Fix vfs2 pipe behavior when splicing to a non-pipe.

Fixes *.sh Java runtime tests, where splice()-ing from a pipe to /dev/zero
would not actually empty the pipe.

There was no guarantee that the data would actually be consumed on a splice
operation unless the output file's implementation of Write/PWrite actually
called VFSPipeFD.CopyIn. Now, whatever bytes are "written" are consumed
regardless of whether CopyIn is called or not.

Furthermore, the number of bytes in the IOSequence for reads is now capped at
the amount of data actually available. Before, splicing to /dev/zero would
always return the requested splice size without taking the actual available
data into account.

This change also refactors the case where an input file is spliced into an
output pipe so that it follows a similar pattern, which is arguably cleaner
anyway.

Updates #3576.

PiperOrigin-RevId: 328843954

2 files changed, 59 insertions, 18 deletions

diff --git a/pkg/sentry/kernel/pipe/pipe.go b/pkg/sentry/kernel/pipe/pipe.go
index 297e8f28f..c410c96aa 100644
--- a/pkg/sentry/kernel/pipe/pipe.go
+++ b/pkg/sentry/kernel/pipe/pipe.go
@@ -200,17 +200,17 @@ type readOps struct {
 //
 // Precondition: this pipe must have readers.
 func (p *Pipe) read(ctx context.Context, ops readOps) (int64, error) {
-	// Don't block for a zero-length read even if the pipe is empty.
-	if ops.left() == 0 {
-		return 0, nil
-	}
-
 	p.mu.Lock()
 	defer p.mu.Unlock()
 	return p.readLocked(ctx, ops)
 }
 
 func (p *Pipe) readLocked(ctx context.Context, ops readOps) (int64, error) {
+	// Don't block for a zero-length read even if the pipe is empty.
+	if ops.left() == 0 {
+		return 0, nil
+	}
+
 	// Is the pipe empty?
 	if p.view.Size() == 0 {
 		if !p.HasWriters() {
@@ -388,6 +388,10 @@ func (p *Pipe) rwReadiness() waiter.EventMask {
 func (p *Pipe) queued() int64 {
 	p.mu.Lock()
 	defer p.mu.Unlock()
+	return p.queuedLocked()
+}
+
+func (p *Pipe) queuedLocked() int64 {
 	return p.view.Size()
 }
 
diff --git a/pkg/sentry/kernel/pipe/vfs.go b/pkg/sentry/kernel/pipe/vfs.go
index 28f998e45..f223d59e1 100644
--- a/pkg/sentry/kernel/pipe/vfs.go
+++ b/pkg/sentry/kernel/pipe/vfs.go
@@ -244,19 +244,57 @@ func (fd *VFSPipeFD) SetPipeSize(size int64) (int64, error) {
 	return fd.pipe.SetFifoSize(size)
 }
 
-// IOSequence returns a useremm.IOSequence that reads up to count bytes from,
-// or writes up to count bytes to, fd.
-func (fd *VFSPipeFD) IOSequence(count int64) usermem.IOSequence {
-	return usermem.IOSequence{
+// SpliceToNonPipe performs a splice operation from fd to a non-pipe file.
+func (fd *VFSPipeFD) SpliceToNonPipe(ctx context.Context, out *vfs.FileDescription, off, count int64) (int64, error) {
+	fd.pipe.mu.Lock()
+	defer fd.pipe.mu.Unlock()
+
+	// Cap the sequence at number of bytes actually available.
+	v := fd.pipe.queuedLocked()
+	if v < count {
+		count = v
+	}
+	src := usermem.IOSequence{
 		IO:    fd,
 		Addrs: usermem.AddrRangeSeqOf(usermem.AddrRange{0, usermem.Addr(count)}),
 	}
+
+	var (
+		n   int64
+		err error
+	)
+	if off == -1 {
+		n, err = out.Write(ctx, src, vfs.WriteOptions{})
+	} else {
+		n, err = out.PWrite(ctx, src, off, vfs.WriteOptions{})
+	}
+	if n > 0 {
+		fd.pipe.view.TrimFront(n)
+	}
+	return n, err
+}
+
+// SpliceFromNonPipe performs a splice operation from a non-pipe file to fd.
+func (fd *VFSPipeFD) SpliceFromNonPipe(ctx context.Context, in *vfs.FileDescription, off, count int64) (int64, error) {
+	fd.pipe.mu.Lock()
+	defer fd.pipe.mu.Unlock()
+
+	dst := usermem.IOSequence{
+		IO:    fd,
+		Addrs: usermem.AddrRangeSeqOf(usermem.AddrRange{0, usermem.Addr(count)}),
+	}
+
+	if off == -1 {
+		return in.Read(ctx, dst, vfs.ReadOptions{})
+	}
+	return in.PRead(ctx, dst, off, vfs.ReadOptions{})
 }
 
-// CopyIn implements usermem.IO.CopyIn.
+// CopyIn implements usermem.IO.CopyIn. Note that it is the caller's
+// responsibility to trim fd.pipe.view after the read is completed.
 func (fd *VFSPipeFD) CopyIn(ctx context.Context, addr usermem.Addr, dst []byte, opts usermem.IOOpts) (int, error) {
 	origCount := int64(len(dst))
-	n, err := fd.pipe.read(ctx, readOps{
+	n, err := fd.pipe.readLocked(ctx, readOps{
 		left: func() int64 {
 			return int64(len(dst))
 		},
@@ -265,7 +303,6 @@ func (fd *VFSPipeFD) CopyIn(ctx context.Context, addr usermem.Addr, dst []byte,
 		},
 		read: func(view *buffer.View) (int64, error) {
 			n, err := view.ReadAt(dst, 0)
-			view.TrimFront(int64(n))
 			return int64(n), err
 		},
 	})
@@ -281,7 +318,7 @@ func (fd *VFSPipeFD) CopyIn(ctx context.Context, addr usermem.Addr, dst []byte,
 // CopyOut implements usermem.IO.CopyOut.
 func (fd *VFSPipeFD) CopyOut(ctx context.Context, addr usermem.Addr, src []byte, opts usermem.IOOpts) (int, error) {
 	origCount := int64(len(src))
-	n, err := fd.pipe.write(ctx, writeOps{
+	n, err := fd.pipe.writeLocked(ctx, writeOps{
 		left: func() int64 {
 			return int64(len(src))
 		},
@@ -305,7 +342,7 @@ func (fd *VFSPipeFD) CopyOut(ctx context.Context, addr usermem.Addr, src []byte,
 // ZeroOut implements usermem.IO.ZeroOut.
 func (fd *VFSPipeFD) ZeroOut(ctx context.Context, addr usermem.Addr, toZero int64, opts usermem.IOOpts) (int64, error) {
 	origCount := toZero
-	n, err := fd.pipe.write(ctx, writeOps{
+	n, err := fd.pipe.writeLocked(ctx, writeOps{
 		left: func() int64 {
 			return toZero
 		},
@@ -326,14 +363,15 @@ func (fd *VFSPipeFD) ZeroOut(ctx context.Context, addr usermem.Addr, toZero int6
 	return n, err
 }
 
-// CopyInTo implements usermem.IO.CopyInTo.
+// CopyInTo implements usermem.IO.CopyInTo. Note that it is the caller's
+// responsibility to trim fd.pipe.view after the read is completed.
 func (fd *VFSPipeFD) CopyInTo(ctx context.Context, ars usermem.AddrRangeSeq, dst safemem.Writer, opts usermem.IOOpts) (int64, error) {
 	count := ars.NumBytes()
 	if count == 0 {
 		return 0, nil
 	}
 	origCount := count
-	n, err := fd.pipe.read(ctx, readOps{
+	n, err := fd.pipe.readLocked(ctx, readOps{
 		left: func() int64 {
 			return count
 		},
@@ -342,7 +380,6 @@ func (fd *VFSPipeFD) CopyInTo(ctx context.Context, ars usermem.AddrRangeSeq, dst
 		},
 		read: func(view *buffer.View) (int64, error) {
 			n, err := view.ReadToSafememWriter(dst, uint64(count))
-			view.TrimFront(int64(n))
 			return int64(n), err
 		},
 	})
@@ -362,7 +399,7 @@ func (fd *VFSPipeFD) CopyOutFrom(ctx context.Context, ars usermem.AddrRangeSeq,
 		return 0, nil
 	}
 	origCount := count
-	n, err := fd.pipe.write(ctx, writeOps{
+	n, err := fd.pipe.writeLocked(ctx, writeOps{
 		left: func() int64 {
 			return count
 		},