diff --git a/nfs_onwrite.go b/nfs_onwrite.go
index 861ef16..8ada93c 100644
--- a/nfs_onwrite.go
+++ b/nfs_onwrite.go
@@ -3,6 +3,7 @@ package nfs
 import (
 	"bytes"
 	"context"
+	"fmt"
 	"io"
 	"math"
 	"os"
@@ -70,6 +71,9 @@ func onWrite(ctx context.Context, w *response, userHandle Handler) error {
 	}
 	if req.Offset > 0 {
 		if _, err := file.Seek(int64(req.Offset), io.SeekStart); err != nil {
+			if closeErr := file.Close(); closeErr != nil {
+				err = fmt.Errorf("%w; closing write file: %v", err, closeErr)
+			}
 			return &NFSStatusError{NFSStatusIO, err}
 		}
 	}
@@ -79,6 +83,9 @@ func onWrite(ctx context.Context, w *response, userHandle Handler) error {
 	}
 	writtenCount, err := file.Write(req.Data[:end])
 	if err != nil {
+		if closeErr := file.Close(); closeErr != nil {
+			err = fmt.Errorf("%w; closing write file: %v", err, closeErr)
+		}
 		Log.Errorf("Error writing: %v", err)
 		return &NFSStatusError{statusFromWriteError(err), err}
 	}
diff --git a/nfs_write_lifetime_test.go b/nfs_write_lifetime_test.go
new file mode 100644
index 0000000..63cd046
--- /dev/null
+++ b/nfs_write_lifetime_test.go
@@ -0,0 +1,158 @@
+package nfs_test
+
+import (
+	"errors"
+	"io"
+	"os"
+	"path/filepath"
+	"sync"
+	"sync/atomic"
+	"syscall"
+	"testing"
+
+	"github.com/go-git/go-billy/v5"
+	"github.com/go-git/go-billy/v5/osfs"
+	nfs "github.com/willscott/go-nfs"
+	"github.com/willscott/go-nfs-client/nfs/rpc"
+)
+
+type writeLifetimeFS struct {
+	billy.Filesystem
+	openErr, seekErr, writeErr, closeErr error
+	mu                                   sync.Mutex
+	files                                []*writeLifetimeFile
+}
+
+func (fs *writeLifetimeFS) OpenFile(name string, flag int, perm os.FileMode) (billy.File, error) {
+	if flag != os.O_RDWR {
+		return fs.Filesystem.OpenFile(name, flag, perm)
+	}
+	if fs.openErr != nil {
+		return nil, fs.openErr
+	}
+	file, err := fs.Filesystem.OpenFile(name, flag, perm)
+	if err != nil {
+		return nil, err
+	}
+	wrapped := &writeLifetimeFile{File: file, owner: fs}
+	fs.mu.Lock()
+	fs.files = append(fs.files, wrapped)
+	fs.mu.Unlock()
+	return wrapped, nil
+}
+
+type writeLifetimeFile struct {
+	billy.File
+	owner  *writeLifetimeFS
+	closes int32
+}
+
+func (f *writeLifetimeFile) Seek(offset int64, whence int) (int64, error) {
+	if f.owner.seekErr != nil {
+		return 0, f.owner.seekErr
+	}
+	return f.File.Seek(offset, whence)
+}
+
+func (f *writeLifetimeFile) Write(data []byte) (int, error) {
+	if f.owner.writeErr != nil {
+		// Preserve a real partial write before injecting the backend failure.
+		n, err := f.File.Write(data[:1])
+		if err != nil {
+			return n, err
+		}
+		return n, f.owner.writeErr
+	}
+	return f.File.Write(data)
+}
+
+func (f *writeLifetimeFile) Close() error {
+	atomic.AddInt32(&f.closes, 1)
+	if err := f.File.Close(); err != nil {
+		return err
+	}
+	// This fault reports an error after the real descriptor has closed; it
+	// does not imply that every billy implementation closes on an error.
+	return f.owner.closeErr
+}
+
+func TestOnWriteClosesOwnedFile(t *testing.T) {
+	for _, tc := range []struct {
+		name                                 string
+		openErr, seekErr, writeErr, closeErr error
+		want                                 nfs.NFSStatus
+	}{
+		{name: "success", want: nfs.NFSStatusOk},
+		{name: "open error", openErr: syscall.EACCES, want: nfs.NFSStatusAccess},
+		{name: "seek error", seekErr: syscall.EIO, want: nfs.NFSStatusIO},
+		{name: "partial write error", writeErr: syscall.EIO, want: nfs.NFSStatusIO},
+		{name: "seek and close errors", seekErr: syscall.EIO, closeErr: syscall.ENOSPC, want: nfs.NFSStatusIO},
+		{name: "write and close errors", writeErr: syscall.ENOSPC, closeErr: syscall.EIO, want: nfs.NFSStatusNoSPC},
+		{name: "close error", closeErr: syscall.EIO, want: nfs.NFSStatusIO},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			root := t.TempDir()
+			if err := os.WriteFile(filepath.Join(root, "payload"), []byte("original"), 0600); err != nil {
+				t.Fatal(err)
+			}
+			fs := &writeLifetimeFS{Filesystem: osfs.New(root), openErr: tc.openErr, seekErr: tc.seekErr, writeErr: tc.writeErr, closeErr: tc.closeErr}
+			target := serveFS(t, fs)
+			// Retain files for the oracle so GC cannot hide a missed Close. On
+			// a failing baseline, reclaim only after the assertions below.
+			t.Cleanup(func() {
+				fs.mu.Lock()
+				defer fs.mu.Unlock()
+				for _, file := range fs.files {
+					if atomic.LoadInt32(&file.closes) == 0 {
+						_ = file.Close()
+					}
+				}
+			})
+			_, handle, err := target.Lookup("/payload")
+			if err != nil {
+				t.Fatal(err)
+			}
+			type writeArgs struct {
+				rpc.Header
+				Handle []byte
+				Offset uint64
+				Count  uint32
+				How    uint32
+				Data   []byte
+			}
+			data := []byte("updated")
+			status := callStatus(t, target, nfs.NFSProcedureWrite, &writeArgs{Header: header(nfs.NFSProcedureWrite), Handle: handle, Offset: 1, Count: uint32(len(data)), How: 0, Data: data})
+			if status != tc.want {
+				t.Errorf("WRITE status: got %v, want %v", status, tc.want)
+			}
+			fs.mu.Lock()
+			opened := append([]*writeLifetimeFile(nil), fs.files...)
+			fs.mu.Unlock()
+			wantOpened := 1
+			if tc.openErr != nil {
+				wantOpened = 0
+			}
+			if len(opened) != wantOpened {
+				t.Fatalf("WRITE opened %d files, want %d", len(opened), wantOpened)
+			}
+			if tc.openErr == nil && tc.seekErr == nil {
+				wantData := "oupdated"
+				if tc.writeErr != nil {
+					wantData = "ouiginal"
+				}
+				actual, err := os.ReadFile(filepath.Join(root, "payload"))
+				if err != nil || string(actual) != wantData {
+					t.Errorf("real file contents: got %q, error %v; want %q", actual, err, wantData)
+				}
+			}
+			for _, file := range opened {
+				if got := atomic.LoadInt32(&file.closes); got != 1 {
+					t.Errorf("WRITE called Close %d times, want exactly 1", got)
+				}
+				if _, err := file.File.Seek(0, io.SeekStart); !errors.Is(err, os.ErrClosed) {
+					t.Errorf("WRITE left its real file open: Seek error %v, want os.ErrClosed", err)
+				}
+			}
+		})
+	}
+}