Port sendfile to vfs2.

And do some refactoring of the wait logic in sendfile/splice/tee.

Updates #1035 #2923

PiperOrigin-RevId: 322815521

3 files changed, 261 insertions, 64 deletions

diff --git a/pkg/sentry/syscalls/linux/vfs2/splice.go b/pkg/sentry/syscalls/linux/vfs2/splice.go
index 945a364a7..63ab11f8c 100644
--- a/pkg/sentry/syscalls/linux/vfs2/splice.go
+++ b/pkg/sentry/syscalls/linux/vfs2/splice.go
@@ -15,12 +15,15 @@
 package vfs2
 
 import (
+	"io"
+
 	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/sentry/arch"
 	"gvisor.dev/gvisor/pkg/sentry/kernel"
 	"gvisor.dev/gvisor/pkg/sentry/kernel/pipe"
 	"gvisor.dev/gvisor/pkg/sentry/vfs"
 	"gvisor.dev/gvisor/pkg/syserror"
+	"gvisor.dev/gvisor/pkg/usermem"
 	"gvisor.dev/gvisor/pkg/waiter"
 )
 
@@ -110,16 +113,20 @@ func Splice(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal
 
 	// Move data.
 	var (
-		n     int64
-		err   error
-		inCh  chan struct{}
-		outCh chan struct{}
+		n   int64
+		err error
 	)
+	dw := dualWaiter{
+		inFile:  inFile,
+		outFile: outFile,
+	}
+	defer dw.destroy()
 	for {
 		// If both input and output are pipes, delegate to the pipe
-		// implementation. Otherwise, exactly one end is a pipe, which we
-		// ensure is consistently ordered after the non-pipe FD's locks by
-		// passing the pipe FD as usermem.IO to the non-pipe end.
+		// implementation. Otherwise, exactly one end is a pipe, which
+		// we ensure is consistently ordered after the non-pipe FD's
+		// locks by passing the pipe FD as usermem.IO to the non-pipe
+		// end.
 		switch {
 		case inIsPipe && outIsPipe:
 			n, err = pipe.Splice(t, outPipeFD, inPipeFD, count)
@@ -137,38 +144,15 @@ func Splice(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal
 			} else {
 				n, err = inFile.Read(t, outPipeFD.IOSequence(count), vfs.ReadOptions{})
 			}
+		default:
+			panic("not possible")
 		}
+
 		if n != 0 || err != syserror.ErrWouldBlock || nonBlock {
 			break
 		}
-
-		// Note that the blocking behavior here is a bit different than the
-		// normal pattern. Because we need to have both data to read and data
-		// to write simultaneously, we actually explicitly block on both of
-		// these cases in turn before returning to the splice operation.
-		if inFile.Readiness(eventMaskRead)&eventMaskRead == 0 {
-			if inCh == nil {
-				inCh = make(chan struct{}, 1)
-				inW, _ := waiter.NewChannelEntry(inCh)
-				inFile.EventRegister(&inW, eventMaskRead)
-				defer inFile.EventUnregister(&inW)
-				continue // Need to refresh readiness.
-			}
-			if err = t.Block(inCh); err != nil {
-				break
-			}
-		}
-		if outFile.Readiness(eventMaskWrite)&eventMaskWrite == 0 {
-			if outCh == nil {
-				outCh = make(chan struct{}, 1)
-				outW, _ := waiter.NewChannelEntry(outCh)
-				outFile.EventRegister(&outW, eventMaskWrite)
-				defer outFile.EventUnregister(&outW)
-				continue // Need to refresh readiness.
-			}
-			if err = t.Block(outCh); err != nil {
-				break
-			}
+		if err = dw.waitForBoth(t); err != nil {
+			break
 		}
 	}
 
@@ -247,45 +231,256 @@ func Tee(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallCo
 
 	// Copy data.
 	var (
-		inCh  chan struct{}
-		outCh chan struct{}
+		n   int64
+		err error
 	)
+	dw := dualWaiter{
+		inFile:  inFile,
+		outFile: outFile,
+	}
+	defer dw.destroy()
 	for {
-		n, err := pipe.Tee(t, outPipeFD, inPipeFD, count)
-		if n != 0 {
-			return uintptr(n), nil, nil
+		n, err = pipe.Tee(t, outPipeFD, inPipeFD, count)
+		if n != 0 || err != syserror.ErrWouldBlock || nonBlock {
+			break
+		}
+		if err = dw.waitForBoth(t); err != nil {
+			break
+		}
+	}
+	if n == 0 {
+		return 0, nil, err
+	}
+	outFile.Dentry().InotifyWithParent(linux.IN_MODIFY, 0, vfs.PathEvent)
+	return uintptr(n), nil, nil
+}
+
+// Sendfile implements linux system call sendfile(2).
+func Sendfile(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	outFD := args[0].Int()
+	inFD := args[1].Int()
+	offsetAddr := args[2].Pointer()
+	count := int64(args[3].SizeT())
+
+	inFile := t.GetFileVFS2(inFD)
+	if inFile == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer inFile.DecRef()
+	if !inFile.IsReadable() {
+		return 0, nil, syserror.EBADF
+	}
+
+	outFile := t.GetFileVFS2(outFD)
+	if outFile == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer outFile.DecRef()
+	if !outFile.IsWritable() {
+		return 0, nil, syserror.EBADF
+	}
+
+	// Verify that the outFile Append flag is not set.
+	if outFile.StatusFlags()&linux.O_APPEND != 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// Verify that inFile is a regular file or block device. This is a
+	// requirement; the same check appears in Linux
+	// (fs/splice.c:splice_direct_to_actor).
+	if stat, err := inFile.Stat(t, vfs.StatOptions{Mask: linux.STATX_TYPE}); err != nil {
+		return 0, nil, err
+	} else if stat.Mask&linux.STATX_TYPE == 0 ||
+		(stat.Mode&linux.S_IFMT != linux.S_IFREG && stat.Mode&linux.S_IFMT != linux.S_IFBLK) {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// Copy offset if it exists.
+	offset := int64(-1)
+	if offsetAddr != 0 {
+		if inFile.Options().DenyPRead {
+			return 0, nil, syserror.ESPIPE
 		}
-		if err != syserror.ErrWouldBlock || nonBlock {
+		if _, err := t.CopyIn(offsetAddr, &offset); err != nil {
 			return 0, nil, err
 		}
+		if offset < 0 {
+			return 0, nil, syserror.EINVAL
+		}
+		if offset+count < 0 {
+			return 0, nil, syserror.EINVAL
+		}
+	}
+
+	// Validate count. This must come after offset checks.
+	if count < 0 {
+		return 0, nil, syserror.EINVAL
+	}
+	if count == 0 {
+		return 0, nil, nil
+	}
+	if count > int64(kernel.MAX_RW_COUNT) {
+		count = int64(kernel.MAX_RW_COUNT)
+	}
 
-		// Note that the blocking behavior here is a bit different than the
-		// normal pattern. Because we need to have both data to read and data
-		// to write simultaneously, we actually explicitly block on both of
-		// these cases in turn before returning to the tee operation.
-		if inFile.Readiness(eventMaskRead)&eventMaskRead == 0 {
-			if inCh == nil {
-				inCh = make(chan struct{}, 1)
-				inW, _ := waiter.NewChannelEntry(inCh)
-				inFile.EventRegister(&inW, eventMaskRead)
-				defer inFile.EventUnregister(&inW)
-				continue // Need to refresh readiness.
+	// Copy data.
+	var (
+		n   int64
+		err error
+	)
+	dw := dualWaiter{
+		inFile:  inFile,
+		outFile: outFile,
+	}
+	defer dw.destroy()
+	outPipeFD, outIsPipe := outFile.Impl().(*pipe.VFSPipeFD)
+	// Reading from input file should never block, since it is regular or
+	// block device. We only need to check if writing to the output file
+	// can block.
+	nonBlock := outFile.StatusFlags()&linux.O_NONBLOCK != 0
+	if outIsPipe {
+		for n < count {
+			var spliceN int64
+			if offset != -1 {
+				spliceN, err = inFile.PRead(t, outPipeFD.IOSequence(count), offset, vfs.ReadOptions{})
+				offset += spliceN
+			} else {
+				spliceN, err = inFile.Read(t, outPipeFD.IOSequence(count), vfs.ReadOptions{})
 			}
-			if err := t.Block(inCh); err != nil {
-				return 0, nil, err
+			n += spliceN
+			if err == syserror.ErrWouldBlock && !nonBlock {
+				err = dw.waitForBoth(t)
+			}
+			if err != nil {
+				break
 			}
 		}
-		if outFile.Readiness(eventMaskWrite)&eventMaskWrite == 0 {
-			if outCh == nil {
-				outCh = make(chan struct{}, 1)
-				outW, _ := waiter.NewChannelEntry(outCh)
-				outFile.EventRegister(&outW, eventMaskWrite)
-				defer outFile.EventUnregister(&outW)
-				continue // Need to refresh readiness.
+	} else {
+		// Read inFile to buffer, then write the contents to outFile.
+		buf := make([]byte, count)
+		for n < count {
+			var readN int64
+			if offset != -1 {
+				readN, err = inFile.PRead(t, usermem.BytesIOSequence(buf), offset, vfs.ReadOptions{})
+				offset += readN
+			} else {
+				readN, err = inFile.Read(t, usermem.BytesIOSequence(buf), vfs.ReadOptions{})
+			}
+			if readN == 0 && err == io.EOF {
+				// We reached the end of the file. Eat the
+				// error and exit the loop.
+				err = nil
+				break
 			}
-			if err := t.Block(outCh); err != nil {
-				return 0, nil, err
+			n += readN
+			if err != nil {
+				break
+			}
+
+			// Write all of the bytes that we read. This may need
+			// multiple write calls to complete.
+			wbuf := buf[:n]
+			for len(wbuf) > 0 {
+				var writeN int64
+				writeN, err = outFile.Write(t, usermem.BytesIOSequence(wbuf), vfs.WriteOptions{})
+				wbuf = wbuf[writeN:]
+				if err == syserror.ErrWouldBlock && !nonBlock {
+					err = dw.waitForOut(t)
+				}
+				if err != nil {
+					// We didn't complete the write. Only
+					// report the bytes that were actually
+					// written, and rewind the offset.
+					notWritten := int64(len(wbuf))
+					n -= notWritten
+					if offset != -1 {
+						offset -= notWritten
+					}
+					break
+				}
+			}
+			if err == syserror.ErrWouldBlock && !nonBlock {
+				err = dw.waitForBoth(t)
 			}
+			if err != nil {
+				break
+			}
+		}
+	}
+
+	if offsetAddr != 0 {
+		// Copy out the new offset.
+		if _, err := t.CopyOut(offsetAddr, offset); err != nil {
+			return 0, nil, err
+		}
+	}
+
+	if n == 0 {
+		return 0, nil, err
+	}
+
+	inFile.Dentry().InotifyWithParent(linux.IN_ACCESS, 0, vfs.PathEvent)
+	outFile.Dentry().InotifyWithParent(linux.IN_MODIFY, 0, vfs.PathEvent)
+	return uintptr(n), nil, nil
+}
+
+// dualWaiter is used to wait on one or both vfs.FileDescriptions. It is not
+// thread-safe, and does not take a reference on the vfs.FileDescriptions.
+//
+// Users must call destroy() when finished.
+type dualWaiter struct {
+	inFile  *vfs.FileDescription
+	outFile *vfs.FileDescription
+
+	inW   waiter.Entry
+	inCh  chan struct{}
+	outW  waiter.Entry
+	outCh chan struct{}
+}
+
+// waitForBoth waits for both dw.inFile and dw.outFile to be ready.
+func (dw *dualWaiter) waitForBoth(t *kernel.Task) error {
+	if dw.inFile.Readiness(eventMaskRead)&eventMaskRead == 0 {
+		if dw.inCh == nil {
+			dw.inW, dw.inCh = waiter.NewChannelEntry(nil)
+			dw.inFile.EventRegister(&dw.inW, eventMaskRead)
+			// We might be ready now. Try again before blocking.
+			return nil
+		}
+		if err := t.Block(dw.inCh); err != nil {
+			return err
+		}
+	}
+	return dw.waitForOut(t)
+}
+
+// waitForOut waits for dw.outfile to be read.
+func (dw *dualWaiter) waitForOut(t *kernel.Task) error {
+	if dw.outFile.Readiness(eventMaskWrite)&eventMaskWrite == 0 {
+		if dw.outCh == nil {
+			dw.outW, dw.outCh = waiter.NewChannelEntry(nil)
+			dw.outFile.EventRegister(&dw.outW, eventMaskWrite)
+			// We might be ready now. Try again before blocking.
+			return nil
 		}
+		if err := t.Block(dw.outCh); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+// destroy cleans up resources help by dw. No more calls to wait* can occur
+// after destroy is called.
+func (dw *dualWaiter) destroy() {
+	if dw.inCh != nil {
+		dw.inFile.EventUnregister(&dw.inW)
+		dw.inCh = nil
+	}
+	if dw.outCh != nil {
+		dw.outFile.EventUnregister(&dw.outW)
+		dw.outCh = nil
 	}
+	dw.inFile = nil
+	dw.outFile = nil
 }
diff --git a/pkg/sentry/syscalls/linux/vfs2/vfs2.go b/pkg/sentry/syscalls/linux/vfs2/vfs2.go
index 8f497ecc7..1b2cfad7d 100644
--- a/pkg/sentry/syscalls/linux/vfs2/vfs2.go
+++ b/pkg/sentry/syscalls/linux/vfs2/vfs2.go
@@ -44,7 +44,7 @@ func Override() {
 	s.Table[23] = syscalls.Supported("select", Select)
 	s.Table[32] = syscalls.Supported("dup", Dup)
 	s.Table[33] = syscalls.Supported("dup2", Dup2)
-	delete(s.Table, 40) // sendfile
+	s.Table[40] = syscalls.Supported("sendfile", Sendfile)
 	s.Table[41] = syscalls.Supported("socket", Socket)
 	s.Table[42] = syscalls.Supported("connect", Connect)
 	s.Table[43] = syscalls.Supported("accept", Accept)
diff --git a/test/syscalls/BUILD b/test/syscalls/BUILD
index c06a75ada..15e5fd223 100644
--- a/test/syscalls/BUILD
+++ b/test/syscalls/BUILD
@@ -640,11 +640,13 @@ syscall_test(
 syscall_test(
     add_overlay = True,
     test = "//test/syscalls/linux:sendfile_socket_test",
+    vfs2 = "True",
 )
 
 syscall_test(
     add_overlay = True,
     test = "//test/syscalls/linux:sendfile_test",
+    vfs2 = "True",
 )
 
 syscall_test(