise.  If
		// the peer can handle more requests it will send an "unchoked" message - which
		// will cause it to get added back to the request queue
		if p.needRequestUpdate == peerUpdateRequestsRemoteRejectReason {
			continue
		}

		existing := t.requestingPeer(req)
		if existing != nil && existing != p {
			// don't steal on cancel - because this is triggered by t.cancelRequest below
			// which means that the cancelled can immediately try to steal back a request
			// it has lost which can lead to circular cancel/add processing
			if p.needRequestUpdate == peerUpdateRequestsPeerCancelReason {
				continue
			}

			// Don't steal from the poor.
			diff := int64(current.Requests.GetCardinality()) + 1 - (int64(existing.uncancelledRequests()) - 1)
			// Steal a request that leaves us with one more request than the existing peer
			// connection if the stealer more recently received a chunk.
			if diff > 1 || (diff == 1 && !p.lastUsefulChunkReceived.After(existing.lastUsefulChunkReceived)) {
				continue
			}
			t.cancelRequest(req)
		}
		more = p.mustRequest(req)
		if !more {
			break
		}
	}
	if !more {
		// This might fail if we incorrectly determine that we can fit up to the maximum allowed
		// requests into the available write buffer space. We don't want that to happen because it
		// makes our peak requests dependent on how much was already in the buffer.
		panic(fmt.Sprintf(
			"couldn't fill apply entire request state [newRequests=%v]",
			current.Requests.GetCardinality()-originalRequestCount))
	}
	newPeakRequests := maxRequests(current.Requests.GetCardinality() - originalRequestCount)
	// log.Printf(
	// 	"requests %v->%v (peak %v->%v) reason %q (peer %v)",
	// 	originalRequestCount, current.Requests.GetCardinality(), p.peakRequests, newPeakRequests, p.needRequestUpdate, p)
	p.peakRequests = newPeakRequests
}

// This could be set to 10s to match the unchoke/request update interval recommended by some
// specifications. I've set it shorter to trigger it more often for testing for now.
const (
	updateRequestsTimerDuration = 3 * time.Second
	enableUpdateRequestsTimer   = false
)