Implement FUSE_READ

Fixes #3206

8 files changed, 433 insertions, 27 deletions

diff --git a/pkg/abi/linux/fuse.go b/pkg/abi/linux/fuse.go
index 4ef0ab9a7..0ece7b756 100644
--- a/pkg/abi/linux/fuse.go
+++ b/pkg/abi/linux/fuse.go
@@ -184,6 +184,13 @@ const (
 	FUSE_KERNEL_MINOR_VERSION = 31
 )
 
+// Constants relevant to FUSE operations.
+const (
+	FUSE_NAME_MAX     = 1024
+	FUSE_PAGE_SIZE    = 4096
+	FUSE_DIRENT_ALIGN = 8
+)
+
 // FUSEInitIn is the request sent by the kernel to the daemon,
 // to negotiate the version and flags.
 //
@@ -392,6 +399,36 @@ type FUSEOpenOut struct {
 
 	// OpenFlag for the opened file.
 	OpenFlag uint32
+}
+
+// FUSE_READ flags, consistent with the ones in include/uapi/linux/fuse.h.
+const (
+	FUSE_READ_LOCKOWNER = 1 << 1
+)
+
+// FUSEReadIn is the request sent by the kernel to the daemon
+// for FUSE_READ.
+//
+// +marshal
+type FUSEReadIn struct {
+	// Fh is the file handle in userspace.
+	Fh uint64
+
+	// Offset is the read offset.
+	Offset uint64
+
+	// Size is the number of bytes to read.
+	Size uint32
+
+	// ReadFlags for this FUSE_READ request.
+	// Currently only contains FUSE_READ_LOCKOWNER.
+	ReadFlags uint32
+
+	// LockOwner is the id of the lock owner if there is one.
+	LockOwner uint64
+
+	// Flags for the underlying file.
+	Flags uint32
 
 	_ uint32
 }
diff --git a/pkg/sentry/fsimpl/fuse/BUILD b/pkg/sentry/fsimpl/fuse/BUILD
index 2d9350d57..a6ee6100d 100644
--- a/pkg/sentry/fsimpl/fuse/BUILD
+++ b/pkg/sentry/fsimpl/fuse/BUILD
@@ -36,7 +36,9 @@ go_library(
         "fusefs.go",
         "init.go",
         "inode_refs.go",
+        "read_write.go",
         "register.go",
+        "regular_file.go",
         "request_list.go",
     ],
     visibility = ["//pkg/sentry:internal"],
@@ -45,6 +47,7 @@ go_library(
         "//pkg/context",
         "//pkg/log",
         "//pkg/refs",
+        "//pkg/safemem",
         "//pkg/sentry/fsimpl/devtmpfs",
         "//pkg/sentry/fsimpl/kernfs",
         "//pkg/sentry/kernel",
diff --git a/pkg/sentry/fsimpl/fuse/connection.go b/pkg/sentry/fsimpl/fuse/connection.go
index 7d3c30116..a6525249d 100644
--- a/pkg/sentry/fsimpl/fuse/connection.go
+++ b/pkg/sentry/fsimpl/fuse/connection.go
@@ -161,6 +161,7 @@ type connection struct {
 	bgLock sync.Mutex
 
 	// 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.
@@ -206,7 +207,7 @@ type connection struct {
 }
 
 // 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)
@@ -216,13 +217,14 @@ func newFUSEConnection(_ context.Context, fd *vfs.FileDescription, maxInFlightRe
 	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,
+		maxRead:             opts.maxRead,
 		maxPages:            fuseDefaultMaxPagesPerReq,
 		initializedChan:     make(chan struct{}),
 		connected:           true,
diff --git a/pkg/sentry/fsimpl/fuse/dev.go b/pkg/sentry/fsimpl/fuse/dev.go
index e2de8e097..fd3592e32 100644
--- a/pkg/sentry/fsimpl/fuse/dev.go
+++ b/pkg/sentry/fsimpl/fuse/dev.go
@@ -401,10 +401,12 @@ func (fd *DeviceFD) sendError(ctx context.Context, errno int32, req *Request) er
 // 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 {
+	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 using information from r.options.
 	}
 
 	return nil
diff --git a/pkg/sentry/fsimpl/fuse/fusefs.go b/pkg/sentry/fsimpl/fuse/fusefs.go
index 4dc8ef993..65e22ba4d 100644
--- a/pkg/sentry/fsimpl/fuse/fusefs.go
+++ b/pkg/sentry/fsimpl/fuse/fusefs.go
@@ -16,6 +16,7 @@
 package fuse
 
 import (
+	"math"
 	"strconv"
 	"sync/atomic"
 
@@ -58,6 +59,11 @@ type filesystemOptions struct {
 	// exist at any time. Any further requests will block when trying to
 	// Call the server.
 	maxActiveRequests uint64
+
+	// maxRead is the max number of bytes to read,
+	// specified as "max_read" in fs parameters.
+	// If not specified by user, use math.MaxUint32 as default value.
+	maxRead uint32
 }
 
 // filesystem implements vfs.FilesystemImpl.
@@ -144,6 +150,21 @@ func (fsType FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt
 	// Set the maxInFlightRequests option.
 	fsopts.maxActiveRequests = maxActiveRequestsDefault
 
+	if maxReadStr, ok := mopts["max_read"]; ok {
+		delete(mopts, "max_read")
+		maxRead, err := strconv.ParseUint(maxReadStr, 10, 32)
+		if err != nil {
+			log.Warningf("%s.GetFilesystem: invalid max_read: max_read=%s", fsType.Name(), maxReadStr)
+			return nil, nil, syserror.EINVAL
+		}
+		if maxRead < fuseMinMaxRead {
+			maxRead = fuseMinMaxRead
+		}
+		fsopts.maxRead = uint32(maxRead)
+	} else {
+		fsopts.maxRead = math.MaxUint32
+	}
+
 	// Check for unparsed options.
 	if len(mopts) != 0 {
 		log.Warningf("%s.GetFilesystem: unknown options: %v", fsType.Name(), mopts)
@@ -179,7 +200,7 @@ func NewFUSEFilesystem(ctx context.Context, devMinor uint32, opts *filesystemOpt
 		opts:     opts,
 	}
 
-	conn, err := newFUSEConnection(ctx, device, opts.maxActiveRequests)
+	conn, err := newFUSEConnection(ctx, device, opts)
 	if err != nil {
 		log.Warningf("fuse.NewFUSEFilesystem: NewFUSEConnection failed with error: %v", err)
 		return nil, syserror.EINVAL
@@ -244,6 +265,7 @@ func (fs *filesystem) newInode(nodeID uint64, attr linux.FUSEAttr) *kernfs.Dentr
 	i := &inode{fs: fs, NodeID: nodeID}
 	creds := auth.Credentials{EffectiveKGID: auth.KGID(attr.UID), EffectiveKUID: auth.KUID(attr.UID)}
 	i.InodeAttrs.Init(&creds, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.FileMode(attr.Mode))
+	atomic.StoreUint64(&i.size, attr.Size)
 	i.OrderedChildren.Init(kernfs.OrderedChildrenOptions{})
 	i.EnableLeakCheck()
 	i.dentry.Init(i)
@@ -269,10 +291,13 @@ func (i *inode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentr
 		fd = &(directoryFD.fileDescription)
 		fdImpl = directoryFD
 	} else {
-		// FOPEN_KEEP_CACHE is the defualt flag for noOpen.
-		fd = &fileDescription{OpenFlag: linux.FOPEN_KEEP_CACHE}
-		fdImpl = fd
+		regularFD := &regularFileFD{}
+		fd = &(regularFD.fileDescription)
+		fdImpl = regularFD
 	}
+	// FOPEN_KEEP_CACHE is the defualt flag for noOpen.
+	fd.OpenFlag = linux.FOPEN_KEEP_CACHE
+
 	// Only send open request when FUSE server support open or is opening a directory.
 	if !i.fs.conn.noOpen || isDir {
 		kernelTask := kernel.TaskFromContext(ctx)
@@ -281,21 +306,25 @@ func (i *inode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentr
 			return nil, syserror.EINVAL
 		}
 
+		// Build the request.
 		var opcode linux.FUSEOpcode
 		if isDir {
 			opcode = linux.FUSE_OPENDIR
 		} else {
 			opcode = linux.FUSE_OPEN
 		}
+
 		in := linux.FUSEOpenIn{Flags: opts.Flags & ^uint32(linux.O_CREAT|linux.O_EXCL|linux.O_NOCTTY)}
 		if !i.fs.conn.atomicOTrunc {
 			in.Flags &= ^uint32(linux.O_TRUNC)
 		}
+
 		req, err := i.fs.conn.NewRequest(auth.CredentialsFromContext(ctx), uint32(kernelTask.ThreadID()), i.NodeID, opcode, &in)
 		if err != nil {
 			return nil, err
 		}
 
+		// Send the request and receive the reply.
 		res, err := i.fs.conn.Call(kernelTask, req)
 		if err != nil {
 			return nil, err
@@ -309,15 +338,17 @@ func (i *inode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentr
 			if err := res.UnmarshalPayload(&out); err != nil {
 				return nil, err
 			}
+
+			// Process the reply.
 			fd.OpenFlag = out.OpenFlag
+			if isDir {
+				fd.OpenFlag &= ^uint32(linux.FOPEN_DIRECT_IO)
+			}
+
 			fd.Fh = out.Fh
 		}
 	}
 
-	if isDir {
-		fd.OpenFlag &= ^uint32(linux.FOPEN_DIRECT_IO)
-	}
-
 	// TODO(gvisor.dev/issue/3234): invalidate mmap after implemented it for FUSE Inode
 	fd.DirectIO = fd.OpenFlag&linux.FOPEN_DIRECT_IO != 0
 	fdOptions := &vfs.FileDescriptionOptions{}
@@ -457,6 +488,16 @@ func (i *inode) Readlink(ctx context.Context, mnt *vfs.Mount) (string, error) {
 	return i.link, nil
 }
 
+// getFUSEAttr returns a linux.FUSEAttr of this inode stored in local cache.
+// TODO(gvisor.dev/issue/3679): Add support for other fields.
+func (i *inode) getFUSEAttr() linux.FUSEAttr {
+	return linux.FUSEAttr{
+		Ino:  i.Ino(),
+		Size: atomic.LoadUint64(&i.size),
+		Mode: uint32(i.Mode()),
+	}
+}
+
 // statFromFUSEAttr makes attributes from linux.FUSEAttr to linux.Statx. The
 // opts.Sync attribute is ignored since the synchronization is handled by the
 // FUSE server.
@@ -510,47 +551,90 @@ func statFromFUSEAttr(attr linux.FUSEAttr, mask, devMinor uint32) linux.Statx {
 	return stat
 }
 
-// Stat implements kernfs.Inode.Stat.
-func (i *inode) Stat(ctx context.Context, fs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) {
-	fusefs := fs.Impl().(*filesystem)
-	conn := fusefs.conn
-	task, creds := kernel.TaskFromContext(ctx), auth.CredentialsFromContext(ctx)
+// getAttr gets the attribute of this inode by issuing a FUSE_GETATTR request
+// or read from local cache.
+// It updates the corresponding attributes if necessary.
+func (i *inode) getAttr(ctx context.Context, fs *vfs.Filesystem, opts vfs.StatOptions) (linux.FUSEAttr, error) {
+	attributeVersion := atomic.LoadUint64(&i.fs.conn.attributeVersion)
+
+	// TODO(gvisor.dev/issue/3679): send the request only if
+	// - invalid local cache for fields specified in the opts.Mask
+	// - forced update
+	// - i.attributeTime expired
+	// If local cache is still valid, return local cache.
+	// Currently we always send a request,
+	// and we always set the metadata with the new result,
+	// unless attributeVersion has changed.
+
+	task := kernel.TaskFromContext(ctx)
 	if task == nil {
 		log.Warningf("couldn't get kernel task from context")
-		return linux.Statx{}, syserror.EINVAL
+		return linux.FUSEAttr{}, syserror.EINVAL
 	}
 
+	creds := auth.CredentialsFromContext(ctx)
+
 	var in linux.FUSEGetAttrIn
 	// We don't set any attribute in the request, because in VFS2 fstat(2) will
 	// finally be translated into vfs.FilesystemImpl.StatAt() (see
 	// pkg/sentry/syscalls/linux/vfs2/stat.go), resulting in the same flow
 	// as stat(2). Thus GetAttrFlags and Fh variable will never be used in VFS2.
-	req, err := conn.NewRequest(creds, uint32(task.ThreadID()), i.NodeID, linux.FUSE_GETATTR, &in)
+	req, err := i.fs.conn.NewRequest(creds, uint32(task.ThreadID()), i.NodeID, linux.FUSE_GETATTR, &in)
 	if err != nil {
-		return linux.Statx{}, err
+		return linux.FUSEAttr{}, err
 	}
 
-	res, err := conn.Call(task, req)
+	res, err := i.fs.conn.Call(task, req)
 	if err != nil {
-		return linux.Statx{}, err
+		return linux.FUSEAttr{}, err
 	}
 	if err := res.Error(); err != nil {
-		return linux.Statx{}, err
+		return linux.FUSEAttr{}, err
 	}
 
 	var out linux.FUSEGetAttrOut
 	if err := res.UnmarshalPayload(&out); err != nil {
-		return linux.Statx{}, err
+		return linux.FUSEAttr{}, err
 	}
 
-	// Set all metadata into kernfs.InodeAttrs.
+	// Local version is newer, return the local one.
+	// Skip the update.
+	if attributeVersion != 0 && atomic.LoadUint64(&i.attributeVersion) > attributeVersion {
+		return i.getFUSEAttr(), nil
+	}
+
+	// Set the metadata of kernfs.InodeAttrs.
 	if err := i.SetStat(ctx, fs, creds, vfs.SetStatOptions{
-		Stat: statFromFUSEAttr(out.Attr, linux.STATX_ALL, fusefs.devMinor),
+		Stat: statFromFUSEAttr(out.Attr, linux.STATX_ALL, i.fs.devMinor),
 	}); err != nil {
+		return linux.FUSEAttr{}, err
+	}
+
+	// Set the size if no error (after SetStat() check).
+	atomic.StoreUint64(&i.size, out.Attr.Size)
+
+	return out.Attr, nil
+}
+
+// reviseAttr attempts to update the attributes for internal purposes
+// by calling getAttr with a pre-specified mask.
+// Used by read, write, lseek.
+func (i *inode) reviseAttr(ctx context.Context) error {
+	// Never need atime for internal purposes.
+	_, err := i.getAttr(ctx, i.fs.VFSFilesystem(), vfs.StatOptions{
+		Mask: linux.STATX_BASIC_STATS &^ linux.STATX_ATIME,
+	})
+	return err
+}
+
+// Stat implements kernfs.Inode.Stat.
+func (i *inode) Stat(ctx context.Context, fs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) {
+	attr, err := i.getAttr(ctx, fs, opts)
+	if err != nil {
 		return linux.Statx{}, err
 	}
 
-	return statFromFUSEAttr(out.Attr, opts.Mask, fusefs.devMinor), nil
+	return statFromFUSEAttr(attr, opts.Mask, i.fs.devMinor), nil
 }
 
 // DecRef implements kernfs.Inode.
diff --git a/pkg/sentry/fsimpl/fuse/init.go b/pkg/sentry/fsimpl/fuse/init.go
index 779c2bd3f..2ff2542b6 100644
--- a/pkg/sentry/fsimpl/fuse/init.go
+++ b/pkg/sentry/fsimpl/fuse/init.go
@@ -29,9 +29,10 @@ 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
diff --git a/pkg/sentry/fsimpl/fuse/read_write.go b/pkg/sentry/fsimpl/fuse/read_write.go
new file mode 100644
index 000000000..4ef8531dc
--- /dev/null
+++ b/pkg/sentry/fsimpl/fuse/read_write.go
@@ -0,0 +1,152 @@
+// 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()
+	}
+}
diff --git a/pkg/sentry/fsimpl/fuse/regular_file.go b/pkg/sentry/fsimpl/fuse/regular_file.go
new file mode 100644
index 000000000..37ce4e268
--- /dev/null
+++ b/pkg/sentry/fsimpl/fuse/regular_file.go
@@ -0,0 +1,125 @@
+// 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); 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
+}