Merge pull request #3934 from avagin:feature/fuse

PiperOrigin-RevId: 332122081

pkg/marshal/marshal_impl_util.go

@@ -44,7 +44,7 @@ func (StubMarshallable) MarshalBytes(dst []byte) {
 
 // UnmarshalBytes implements Marshallable.UnmarshalBytes.
 func (StubMarshallable) UnmarshalBytes(src []byte) {
-	panic("Please implement your own UnMarshalBytes function")
+	panic("Please implement your own UnmarshalBytes function")
 }
 
 // Packed implements Marshallable.Packed.

pkg/sentry/fs/g3doc/fuse.md

@@ -95,6 +95,52 @@ ops can be implemented in parallel.
 -   Implement the remaining FUSE ops and decide if we can omit rarely used
     operations like ioctl.
 
+### Design Details
+
+#### Lifecycle for a FUSE Request
+
+-   User invokes a syscall
+-   Sentry prepares corresponding request
+    -   If FUSE device is available
+        -   Write the request in binary
+    -   If FUSE device is full
+        -   Kernel task blocked until available
+-   Sentry notifies the readers of fuse device that it's ready for read
+-   FUSE daemon reads the request and processes it
+-   Sentry waits until a reply is written to the FUSE device
+    -   but returns directly for async requests
+-   FUSE daemon writes to the fuse device
+-   Sentry processes the reply
+    -   For sync requests, unblock blocked kernel task
+    -   For async requests, execute pre-specified callback if any
+-   Sentry returns the syscall to the user
+
+#### Channels and Queues for Requests in Different Stages
+
+`connection.initializedChan`
+
+-   a channel that the requests issued before connection initialization blocks
+    on.
+
+`fd.queue`
+
+-   a queue of requests that haven’t been read by the FUSE daemon yet.
+
+`fd.completions`
+
+-   a map of the requests that have been prepared but not yet received a
+    response, including the ones on the `fd.queue`.
+
+`fd.waitQueue`
+
+-   a queue of waiters that is waiting for the fuse device fd to be available,
+    such as the FUSE daemon.
+
+`fd.fullQueueCh`
+
+-   a channel that the kernel task will be blocked on when the fd is not
+    available.
+
 # Appendix
 
 ## FUSE Protocol

pkg/sentry/fsimpl/fuse/BUILD

@@ -30,12 +30,17 @@ go_library(
     name = "fuse",
     srcs = [
         "connection.go",
+        "connection_control.go",
         "dev.go",
+        "directory.go",
+        "file.go",
         "fusefs.go",
-        "init.go",
         "inode_refs.go",
+        "read_write.go",
         "register.go",
+        "regular_file.go",
         "request_list.go",
+        "request_response.go",
     ],
     visibility = ["//pkg/sentry:internal"],
     deps = [
@@ -44,6 +49,7 @@ go_library(
         "//pkg/log",
         "//pkg/marshal",
         "//pkg/refs",
+        "//pkg/safemem",
         "//pkg/sentry/fsimpl/devtmpfs",
         "//pkg/sentry/fsimpl/kernfs",
         "//pkg/sentry/kernel",
@@ -60,7 +66,11 @@ go_library(
 go_test(
     name = "fuse_test",
     size = "small",
-    srcs = ["dev_test.go"],
+    srcs = [
+        "connection_test.go",
+        "dev_test.go",
+        "utils_test.go",
+    ],
     library = ":fuse",
     deps = [
         "//pkg/abi/linux",

pkg/sentry/fsimpl/fuse/connection.go

@@ -15,31 +15,17 @@
 package fuse
 
 import (
-	"errors"
-	"fmt"
 	"sync"
-	"sync/atomic"
-	"syscall"
 
 	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/context"
 	"gvisor.dev/gvisor/pkg/log"
-	"gvisor.dev/gvisor/pkg/marshal"
 	"gvisor.dev/gvisor/pkg/sentry/kernel"
-	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
 	"gvisor.dev/gvisor/pkg/sentry/vfs"
 	"gvisor.dev/gvisor/pkg/syserror"
 	"gvisor.dev/gvisor/pkg/waiter"
 )
 
-// maxActiveRequestsDefault is the default setting controlling the upper bound
-// on the number of active requests at any given time.
-const maxActiveRequestsDefault = 10000
-
-// Ordinary requests have even IDs, while interrupts IDs are odd.
-// Used to increment the unique ID for each FUSE request.
-var reqIDStep uint64 = 2
-
 const (
 	// fuseDefaultMaxBackground is the default value for MaxBackground.
 	fuseDefaultMaxBackground = 12
@@ -52,43 +38,36 @@ const (
 	fuseDefaultMaxPagesPerReq = 32
 )
 
-// Request represents a FUSE operation request that hasn't been sent to the
-// server yet.
-//
-// +stateify savable
-type Request struct {
-	requestEntry
-
-	id   linux.FUSEOpID
-	hdr  *linux.FUSEHeaderIn
-	data []byte
-}
-
-// Response represents an actual response from the server, including the
-// response payload.
-//
-// +stateify savable
-type Response struct {
-	opcode linux.FUSEOpcode
-	hdr    linux.FUSEHeaderOut
-	data   []byte
-}
-
 // connection is the struct by which the sentry communicates with the FUSE server daemon.
+// Lock order:
+// - conn.fd.mu
+// - conn.mu
+// - conn.asyncMu
 type connection struct {
 	fd *DeviceFD
 
+	// mu protects access to struct memebers.
+	mu sync.Mutex
+
+	// attributeVersion is the version of connection's attributes.
+	attributeVersion uint64
+
+	// We target FUSE 7.23.
 	// The following FUSE_INIT flags are currently unsupported by this implementation:
-	// - FUSE_ATOMIC_O_TRUNC: requires open(..., O_TRUNC)
 	// - FUSE_EXPORT_SUPPORT
-	// - FUSE_HANDLE_KILLPRIV
 	// - FUSE_POSIX_LOCKS: requires POSIX locks
 	// - FUSE_FLOCK_LOCKS: requires POSIX locks
 	// - FUSE_AUTO_INVAL_DATA: requires page caching eviction
-	// - FUSE_EXPLICIT_INVAL_DATA: requires page caching eviction
 	// - FUSE_DO_READDIRPLUS/FUSE_READDIRPLUS_AUTO: requires FUSE_READDIRPLUS implementation
 	// - FUSE_ASYNC_DIO
-	// - FUSE_POSIX_ACL: affects defaultPermissions, posixACL, xattr handler
+	// - FUSE_PARALLEL_DIROPS (7.25)
+	// - FUSE_HANDLE_KILLPRIV (7.26)
+	// - FUSE_POSIX_ACL: affects defaultPermissions, posixACL, xattr handler (7.26)
+	// - FUSE_ABORT_ERROR (7.27)
+	// - FUSE_CACHE_SYMLINKS (7.28)
+	// - FUSE_NO_OPENDIR_SUPPORT (7.29)
+	// - FUSE_EXPLICIT_INVAL_DATA: requires page caching eviction (7.30)
+	// - FUSE_MAP_ALIGNMENT (7.31)
 
 	// initialized after receiving FUSE_INIT reply.
 	// Until it's set, suspend sending FUSE requests.
@@ -98,10 +77,6 @@ type connection struct {
 	// initializedChan is used to block requests before initialization.
 	initializedChan chan struct{}
 
-	// blocked when there are too many outstading backgrounds requests (NumBackground == MaxBackground).
-	// TODO(gvisor.dev/issue/3185): update the numBackground accordingly; use a channel to block.
-	blocked bool
-
 	// connected (connection established) when a new FUSE file system is created.
 	// Set to false when:
 	//   umount,
@@ -109,48 +84,55 @@ type connection struct {
 	//   device release.
 	connected bool
 
-	// aborted via sysfs.
-	// TODO(gvisor.dev/issue/3185): abort all queued requests.
-	aborted bool
-
 	// connInitError if FUSE_INIT encountered error (major version mismatch).
 	// Only set in INIT.
 	connInitError bool
 
 	// connInitSuccess if FUSE_INIT is successful.
 	// Only set in INIT.
-	// Used for destory.
+	// Used for destory (not yet implemented).
 	connInitSuccess bool
 
-	// TODO(gvisor.dev/issue/3185): All the queue logic are working in progress.
+	// aborted via sysfs, and will send ECONNABORTED to read after disconnection (instead of ENODEV).
+	// Set only if abortErr is true and via fuse control fs (not yet implemented).
+	// TODO(gvisor.dev/issue/3525): set this to true when user aborts.
+	aborted bool
 
-	// NumberBackground is the number of requests in the background.
-	numBackground uint16
-
-	// congestionThreshold for NumBackground.
-	// Negotiated in FUSE_INIT.
-	congestionThreshold uint16
-
-	// maxBackground is the maximum number of NumBackground.
-	// Block connection when it is reached.
-	// Negotiated in FUSE_INIT.
-	maxBackground uint16
-
-	// numActiveBackground is the number of requests in background and has being marked as active.
-	numActiveBackground uint16
-
-	// numWating is the number of requests waiting for completion.
+	// numWating is the number of requests waiting to be
+	// sent to FUSE device or being processed by FUSE daemon.
 	numWaiting uint32
 
-	// TODO(gvisor.dev/issue/3185): BgQueue
-	// some queue for background queued requests.
+	// Terminology note:
+	//
+	// - `asyncNumMax` is the `MaxBackground` in the FUSE_INIT_IN struct.
+	//
+	// - `asyncCongestionThreshold` is the `CongestionThreshold` in the FUSE_INIT_IN struct.
+	//
+	// We call the "background" requests in unix term as async requests.
+	// The "async requests" in unix term is our async requests that expect a reply,
+	// i.e. `!request.noReply`
 
-	// bgLock protects:
-	// MaxBackground, CongestionThreshold, NumBackground,
-	// NumActiveBackground, BgQueue, Blocked.
-	bgLock sync.Mutex
+	// asyncMu protects the async request fields.
+	asyncMu sync.Mutex
+
+	// asyncNum is the number of async requests.
+	// Protected by asyncMu.
+	asyncNum uint16
+
+	// asyncCongestionThreshold the number of async requests.
+	// Negotiated in FUSE_INIT as "CongestionThreshold".
+	// TODO(gvisor.dev/issue/3529): add congestion control.
+	// Protected by asyncMu.
+	asyncCongestionThreshold uint16
+
+	// asyncNumMax is the maximum number of asyncNum.
+	// Connection blocks the async requests when it is reached.
+	// Negotiated in FUSE_INIT as "MaxBackground".
+	// Protected by asyncMu.
+	asyncNumMax uint16
 
 	// maxRead is the maximum size of a read buffer in in bytes.
+	// Initialized from a fuse fs parameter.
 	maxRead uint32
 
 	// maxWrite is the maximum size of a write buffer in bytes.
@@ -165,23 +147,20 @@ type connection struct {
 	// Negotiated and only set in INIT.
 	minor uint32
 
+	// atomicOTrunc is true when FUSE does not send a separate SETATTR request
+	// before open with O_TRUNC flag.
+	// Negotiated and only set in INIT.
+	atomicOTrunc bool
+
 	// asyncRead if read pages asynchronously.
 	// Negotiated and only set in INIT.
 	asyncRead bool
 
-	// abortErr is true if kernel need to return an unique read error after abort.
-	// Negotiated and only set in INIT.
-	abortErr bool
-
 	// writebackCache is true for write-back cache policy,
 	// false for write-through policy.
 	// Negotiated and only set in INIT.
 	writebackCache bool
 
-	// cacheSymlinks if filesystem needs to cache READLINK responses in page cache.
-	// Negotiated and only set in INIT.
-	cacheSymlinks bool
-
 	// bigWrites if doing multi-page cached writes.
 	// Negotiated and only set in INIT.
 	bigWrites bool
@@ -189,116 +168,70 @@ type connection struct {
 	// dontMask if filestestem does not apply umask to creation modes.
 	// Negotiated in INIT.
 	dontMask bool
+
+	// noOpen if FUSE server doesn't support open operation.
+	// This flag only influence performance, not correctness of the program.
+	noOpen bool
 }
 
 // newFUSEConnection creates a FUSE connection to fd.
-func newFUSEConnection(_ context.Context, fd *vfs.FileDescription, maxInFlightRequests uint64) (*connection, error) {
+func newFUSEConnection(_ context.Context, fd *vfs.FileDescription, opts *filesystemOptions) (*connection, error) {
 	// Mark the device as ready so it can be used. /dev/fuse can only be used if the FD was used to
 	// mount a FUSE filesystem.
 	fuseFD := fd.Impl().(*DeviceFD)
-	fuseFD.mounted = true
 
 	// Create the writeBuf for the header to be stored in.
 	hdrLen := uint32((*linux.FUSEHeaderOut)(nil).SizeBytes())
 	fuseFD.writeBuf = make([]byte, hdrLen)
 	fuseFD.completions = make(map[linux.FUSEOpID]*futureResponse)
-	fuseFD.fullQueueCh = make(chan struct{}, maxInFlightRequests)
+	fuseFD.fullQueueCh = make(chan struct{}, opts.maxActiveRequests)
 	fuseFD.writeCursor = 0
 
 	return &connection{
-		fd:                  fuseFD,
-		maxBackground:       fuseDefaultMaxBackground,
-		congestionThreshold: fuseDefaultCongestionThreshold,
-		maxPages:            fuseDefaultMaxPagesPerReq,
-		initializedChan:     make(chan struct{}),
-		connected:           true,
+		fd:                       fuseFD,
+		asyncNumMax:              fuseDefaultMaxBackground,
+		asyncCongestionThreshold: fuseDefaultCongestionThreshold,
+		maxRead:                  opts.maxRead,
+		maxPages:                 fuseDefaultMaxPagesPerReq,
+		initializedChan:          make(chan struct{}),
+		connected:                true,
 	}, nil
 }
 
-// SetInitialized atomically sets the connection as initialized.
-func (conn *connection) SetInitialized() {
-	// Unblock the requests sent before INIT.
-	close(conn.initializedChan)
-
-	// Close the channel first to avoid the non-atomic situation
-	// where conn.initialized is true but there are
-	// tasks being blocked on the channel.
-	// And it prevents the newer tasks from gaining
-	// unnecessary higher chance to be issued before the blocked one.
-
-	atomic.StoreInt32(&(conn.initialized), int32(1))
+// CallAsync makes an async (aka background) request.
+// It's a simple wrapper around Call().
+func (conn *connection) CallAsync(t *kernel.Task, r *Request) error {
+	r.async = true
+	_, err := conn.Call(t, r)
+	return err
 }
 
-// IsInitialized atomically check if the connection is initialized.
-// pairs with SetInitialized().
-func (conn *connection) Initialized() bool {
-	return atomic.LoadInt32(&(conn.initialized)) != 0
-}
-
-// NewRequest creates a new request that can be sent to the FUSE server.
-func (conn *connection) NewRequest(creds *auth.Credentials, pid uint32, ino uint64, opcode linux.FUSEOpcode, payload marshal.Marshallable) (*Request, error) {
-	conn.fd.mu.Lock()
-	defer conn.fd.mu.Unlock()
-	conn.fd.nextOpID += linux.FUSEOpID(reqIDStep)
-
-	hdrLen := (*linux.FUSEHeaderIn)(nil).SizeBytes()
-	hdr := linux.FUSEHeaderIn{
-		Len:    uint32(hdrLen + payload.SizeBytes()),
-		Opcode: opcode,
-		Unique: conn.fd.nextOpID,
-		NodeID: ino,
-		UID:    uint32(creds.EffectiveKUID),
-		GID:    uint32(creds.EffectiveKGID),
-		PID:    pid,
-	}
-
-	buf := make([]byte, hdr.Len)
-	hdr.MarshalUnsafe(buf[:hdrLen])
-	payload.MarshalUnsafe(buf[hdrLen:])
-
-	return &Request{
-		id:   hdr.Unique,
-		hdr:  &hdr,
-		data: buf,
-	}, nil
-}
-
-// Call makes a request to the server and blocks the invoking task until a
-// server responds with a response. Task should never be nil.
-// Requests will not be sent before the connection is initialized.
-// For async tasks, use CallAsync().
+// Call makes a request to the server.
+// Block before the connection is initialized.
+// When the Request is FUSE_INIT, it will not be blocked before initialization.
+// Task should never be nil.
+//
+// For a sync request, it blocks the invoking task until
+// a server responds with a response.
+//
+// For an async request (that do not expect a response immediately),
+// it returns directly unless being blocked either before initialization
+// or when there are too many async requests ongoing.
+//
+// Example for async request:
+// init, readahead, write, async read/write, fuse_notify_reply,
+// non-sync release, interrupt, forget.
+//
+// The forget request does not have a reply,
+// as documented in include/uapi/linux/fuse.h:FUSE_FORGET.
 func (conn *connection) Call(t *kernel.Task, r *Request) (*Response, error) {
 	// Block requests sent before connection is initalized.
-	if !conn.Initialized() {
+	if !conn.Initialized() && r.hdr.Opcode != linux.FUSE_INIT {
 		if err := t.Block(conn.initializedChan); err != nil {
 			return nil, err
 		}
 	}
 
-	return conn.call(t, r)
-}
-
-// CallAsync makes an async (aka background) request.
-// Those requests either do not expect a response (e.g. release) or
-// the response should be handled by others (e.g. init).
-// Return immediately unless the connection is blocked (before initialization).
-// Async call example: init, release, forget, aio, interrupt.
-// When the Request is FUSE_INIT, it will not be blocked before initialization.
-func (conn *connection) CallAsync(t *kernel.Task, r *Request) error {
-	// Block requests sent before connection is initalized.
-	if !conn.Initialized() && r.hdr.Opcode != linux.FUSE_INIT {
-		if err := t.Block(conn.initializedChan); err != nil {
-			return err
-		}
-	}
-
-	// This should be the only place that invokes call() with a nil task.
-	_, err := conn.call(nil, r)
-	return err
-}
-
-// call makes a call without blocking checks.
-func (conn *connection) call(t *kernel.Task, r *Request) (*Response, error) {
 	if !conn.connected {
 		return nil, syserror.ENOTCONN
 	}
@@ -315,31 +248,6 @@ func (conn *connection) call(t *kernel.Task, r *Request) (*Response, error) {
 	return fut.resolve(t)
 }
 
-// Error returns the error of the FUSE call.
-func (r *Response) Error() error {
-	errno := r.hdr.Error
-	if errno >= 0 {
-		return nil
-	}
-
-	sysErrNo := syscall.Errno(-errno)
-	return error(sysErrNo)
-}
-
-// UnmarshalPayload unmarshals the response data into m.
-func (r *Response) UnmarshalPayload(m marshal.Marshallable) error {
-	hdrLen := r.hdr.SizeBytes()
-	haveDataLen := r.hdr.Len - uint32(hdrLen)
-	wantDataLen := uint32(m.SizeBytes())
-
-	if haveDataLen < wantDataLen {
-		return fmt.Errorf("payload too small. Minimum data lenth required: %d,  but got data length %d", wantDataLen, haveDataLen)
-	}
-
-	m.UnmarshalUnsafe(r.data[hdrLen:])
-	return nil
-}
-
 // callFuture makes a request to the server and returns a future response.
 // Call resolve() when the response needs to be fulfilled.
 func (conn *connection) callFuture(t *kernel.Task, r *Request) (*futureResponse, error) {
@@ -358,11 +266,6 @@ func (conn *connection) callFuture(t *kernel.Task, r *Request) (*futureResponse,
 	// if there are always too many ongoing requests all the time. The
 	// supported maxActiveRequests setting should be really high to avoid this.
 	for conn.fd.numActiveRequests == conn.fd.fs.opts.maxActiveRequests {
-		if t == nil {
-			// Since there is no task that is waiting. We must error out.
-			return nil, errors.New("FUSE request queue full")
-		}
-
 		log.Infof("Blocking request %v from being queued. Too many active requests: %v",
 			r.id, conn.fd.numActiveRequests)
 		conn.fd.mu.Unlock()
@@ -378,9 +281,19 @@ func (conn *connection) callFuture(t *kernel.Task, r *Request) (*futureResponse,
 
 // callFutureLocked makes a request to the server and returns a future response.
 func (conn *connection) callFutureLocked(t *kernel.Task, r *Request) (*futureResponse, error) {
+	// Check connected again holding conn.mu.
+	conn.mu.Lock()
+	if !conn.connected {
+		conn.mu.Unlock()
+		// we checked connected before,
+		// this must be due to aborted connection.
+		return nil, syserror.ECONNABORTED
+	}
+	conn.mu.Unlock()
+
 	conn.fd.queue.PushBack(r)
-	conn.fd.numActiveRequests += 1
-	fut := newFutureResponse(r.hdr.Opcode)
+	conn.fd.numActiveRequests++
+	fut := newFutureResponse(r)
 	conn.fd.completions[r.id] = fut
 
 	// Signal the readers that there is something to read.
@@ -388,50 +301,3 @@ func (conn *connection) callFutureLocked(t *kernel.Task, r *Request) (*futureRes
 
 	return fut, nil
 }
-
-// futureResponse represents an in-flight request, that may or may not have
-// completed yet. Convert it to a resolved Response by calling Resolve, but note
-// that this may block.
-//
-// +stateify savable
-type futureResponse struct {
-	opcode linux.FUSEOpcode
-	ch     chan struct{}
-	hdr    *linux.FUSEHeaderOut
-	data   []byte
-}
-
-// newFutureResponse creates a future response to a FUSE request.
-func newFutureResponse(opcode linux.FUSEOpcode) *futureResponse {
-	return &futureResponse{
-		opcode: opcode,
-		ch:     make(chan struct{}),
-	}
-}
-
-// resolve blocks the task until the server responds to its corresponding request,
-// then returns a resolved response.
-func (f *futureResponse) resolve(t *kernel.Task) (*Response, error) {
-	// If there is no Task associated with this request  - then we don't try to resolve
-	// the response.  Instead, the task writing the response (proxy to the server) will
-	// process the response on our behalf.
-	if t == nil {
-		log.Infof("fuse.Response.resolve: Not waiting on a response from server.")
-		return nil, nil
-	}
-
-	if err := t.Block(f.ch); err != nil {
-		return nil, err
-	}
-
-	return f.getResponse(), nil
-}
-
-// getResponse creates a Response from the data the futureResponse has.
-func (f *futureResponse) getResponse() *Response {
-	return &Response{
-		opcode: f.opcode,
-		hdr:    *f.hdr,
-		data:   f.data,
-	}
-}

pkg/sentry/fsimpl/fuse/connection_control.go

@@ -15,7 +15,11 @@
 package fuse
 
 import (
+	"sync/atomic"
+	"syscall"
+
 	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
 	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
 )
 
@@ -29,9 +33,10 @@ const (
 	// Follow the same behavior as unix fuse implementation.
 	fuseMaxTimeGranNs = 1000000000
 
-	// Minimum value for MaxWrite.
+	// Minimum value for MaxWrite and MaxRead.
 	// Follow the same behavior as unix fuse implementation.
 	fuseMinMaxWrite = 4096
+	fuseMinMaxRead  = 4096
 
 	// Temporary default value for max readahead, 128kb.
 	fuseDefaultMaxReadahead = 131072
@@ -49,6 +54,26 @@ var (
 	MaxUserCongestionThreshold uint16 = fuseDefaultCongestionThreshold
 )
 
+// SetInitialized atomically sets the connection as initialized.
+func (conn *connection) SetInitialized() {
+	// Unblock the requests sent before INIT.
+	close(conn.initializedChan)
+
+	// Close the channel first to avoid the non-atomic situation
+	// where conn.initialized is true but there are
+	// tasks being blocked on the channel.
+	// And it prevents the newer tasks from gaining
+	// unnecessary higher chance to be issued before the blocked one.
+
+	atomic.StoreInt32(&(conn.initialized), int32(1))
+}
+
+// IsInitialized atomically check if the connection is initialized.
+// pairs with SetInitialized().
+func (conn *connection) Initialized() bool {
+	return atomic.LoadInt32(&(conn.initialized)) != 0
+}
+
 // InitSend sends a FUSE_INIT request.
 func (conn *connection) InitSend(creds *auth.Credentials, pid uint32) error {
 	in := linux.FUSEInitIn{
@@ -70,29 +95,31 @@ func (conn *connection) InitSend(creds *auth.Credentials, pid uint32) error {
 }
 
 // InitRecv receives a FUSE_INIT reply and process it.
+//
+// Preconditions: conn.asyncMu must not be held if minor verion is newer than 13.
 func (conn *connection) InitRecv(res *Response, hasSysAdminCap bool) error {
 	if err := res.Error(); err != nil {
 		return err
 	}
 
-	var out linux.FUSEInitOut
-	if err := res.UnmarshalPayload(&out); err != nil {
+	initRes := fuseInitRes{initLen: res.DataLen()}
+	if err := res.UnmarshalPayload(&initRes); err != nil {
 		return err
 	}
 
-	return conn.initProcessReply(&out, hasSysAdminCap)
+	return conn.initProcessReply(&initRes.initOut, hasSysAdminCap)
 }
 
 // Process the FUSE_INIT reply from the FUSE server.
+// It tries to acquire the conn.asyncMu lock if minor version is newer than 13.
 func (conn *connection) initProcessReply(out *linux.FUSEInitOut, hasSysAdminCap bool) error {
+	// No matter error or not, always set initialzied.
+	// to unblock the blocked requests.
+	defer conn.SetInitialized()
+
 	// No support for old major fuse versions.
 	if out.Major != linux.FUSE_KERNEL_VERSION {
 		conn.connInitError = true
-
-		// Set the connection as initialized and unblock the blocked requests
-		// (i.e. return error for them).
-		conn.SetInitialized()
-
 		return nil
 	}
 
@@ -100,29 +127,14 @@ func (conn *connection) initProcessReply(out *linux.FUSEInitOut, hasSysAdminCap
 	conn.connInitSuccess = true
 	conn.minor = out.Minor
 
-	// No support for limits before minor version 13.
-	if out.Minor >= 13 {
-		conn.bgLock.Lock()
-
-		if out.MaxBackground > 0 {
-			conn.maxBackground = out.MaxBackground
-
-			if !hasSysAdminCap &&
-				conn.maxBackground > MaxUserBackgroundRequest {
-				conn.maxBackground = MaxUserBackgroundRequest
-			}
-		}
-
-		if out.CongestionThreshold > 0 {
-			conn.congestionThreshold = out.CongestionThreshold
-
-			if !hasSysAdminCap &&
-				conn.congestionThreshold > MaxUserCongestionThreshold {
-				conn.congestionThreshold = MaxUserCongestionThreshold
-			}
-		}
-
-		conn.bgLock.Unlock()
+	// No support for negotiating MaxWrite before minor version 5.
+	if out.Minor >= 5 {
+		conn.maxWrite = out.MaxWrite
+	} else {
+		conn.maxWrite = fuseMinMaxWrite
+	}
+	if conn.maxWrite < fuseMinMaxWrite {
+		conn.maxWrite = fuseMinMaxWrite
 	}
 
 	// No support for the following flags before minor version 6.
@@ -131,8 +143,6 @@ func (conn *connection) initProcessReply(out *linux.FUSEInitOut, hasSysAdminCap
 		conn.bigWrites = out.Flags&linux.FUSE_BIG_WRITES != 0
 		conn.dontMask = out.Flags&linux.FUSE_DONT_MASK != 0
 		conn.writebackCache = out.Flags&linux.FUSE_WRITEBACK_CACHE != 0
-		conn.cacheSymlinks = out.Flags&linux.FUSE_CACHE_SYMLINKS != 0
-		conn.abortErr = out.Flags&linux.FUSE_ABORT_ERROR != 0
 
 		// TODO(gvisor.dev/issue/3195): figure out how to use TimeGran (0 < TimeGran <= fuseMaxTimeGranNs).
 
@@ -148,19 +158,90 @@ func (conn *connection) initProcessReply(out *linux.FUSEInitOut, hasSysAdminCap
 		}
 	}
 
-	// No support for negotiating MaxWrite before minor version 5.
-	if out.Minor >= 5 {
-		conn.maxWrite = out.MaxWrite
-	} else {
-		conn.maxWrite = fuseMinMaxWrite
-	}
-	if conn.maxWrite < fuseMinMaxWrite {
-		conn.maxWrite = fuseMinMaxWrite
-	}
+	// No support for limits before minor version 13.
+	if out.Minor >= 13 {
+		conn.asyncMu.Lock()
 
-	// Set connection as initialized and unblock the requests
-	// issued before init.
-	conn.SetInitialized()
+		if out.MaxBackground > 0 {
+			conn.asyncNumMax = out.MaxBackground
+
+			if !hasSysAdminCap &&
+				conn.asyncNumMax > MaxUserBackgroundRequest {
+				conn.asyncNumMax = MaxUserBackgroundRequest
+			}
+		}
+
+		if out.CongestionThreshold > 0 {
+			conn.asyncCongestionThreshold = out.CongestionThreshold
+
+			if !hasSysAdminCap &&
+				conn.asyncCongestionThreshold > MaxUserCongestionThreshold {
+				conn.asyncCongestionThreshold = MaxUserCongestionThreshold
+			}
+		}
+
+		conn.asyncMu.Unlock()
+	}
 
 	return nil
 }
+
+// Abort this FUSE connection.
+// It tries to acquire conn.fd.mu, conn.lock, conn.bgLock in order.
+// All possible requests waiting or blocking will be aborted.
+//
+// Preconditions: conn.fd.mu is locked.
+func (conn *connection) Abort(ctx context.Context) {
+	conn.mu.Lock()
+	conn.asyncMu.Lock()
+
+	if !conn.connected {
+		conn.asyncMu.Unlock()
+		conn.mu.Unlock()
+		conn.fd.mu.Unlock()
+		return
+	}
+
+	conn.connected = false
+
+	// Empty the `fd.queue` that holds the requests
+	// not yet read by the FUSE daemon yet.
+	// These are a subset of the requests in `fuse.completion` map.
+	for !conn.fd.queue.Empty() {
+		req := conn.fd.queue.Front()
+		conn.fd.queue.Remove(req)
+	}
+
+	var terminate []linux.FUSEOpID
+
+	// 2. Collect the requests have not been sent to FUSE daemon,
+	// or have not received a reply.
+	for unique := range conn.fd.completions {
+		terminate = append(terminate, unique)
+	}
+
+	// Release locks to avoid deadlock.
+	conn.asyncMu.Unlock()
+	conn.mu.Unlock()
+
+	// 1. The requets blocked before initialization.
+	// Will reach call() `connected` check and return.
+	if !conn.Initialized() {
+		conn.SetInitialized()
+	}
+
+	// 2. Terminate the requests collected above.
+	// Set ECONNABORTED error.
+	// sendError() will remove them from `fd.completion` map.
+	// Will enter the path of a normally received error.
+	for _, toTerminate := range terminate {
+		conn.fd.sendError(ctx, -int32(syscall.ECONNABORTED), toTerminate)
+	}
+
+	// 3. The requests not yet written to FUSE device.
+	// Early terminate.
+	// Will reach callFutureLocked() `connected` check and return.
+	close(conn.fd.fullQueueCh)
+
+	// TODO(gvisor.dev/issue/3528): Forget all pending forget reqs.
+}

pkg/sentry/fsimpl/fuse/connection_test.go

@@ -0,0 +1,117 @@
+// Copyright 2020 The gVisor Authors.
+//
+// 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.
+
+package fuse
+
+import (
+	"math/rand"
+	"syscall"
+	"testing"
+
+	"gvisor.dev/gvisor/pkg/sentry/kernel"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/syserror"
+)
+
+// TestConnectionInitBlock tests if initialization
+// correctly blocks and unblocks the connection.
+// Since it's unfeasible to test kernelTask.Block() in unit test,
+// the code in Call() are not tested here.
+func TestConnectionInitBlock(t *testing.T) {
+	s := setup(t)
+	defer s.Destroy()
+
+	k := kernel.KernelFromContext(s.Ctx)
+
+	conn, _, err := newTestConnection(s, k, maxActiveRequestsDefault)
+	if err != nil {
+		t.Fatalf("newTestConnection: %v", err)
+	}
+
+	select {
+	case <-conn.initializedChan:
+		t.Fatalf("initializedChan should be blocking before SetInitialized")
+	default:
+	}
+
+	conn.SetInitialized()
+
+	select {
+	case <-conn.initializedChan:
+	default:
+		t.Fatalf("initializedChan should not be blocking after SetInitialized")
+	}
+}
+
+func TestConnectionAbort(t *testing.T) {
+	s := setup(t)
+	defer s.Destroy()
+
+	k := kernel.KernelFromContext(s.Ctx)
+	creds := auth.CredentialsFromContext(s.Ctx)
+	task := kernel.TaskFromContext(s.Ctx)
+
+	const numRequests uint64 = 256
+
+	conn, _, err := newTestConnection(s, k, numRequests)
+	if err != nil {
+		t.Fatalf("newTestConnection: %v", err)
+	}
+
+	testObj := &testPayload{
+		data: rand.Uint32(),
+	}
+
+	var futNormal []*futureResponse
+
+	for i := 0; i < int(numRequests); i++ {
+		req, err := conn.NewRequest(creds, uint32(i), uint64(i), 0, testObj)
+		if err != nil {
+			t.Fatalf("NewRequest creation failed: %v", err)
+		}
+		fut, err := conn.callFutureLocked(task, req)
+		if err != nil {
+			t.Fatalf("callFutureLocked failed: %v", err)
+		}
+		futNormal = append(futNormal, fut)
+	}
+
+	conn.Abort(s.Ctx)
+
+	// Abort should unblock the initialization channel.
+	// Note: no test requests are actually blocked on `conn.initializedChan`.
+	select {
+	case <-conn.initializedChan:
+	default:
+		t.Fatalf("initializedChan should not be blocking after SetInitialized")
+	}
+
+	// Abort will return ECONNABORTED error to unblocked requests.
+	for _, fut := range futNormal {
+		if fut.getResponse().hdr.Error != -int32(syscall.ECONNABORTED) {
+			t.Fatalf("Incorrect error code received for aborted connection: %v", fut.getResponse().hdr.Error)
+		}
+	}
+
+	// After abort, Call() should return directly with ENOTCONN.
+	req, err := conn.NewRequest(creds, 0, 0, 0, testObj)
+	if err != nil {
+		t.Fatalf("NewRequest creation failed: %v", err)
+	}
+	_, err = conn.Call(task, req)
+	if err != syserror.ENOTCONN {
+		t.Fatalf("Incorrect error code received for Call() after connection aborted")
+	}
+
+}

pkg/sentry/fsimpl/fuse/dev.go

@@ -19,7 +19,6 @@ import (
 
 	"gvisor.dev/gvisor/pkg/abi/linux"
 	"gvisor.dev/gvisor/pkg/context"
-	"gvisor.dev/gvisor/pkg/log"
 	"gvisor.dev/gvisor/pkg/sentry/kernel"
 	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
 	"gvisor.dev/gvisor/pkg/sentry/vfs"
@@ -56,9 +55,6 @@ type DeviceFD struct {
 	vfs.DentryMetadataFileDescriptionImpl
 	vfs.NoLockFD
 
-	// mounted specifies whether a FUSE filesystem was mounted using the DeviceFD.
-	mounted bool
-
 	// nextOpID is used to create new requests.
 	nextOpID linux.FUSEOpID
 
@@ -100,13 +96,15 @@ type DeviceFD struct {
 
 // Release implements vfs.FileDescriptionImpl.Release.
 func (fd *DeviceFD) Release(context.Context) {
-	fd.fs.conn.connected = false
+	if fd.fs != nil {
+		fd.fs.conn.connected = false
+	}
 }
 
 // PRead implements vfs.FileDescriptionImpl.PRead.
 func (fd *DeviceFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) {
 	// Operations on /dev/fuse don't make sense until a FUSE filesystem is mounted.
-	if !fd.mounted {
+	if fd.fs == nil {
 		return 0, syserror.EPERM
 	}
 
@@ -116,10 +114,16 @@ func (fd *DeviceFD) PRead(ctx context.Context, dst usermem.IOSequence, offset in
 // Read implements vfs.FileDescriptionImpl.Read.
 func (fd *DeviceFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) {
 	// Operations on /dev/fuse don't make sense until a FUSE filesystem is mounted.
-	if !fd.mounted {
+	if fd.fs == nil {
 		return 0, syserror.EPERM
 	}
 
+	// Return ENODEV if the filesystem is umounted.
+	if fd.fs.umounted {
+		// TODO(gvisor.dev/issue/3525): return ECONNABORTED if aborted via fuse control fs.
+		return 0, syserror.ENODEV
+	}
+
 	// We require that any Read done on this filesystem have a sane minimum
 	// read buffer. It must have the capacity for the fixed parts of any request
 	// header (Linux uses the request header and the FUSEWriteIn header for this
@@ -143,58 +147,82 @@ func (fd *DeviceFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.R
 }
 
 // readLocked implements the reading of the fuse device while locked with DeviceFD.mu.
+//
+// Preconditions: dst is large enough for any reasonable request.
 func (fd *DeviceFD) readLocked(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) {
-	if fd.queue.Empty() {
+	var req *Request
+
+	// Find the first valid request.
+	// For the normal case this loop only execute once.
+	for !fd.queue.Empty() {
+		req = fd.queue.Front()
+
+		if int64(req.hdr.Len)+int64(len(req.payload)) <= dst.NumBytes() {
+			break
+		}
+
+		// The request is too large. Cannot process it. All requests must be smaller than the
+		// negotiated size as specified by Connection.MaxWrite set as part of the FUSE_INIT
+		// handshake.
+		errno := -int32(syscall.EIO)
+		if req.hdr.Opcode == linux.FUSE_SETXATTR {
+			errno = -int32(syscall.E2BIG)
+		}
+
+		// Return the error to the calling task.
+		if err := fd.sendError(ctx, errno, req.hdr.Unique); err != nil {
+			return 0, err
+		}
+
+		// We're done with this request.
+		fd.queue.Remove(req)
+		req = nil
+	}
+
+	if req == nil {
 		return 0, syserror.ErrWouldBlock
 	}
 
-	var readCursor uint32
-	var bytesRead int64
-	for {
-		req := fd.queue.Front()
-		if dst.NumBytes() < int64(req.hdr.Len) {
-			// The request is too large. Cannot process it. All requests must be smaller than the
-			// negotiated size as specified by Connection.MaxWrite set as part of the FUSE_INIT
-			// handshake.
-			errno := -int32(syscall.EIO)
-			if req.hdr.Opcode == linux.FUSE_SETXATTR {
-				errno = -int32(syscall.E2BIG)
-			}
+	// We already checked the size: dst must be able to fit the whole request.
+	// Now we write the marshalled header, the payload,
+	// and the potential additional payload
+	// to the user memory IOSequence.
 
-			// Return the error to the calling task.
-			if err := fd.sendError(ctx, errno, req); err != nil {
-				return 0, err
-			}
+	n, err := dst.CopyOut(ctx, req.data)
+	if err != nil {
+		return 0, err
+	}
+	if n != len(req.data) {
+		return 0, syserror.EIO
+	}
 
-			// We're done with this request.
-			fd.queue.Remove(req)
-
-			// Restart the read as this request was invalid.
-			log.Warningf("fuse.DeviceFD.Read: request found was too large. Restarting read.")
-			return fd.readLocked(ctx, dst, opts)
-		}
-
-		n, err := dst.CopyOut(ctx, req.data[readCursor:])
+	if req.hdr.Opcode == linux.FUSE_WRITE {
+		written, err := dst.DropFirst(n).CopyOut(ctx, req.payload)
 		if err != nil {
 			return 0, err
 		}
-		readCursor += uint32(n)
-		bytesRead += int64(n)
-
-		if readCursor >= req.hdr.Len {
-			// Fully done with this req, remove it from the queue.
-			fd.queue.Remove(req)
-			break
+		if written != len(req.payload) {
+			return 0, syserror.EIO
 		}
+		n += int(written)
 	}
 
-	return bytesRead, nil
+	// Fully done with this req, remove it from the queue.
+	fd.queue.Remove(req)
+
+	// Remove noReply ones from map of requests expecting a reply.
+	if req.noReply {
+		fd.numActiveRequests -= 1
+		delete(fd.completions, req.hdr.Unique)
+	}
+
+	return int64(n), nil
 }
 
 // PWrite implements vfs.FileDescriptionImpl.PWrite.
 func (fd *DeviceFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) {
 	// Operations on /dev/fuse don't make sense until a FUSE filesystem is mounted.
-	if !fd.mounted {
+	if fd.fs == nil {
 		return 0, syserror.EPERM
 	}
 
@@ -211,10 +239,15 @@ func (fd *DeviceFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.
 // writeLocked implements writing to the fuse device while locked with DeviceFD.mu.
 func (fd *DeviceFD) writeLocked(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) {
 	// Operations on /dev/fuse don't make sense until a FUSE filesystem is mounted.
-	if !fd.mounted {
+	if fd.fs == nil {
 		return 0, syserror.EPERM
 	}
 
+	// Return ENODEV if the filesystem is umounted.
+	if fd.fs.umounted {
+		return 0, syserror.ENODEV
+	}
+
 	var cn, n int64
 	hdrLen := uint32((*linux.FUSEHeaderOut)(nil).SizeBytes())
 
@@ -276,7 +309,8 @@ func (fd *DeviceFD) writeLocked(ctx context.Context, src usermem.IOSequence, opt
 
 			fut, ok := fd.completions[hdr.Unique]
 			if !ok {
-				// Server sent us a response for a request we never sent?
+				// Server sent us a response for a request we never sent,
+				// or for which we already received a reply (e.g. aborted), an unlikely event.
 				return 0, syserror.EINVAL
 			}
 
@@ -307,8 +341,23 @@ func (fd *DeviceFD) writeLocked(ctx context.Context, src usermem.IOSequence, opt
 
 // Readiness implements vfs.FileDescriptionImpl.Readiness.
 func (fd *DeviceFD) Readiness(mask waiter.EventMask) waiter.EventMask {
+	fd.mu.Lock()
+	defer fd.mu.Unlock()
+	return fd.readinessLocked(mask)
+}
+
+// readinessLocked implements checking the readiness of the fuse device while
+// locked with DeviceFD.mu.
+func (fd *DeviceFD) readinessLocked(mask waiter.EventMask) waiter.EventMask {
 	var ready waiter.EventMask
-	ready |= waiter.EventOut // FD is always writable
+
+	if fd.fs.umounted {
+		ready |= waiter.EventErr
+		return ready & mask
+	}
+
+	// FD is always writable.
+	ready |= waiter.EventOut
 	if !fd.queue.Empty() {
 		// Have reqs available, FD is readable.
 		ready |= waiter.EventIn
@@ -330,7 +379,7 @@ func (fd *DeviceFD) EventUnregister(e *waiter.Entry) {
 // Seek implements vfs.FileDescriptionImpl.Seek.
 func (fd *DeviceFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) {
 	// Operations on /dev/fuse don't make sense until a FUSE filesystem is mounted.
-	if !fd.mounted {
+	if fd.fs == nil {
 		return 0, syserror.EPERM
 	}
 
@@ -338,59 +387,59 @@ func (fd *DeviceFD) Seek(ctx context.Context, offset int64, whence int32) (int64
 }
 
 // sendResponse sends a response to the waiting task (if any).
+//
+// Preconditions: fd.mu must be held.
 func (fd *DeviceFD) sendResponse(ctx context.Context, fut *futureResponse) error {
-	// See if the running task need to perform some action before returning.
-	// Since we just finished writing the future, we can be sure that
-	// getResponse generates a populated response.
-	if err := fd.noReceiverAction(ctx, fut.getResponse()); err != nil {
-		return err
-	}
+	// Signal the task waiting on a response if any.
+	defer close(fut.ch)
 
 	// Signal that the queue is no longer full.
 	select {
 	case fd.fullQueueCh <- struct{}{}:
 	default:
 	}
-	fd.numActiveRequests -= 1
+	fd.numActiveRequests--
+
+	if fut.async {
+		return fd.asyncCallBack(ctx, fut.getResponse())
+	}
 
-	// Signal the task waiting on a response.
-	close(fut.ch)
 	return nil
 }
 
-// sendError sends an error response to the waiting task (if any).
-func (fd *DeviceFD) sendError(ctx context.Context, errno int32, req *Request) error {
+// sendError sends an error response to the waiting task (if any) by calling sendResponse().
+//
+// Preconditions: fd.mu must be held.
+func (fd *DeviceFD) sendError(ctx context.Context, errno int32, unique linux.FUSEOpID) error {
 	// Return the error to the calling task.
 	outHdrLen := uint32((*linux.FUSEHeaderOut)(nil).SizeBytes())
 	respHdr := linux.FUSEHeaderOut{
 		Len:    outHdrLen,
 		Error:  errno,
-		Unique: req.hdr.Unique,
+		Unique: unique,
 	}
 
 	fut, ok := fd.completions[respHdr.Unique]
 	if !ok {
-		// Server sent us a response for a request we never sent?
+		// A response for a request we never sent,
+		// or for which we already received a reply (e.g. aborted).
 		return syserror.EINVAL
 	}
 	delete(fd.completions, respHdr.Unique)
 
 	fut.hdr = &respHdr
-	if err := fd.sendResponse(ctx, fut); err != nil {
-		return err
-	}
-
-	return nil
+	return fd.sendResponse(ctx, fut)
 }
 
-// noReceiverAction has the calling kernel.Task do some action if its known that no
-// receiver is going to be waiting on the future channel. This is to be used by:
-// FUSE_INIT.
-func (fd *DeviceFD) noReceiverAction(ctx context.Context, r *Response) error {
-	if r.opcode == linux.FUSE_INIT {
+// asyncCallBack executes pre-defined callback function for async requests.
+// Currently used by: FUSE_INIT.
+func (fd *DeviceFD) asyncCallBack(ctx context.Context, r *Response) error {
+	switch r.opcode {
+	case linux.FUSE_INIT:
 		creds := auth.CredentialsFromContext(ctx)
 		rootUserNs := kernel.KernelFromContext(ctx).RootUserNamespace()
 		return fd.fs.conn.InitRecv(r, creds.HasCapabilityIn(linux.CAP_SYS_ADMIN, rootUserNs))
+		// TODO(gvisor.dev/issue/3247): support async read: correctly process the response.
 	}
 
 	return nil

pkg/sentry/fsimpl/fuse/dev_test.go

@@ -16,12 +16,10 @@ package fuse
 
 import (
 	"fmt"
-	"io"
 	"math/rand"
 	"testing"
 
 	"gvisor.dev/gvisor/pkg/abi/linux"
-	"gvisor.dev/gvisor/pkg/marshal"
 	"gvisor.dev/gvisor/pkg/sentry/fsimpl/testutil"
 	"gvisor.dev/gvisor/pkg/sentry/kernel"
 	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
@@ -35,10 +33,6 @@ import (
 // will simply echo the payload back with the appropriate headers.
 const echoTestOpcode linux.FUSEOpcode = 1000
 
-type testPayload struct {
-	data uint32
-}
-
 // TestFUSECommunication tests that the communication layer between the Sentry and the
 // FUSE server daemon works as expected.
 func TestFUSECommunication(t *testing.T) {
@@ -327,102 +321,3 @@ func fuseServerRun(t *testing.T, s *testutil.System, k *kernel.Kernel, fd *vfs.F
 		}
 	}
 }
-
-func setup(t *testing.T) *testutil.System {
-	k, err := testutil.Boot()
-	if err != nil {
-		t.Fatalf("Error creating kernel: %v", err)
-	}
-
-	ctx := k.SupervisorContext()
-	creds := auth.CredentialsFromContext(ctx)
-
-	k.VFS().MustRegisterFilesystemType(Name, &FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
-		AllowUserList:  true,
-		AllowUserMount: true,
-	})
-
-	mntns, err := k.VFS().NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.MountOptions{})
-	if err != nil {
-		t.Fatalf("NewMountNamespace(): %v", err)
-	}
-
-	return testutil.NewSystem(ctx, t, k.VFS(), mntns)
-}
-
-// newTestConnection creates a fuse connection that the sentry can communicate with
-// and the FD for the server to communicate with.
-func newTestConnection(system *testutil.System, k *kernel.Kernel, maxActiveRequests uint64) (*connection, *vfs.FileDescription, error) {
-	vfsObj := &vfs.VirtualFilesystem{}
-	fuseDev := &DeviceFD{}
-
-	if err := vfsObj.Init(system.Ctx); err != nil {
-		return nil, nil, err
-	}
-
-	vd := vfsObj.NewAnonVirtualDentry("genCountFD")
-	defer vd.DecRef(system.Ctx)
-	if err := fuseDev.vfsfd.Init(fuseDev, linux.O_RDWR|linux.O_CREAT, vd.Mount(), vd.Dentry(), &vfs.FileDescriptionOptions{}); err != nil {
-		return nil, nil, err
-	}
-
-	fsopts := filesystemOptions{
-		maxActiveRequests: maxActiveRequests,
-	}
-	fs, err := NewFUSEFilesystem(system.Ctx, 0, &fsopts, &fuseDev.vfsfd)
-	if err != nil {
-		return nil, nil, err
-	}
-
-	return fs.conn, &fuseDev.vfsfd, nil
-}
-
-// SizeBytes implements marshal.Marshallable.SizeBytes.
-func (t *testPayload) SizeBytes() int {
-	return 4
-}
-
-// MarshalBytes implements marshal.Marshallable.MarshalBytes.
-func (t *testPayload) MarshalBytes(dst []byte) {
-	usermem.ByteOrder.PutUint32(dst[:4], t.data)
-}
-
-// UnmarshalBytes implements marshal.Marshallable.UnmarshalBytes.
-func (t *testPayload) UnmarshalBytes(src []byte) {
-	*t = testPayload{data: usermem.ByteOrder.Uint32(src[:4])}
-}
-
-// Packed implements marshal.Marshallable.Packed.
-func (t *testPayload) Packed() bool {
-	return true
-}
-
-// MarshalUnsafe implements marshal.Marshallable.MarshalUnsafe.
-func (t *testPayload) MarshalUnsafe(dst []byte) {
-	t.MarshalBytes(dst)
-}
-
-// UnmarshalUnsafe implements marshal.Marshallable.UnmarshalUnsafe.
-func (t *testPayload) UnmarshalUnsafe(src []byte) {
-	t.UnmarshalBytes(src)
-}
-
-// CopyOutN implements marshal.Marshallable.CopyOutN.
-func (t *testPayload) CopyOutN(cc marshal.CopyContext, addr usermem.Addr, limit int) (int, error) {
-	panic("not implemented")
-}
-
-// CopyOut implements marshal.Marshallable.CopyOut.
-func (t *testPayload) CopyOut(cc marshal.CopyContext, addr usermem.Addr) (int, error) {
-	panic("not implemented")
-}
-
-// CopyIn implements marshal.Marshallable.CopyIn.
-func (t *testPayload) CopyIn(cc marshal.CopyContext, addr usermem.Addr) (int, error) {
-	panic("not implemented")
-}
-
-// WriteTo implements io.WriterTo.WriteTo.
-func (t *testPayload) WriteTo(w io.Writer) (int64, error) {
-	panic("not implemented")
-}

pkg/sentry/fsimpl/fuse/directory.go

@@ -0,0 +1,105 @@
+// Copyright 2020 The gVisor Authors.
+//
+// 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.
+
+package fuse
+
+import (
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/sentry/kernel"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/syserror"
+	"gvisor.dev/gvisor/pkg/usermem"
+)
+
+type directoryFD struct {
+	fileDescription
+}
+
+// Allocate implements directoryFD.Allocate.
+func (*directoryFD) Allocate(ctx context.Context, mode, offset, length uint64) error {
+	return syserror.EISDIR
+}
+
+// PRead implements FileDescriptionImpl.PRead.
+func (*directoryFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) {
+	return 0, syserror.EISDIR
+}
+
+// Read implements FileDescriptionImpl.Read.
+func (*directoryFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) {
+	return 0, syserror.EISDIR
+}
+
+// PWrite implements FileDescriptionImpl.PWrite.
+func (*directoryFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) {
+	return 0, syserror.EISDIR
+}
+
+// Write implements FileDescriptionImpl.Write.
+func (*directoryFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) {
+	return 0, syserror.EISDIR
+}
+
+// IterDirents implements FileDescriptionImpl.IterDirents.
+func (dir *directoryFD) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback) error {
+	fusefs := dir.inode().fs
+	task, creds := kernel.TaskFromContext(ctx), auth.CredentialsFromContext(ctx)
+
+	in := linux.FUSEReadIn{
+		Fh:     dir.Fh,
+		Offset: uint64(atomic.LoadInt64(&dir.off)),
+		Size:   linux.FUSE_PAGE_SIZE,
+		Flags:  dir.statusFlags(),
+	}
+
+	// TODO(gVisor.dev/issue/3404): Support FUSE_READDIRPLUS.
+	req, err := fusefs.conn.NewRequest(creds, uint32(task.ThreadID()), dir.inode().nodeID, linux.FUSE_READDIR, &in)
+	if err != nil {
+		return err
+	}
+
+	res, err := fusefs.conn.Call(task, req)
+	if err != nil {
+		return err
+	}
+	if err := res.Error(); err != nil {
+		return err
+	}
+
+	var out linux.FUSEDirents
+	if err := res.UnmarshalPayload(&out); err != nil {
+		return err
+	}
+
+	for _, fuseDirent := range out.Dirents {
+		nextOff := int64(fuseDirent.Meta.Off)
+		dirent := vfs.Dirent{
+			Name:    fuseDirent.Name,
+			Type:    uint8(fuseDirent.Meta.Type),
+			Ino:     fuseDirent.Meta.Ino,
+			NextOff: nextOff,
+		}
+
+		if err := callback.Handle(dirent); err != nil {
+			return err
+		}
+		atomic.StoreInt64(&dir.off, nextOff)
+	}
+
+	return nil
+}

pkg/sentry/fsimpl/fuse/file.go

@@ -0,0 +1,133 @@
+// Copyright 2020 The gVisor Authors.
+//
+// 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.
+
+package fuse
+
+import (
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs"
+	"gvisor.dev/gvisor/pkg/sentry/kernel"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/usermem"
+)
+
+// fileDescription implements vfs.FileDescriptionImpl for fuse.
+type fileDescription struct {
+	vfsfd vfs.FileDescription
+	vfs.FileDescriptionDefaultImpl
+	vfs.DentryMetadataFileDescriptionImpl
+	vfs.NoLockFD
+
+	// the file handle used in userspace.
+	Fh uint64
+
+	// Nonseekable is indicate cannot perform seek on a file.
+	Nonseekable bool
+
+	// DirectIO suggest fuse to use direct io operation.
+	DirectIO bool
+
+	// OpenFlag is the flag returned by open.
+	OpenFlag uint32
+
+	// off is the file offset.
+	off int64
+}
+
+func (fd *fileDescription) dentry() *kernfs.Dentry {
+	return fd.vfsfd.Dentry().Impl().(*kernfs.Dentry)
+}
+
+func (fd *fileDescription) inode() *inode {
+	return fd.dentry().Inode().(*inode)
+}
+
+func (fd *fileDescription) filesystem() *vfs.Filesystem {
+	return fd.vfsfd.VirtualDentry().Mount().Filesystem()
+}
+
+func (fd *fileDescription) statusFlags() uint32 {
+	return fd.vfsfd.StatusFlags()
+}
+
+// Release implements vfs.FileDescriptionImpl.Release.
+func (fd *fileDescription) Release(ctx context.Context) {
+	// no need to release if FUSE server doesn't implement Open.
+	conn := fd.inode().fs.conn
+	if conn.noOpen {
+		return
+	}
+
+	in := linux.FUSEReleaseIn{
+		Fh:    fd.Fh,
+		Flags: fd.statusFlags(),
+	}
+	// TODO(gvisor.dev/issue/3245): add logic when we support file lock owner.
+	var opcode linux.FUSEOpcode
+	if fd.inode().Mode().IsDir() {
+		opcode = linux.FUSE_RELEASEDIR
+	} else {
+		opcode = linux.FUSE_RELEASE
+	}
+	kernelTask := kernel.TaskFromContext(ctx)
+	// ignoring errors and FUSE server reply is analogous to Linux's behavior.
+	req, err := conn.NewRequest(auth.CredentialsFromContext(ctx), uint32(kernelTask.ThreadID()), fd.inode().nodeID, opcode, &in)
+	if err != nil {
+		// No way to invoke Call() with an errored request.
+		return
+	}
+	// The reply will be ignored since no callback is defined in asyncCallBack().
+	conn.CallAsync(kernelTask, req)
+}
+
+// PRead implements vfs.FileDescriptionImpl.PRead.
+func (fd *fileDescription) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) {
+	return 0, nil
+}
+
+// Read implements vfs.FileDescriptionImpl.Read.
+func (fd *fileDescription) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) {
+	return 0, nil
+}
+
+// PWrite implements vfs.FileDescriptionImpl.PWrite.
+func (fd *fileDescription) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) {
+	return 0, nil
+}
+
+// Write implements vfs.FileDescriptionImpl.Write.
+func (fd *fileDescription) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) {
+	return 0, nil
+}
+
+// Seek implements vfs.FileDescriptionImpl.Seek.
+func (fd *fileDescription) Seek(ctx context.Context, offset int64, whence int32) (int64, error) {
+	return 0, nil
+}
+
+// Stat implements vfs.FileDescriptionImpl.Stat.
+func (fd *fileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) {
+	fs := fd.filesystem()
+	inode := fd.inode()
+	return inode.Stat(ctx, fs, opts)
+}
+
+// SetStat implements vfs.FileDescriptionImpl.SetStat.
+func (fd *fileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error {
+	fs := fd.filesystem()
+	creds := auth.CredentialsFromContext(ctx)
+	return fd.inode().setAttr(ctx, fs, creds, opts, true, fd.Fh)
+}

pkg/sentry/fsimpl/fuse/read_write.go

@@ -0,0 +1,242 @@
+// Copyright 2020 The gVisor Authors.
+//
+// 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.
+
+package fuse
+
+import (
+	"io"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/log"
+	"gvisor.dev/gvisor/pkg/sentry/kernel"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/syserror"
+	"gvisor.dev/gvisor/pkg/usermem"
+)
+
+// ReadInPages sends FUSE_READ requests for the size after round it up to
+// a multiple of page size, blocks on it for reply, processes the reply
+// and returns the payload (or joined payloads) as a byte slice.
+// This is used for the general purpose reading.
+// We do not support direct IO (which read the exact number of bytes)
+// at this moment.
+func (fs *filesystem) ReadInPages(ctx context.Context, fd *regularFileFD, off uint64, size uint32) ([][]byte, uint32, error) {
+	attributeVersion := atomic.LoadUint64(&fs.conn.attributeVersion)
+
+	t := kernel.TaskFromContext(ctx)
+	if t == nil {
+		log.Warningf("fusefs.Read: couldn't get kernel task from context")
+		return nil, 0, syserror.EINVAL
+	}
+
+	// Round up to a multiple of page size.
+	readSize, _ := usermem.PageRoundUp(uint64(size))
+
+	// One request cannnot exceed either maxRead or maxPages.
+	maxPages := fs.conn.maxRead >> usermem.PageShift
+	if maxPages > uint32(fs.conn.maxPages) {
+		maxPages = uint32(fs.conn.maxPages)
+	}
+
+	var outs [][]byte
+	var sizeRead uint32
+
+	// readSize is a multiple of usermem.PageSize.
+	// Always request bytes as a multiple of pages.
+	pagesRead, pagesToRead := uint32(0), uint32(readSize>>usermem.PageShift)
+
+	// Reuse the same struct for unmarshalling to avoid unnecessary memory allocation.
+	in := linux.FUSEReadIn{
+		Fh:        fd.Fh,
+		LockOwner: 0, // TODO(gvisor.dev/issue/3245): file lock
+		ReadFlags: 0, // TODO(gvisor.dev/issue/3245): |= linux.FUSE_READ_LOCKOWNER
+		Flags:     fd.statusFlags(),
+	}
+
+	// This loop is intended for fragmented read where the bytes to read is
+	// larger than either the maxPages or maxRead.
+	// For the majority of reads with normal size, this loop should only
+	// execute once.
+	for pagesRead < pagesToRead {
+		pagesCanRead := pagesToRead - pagesRead
+		if pagesCanRead > maxPages {
+			pagesCanRead = maxPages
+		}
+
+		in.Offset = off + (uint64(pagesRead) << usermem.PageShift)
+		in.Size = pagesCanRead << usermem.PageShift
+
+		req, err := fs.conn.NewRequest(auth.CredentialsFromContext(ctx), uint32(t.ThreadID()), fd.inode().nodeID, linux.FUSE_READ, &in)
+		if err != nil {
+			return nil, 0, err
+		}
+
+		// TODO(gvisor.dev/issue/3247): support async read.
+
+		res, err := fs.conn.Call(t, req)
+		if err != nil {
+			return nil, 0, err
+		}
+		if err := res.Error(); err != nil {
+			return nil, 0, err
+		}
+
+		// Not enough bytes in response,
+		// either we reached EOF,
+		// or the FUSE server sends back a response
+		// that cannot even fit the hdr.
+		if len(res.data) <= res.hdr.SizeBytes() {
+			// We treat both case as EOF here for now
+			// since there is no reliable way to detect
+			// the over-short hdr case.
+			break
+		}
+
+		// Directly using the slice to avoid extra copy.
+		out := res.data[res.hdr.SizeBytes():]
+
+		outs = append(outs, out)
+		sizeRead += uint32(len(out))
+
+		pagesRead += pagesCanRead
+	}
+
+	defer fs.ReadCallback(ctx, fd, off, size, sizeRead, attributeVersion)
+
+	// No bytes returned: offset >= EOF.
+	if len(outs) == 0 {
+		return nil, 0, io.EOF
+	}
+
+	return outs, sizeRead, nil
+}
+
+// ReadCallback updates several information after receiving a read response.
+// Due to readahead, sizeRead can be larger than size.
+func (fs *filesystem) ReadCallback(ctx context.Context, fd *regularFileFD, off uint64, size uint32, sizeRead uint32, attributeVersion uint64) {
+	// TODO(gvisor.dev/issue/3247): support async read.
+	// If this is called by an async read, correctly process it.
+	// May need to update the signature.
+
+	i := fd.inode()
+	// TODO(gvisor.dev/issue/1193): Invalidate or update atime.
+
+	// Reached EOF.
+	if sizeRead < size {
+		// TODO(gvisor.dev/issue/3630): If we have writeback cache, then we need to fill this hole.
+		// Might need to update the buf to be returned from the Read().
+
+		// Update existing size.
+		newSize := off + uint64(sizeRead)
+		fs.conn.mu.Lock()
+		if attributeVersion == i.attributeVersion && newSize < atomic.LoadUint64(&i.size) {
+			fs.conn.attributeVersion++
+			i.attributeVersion = i.fs.conn.attributeVersion
+			atomic.StoreUint64(&i.size, newSize)
+		}
+		fs.conn.mu.Unlock()
+	}
+}
+
+// Write sends FUSE_WRITE requests and return the bytes
+// written according to the response.
+//
+// Preconditions: len(data) == size.
+func (fs *filesystem) Write(ctx context.Context, fd *regularFileFD, off uint64, size uint32, data []byte) (uint32, error) {
+	t := kernel.TaskFromContext(ctx)
+	if t == nil {
+		log.Warningf("fusefs.Read: couldn't get kernel task from context")
+		return 0, syserror.EINVAL
+	}
+
+	// One request cannnot exceed either maxWrite or maxPages.
+	maxWrite := uint32(fs.conn.maxPages) << usermem.PageShift
+	if maxWrite > fs.conn.maxWrite {
+		maxWrite = fs.conn.maxWrite
+	}
+
+	// Reuse the same struct for unmarshalling to avoid unnecessary memory allocation.
+	in := linux.FUSEWriteIn{
+		Fh: fd.Fh,
+		// TODO(gvisor.dev/issue/3245): file lock
+		LockOwner: 0,
+		// TODO(gvisor.dev/issue/3245): |= linux.FUSE_READ_LOCKOWNER
+		// TODO(gvisor.dev/issue/3237): |= linux.FUSE_WRITE_CACHE (not added yet)
+		WriteFlags: 0,
+		Flags:      fd.statusFlags(),
+	}
+
+	var written uint32
+
+	// This loop is intended for fragmented write where the bytes to write is
+	// larger than either the maxWrite or maxPages or when bigWrites is false.
+	// Unless a small value for max_write is explicitly used, this loop
+	// is expected to execute only once for the majority of the writes.
+	for written < size {
+		toWrite := size - written
+
+		// Limit the write size to one page.
+		// Note that the bigWrites flag is obsolete,
+		// latest libfuse always sets it on.
+		if !fs.conn.bigWrites && toWrite > usermem.PageSize {
+			toWrite = usermem.PageSize
+		}
+
+		// Limit the write size to maxWrite.
+		if toWrite > maxWrite {
+			toWrite = maxWrite
+		}
+
+		in.Offset = off + uint64(written)
+		in.Size = toWrite
+
+		req, err := fs.conn.NewRequest(auth.CredentialsFromContext(ctx), uint32(t.ThreadID()), fd.inode().nodeID, linux.FUSE_WRITE, &in)
+		if err != nil {
+			return 0, err
+		}
+
+		req.payload = data[written : written+toWrite]
+
+		// TODO(gvisor.dev/issue/3247): support async write.
+
+		res, err := fs.conn.Call(t, req)
+		if err != nil {
+			return 0, err
+		}
+		if err := res.Error(); err != nil {
+			return 0, err
+		}
+
+		out := linux.FUSEWriteOut{}
+		if err := res.UnmarshalPayload(&out); err != nil {
+			return 0, err
+		}
+
+		// Write more than requested? EIO.
+		if out.Size > toWrite {
+			return 0, syserror.EIO
+		}
+
+		written += out.Size
+
+		// Break if short write. Not necessarily an error.
+		if out.Size != toWrite {
+			break
+		}
+	}
+
+	return written, nil
+}

pkg/sentry/fsimpl/fuse/regular_file.go

@@ -0,0 +1,230 @@
+// Copyright 2020 The gVisor Authors.
+//
+// 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.
+
+package fuse
+
+import (
+	"io"
+	"math"
+	"sync"
+	"sync/atomic"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/syserror"
+	"gvisor.dev/gvisor/pkg/usermem"
+)
+
+type regularFileFD struct {
+	fileDescription
+
+	// off is the file offset.
+	off int64
+	// offMu protects off.
+	offMu sync.Mutex
+}
+
+// PRead implements vfs.FileDescriptionImpl.PRead.
+func (fd *regularFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) {
+	if offset < 0 {
+		return 0, syserror.EINVAL
+	}
+
+	// Check that flags are supported.
+	//
+	// TODO(gvisor.dev/issue/2601): Support select preadv2 flags.
+	if opts.Flags&^linux.RWF_HIPRI != 0 {
+		return 0, syserror.EOPNOTSUPP
+	}
+
+	size := dst.NumBytes()
+	if size == 0 {
+		// Early return if count is 0.
+		return 0, nil
+	} else if size > math.MaxUint32 {
+		// FUSE only supports uint32 for size.
+		// Overflow.
+		return 0, syserror.EINVAL
+	}
+
+	// TODO(gvisor.dev/issue/3678): Add direct IO support.
+
+	inode := fd.inode()
+
+	// Reading beyond EOF, update file size if outdated.
+	if uint64(offset+size) > atomic.LoadUint64(&inode.size) {
+		if err := inode.reviseAttr(ctx, linux.FUSE_GETATTR_FH, fd.Fh); err != nil {
+			return 0, err
+		}
+		// If the offset after update is still too large, return error.
+		if uint64(offset) >= atomic.LoadUint64(&inode.size) {
+			return 0, io.EOF
+		}
+	}
+
+	// Truncate the read with updated file size.
+	fileSize := atomic.LoadUint64(&inode.size)
+	if uint64(offset+size) > fileSize {
+		size = int64(fileSize) - offset
+	}
+
+	buffers, n, err := inode.fs.ReadInPages(ctx, fd, uint64(offset), uint32(size))
+	if err != nil {
+		return 0, err
+	}
+
+	// TODO(gvisor.dev/issue/3237): support indirect IO (e.g. caching),
+	// store the bytes that were read ahead.
+
+	// Update the number of bytes to copy for short read.
+	if n < uint32(size) {
+		size = int64(n)
+	}
+
+	// Copy the bytes read to the dst.
+	// This loop is intended for fragmented reads.
+	// For the majority of reads, this loop only execute once.
+	var copied int64
+	for _, buffer := range buffers {
+		toCopy := int64(len(buffer))
+		if copied+toCopy > size {
+			toCopy = size - copied
+		}
+		cp, err := dst.DropFirst64(copied).CopyOut(ctx, buffer[:toCopy])
+		if err != nil {
+			return 0, err
+		}
+		if int64(cp) != toCopy {
+			return 0, syserror.EIO
+		}
+		copied += toCopy
+	}
+
+	return copied, nil
+}
+
+// Read implements vfs.FileDescriptionImpl.Read.
+func (fd *regularFileFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) {
+	fd.offMu.Lock()
+	n, err := fd.PRead(ctx, dst, fd.off, opts)
+	fd.off += n
+	fd.offMu.Unlock()
+	return n, err
+}
+
+// PWrite implements vfs.FileDescriptionImpl.PWrite.
+func (fd *regularFileFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) {
+	n, _, err := fd.pwrite(ctx, src, offset, opts)
+	return n, err
+}
+
+// Write implements vfs.FileDescriptionImpl.Write.
+func (fd *regularFileFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) {
+	fd.offMu.Lock()
+	n, off, err := fd.pwrite(ctx, src, fd.off, opts)
+	fd.off = off
+	fd.offMu.Unlock()
+	return n, err
+}
+
+// pwrite returns the number of bytes written, final offset and error. The
+// final offset should be ignored by PWrite.
+func (fd *regularFileFD) pwrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (written, finalOff int64, err error) {
+	if offset < 0 {
+		return 0, offset, syserror.EINVAL
+	}
+
+	// Check that flags are supported.
+	//
+	// TODO(gvisor.dev/issue/2601): Support select preadv2 flags.
+	if opts.Flags&^linux.RWF_HIPRI != 0 {
+		return 0, offset, syserror.EOPNOTSUPP
+	}
+
+	inode := fd.inode()
+	inode.metadataMu.Lock()
+	defer inode.metadataMu.Unlock()
+
+	// If the file is opened with O_APPEND, update offset to file size.
+	// Note: since our Open() implements the interface of kernfs,
+	// and kernfs currently does not support O_APPEND, this will never
+	// be true before we switch out from kernfs.
+	if fd.vfsfd.StatusFlags()&linux.O_APPEND != 0 {
+		// Locking inode.metadataMu is sufficient for reading size
+		offset = int64(inode.size)
+	}
+
+	srclen := src.NumBytes()
+
+	if srclen > math.MaxUint32 {
+		// FUSE only supports uint32 for size.
+		// Overflow.
+		return 0, offset, syserror.EINVAL
+	}
+	if end := offset + srclen; end < offset {
+		// Overflow.
+		return 0, offset, syserror.EINVAL
+	}
+
+	srclen, err = vfs.CheckLimit(ctx, offset, srclen)
+	if err != nil {
+		return 0, offset, err
+	}
+
+	if srclen == 0 {
+		// Return before causing any side effects.
+		return 0, offset, nil
+	}
+
+	src = src.TakeFirst64(srclen)
+
+	// TODO(gvisor.dev/issue/3237): Add cache support:
+	// buffer cache. Ideally we write from src to our buffer cache first.
+	// The slice passed to fs.Write() should be a slice from buffer cache.
+	data := make([]byte, srclen)
+	// Reason for making a copy here: connection.Call() blocks on kerneltask,
+	// which in turn acquires mm.activeMu lock. Functions like CopyInTo() will
+	// attemp to acquire the mm.activeMu lock as well -> deadlock.
+	// We must finish reading from the userspace memory before
+	// t.Block() deactivates it.
+	cp, err := src.CopyIn(ctx, data)
+	if err != nil {
+		return 0, offset, err
+	}
+	if int64(cp) != srclen {
+		return 0, offset, syserror.EIO
+	}
+
+	n, err := fd.inode().fs.Write(ctx, fd, uint64(offset), uint32(srclen), data)
+	if err != nil {
+		return 0, offset, err
+	}
+
+	if n == 0 {
+		// We have checked srclen != 0 previously.
+		// If err == nil, then it's a short write and we return EIO.
+		return 0, offset, syserror.EIO
+	}
+
+	written = int64(n)
+	finalOff = offset + written
+
+	if finalOff > int64(inode.size) {
+		atomic.StoreUint64(&inode.size, uint64(finalOff))
+		atomic.AddUint64(&inode.fs.conn.attributeVersion, 1)
+	}
+
+	return
+}

pkg/sentry/fsimpl/fuse/request_response.go

@@ -0,0 +1,229 @@
+// Copyright 2020 The gVisor Authors.
+//
+// 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.
+
+package fuse
+
+import (
+	"fmt"
+	"syscall"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/marshal"
+	"gvisor.dev/gvisor/pkg/sentry/kernel"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/usermem"
+)
+
+// fuseInitRes is a variable-length wrapper of linux.FUSEInitOut. The FUSE
+// server may implement an older version of FUSE protocol, which contains a
+// linux.FUSEInitOut with less attributes.
+//
+// Dynamically-sized objects cannot be marshalled.
+type fuseInitRes struct {
+	marshal.StubMarshallable
+
+	// initOut contains the response from the FUSE server.
+	initOut linux.FUSEInitOut
+
+	// initLen is the total length of bytes of the response.
+	initLen uint32
+}
+
+// UnmarshalBytes deserializes src to the initOut attribute in a fuseInitRes.
+func (r *fuseInitRes) UnmarshalBytes(src []byte) {
+	out := &r.initOut
+
+	// Introduced before FUSE kernel version 7.13.
+	out.Major = uint32(usermem.ByteOrder.Uint32(src[:4]))
+	src = src[4:]
+	out.Minor = uint32(usermem.ByteOrder.Uint32(src[:4]))
+	src = src[4:]
+	out.MaxReadahead = uint32(usermem.ByteOrder.Uint32(src[:4]))
+	src = src[4:]
+	out.Flags = uint32(usermem.ByteOrder.Uint32(src[:4]))
+	src = src[4:]
+	out.MaxBackground = uint16(usermem.ByteOrder.Uint16(src[:2]))
+	src = src[2:]
+	out.CongestionThreshold = uint16(usermem.ByteOrder.Uint16(src[:2]))
+	src = src[2:]
+	out.MaxWrite = uint32(usermem.ByteOrder.Uint32(src[:4]))
+	src = src[4:]
+
+	// Introduced in FUSE kernel version 7.23.
+	if len(src) >= 4 {
+		out.TimeGran = uint32(usermem.ByteOrder.Uint32(src[:4]))
+		src = src[4:]
+	}
+	// Introduced in FUSE kernel version 7.28.
+	if len(src) >= 2 {
+		out.MaxPages = uint16(usermem.ByteOrder.Uint16(src[:2]))
+		src = src[2:]
+	}
+}
+
+// SizeBytes is the size of the payload of the FUSE_INIT response.
+func (r *fuseInitRes) SizeBytes() int {
+	return int(r.initLen)
+}
+
+// Ordinary requests have even IDs, while interrupts IDs are odd.
+// Used to increment the unique ID for each FUSE request.
+var reqIDStep uint64 = 2
+
+// Request represents a FUSE operation request that hasn't been sent to the
+// server yet.
+//
+// +stateify savable
+type Request struct {
+	requestEntry
+
+	id   linux.FUSEOpID
+	hdr  *linux.FUSEHeaderIn
+	data []byte
+
+	// payload for this request: extra bytes to write after
+	// the data slice. Used by FUSE_WRITE.
+	payload []byte
+
+	// If this request is async.
+	async bool
+	// If we don't care its response.
+	// Manually set by the caller.
+	noReply bool
+}
+
+// NewRequest creates a new request that can be sent to the FUSE server.
+func (conn *connection) NewRequest(creds *auth.Credentials, pid uint32, ino uint64, opcode linux.FUSEOpcode, payload marshal.Marshallable) (*Request, error) {
+	conn.fd.mu.Lock()
+	defer conn.fd.mu.Unlock()
+	conn.fd.nextOpID += linux.FUSEOpID(reqIDStep)
+
+	hdrLen := (*linux.FUSEHeaderIn)(nil).SizeBytes()
+	hdr := linux.FUSEHeaderIn{
+		Len:    uint32(hdrLen + payload.SizeBytes()),
+		Opcode: opcode,
+		Unique: conn.fd.nextOpID,
+		NodeID: ino,
+		UID:    uint32(creds.EffectiveKUID),
+		GID:    uint32(creds.EffectiveKGID),
+		PID:    pid,
+	}
+
+	buf := make([]byte, hdr.Len)
+
+	// TODO(gVisor.dev/issue/3698): Use the unsafe version once go_marshal is safe to use again.
+	hdr.MarshalBytes(buf[:hdrLen])
+	payload.MarshalBytes(buf[hdrLen:])
+
+	return &Request{
+		id:   hdr.Unique,
+		hdr:  &hdr,
+		data: buf,
+	}, nil
+}
+
+// futureResponse represents an in-flight request, that may or may not have
+// completed yet. Convert it to a resolved Response by calling Resolve, but note
+// that this may block.
+//
+// +stateify savable
+type futureResponse struct {
+	opcode linux.FUSEOpcode
+	ch     chan struct{}
+	hdr    *linux.FUSEHeaderOut
+	data   []byte
+
+	// If this request is async.
+	async bool
+}
+
+// newFutureResponse creates a future response to a FUSE request.
+func newFutureResponse(req *Request) *futureResponse {
+	return &futureResponse{
+		opcode: req.hdr.Opcode,
+		ch:     make(chan struct{}),
+		async:  req.async,
+	}
+}
+
+// resolve blocks the task until the server responds to its corresponding request,
+// then returns a resolved response.
+func (f *futureResponse) resolve(t *kernel.Task) (*Response, error) {
+	// Return directly for async requests.
+	if f.async {
+		return nil, nil
+	}
+
+	if err := t.Block(f.ch); err != nil {
+		return nil, err
+	}
+
+	return f.getResponse(), nil
+}
+
+// getResponse creates a Response from the data the futureResponse has.
+func (f *futureResponse) getResponse() *Response {
+	return &Response{
+		opcode: f.opcode,
+		hdr:    *f.hdr,
+		data:   f.data,
+	}
+}
+
+// Response represents an actual response from the server, including the
+// response payload.
+//
+// +stateify savable
+type Response struct {
+	opcode linux.FUSEOpcode
+	hdr    linux.FUSEHeaderOut
+	data   []byte
+}
+
+// Error returns the error of the FUSE call.
+func (r *Response) Error() error {
+	errno := r.hdr.Error
+	if errno >= 0 {
+		return nil
+	}
+
+	sysErrNo := syscall.Errno(-errno)
+	return error(sysErrNo)
+}
+
+// DataLen returns the size of the response without the header.
+func (r *Response) DataLen() uint32 {
+	return r.hdr.Len - uint32(r.hdr.SizeBytes())
+}
+
+// UnmarshalPayload unmarshals the response data into m.
+func (r *Response) UnmarshalPayload(m marshal.Marshallable) error {
+	hdrLen := r.hdr.SizeBytes()
+	haveDataLen := r.hdr.Len - uint32(hdrLen)
+	wantDataLen := uint32(m.SizeBytes())
+
+	if haveDataLen < wantDataLen {
+		return fmt.Errorf("payload too small. Minimum data lenth required: %d,  but got data length %d", wantDataLen, haveDataLen)
+	}
+
+	// The response data is empty unless there is some payload. And so, doesn't
+	// need to be unmarshalled.
+	if r.data == nil {
+		return nil
+	}
+
+	// TODO(gVisor.dev/issue/3698): Use the unsafe version once go_marshal is safe to use again.
+	m.UnmarshalBytes(r.data[hdrLen:])
+	return nil
+}

pkg/sentry/fsimpl/fuse/utils_test.go

@@ -0,0 +1,132 @@
+// Copyright 2020 The gVisor Authors.
+//
+// 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.
+
+package fuse
+
+import (
+	"io"
+	"testing"
+
+	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/marshal"
+	"gvisor.dev/gvisor/pkg/sentry/fsimpl/testutil"
+	"gvisor.dev/gvisor/pkg/sentry/kernel"
+	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.dev/gvisor/pkg/sentry/vfs"
+	"gvisor.dev/gvisor/pkg/usermem"
+)
+
+func setup(t *testing.T) *testutil.System {
+	k, err := testutil.Boot()
+	if err != nil {
+		t.Fatalf("Error creating kernel: %v", err)
+	}
+
+	ctx := k.SupervisorContext()
+	creds := auth.CredentialsFromContext(ctx)
+
+	k.VFS().MustRegisterFilesystemType(Name, &FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{
+		AllowUserList:  true,
+		AllowUserMount: true,
+	})
+
+	mntns, err := k.VFS().NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.MountOptions{})
+	if err != nil {
+		t.Fatalf("NewMountNamespace(): %v", err)
+	}
+
+	return testutil.NewSystem(ctx, t, k.VFS(), mntns)
+}
+
+// newTestConnection creates a fuse connection that the sentry can communicate with
+// and the FD for the server to communicate with.
+func newTestConnection(system *testutil.System, k *kernel.Kernel, maxActiveRequests uint64) (*connection, *vfs.FileDescription, error) {
+	vfsObj := &vfs.VirtualFilesystem{}
+	fuseDev := &DeviceFD{}
+
+	if err := vfsObj.Init(system.Ctx); err != nil {
+		return nil, nil, err
+	}
+
+	vd := vfsObj.NewAnonVirtualDentry("genCountFD")
+	defer vd.DecRef(system.Ctx)
+	if err := fuseDev.vfsfd.Init(fuseDev, linux.O_RDWR|linux.O_CREAT, vd.Mount(), vd.Dentry(), &vfs.FileDescriptionOptions{}); err != nil {
+		return nil, nil, err
+	}
+
+	fsopts := filesystemOptions{
+		maxActiveRequests: maxActiveRequests,
+	}
+	fs, err := newFUSEFilesystem(system.Ctx, 0, &fsopts, &fuseDev.vfsfd)
+	if err != nil {
+		return nil, nil, err
+	}
+
+	return fs.conn, &fuseDev.vfsfd, nil
+}
+
+type testPayload struct {
+	marshal.StubMarshallable
+	data uint32
+}
+
+// SizeBytes implements marshal.Marshallable.SizeBytes.
+func (t *testPayload) SizeBytes() int {
+	return 4
+}
+
+// MarshalBytes implements marshal.Marshallable.MarshalBytes.
+func (t *testPayload) MarshalBytes(dst []byte) {
+	usermem.ByteOrder.PutUint32(dst[:4], t.data)
+}
+
+// UnmarshalBytes implements marshal.Marshallable.UnmarshalBytes.
+func (t *testPayload) UnmarshalBytes(src []byte) {
+	*t = testPayload{data: usermem.ByteOrder.Uint32(src[:4])}
+}
+
+// Packed implements marshal.Marshallable.Packed.
+func (t *testPayload) Packed() bool {
+	return true
+}
+
+// MarshalUnsafe implements marshal.Marshallable.MarshalUnsafe.
+func (t *testPayload) MarshalUnsafe(dst []byte) {
+	t.MarshalBytes(dst)
+}
+
+// UnmarshalUnsafe implements marshal.Marshallable.UnmarshalUnsafe.
+func (t *testPayload) UnmarshalUnsafe(src []byte) {
+	t.UnmarshalBytes(src)
+}
+
+// CopyOutN implements marshal.Marshallable.CopyOutN.
+func (t *testPayload) CopyOutN(task marshal.CopyContext, addr usermem.Addr, limit int) (int, error) {
+	panic("not implemented")
+}
+
+// CopyOut implements marshal.Marshallable.CopyOut.
+func (t *testPayload) CopyOut(task marshal.CopyContext, addr usermem.Addr) (int, error) {
+	panic("not implemented")
+}
+
+// CopyIn implements marshal.Marshallable.CopyIn.
+func (t *testPayload) CopyIn(task marshal.CopyContext, addr usermem.Addr) (int, error) {
+	panic("not implemented")
+}
+
+// WriteTo implements io.WriterTo.WriteTo.
+func (t *testPayload) WriteTo(w io.Writer) (int64, error) {
+	panic("not implemented")
+}

pkg/sentry/fsimpl/kernfs/filesystem.go

@@ -140,7 +140,7 @@ func (fs *Filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.Vir
 		}
 		// Reference on childVFSD dropped by a corresponding Valid.
 		child = childVFSD.Impl().(*Dentry)
-		parent.insertChildLocked(name, child)
+		parent.InsertChildLocked(name, child)
 	}
 	return child, nil
 }
@@ -548,7 +548,7 @@ func (fs *Filesystem) ReadlinkAt(ctx context.Context, rp *vfs.ResolvingPath) (st
 	if !d.Impl().(*Dentry).isSymlink() {
 		return "", syserror.EINVAL
 	}
-	return inode.Readlink(ctx)
+	return inode.Readlink(ctx, rp.Mount())
 }
 
 // RenameAt implements vfs.FilesystemImpl.RenameAt.
@@ -657,6 +657,10 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa
 func (fs *Filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error {
 	fs.mu.Lock()
 	defer fs.mu.Unlock()
+
+	// Store the name before walkExistingLocked as rp will be advanced past the
+	// name in the following call.
+	name := rp.Component()
 	vfsd, inode, err := fs.walkExistingLocked(ctx, rp)
 	fs.processDeferredDecRefsLocked(ctx)
 	if err != nil {
@@ -686,7 +690,8 @@ func (fs *Filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error
 	if err := virtfs.PrepareDeleteDentry(mntns, vfsd); err != nil {
 		return err
 	}
-	if err := parentDentry.inode.RmDir(ctx, rp.Component(), vfsd); err != nil {
+
+	if err := parentDentry.inode.RmDir(ctx, name, vfsd); err != nil {
 		virtfs.AbortDeleteDentry(vfsd)
 		return err
 	}
@@ -765,6 +770,10 @@ func (fs *Filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, targ
 func (fs *Filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error {
 	fs.mu.Lock()
 	defer fs.mu.Unlock()
+
+	// Store the name before walkExistingLocked as rp will be advanced past the
+	// name in the following call.
+	name := rp.Component()
 	vfsd, _, err := fs.walkExistingLocked(ctx, rp)
 	fs.processDeferredDecRefsLocked(ctx)
 	if err != nil {
@@ -790,7 +799,7 @@ func (fs *Filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error
 	if err := virtfs.PrepareDeleteDentry(mntns, vfsd); err != nil {
 		return err
 	}
-	if err := parentDentry.inode.Unlink(ctx, rp.Component(), vfsd); err != nil {
+	if err := parentDentry.inode.Unlink(ctx, name, vfsd); err != nil {
 		virtfs.AbortDeleteDentry(vfsd)
 		return err
 	}

pkg/sentry/fsimpl/kernfs/inode_impl_util.go

@@ -172,7 +172,7 @@ func (InodeNoDynamicLookup) Valid(ctx context.Context) bool {
 type InodeNotSymlink struct{}
 
 // Readlink implements Inode.Readlink.
-func (InodeNotSymlink) Readlink(context.Context) (string, error) {
+func (InodeNotSymlink) Readlink(context.Context, *vfs.Mount) (string, error) {
 	return "", syserror.EINVAL
 }
 
@@ -256,6 +256,13 @@ func (a *InodeAttrs) Stat(context.Context, *vfs.Filesystem, vfs.StatOptions) (li
 
 // SetStat implements Inode.SetStat.
 func (a *InodeAttrs) SetStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error {
+	return a.SetInodeStat(ctx, fs, creds, opts)
+}
+
+// SetInodeStat sets the corresponding attributes from opts to InodeAttrs.
+// This function can be used by other kernfs-based filesystem implementation to
+// sets the unexported attributes into kernfs.InodeAttrs.
+func (a *InodeAttrs) SetInodeStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error {
 	if opts.Stat.Mask == 0 {
 		return nil
 	}

pkg/sentry/fsimpl/kernfs/kernfs.go

@@ -60,6 +60,7 @@ import (
 	"gvisor.dev/gvisor/pkg/sentry/kernel/auth"
 	"gvisor.dev/gvisor/pkg/sentry/vfs"
 	"gvisor.dev/gvisor/pkg/sync"
+	"gvisor.dev/gvisor/pkg/syserror"
 )
 
 // Filesystem mostly implements vfs.FilesystemImpl for a generic in-memory
@@ -246,15 +247,15 @@ func (d *Dentry) OnZeroWatches(context.Context) {}
 // Precondition: d must represent a directory inode.
 func (d *Dentry) InsertChild(name string, child *Dentry) {
 	d.dirMu.Lock()
-	d.insertChildLocked(name, child)
+	d.InsertChildLocked(name, child)
 	d.dirMu.Unlock()
 }
 
-// insertChildLocked is equivalent to InsertChild, with additional
+// InsertChildLocked is equivalent to InsertChild, with additional
 // preconditions.
 //
 // Precondition: d.dirMu must be locked.
-func (d *Dentry) insertChildLocked(name string, child *Dentry) {
+func (d *Dentry) InsertChildLocked(name string, child *Dentry) {
 	if !d.isDir() {
 		panic(fmt.Sprintf("InsertChild called on non-directory Dentry: %+v.", d))
 	}
@@ -267,6 +268,36 @@ func (d *Dentry) insertChildLocked(name string, child *Dentry) {
 	d.children[name] = child
 }
 
+// RemoveChild removes child from the vfs dentry cache. This does not update the
+// directory inode or modify the inode to be unlinked. So calling this on its own
+// isn't sufficient to remove a child from a directory.
+//
+// Precondition: d must represent a directory inode.
+func (d *Dentry) RemoveChild(name string, child *vfs.Dentry) error {
+	d.dirMu.Lock()
+	defer d.dirMu.Unlock()
+	return d.RemoveChildLocked(name, child)
+}
+
+// RemoveChildLocked is equivalent to RemoveChild, with additional
+// preconditions.
+//
+// Precondition: d.dirMu must be locked.
+func (d *Dentry) RemoveChildLocked(name string, child *vfs.Dentry) error {
+	if !d.isDir() {
+		panic(fmt.Sprintf("RemoveChild called on non-directory Dentry: %+v.", d))
+	}
+	c, ok := d.children[name]
+	if !ok {
+		return syserror.ENOENT
+	}
+	if &c.vfsd != child {
+		panic(fmt.Sprintf("Dentry hashed into inode doesn't match what vfs thinks! Child: %+v, vfs: %+v", c, child))
+	}
+	delete(d.children, name)
+	return nil
+}
+
 // Inode returns the dentry's inode.
 func (d *Dentry) Inode() Inode {
 	return d.inode
@@ -425,7 +456,7 @@ type inodeDynamicLookup interface {
 	Valid(ctx context.Context) bool
 
 	// IterDirents is used to iterate over dynamically created entries. It invokes
-	// cb on each entry in the directory represented by the FileDescription.
+	// cb on each entry in the directory represented by the Inode.
 	// 'offset' is the offset for the entire IterDirents call, which may include
 	// results from the caller (e.g. "." and ".."). 'relOffset' is the offset
 	// inside the entries returned by this IterDirents invocation. In other words,
@@ -437,7 +468,7 @@ type inodeDynamicLookup interface {
 type inodeSymlink interface {
 	// Readlink returns the target of a symbolic link. If an inode is not a
 	// symlink, the implementation should return EINVAL.
-	Readlink(ctx context.Context) (string, error)
+	Readlink(ctx context.Context, mnt *vfs.Mount) (string, error)
 
 	// Getlink returns the target of a symbolic link, as used by path
 	// resolution:

pkg/sentry/fsimpl/kernfs/symlink.go

@@ -52,7 +52,7 @@ func (s *StaticSymlink) Init(creds *auth.Credentials, devMajor uint32, devMinor
 }
 
 // Readlink implements Inode.
-func (s *StaticSymlink) Readlink(_ context.Context) (string, error) {
+func (s *StaticSymlink) Readlink(_ context.Context, _ *vfs.Mount) (string, error) {
 	return s.target, nil
 }