diff --git a/file_system.go b/file_system.go
index fc6259d..0588966 100644
--- a/file_system.go
+++ b/file_system.go
@@ -69,10 +69,32 @@ type FileSystem interface {
 
 	// Create a directory inode as a child of an existing directory inode. The
 	// kernel sends this in response to a mkdir(2) call.
+	//
+	// The kernel appears to verify the name doesn't already exist (mkdir calls
+	// mkdirat calls user_path_create calls filename_create, which verifies:
+	// http://goo.gl/FZpLu5). But volatile file systems and paranoid non-volatile
+	// file systems should check for the reasons described below on CreateFile.
 	MkDir(
 		ctx context.Context,
 		req *MkDirRequest) (*MkDirResponse, error)
 
+	// Create a file inode and open it.
+	//
+	// The kernel calls this method when the user asks to open a file with the
+	// O_CREAT flag and the kernel has observed that the file doesn't exist. (See
+	// for example lookup_open, http://goo.gl/PlqE9d).
+	//
+	// However it's impossible to tell for sure that all kernels make this check
+	// in all cases and the official fuse documentation is less than encouraging
+	// (" the file does not exist, first create it with the specified mode, and
+	// then open it"). Therefore file systems would be smart to be paranoid and
+	// check themselves, returning EEXIST when the file already exists. This of
+	// course particularly applies to file systems that are volatile from the
+	// kernel's point of view.
+	CreateFile(
+		ctx context.Context,
+		req *CreateFileRequest) (*CreateFileResponse, error)
+
 	///////////////////////////////////
 	// Inode destruction
 	///////////////////////////////////
@@ -128,11 +150,45 @@ type FileSystem interface {
 		ctx context.Context,
 		req *OpenFileRequest) (*OpenFileResponse, error)
 
-	// Read data from a file previously opened with OpenFile.
+	// Read data from a file previously opened with CreateFile or OpenFile.
+	//
+	// Note that this method is not called for every call to read(2) by the end
+	// user; some reads may be served by the page cache. See notes on Write for
+	// more.
 	ReadFile(
 		ctx context.Context,
 		req *ReadFileRequest) (*ReadFileResponse, error)
 
+	// Write data to a file previously opened with CreateFile or OpenFile.
+	//
+	// When the user writes data using write(2), the write goes into the page
+	// cache and the page is marked dirty. Later the kernel may write back the
+	// page via the FUSE VFS layer, causing this method to be called:
+	//
+	//  *  The kernel calls address_space_operations::writepage when a dirty page
+	//     needs to be written to backing store (see vfs.txt). Fuse sets this to
+	//     fuse_writepage (see file.c).
+	//
+	//  *  fuse_writepage calls fuse_writepage_locked.
+	//
+	//  *  fuse_writepage_locked makes a write request to the userspace server.
+	//
+	// Note that writes *will* be received before a call to Flush when closing
+	// the file descriptor to which they were written:
+	//
+	//  *  fuse_flush calls write_inode_now, which appears to start a writeback
+	//     in the background (it talks about a "flusher thread").
+	//
+	//  *  fuse_flush then calls fuse_sync_writes, which "[waits] for all pending
+	//     writepages on the inode to finish".
+	//
+	//  *  Only then does fuse_flush finally send the flush request.
+	//
+	// TODO(jacobsa): Add links for all of the references above.
+	WriteFile(
+		ctx context.Context,
+		req *WriteFileRequest) (*WriteFileResponse, error)
+
 	// Release a previously-minted file handle. The kernel calls this when there
 	// are no more references to an open file: all file descriptors are closed
 	// and all memory mappings are unmapped.
@@ -235,8 +291,8 @@ type RequestHeader struct {
 }
 
 // Information about a child inode within its parent directory. Shared by the
-// responses for LookUpInode, MkDir, etc. Consumed by the kernel in order to
-// set up a dcache entry.
+// responses for LookUpInode, MkDir, CreateFile, etc. Consumed by the kernel in
+// order to set up a dcache entry.
 type ChildInodeEntry struct {
 	// The ID of the child inode. The file system must ensure that the returned
 	// inode ID remains valid until a later call to ForgetInode.
@@ -247,6 +303,16 @@ type ChildInodeEntry struct {
 	Generation GenerationNumber
 
 	// Current attributes for the child inode.
+	//
+	// When creating a new inode, the file system is responsible for initializing
+	// and recording (where supported) attributes like time information,
+	// ownership information, etc.
+	//
+	// Ownership information in particular must be set to something reasonable or
+	// by default root will own everything and unprivileged users won't be able
+	// to do anything useful. In traditional file systems in the kernel, the
+	// function inode_init_owner (http://goo.gl/5qavg8) contains the
+	// standards-compliant logic for this.
 	Attributes InodeAttributes
 
 	// The FUSE VFS layer in the kernel maintains a cache of file attributes,
@@ -377,18 +443,42 @@ type MkDirRequest struct {
 
 type MkDirResponse struct {
 	// Information about the inode that was created.
-	//
-	// The file system is responsible for initializing and recording (where
-	// supported) attributes like time information, ownership information, etc.
-	//
-	// Ownership information in particular must be set to something reasonable or
-	// by default root will own everything and unprivileged users won't be able
-	// to do anything useful. In traditional file systems in the kernel, the
-	// function inode_init_owner (http://goo.gl/5qavg8) contains the
-	// standards-compliant logic for this.
 	Entry ChildInodeEntry
 }
 
+type CreateFileRequest struct {
+	Header RequestHeader
+
+	// The ID of parent directory inode within which to create the child file.
+	Parent InodeID
+
+	// The name of the child to create, and the mode with which to create it.
+	Name string
+	Mode os.FileMode
+
+	// Flags for the open operation.
+	Flags bazilfuse.OpenFlags
+}
+
+type CreateFileResponse struct {
+	// Information about the inode that was created.
+	Entry ChildInodeEntry
+
+	// An opaque ID that will be echoed in follow-up calls for this file using
+	// the same struct file in the kernel. In practice this usually means
+	// follow-up calls using the file descriptor returned by open(2).
+	//
+	// The handle may be supplied to the following methods:
+	//
+	//  *  ReadFile
+	//  *  WriteFile
+	//  *  ReleaseFileHandle
+	//
+	// The file system must ensure this ID remains valid until a later call to
+	// ReleaseFileHandle.
+	Handle HandleID
+}
+
 type RmDirRequest struct {
 	Header RequestHeader
 
@@ -547,6 +637,7 @@ type OpenFileResponse struct {
 	// The handle may be supplied to the following methods:
 	//
 	//  *  ReadFile
+	//  *  WriteFile
 	//  *  ReleaseFileHandle
 	//
 	// The file system must ensure this ID remains valid until a later call to
@@ -558,7 +649,7 @@ type ReadFileRequest struct {
 	Header RequestHeader
 
 	// The file inode that we are reading, and the handle previously returned by
-	// OpenFile when opening that inode.
+	// CreateFile or OpenFile when opening that inode.
 	Inode  InodeID
 	Handle HandleID
 
@@ -579,6 +670,48 @@ type ReadFileResponse struct {
 	Data []byte
 }
 
+type WriteFileRequest struct {
+	Header RequestHeader
+
+	// The file inode that we are modifying, and the handle previously returned
+	// by CreateFile or OpenFile when opening that inode.
+	Inode  InodeID
+	Handle HandleID
+
+	// The offset at which to write the data below.
+	//
+	// The man page for pwrite(2) implies that aside from changing the file
+	// handle's offset, using pwrite is equivalent to using lseek(2) and then
+	// write(2). The man page for lseek(2) says the following:
+	//
+	// "The lseek() function allows the file offset to be set beyond the end of
+	// the file (but this does not change the size of the file). If data is later
+	// written at this point, subsequent reads of the data in the gap (a "hole")
+	// return null bytes (aq\0aq) until data is actually written into the gap."
+	//
+	// It is therefore reasonable to assume that the kernel is looking for
+	// the following semantics:
+	//
+	// *   If the offset is less than or equal to the current size, extend the
+	//     file as necessary to fit any data that goes past the end of the file.
+	//
+	// *   If the offset is greater than the current size, extend the file
+	//     with null bytes until it is not, then do the above.
+	//
+	Offset int64
+
+	// The data to write.
+	//
+	// The FUSE documentation requires that exactly the number of bytes supplied
+	// be written, except on error (http://goo.gl/KUpwwn). This appears to be
+	// because it uses file mmapping machinery (http://goo.gl/SGxnaN) to write a
+	// page at a time.
+	Data []byte
+}
+
+type WriteFileResponse struct {
+}
+
 type ReleaseFileHandleRequest struct {
 	Header RequestHeader
 
diff --git a/fuseutil/not_implemented_file_system.go b/fuseutil/not_implemented_file_system.go
index 55697b9..7cf7e2d 100644
--- a/fuseutil/not_implemented_file_system.go
+++ b/fuseutil/not_implemented_file_system.go
@@ -57,6 +57,12 @@ func (fs *NotImplementedFileSystem) MkDir(
 	return nil, fuse.ENOSYS
 }
 
+func (fs *NotImplementedFileSystem) CreateFile(
+	ctx context.Context,
+	req *fuse.CreateFileRequest) (*fuse.CreateFileResponse, error) {
+	return nil, fuse.ENOSYS
+}
+
 func (fs *NotImplementedFileSystem) RmDir(
 	ctx context.Context,
 	req *fuse.RmDirRequest) (*fuse.RmDirResponse, error) {
@@ -93,6 +99,12 @@ func (fs *NotImplementedFileSystem) ReadFile(
 	return nil, fuse.ENOSYS
 }
 
+func (fs *NotImplementedFileSystem) WriteFile(
+	ctx context.Context,
+	req *fuse.WriteFileRequest) (*fuse.WriteFileResponse, error) {
+	return nil, fuse.ENOSYS
+}
+
 func (fs *NotImplementedFileSystem) ReleaseFileHandle(
 	ctx context.Context,
 	req *fuse.ReleaseFileHandleRequest) (*fuse.ReleaseFileHandleResponse, error) {
diff --git a/samples/memfs/fs.go b/samples/memfs/fs.go
index eceed2b..b0fe38c 100644
--- a/samples/memfs/fs.go
+++ b/samples/memfs/fs.go
@@ -279,7 +279,7 @@ func (fs *memFS) MkDir(
 	parent := fs.getInodeForModifyingOrDie(req.Parent)
 	defer parent.mu.Unlock()
 
-	// Set up attributes from the child, using the credientials of the calling
+	// Set up attributes from the child, using the credentials of the calling
 	// process as owner (matching inode_init_owner, cf. http://goo.gl/5qavg8).
 	now := fs.clock.Now()
 	childAttrs := fuse.InodeAttributes{
@@ -311,6 +311,52 @@ func (fs *memFS) MkDir(
 	return
 }
 
+func (fs *memFS) CreateFile(
+	ctx context.Context,
+	req *fuse.CreateFileRequest) (resp *fuse.CreateFileResponse, err error) {
+	resp = &fuse.CreateFileResponse{}
+
+	fs.mu.Lock()
+	defer fs.mu.Unlock()
+
+	// Grab the parent, which we will update shortly.
+	parent := fs.getInodeForModifyingOrDie(req.Parent)
+	defer parent.mu.Unlock()
+
+	// Set up attributes from the child, using the credentials of the calling
+	// process as owner (matching inode_init_owner, cf. http://goo.gl/5qavg8).
+	now := fs.clock.Now()
+	childAttrs := fuse.InodeAttributes{
+		Mode:   req.Mode,
+		Atime:  now,
+		Mtime:  now,
+		Ctime:  now,
+		Crtime: now,
+		Uid:    req.Header.Uid,
+		Gid:    req.Header.Gid,
+	}
+
+	// Allocate a child.
+	childID, child := fs.allocateInode(childAttrs)
+	defer child.mu.Unlock()
+
+	// Add an entry in the parent.
+	parent.AddChild(childID, req.Name, fuseutil.DT_File)
+
+	// Fill in the response entry.
+	resp.Entry.Child = childID
+	resp.Entry.Attributes = child.attributes
+
+	// We don't spontaneously mutate, so the kernel can cache as long as it wants
+	// (since it also handles invalidation).
+	resp.Entry.AttributesExpiration = fs.clock.Now().Add(365 * 24 * time.Hour)
+	resp.Entry.EntryExpiration = resp.Entry.EntryExpiration
+
+	// We have nothing interesting to put in the Handle field.
+
+	return
+}
+
 func (fs *memFS) RmDir(
 	ctx context.Context,
 	req *fuse.RmDirRequest) (resp *fuse.RmDirResponse, err error) {
@@ -391,3 +437,42 @@ func (fs *memFS) ReadDir(
 
 	return
 }
+
+func (fs *memFS) OpenFile(
+	ctx context.Context,
+	req *fuse.OpenFileRequest) (resp *fuse.OpenFileResponse, err error) {
+	resp = &fuse.OpenFileResponse{}
+
+	fs.mu.RLock()
+	defer fs.mu.RUnlock()
+
+	// We don't mutate spontaneosuly, so if the VFS layer has asked for an
+	// inode that doesn't exist, something screwed up earlier (a lookup, a
+	// cache invalidation, etc.).
+	inode := fs.getInodeForReadingOrDie(req.Inode)
+	defer inode.mu.RUnlock()
+
+	if inode.dir {
+		panic("Found directory.")
+	}
+
+	return
+}
+
+func (fs *memFS) WriteFile(
+	ctx context.Context,
+	req *fuse.WriteFileRequest) (resp *fuse.WriteFileResponse, err error) {
+	resp = &fuse.WriteFileResponse{}
+
+	fs.mu.RLock()
+	defer fs.mu.RUnlock()
+
+	// Find the inode in question.
+	inode := fs.getInodeForModifyingOrDie(req.Inode)
+	defer inode.mu.Unlock()
+
+	// Serve the request.
+	_, err = inode.WriteAt(req.Data, req.Offset)
+
+	return
+}
diff --git a/samples/memfs/inode.go b/samples/memfs/inode.go
index 163a22f..647614c 100644
--- a/samples/memfs/inode.go
+++ b/samples/memfs/inode.go
@@ -54,6 +54,7 @@ type inode struct {
 	// INVARIANT: No non-permission mode bits are set besides os.ModeDir
 	// INVARIANT: If dir, then os.ModeDir is set
 	// INVARIANT: If !dir, then os.ModeDir is not set
+	// INVARIANT: attributes.Size == len(contents)
 	attributes fuse.InodeAttributes // GUARDED_BY(mu)
 
 	// For directories, entries describing the children of the directory. Unused
@@ -142,6 +143,15 @@ func (inode *inode) checkInvariants() {
 			panic("Non-nil entries in a file.")
 		}
 	}
+
+	// Check the size.
+	if inode.attributes.Size != uint64(len(inode.contents)) {
+		panic(
+			fmt.Sprintf(
+				"Unexpected size: %v vs. %v",
+				inode.attributes.Size,
+				len(inode.contents)))
+	}
 }
 
 // Return the index of the child within inode.entries, if it exists.
@@ -279,3 +289,31 @@ func (inode *inode) ReadDir(offset int, size int) (data []byte, err error) {
 
 	return
 }
+
+// Write to the file's contents. See documentation for ioutil.WriterAt.
+//
+// REQUIRES: !inode.dir
+// EXCLUSIVE_LOCKS_REQUIRED(inode.mu)
+func (inode *inode) WriteAt(p []byte, off int64) (n int, err error) {
+	if inode.dir {
+		panic("WriteAt called on directory.")
+	}
+
+	// Ensure that the contents slice is long enough.
+	newLen := int(off) + len(p)
+	if len(inode.contents) < newLen {
+		padding := make([]byte, newLen-len(inode.contents))
+		inode.contents = append(inode.contents, padding...)
+		inode.attributes.Size = uint64(newLen)
+	}
+
+	// Copy in the data.
+	n = copy(inode.contents[off:], p)
+
+	// Sanity check.
+	if n != len(p) {
+		panic(fmt.Sprintf("Unexpected short copy: %v", n))
+	}
+
+	return
+}
diff --git a/samples/memfs/memfs_test.go b/samples/memfs/memfs_test.go
index 1b848ef..8b2afd9 100644
--- a/samples/memfs/memfs_test.go
+++ b/samples/memfs/memfs_test.go
@@ -280,7 +280,7 @@ func (t *MemFSTest) Mkdir_IntermediateIsFile() {
 	err = os.Mkdir(dirName, 0754)
 
 	AssertNe(nil, err)
-	ExpectThat(err, Error(HasSubstr("TODO")))
+	ExpectThat(err, Error(HasSubstr("not a directory")))
 }
 
 func (t *MemFSTest) Mkdir_IntermediateIsNonExistent() {
@@ -309,7 +309,45 @@ func (t *MemFSTest) Mkdir_PermissionDenied() {
 }
 
 func (t *MemFSTest) CreateNewFile_InRoot() {
-	AssertTrue(false, "TODO")
+	var err error
+	var fi os.FileInfo
+	var stat *syscall.Stat_t
+
+	fileName := path.Join(t.mfs.Dir(), "foo")
+	const contents = "Hello\x00world"
+
+	// Write a file.
+	createTime := t.clock.Now()
+	err = ioutil.WriteFile(fileName, []byte(contents), 0400)
+	AssertEq(nil, err)
+
+	// Simulate time advancing.
+	t.clock.AdvanceTime(time.Second)
+
+	// Stat it.
+	fi, err = os.Stat(fileName)
+	stat = fi.Sys().(*syscall.Stat_t)
+
+	AssertEq(nil, err)
+	ExpectEq("foo", fi.Name())
+	ExpectEq(len(contents), fi.Size())
+	ExpectEq(0400, fi.Mode())
+	ExpectEq(0, fi.ModTime().Sub(createTime))
+	ExpectFalse(fi.IsDir())
+
+	ExpectNe(0, stat.Ino)
+	ExpectEq(1, stat.Nlink)
+	ExpectEq(currentUid(), stat.Uid)
+	ExpectEq(currentGid(), stat.Gid)
+	ExpectEq(len(contents), stat.Size)
+	ExpectEq(0, timespecToTime(stat.Atimespec).Sub(createTime))
+	ExpectEq(0, timespecToTime(stat.Mtimespec).Sub(createTime))
+	ExpectEq(0, timespecToTime(stat.Ctimespec).Sub(createTime))
+
+	// Read it back.
+	slice, err := ioutil.ReadFile(fileName)
+	AssertEq(nil, err)
+	ExpectEq(contents, string(slice))
 }
 
 func (t *MemFSTest) CreateNewFile_InSubDir() {
@@ -336,6 +374,10 @@ func (t *MemFSTest) UnlinkFile_NonExistent() {
 	AssertTrue(false, "TODO")
 }
 
+func (t *MemFSTest) UnlinkFile_StillOpen() {
+	AssertTrue(false, "TODO")
+}
+
 func (t *MemFSTest) Rmdir_NonEmpty() {
 	var err error
 
@@ -468,3 +510,11 @@ func (t *MemFSTest) CaseSensitive() {
 		AssertThat(err, Error(HasSubstr("no such file or directory")))
 	}
 }
+
+func (t *MemFSTest) FileReadsAndWrites() {
+	AssertTrue(false, "TODO")
+}
+
+func (t *MemFSTest) FileReadsAndWrites_BeyondEOF() {
+	AssertTrue(false, "TODO")
+}
diff --git a/samples/memfs/posix_test.go b/samples/memfs/posix_test.go
new file mode 100644
index 0000000..f48be48
--- /dev/null
+++ b/samples/memfs/posix_test.go
@@ -0,0 +1,253 @@
+// Copyright 2015 Google Inc. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Tests for the behavior of os.File objects on plain old posix file systems,
+// for use in verifying the intended behavior of memfs.
+
+package memfs_test
+
+import (
+	"io"
+	"io/ioutil"
+	"os"
+	"path"
+	"testing"
+
+	. "github.com/jacobsa/ogletest"
+)
+
+func TestPosix(t *testing.T) { RunTests(t) }
+
+////////////////////////////////////////////////////////////////////////
+// Helpers
+////////////////////////////////////////////////////////////////////////
+
+func getFileOffset(f *os.File) (offset int64, err error) {
+	const relativeToCurrent = 1
+	offset, err = f.Seek(0, relativeToCurrent)
+	return
+}
+
+////////////////////////////////////////////////////////////////////////
+// Boilerplate
+////////////////////////////////////////////////////////////////////////
+
+type PosixTest struct {
+	// A temporary directory.
+	dir string
+
+	// Files to close when tearing down. Nil entries are skipped.
+	toClose []io.Closer
+}
+
+var _ SetUpInterface = &PosixTest{}
+var _ TearDownInterface = &PosixTest{}
+
+func init() { RegisterTestSuite(&PosixTest{}) }
+
+func (t *PosixTest) SetUp(ti *TestInfo) {
+	var err error
+
+	// Create a temporary directory.
+	t.dir, err = ioutil.TempDir("", "posix_test")
+	if err != nil {
+		panic(err)
+	}
+}
+
+func (t *PosixTest) TearDown() {
+	// Close any files we opened.
+	for _, c := range t.toClose {
+		if c == nil {
+			continue
+		}
+
+		err := c.Close()
+		if err != nil {
+			panic(err)
+		}
+	}
+
+	// Remove the temporary directory.
+	err := os.RemoveAll(t.dir)
+	if err != nil {
+		panic(err)
+	}
+}
+
+////////////////////////////////////////////////////////////////////////
+// Test functions
+////////////////////////////////////////////////////////////////////////
+
+func (t *PosixTest) WriteOverlapsEndOfFile() {
+	var err error
+	var n int
+
+	// Create a file.
+	f, err := os.Create(path.Join(t.dir, "foo"))
+	t.toClose = append(t.toClose, f)
+	AssertEq(nil, err)
+
+	// Make it 4 bytes long.
+	err = f.Truncate(4)
+	AssertEq(nil, err)
+
+	// Write the range [2, 6).
+	n, err = f.WriteAt([]byte("taco"), 2)
+	AssertEq(nil, err)
+	AssertEq(4, n)
+
+	// Read the full contents of the file.
+	contents, err := ioutil.ReadAll(f)
+	AssertEq(nil, err)
+	ExpectEq("\x00\x00taco", string(contents))
+}
+
+func (t *PosixTest) WriteStartsAtEndOfFile() {
+	var err error
+	var n int
+
+	// Create a file.
+	f, err := os.Create(path.Join(t.dir, "foo"))
+	t.toClose = append(t.toClose, f)
+	AssertEq(nil, err)
+
+	// Make it 2 bytes long.
+	err = f.Truncate(2)
+	AssertEq(nil, err)
+
+	// Write the range [2, 6).
+	n, err = f.WriteAt([]byte("taco"), 2)
+	AssertEq(nil, err)
+	AssertEq(4, n)
+
+	// Read the full contents of the file.
+	contents, err := ioutil.ReadAll(f)
+	AssertEq(nil, err)
+	ExpectEq("\x00\x00taco", string(contents))
+}
+
+func (t *PosixTest) WriteStartsPastEndOfFile() {
+	var err error
+	var n int
+
+	// Create a file.
+	f, err := os.Create(path.Join(t.dir, "foo"))
+	t.toClose = append(t.toClose, f)
+	AssertEq(nil, err)
+
+	// Write the range [2, 6).
+	n, err = f.WriteAt([]byte("taco"), 2)
+	AssertEq(nil, err)
+	AssertEq(4, n)
+
+	// Read the full contents of the file.
+	contents, err := ioutil.ReadAll(f)
+	AssertEq(nil, err)
+	ExpectEq("\x00\x00taco", string(contents))
+}
+
+func (t *PosixTest) WriteAtDoesntChangeOffset_NotAppendMode() {
+	var err error
+	var n int
+
+	// Create a file.
+	f, err := os.Create(path.Join(t.dir, "foo"))
+	t.toClose = append(t.toClose, f)
+	AssertEq(nil, err)
+
+	// Make it 16 bytes long.
+	err = f.Truncate(16)
+	AssertEq(nil, err)
+
+	// Seek to offset 4.
+	_, err = f.Seek(4, 0)
+	AssertEq(nil, err)
+
+	// Write the range [10, 14).
+	n, err = f.WriteAt([]byte("taco"), 2)
+	AssertEq(nil, err)
+	AssertEq(4, n)
+
+	// We should still be at offset 4.
+	offset, err := getFileOffset(f)
+	AssertEq(nil, err)
+	ExpectEq(4, offset)
+}
+
+func (t *PosixTest) WriteAtDoesntChangeOffset_AppendMode() {
+	var err error
+	var n int
+
+	// Create a file in append mode.
+	f, err := os.OpenFile(
+		path.Join(t.dir, "foo"),
+		os.O_RDWR|os.O_APPEND|os.O_CREATE,
+		0600)
+
+	t.toClose = append(t.toClose, f)
+	AssertEq(nil, err)
+
+	// Make it 16 bytes long.
+	err = f.Truncate(16)
+	AssertEq(nil, err)
+
+	// Seek to offset 4.
+	_, err = f.Seek(4, 0)
+	AssertEq(nil, err)
+
+	// Write the range [10, 14).
+	n, err = f.WriteAt([]byte("taco"), 2)
+	AssertEq(nil, err)
+	AssertEq(4, n)
+
+	// We should still be at offset 4.
+	offset, err := getFileOffset(f)
+	AssertEq(nil, err)
+	ExpectEq(4, offset)
+}
+
+func (t *PosixTest) ReadsPastEndOfFile() {
+	var err error
+	var n int
+	buf := make([]byte, 1024)
+
+	// Create a file.
+	f, err := os.Create(path.Join(t.dir, "foo"))
+	t.toClose = append(t.toClose, f)
+	AssertEq(nil, err)
+
+	// Give it some contents.
+	n, err = f.Write([]byte("taco"))
+	AssertEq(nil, err)
+	AssertEq(4, n)
+
+	// Read a range overlapping EOF.
+	n, err = f.ReadAt(buf[:4], 2)
+	AssertEq(io.EOF, err)
+	ExpectEq(2, n)
+	ExpectEq("co", string(buf[:n]))
+
+	// Read a range starting at EOF.
+	n, err = f.ReadAt(buf[:4], 4)
+	AssertEq(io.EOF, err)
+	ExpectEq(0, n)
+	ExpectEq("", string(buf[:n]))
+
+	// Read a range starting past EOF.
+	n, err = f.ReadAt(buf[:4], 100)
+	AssertEq(io.EOF, err)
+	ExpectEq(0, n)
+	ExpectEq("", string(buf[:n]))
+}
diff --git a/server.go b/server.go
index 14c8f15..cc343b9 100644
--- a/server.go
+++ b/server.go
@@ -195,6 +195,35 @@ func (s *server) handleFuseRequest(fuseReq bazilfuse.Request) {
 		s.logger.Println("Responding:", fuseResp)
 		typed.Respond(fuseResp)
 
+	case *bazilfuse.CreateRequest:
+		// Convert the request.
+		req := &CreateFileRequest{
+			Header: convertHeader(typed.Header),
+			Parent: InodeID(typed.Header.Node),
+			Name:   typed.Name,
+			Mode:   typed.Mode,
+			Flags:  typed.Flags,
+		}
+
+		// Call the file system.
+		resp, err := s.fs.CreateFile(ctx, req)
+		if err != nil {
+			s.logger.Println("Responding:", err)
+			typed.RespondError(err)
+			return
+		}
+
+		// Convert the response.
+		fuseResp := &bazilfuse.CreateResponse{
+			OpenResponse: bazilfuse.OpenResponse{
+				Handle: bazilfuse.HandleID(resp.Handle),
+			},
+		}
+		convertChildInodeEntry(s.clock, &resp.Entry, &fuseResp.LookupResponse)
+
+		s.logger.Println("Responding:", fuseResp)
+		typed.Respond(fuseResp)
+
 	case *bazilfuse.RemoveRequest:
 		// We don't yet support files.
 		if !typed.Dir {
@@ -326,6 +355,32 @@ func (s *server) handleFuseRequest(fuseReq bazilfuse.Request) {
 			typed.Respond(fuseResp)
 		}
 
+	case *bazilfuse.WriteRequest:
+		// Convert the request.
+		req := &WriteFileRequest{
+			Header: convertHeader(typed.Header),
+			Inode:  InodeID(typed.Header.Node),
+			Handle: HandleID(typed.Handle),
+			Data:   typed.Data,
+			Offset: typed.Offset,
+		}
+
+		// Call the file system.
+		_, err := s.fs.WriteFile(ctx, req)
+		if err != nil {
+			s.logger.Println("Responding:", err)
+			typed.RespondError(err)
+			return
+		}
+
+		// Convert the response.
+		fuseResp := &bazilfuse.WriteResponse{
+			Size: len(typed.Data),
+		}
+
+		s.logger.Println("Responding:", fuseResp)
+		typed.Respond(fuseResp)
+
 	default:
 		s.logger.Println("Unhandled type. Returning ENOSYS.")
 		typed.RespondError(ENOSYS)