X-Git-Url: http://www.git.stargrave.org/?a=blobdiff_plain;f=reader.go;h=4b20206cbf2ecc675ccde147cecc278ebc688040;hb=HEAD;hp=07eefc87db05e67dccc952f24f245e9dc81966aa;hpb=6dd3b9c12c523818b35ef513942bb828a1b18dbe;p=btrtrc.git

diff --git a/reader.go b/reader.go
index 07eefc87..4b20206c 100644
--- a/reader.go
+++ b/reader.go
@@ -3,20 +3,28 @@ package torrent
 import (
 	"context"
 	"errors"
+	"fmt"
 	"io"
-	"log"
 	"sync"
 
-	"github.com/anacrolix/missinggo"
-	"github.com/anacrolix/torrent/peer_protocol"
+	"github.com/anacrolix/log"
+	"github.com/anacrolix/missinggo/v2"
 )
 
+// Accesses Torrent data via a Client. Reads block until the data is available. Seeks and readahead
+// also drive Client behaviour. Not safe for concurrent use.
 type Reader interface {
-	io.Reader
-	io.Seeker
-	io.Closer
+	io.ReadSeekCloser
 	missinggo.ReadContexter
+	// Configure the number of bytes ahead of a read that should also be prioritized in preparation
+	// for further reads. Overridden by non-nil readahead func, see SetReadaheadFunc.
 	SetReadahead(int64)
+	// If non-nil, the provided function is called when the implementation needs to know the
+	// readahead for the current reader. Calls occur during Reads and Seeks, and while the Client is
+	// locked.
+	SetReadaheadFunc(ReadaheadFunc)
+	// Don't wait for pieces to complete and be verified. Read calls return as soon as they can when
+	// the underlying chunks become available.
 	SetResponsive()
 }
 
@@ -25,33 +33,41 @@ type pieceRange struct {
 	begin, end pieceIndex
 }
 
-// Accesses Torrent data via a Client. Reads block until the data is
-// available. Seeks and readahead also drive Client behaviour.
+type ReadaheadContext struct {
+	ContiguousReadStartPos int64
+	CurrentPos             int64
+}
+
+// Returns the desired readahead for a Reader.
+type ReadaheadFunc func(ReadaheadContext) int64
+
 type reader struct {
-	t          *Torrent
-	responsive bool
-	// Adjust the read/seek window to handle Readers locked to File extents
-	// and the like.
+	t *Torrent
+	// Adjust the read/seek window to handle Readers locked to File extents and the like.
 	offset, length int64
-	// Ensure operations that change the position are exclusive, like Read()
-	// and Seek().
-	opMu sync.Mutex
 
-	// Required when modifying pos and readahead, or reading them without
-	// opMu.
-	mu        sync.Locker
-	pos       int64
-	readahead int64
-	// The cached piece range this reader wants downloaded. The zero value
-	// corresponds to nothing. We cache this so that changes can be detected,
-	// and bubbled up to the Torrent only as required.
+	// Function to dynamically calculate readahead. If nil, readahead is static.
+	readaheadFunc ReadaheadFunc
+
+	// Required when modifying pos and readahead.
+	mu sync.Locker
+
+	readahead, pos int64
+	// Position that reads have continued contiguously from.
+	contiguousReadStartPos int64
+	// The cached piece range this reader wants downloaded. The zero value corresponds to nothing.
+	// We cache this so that changes can be detected, and bubbled up to the Torrent only as
+	// required.
 	pieces pieceRange
+
+	// Reads have been initiated since the last seek. This is used to prevent readaheads occurring
+	// after a seek or with a new reader at the starting position.
+	reading    bool
+	responsive bool
 }
 
-var _ io.ReadCloser = &reader{}
+var _ io.ReadSeekCloser = (*reader)(nil)
 
-// Don't wait for pieces to complete and be verified. Read calls return as
-// soon as they can when the underlying chunks become available.
 func (r *reader) SetResponsive() {
 	r.responsive = true
 	r.t.cl.event.Broadcast()
@@ -63,29 +79,19 @@ func (r *reader) SetNonResponsive() {
 	r.t.cl.event.Broadcast()
 }
 
-// Configure the number of bytes ahead of a read that should also be
-// prioritized in preparation for further reads.
 func (r *reader) SetReadahead(readahead int64) {
 	r.mu.Lock()
 	r.readahead = readahead
-	r.mu.Unlock()
-	r.t.cl.lock()
-	defer r.t.cl.unlock()
+	r.readaheadFunc = nil
 	r.posChanged()
+	r.mu.Unlock()
 }
 
-func (r *reader) readable(off int64) (ret bool) {
-	if r.t.closed.IsSet() {
-		return true
-	}
-	req, ok := r.t.offsetRequest(r.torrentOffset(off))
-	if !ok {
-		panic(off)
-	}
-	if r.responsive {
-		return r.t.haveChunk(req)
-	}
-	return r.t.pieceComplete(pieceIndex(req.Index))
+func (r *reader) SetReadaheadFunc(f ReadaheadFunc) {
+	r.mu.Lock()
+	r.readaheadFunc = f
+	r.posChanged()
+	r.mu.Unlock()
 }
 
 // How many bytes are available to read. Max is the most we could require.
@@ -96,6 +102,9 @@ func (r *reader) available(off, max int64) (ret int64) {
 		if !ok {
 			break
 		}
+		if !r.responsive && !r.t.pieceComplete(pieceIndex(req.Index)) {
+			break
+		}
 		if !r.t.haveChunk(req) {
 			break
 		}
@@ -111,21 +120,23 @@ func (r *reader) available(off, max int64) (ret int64) {
 	return
 }
 
-func (r *reader) waitReadable(off int64) {
-	// We may have been sent back here because we were told we could read but
-	// it failed.
-	r.t.cl.event.Wait()
-}
-
-// Calculates the pieces this reader wants downloaded, ignoring the cached
-// value at r.pieces.
+// Calculates the pieces this reader wants downloaded, ignoring the cached value at r.pieces.
 func (r *reader) piecesUncached() (ret pieceRange) {
 	ra := r.readahead
+	if r.readaheadFunc != nil {
+		ra = r.readaheadFunc(ReadaheadContext{
+			ContiguousReadStartPos: r.contiguousReadStartPos,
+			CurrentPos:             r.pos,
+		})
+	}
 	if ra < 1 {
 		// Needs to be at least 1, because [x, x) means we don't want
 		// anything.
 		ra = 1
 	}
+	if !r.reading {
+		ra = 0
+	}
 	if ra > r.length-r.pos {
 		ra = r.length - r.pos
 	}
@@ -138,40 +149,27 @@ func (r *reader) Read(b []byte) (n int, err error) {
 }
 
 func (r *reader) ReadContext(ctx context.Context, b []byte) (n int, err error) {
-	// This is set under the Client lock if the Context is canceled.
-	var ctxErr error
-	if ctx.Done() != nil {
-		ctx, cancel := context.WithCancel(ctx)
-		// Abort the goroutine when the function returns.
-		defer cancel()
-		go func() {
-			<-ctx.Done()
-			r.t.cl.lock()
-			ctxErr = ctx.Err()
-			r.t.tickleReaders()
-			r.t.cl.unlock()
-		}()
-	}
-	// Hmmm, if a Read gets stuck, this means you can't change position for
-	// other purposes. That seems reasonable, but unusual.
-	r.opMu.Lock()
-	defer r.opMu.Unlock()
-	for len(b) != 0 {
-		var n1 int
-		n1, err = r.readOnceAt(b, r.pos, &ctxErr)
-		if n1 == 0 {
-			if err == nil {
-				panic("expected error")
-			}
-			break
-		}
-		b = b[n1:]
-		n += n1
+	if len(b) > 0 {
+		r.reading = true
+		// TODO: Rework reader piece priorities so we don't have to push updates in to the Client
+		// and take the lock here.
 		r.mu.Lock()
-		r.pos += int64(n1)
 		r.posChanged()
 		r.mu.Unlock()
 	}
+	n, err = r.readOnceAt(ctx, b, r.pos)
+	if n == 0 {
+		if err == nil && len(b) > 0 {
+			panic("expected error")
+		} else {
+			return
+		}
+	}
+
+	r.mu.Lock()
+	r.pos += int64(n)
+	r.posChanged()
+	r.mu.Unlock()
 	if r.pos >= r.length {
 		err = io.EOF
 	} else if err == io.EOF {
@@ -180,62 +178,112 @@ func (r *reader) ReadContext(ctx context.Context, b []byte) (n int, err error) {
 	return
 }
 
-// Wait until some data should be available to read. Tickles the client if it
-// isn't. Returns how much should be readable without blocking.
-func (r *reader) waitAvailable(pos, wanted int64, ctxErr *error) (avail int64) {
-	r.t.cl.lock()
-	defer r.t.cl.unlock()
-	for !r.readable(pos) && *ctxErr == nil {
-		r.waitReadable(pos)
+var closedChan = make(chan struct{})
+
+func init() {
+	close(closedChan)
+}
+
+// Wait until some data should be available to read. Tickles the client if it isn't. Returns how
+// much should be readable without blocking.
+func (r *reader) waitAvailable(ctx context.Context, pos, wanted int64, wait bool) (avail int64, err error) {
+	t := r.t
+	for {
+		r.t.cl.rLock()
+		avail = r.available(pos, wanted)
+		readerCond := t.piece(int((r.offset + pos) / t.info.PieceLength)).readerCond.Signaled()
+		r.t.cl.rUnlock()
+		if avail != 0 {
+			return
+		}
+		var dontWait <-chan struct{}
+		if !wait || wanted == 0 {
+			dontWait = closedChan
+		}
+		select {
+		case <-r.t.closed.Done():
+			err = errors.New("torrent closed")
+			return
+		case <-ctx.Done():
+			err = ctx.Err()
+			return
+		case <-r.t.dataDownloadDisallowed.On():
+			err = errors.New("torrent data downloading disabled")
+		case <-r.t.networkingEnabled.Off():
+			err = errors.New("torrent networking disabled")
+			return
+		case <-dontWait:
+			return
+		case <-readerCond:
+		}
 	}
-	return r.available(pos, wanted)
 }
 
+// Adds the reader's torrent offset to the reader object offset (for example the reader might be
+// constrainted to a particular file within the torrent).
 func (r *reader) torrentOffset(readerPos int64) int64 {
 	return r.offset + readerPos
 }
 
 // Performs at most one successful read to torrent storage.
-func (r *reader) readOnceAt(b []byte, pos int64, ctxErr *error) (n int, err error) {
+func (r *reader) readOnceAt(ctx context.Context, b []byte, pos int64) (n int, err error) {
 	if pos >= r.length {
 		err = io.EOF
 		return
 	}
 	for {
-		avail := r.waitAvailable(pos, int64(len(b)), ctxErr)
+		var avail int64
+		avail, err = r.waitAvailable(ctx, pos, int64(len(b)), n == 0)
 		if avail == 0 {
-			if r.t.closed.IsSet() {
-				err = errors.New("torrent closed")
-				return
-			}
-			if *ctxErr != nil {
-				err = *ctxErr
-				return
-			}
+			return
 		}
-		pi := peer_protocol.Integer(r.torrentOffset(pos) / r.t.info.PieceLength)
-		ip := r.t.info.Piece(int(pi))
-		po := r.torrentOffset(pos) % r.t.info.PieceLength
-		b1 := missinggo.LimitLen(b, ip.Length()-po, avail)
+		firstPieceIndex := pieceIndex(r.torrentOffset(pos) / r.t.info.PieceLength)
+		firstPieceOffset := r.torrentOffset(pos) % r.t.info.PieceLength
+		b1 := missinggo.LimitLen(b, avail)
 		n, err = r.t.readAt(b1, r.torrentOffset(pos))
 		if n != 0 {
 			err = nil
 			return
 		}
+		if r.t.closed.IsSet() {
+			err = fmt.Errorf("reading from closed torrent: %w", err)
+			return
+		}
 		r.t.cl.lock()
-		// TODO: Just reset pieces in the readahead window. This might help
-		// prevent thrashing with small caches and file and piece priorities.
-		log.Printf("error reading torrent %q piece %d offset %d, %d bytes: %s", r.t, pi, po, len(b1), err)
-		r.t.updateAllPieceCompletions()
-		r.t.updateAllPiecePriorities()
-		r.t.cl.unlock()
+		// I think there's a panic here caused by the Client being closed before obtaining this
+		// lock. TestDropTorrentWithMmapStorageWhileHashing seems to tickle occasionally in CI.
+		func() {
+			// Just add exceptions already.
+			defer r.t.cl.unlock()
+			if r.t.closed.IsSet() {
+				// Can't update because Torrent's piece order is removed from Client.
+				return
+			}
+			// TODO: Just reset pieces in the readahead window. This might help
+			// prevent thrashing with small caches and file and piece priorities.
+			r.log(log.Fstr("error reading torrent %s piece %d offset %d, %d bytes: %v",
+				r.t.infoHash.HexString(), firstPieceIndex, firstPieceOffset, len(b1), err))
+			if !r.t.updatePieceCompletion(firstPieceIndex) {
+				r.log(log.Fstr("piece %d completion unchanged", firstPieceIndex))
+			}
+			// Update the rest of the piece completions in the readahead window, without alerting to
+			// changes (since only the first piece, the one above, could have generated the read error
+			// we're currently handling).
+			if r.pieces.begin != firstPieceIndex {
+				panic(fmt.Sprint(r.pieces.begin, firstPieceIndex))
+			}
+			for index := r.pieces.begin + 1; index < r.pieces.end; index++ {
+				r.t.updatePieceCompletion(index)
+			}
+		}()
 	}
 }
 
+// Hodor
 func (r *reader) Close() error {
 	r.t.cl.lock()
-	defer r.t.cl.unlock()
 	r.t.deleteReader(r)
+	r.t.cl.unlock()
 	return nil
 }
 
@@ -250,24 +298,35 @@ func (r *reader) posChanged() {
 	r.t.readerPosChanged(from, to)
 }
 
-func (r *reader) Seek(off int64, whence int) (ret int64, err error) {
-	r.opMu.Lock()
-	defer r.opMu.Unlock()
-
-	r.mu.Lock()
-	defer r.mu.Unlock()
+func (r *reader) Seek(off int64, whence int) (newPos int64, err error) {
 	switch whence {
 	case io.SeekStart:
-		r.pos = off
+		newPos = off
+		r.mu.Lock()
 	case io.SeekCurrent:
-		r.pos += off
+		r.mu.Lock()
+		newPos = r.pos + off
 	case io.SeekEnd:
-		r.pos = r.length + off
+		newPos = r.length + off
+		r.mu.Lock()
 	default:
-		err = errors.New("bad whence")
+		return 0, errors.New("bad whence")
 	}
-	ret = r.pos
-
-	r.posChanged()
+	if newPos != r.pos {
+		r.reading = false
+		r.pos = newPos
+		r.contiguousReadStartPos = newPos
+		r.posChanged()
+	}
+	r.mu.Unlock()
 	return
 }
+
+func (r *reader) log(m log.Msg) {
+	r.t.logger.LogLevel(log.Debug, m.Skip(1))
+}
+
+// Implementation inspired by https://news.ycombinator.com/item?id=27019613.
+func defaultReadaheadFunc(r ReadaheadContext) int64 {
+	return r.CurrentPos - r.ContiguousReadStartPos
+}