From 694c1bf9dbea75ebdafcf075fa92dd18ddb7f0f6 Mon Sep 17 00:00:00 2001 From: Vitaliy Filippov Date: Wed, 25 Aug 2021 11:05:53 +0300 Subject: [PATCH] Introduce VectoredReadOp Read requests can now take vectored responses from the filesystem implementation and send them to FUSE device via the writev() system call. This allows file systems to send data without copying it into the library-provided buffer if the data is already in memory. The change also speeds up normal ReadFileOps as a side effect because it removes extra memory allocations. --- connection.go | 16 ++-- conversions.go | 71 ++++++++------- debug.go | 5 ++ fuseops/ops.go | 31 +++++++ fuseutil/file_system.go | 4 + fuseutil/not_implemented_file_system.go | 6 ++ internal/buffer/in_message.go | 10 +++ internal/buffer/out_message.go | 109 ++++++++++-------------- internal/buffer/out_message_test.go | 41 +++++---- mount_config.go | 5 ++ writev.go | 29 +++++++ 11 files changed, 209 insertions(+), 118 deletions(-) create mode 100644 writev.go diff --git a/connection.go b/connection.go index 885ab49..7d08cc0 100644 --- a/connection.go +++ b/connection.go @@ -338,7 +338,7 @@ func (c *Connection) readMessage() (*buffer.InMessage, error) { // Loop past transient errors. for { - // Attempt a reaed. + // Attempt a read. err := m.Init(c.dev) // Special cases: @@ -405,7 +405,7 @@ func (c *Connection) ReadOp() (_ context.Context, op interface{}, _ error) { // Convert the message to an op. outMsg := c.getOutMessage() - op, err = convertInMessage(inMsg, outMsg, c.protocol) + op, err = convertInMessage(&c.cfg, inMsg, outMsg, c.protocol) if err != nil { c.putOutMessage(outMsg) return nil, nil, fmt.Errorf("convertInMessage: %v", err) @@ -510,10 +510,16 @@ func (c *Connection) Reply(ctx context.Context, opErr error) { noResponse := c.kernelResponse(outMsg, inMsg.Header().Unique, op, opErr) if !noResponse { - err := c.writeMessage(outMsg.Bytes()) - if err != nil && c.errorLogger != nil { - c.errorLogger.Printf("writeMessage: %v %v", err, outMsg.Bytes()) + var err error + if outMsg.Sglist != nil { + _, err = writev(int(c.dev.Fd()), outMsg.Sglist) + } else { + err = c.writeMessage(outMsg.OutHeaderBytes()) } + if err != nil && c.errorLogger != nil { + c.errorLogger.Printf("writeMessage: %v %v", err, outMsg.OutHeaderBytes()) + } + outMsg.Sglist = nil } } diff --git a/conversions.go b/conversions.go index 6e94f97..6257b48 100644 --- a/conversions.go +++ b/conversions.go @@ -38,6 +38,7 @@ import ( // // The caller is responsible for arranging for the message to be destroyed. func convertInMessage( + config *MountConfig, inMsg *buffer.InMessage, outMsg *buffer.OutMessage, protocol fusekernel.Protocol) (o interface{}, err error) { @@ -284,24 +285,28 @@ func convertInMessage( return nil, errors.New("Corrupt OpRead") } - to := &fuseops.ReadFileOp{ - Inode: fuseops.InodeID(inMsg.Header().Nodeid), - Handle: fuseops.HandleID(in.Fh), - Offset: int64(in.Offset), - OpContext: fuseops.OpContext{Pid: inMsg.Header().Pid}, + if !config.UseVectoredRead { + // Use part of the incoming message storage as the read buffer + buf := inMsg.GetFree(int(in.Size)) + to := &fuseops.ReadFileOp{ + Inode: fuseops.InodeID(inMsg.Header().Nodeid), + Handle: fuseops.HandleID(in.Fh), + Offset: int64(in.Offset), + Dst: buf, + OpContext: fuseops.OpContext{Pid: inMsg.Header().Pid}, + } + o = to + } else { + // Don't allocate any buffers when zero-copy is used + to := &fuseops.VectoredReadOp{ + Inode: fuseops.InodeID(inMsg.Header().Nodeid), + Handle: fuseops.HandleID(in.Fh), + Offset: int64(in.Offset), + Size: int64(in.Size), + OpContext: fuseops.OpContext{Pid: inMsg.Header().Pid}, + } + o = to } - o = to - - readSize := int(in.Size) - p := outMsg.GrowNoZero(readSize) - if p == nil { - return nil, fmt.Errorf("Can't grow for %d-byte read", readSize) - } - - sh := (*reflect.SliceHeader)(unsafe.Pointer(&to.Dst)) - sh.Data = uintptr(p) - sh.Len = readSize - sh.Cap = readSize case fusekernel.OpReaddir: in := (*fusekernel.ReadIn)(inMsg.Consume(fusekernel.ReadInSize(protocol))) @@ -318,7 +323,7 @@ func convertInMessage( o = to readSize := int(in.Size) - p := outMsg.GrowNoZero(readSize) + p := outMsg.Grow(readSize) if p == nil { return nil, fmt.Errorf("Can't grow for %d-byte read", readSize) } @@ -489,15 +494,19 @@ func convertInMessage( o = to readSize := int(in.Size) - p := outMsg.GrowNoZero(readSize) - if p == nil { - return nil, fmt.Errorf("Can't grow for %d-byte read", readSize) - } + if readSize > 0 { + p := outMsg.Grow(readSize) + if p == nil { + return nil, fmt.Errorf("Can't grow for %d-byte read", readSize) + } - sh := (*reflect.SliceHeader)(unsafe.Pointer(&to.Dst)) - sh.Data = uintptr(p) - sh.Len = readSize - sh.Cap = readSize + sh := (*reflect.SliceHeader)(unsafe.Pointer(&to.Dst)) + sh.Data = uintptr(p) + sh.Len = readSize + sh.Cap = readSize + } else { + to.Dst = nil + } case fusekernel.OpListxattr: type input fusekernel.ListxattrIn @@ -514,7 +523,7 @@ func convertInMessage( readSize := int(in.Size) if readSize != 0 { - p := outMsg.GrowNoZero(readSize) + p := outMsg.Grow(readSize) if p == nil { return nil, fmt.Errorf("Can't grow for %d-byte read", readSize) } @@ -718,9 +727,11 @@ func (c *Connection) kernelResponseForOp( } case *fuseops.ReadFileOp: - // convertInMessage already set up the destination buffer to be at the end - // of the out message. We need only shrink to the right size based on how - // much the user read. + m.Append(o.Dst) + m.ShrinkTo(buffer.OutMessageHeaderSize + o.BytesRead) + + case *fuseops.VectoredReadOp: + m.Append(o.Data...) m.ShrinkTo(buffer.OutMessageHeaderSize + o.BytesRead) case *fuseops.WriteFileOp: diff --git a/debug.go b/debug.go index 2a4dcc0..8b0d0e1 100644 --- a/debug.go +++ b/debug.go @@ -97,6 +97,11 @@ func describeRequest(op interface{}) (s string) { addComponent("offset %d", typed.Offset) addComponent("%d bytes", len(typed.Dst)) + case *fuseops.VectoredReadOp: + addComponent("handle %d", typed.Handle) + addComponent("offset %d", typed.Offset) + addComponent("%d bytes", typed.Size) + case *fuseops.WriteFileOp: addComponent("handle %d", typed.Handle) addComponent("offset %d", typed.Offset) diff --git a/fuseops/ops.go b/fuseops/ops.go index 353c70b..b8bf2b8 100644 --- a/fuseops/ops.go +++ b/fuseops/ops.go @@ -654,6 +654,37 @@ type ReadFileOp struct { OpContext OpContext } +// Vectored read - same as ReadFileOp, but the buffer isn't provided by the library. +// The file system returns a list of slices instead. +type VectoredReadOp struct { + // The file inode that we are reading, and the handle previously returned by + // CreateFile or OpenFile when opening that inode. + Inode InodeID + Handle HandleID + + // The offset within the file at which to read. + Offset int64 + + // The size of the read. + Size int64 + + // Set by the file system: data to send back to the client. + Data [][]byte + + // Set by the file system: the number of bytes read. + // + // The FUSE documentation requires that exactly the requested number of bytes + // be returned, except in the case of EOF or error (http://goo.gl/ZgfBkF). + // This appears to be because it uses file mmapping machinery + // (http://goo.gl/SGxnaN) to read a page at a time. It appears to understand + // where EOF is by checking the inode size (http://goo.gl/0BkqKD), returned + // by a previous call to LookUpInode, GetInodeAttributes, etc. + // + // If direct IO is enabled, semantics should match those of read(2). + BytesRead int + OpContext OpContext +} + // Write data to a file previously opened with CreateFile or OpenFile. // // When the user writes data using write(2), the write goes into the page diff --git a/fuseutil/file_system.go b/fuseutil/file_system.go index 5eb8bca..711a4b4 100644 --- a/fuseutil/file_system.go +++ b/fuseutil/file_system.go @@ -52,6 +52,7 @@ type FileSystem interface { ReleaseDirHandle(context.Context, *fuseops.ReleaseDirHandleOp) error OpenFile(context.Context, *fuseops.OpenFileOp) error ReadFile(context.Context, *fuseops.ReadFileOp) error + VectoredRead(context.Context, *fuseops.VectoredReadOp) error WriteFile(context.Context, *fuseops.WriteFileOp) error SyncFile(context.Context, *fuseops.SyncFileOp) error FlushFile(context.Context, *fuseops.FlushFileOp) error @@ -190,6 +191,9 @@ func (s *fileSystemServer) handleOp( case *fuseops.ReadFileOp: err = s.fs.ReadFile(ctx, typed) + case *fuseops.VectoredReadOp: + err = s.fs.VectoredRead(ctx, typed) + case *fuseops.WriteFileOp: err = s.fs.WriteFile(ctx, typed) diff --git a/fuseutil/not_implemented_file_system.go b/fuseutil/not_implemented_file_system.go index 4d29cfa..9aa59d2 100644 --- a/fuseutil/not_implemented_file_system.go +++ b/fuseutil/not_implemented_file_system.go @@ -138,6 +138,12 @@ func (fs *NotImplementedFileSystem) ReadFile( return fuse.ENOSYS } +func (fs *NotImplementedFileSystem) VectoredRead( + ctx context.Context, + op *fuseops.VectoredReadOp) error { + return fuse.ENOSYS +} + func (fs *NotImplementedFileSystem) WriteFile( ctx context.Context, op *fuseops.WriteFileOp) error { diff --git a/internal/buffer/in_message.go b/internal/buffer/in_message.go index 64df2c6..cb48816 100644 --- a/internal/buffer/in_message.go +++ b/internal/buffer/in_message.go @@ -42,6 +42,7 @@ func init() { type InMessage struct { remaining []byte storage []byte + size int } // NewInMessage creates a new InMessage with its storage initialized. @@ -66,6 +67,7 @@ func (m *InMessage) Init(r io.Reader) error { return fmt.Errorf("Unexpectedly read only %d bytes.", n) } + m.size = n m.remaining = m.storage[headerSize:n] // Check the header's length. @@ -114,3 +116,11 @@ func (m *InMessage) ConsumeBytes(n uintptr) []byte { return b } + +// Get the next n bytes after the message to use them as a temporary buffer +func (m *InMessage) GetFree(n int) []byte { + if n <= 0 || n > len(m.storage)-m.size { + return nil + } + return m.storage[m.size : m.size+n] +} diff --git a/internal/buffer/out_message.go b/internal/buffer/out_message.go index 7e89afb..c8036cb 100644 --- a/internal/buffer/out_message.go +++ b/internal/buffer/out_message.go @@ -16,7 +16,6 @@ package buffer import ( "fmt" - "log" "reflect" "unsafe" @@ -33,30 +32,15 @@ const OutMessageHeaderSize = int(unsafe.Sizeof(fusekernel.OutHeader{})) // // Must be initialized with Reset. type OutMessage struct { - // The offset into payload to which we're currently writing. - payloadOffset int - - header fusekernel.OutHeader - payload [MaxReadSize]byte -} - -// Make sure that the header and payload are contiguous. -func init() { - a := unsafe.Offsetof(OutMessage{}.header) + uintptr(OutMessageHeaderSize) - b := unsafe.Offsetof(OutMessage{}.payload) - - if a != b { - log.Panicf( - "header ends at offset %d, but payload starts at offset %d", - a, b) - } + header fusekernel.OutHeader + Sglist [][]byte } // Reset resets m so that it's ready to be used again. Afterward, the contents // are solely a zeroed fusekernel.OutHeader struct. func (m *OutMessage) Reset() { - m.payloadOffset = 0 m.header = fusekernel.OutHeader{} + m.Sglist = nil } // OutHeader returns a pointer to the header at the start of the message. @@ -64,30 +48,12 @@ func (m *OutMessage) OutHeader() *fusekernel.OutHeader { return &m.header } -// Grow grows m's buffer by the given number of bytes, returning a pointer to -// the start of the new segment, which is guaranteed to be zeroed. If there is -// insufficient space, it returns nil. +// Grow adds a new buffer of bytes to the message, returning a pointer to +// the start of the new segment, which is guaranteed to be zeroed. func (m *OutMessage) Grow(n int) unsafe.Pointer { - p := m.GrowNoZero(n) - if p != nil { - jacobsa_fuse_memclr(p, uintptr(n)) - } - - return p -} - -// GrowNoZero is equivalent to Grow, except the new segment is not zeroed. Use -// with caution! -func (m *OutMessage) GrowNoZero(n int) unsafe.Pointer { - // Will we overflow the buffer? - o := m.payloadOffset - if len(m.payload)-o < n { - return nil - } - - p := unsafe.Pointer(uintptr(unsafe.Pointer(&m.payload)) + uintptr(o)) - m.payloadOffset = o + n - + b := make([]byte, n) + m.Append(b) + p := unsafe.Pointer(&b[0]) return p } @@ -100,51 +66,62 @@ func (m *OutMessage) ShrinkTo(n int) { n, m.Len())) } - - m.payloadOffset = n - OutMessageHeaderSize + if n == OutMessageHeaderSize { + m.Sglist = nil + } else { + i := 1 + n -= OutMessageHeaderSize + for len(m.Sglist) > i && n >= len(m.Sglist[i]) { + n -= len(m.Sglist[i]) + i++ + } + if n > 0 { + m.Sglist[i] = m.Sglist[i][0:n] + i++ + } + m.Sglist = m.Sglist[0:i] + } } // Append is equivalent to growing by len(src), then copying src over the new // segment. Int panics if there is not enough room available. -func (m *OutMessage) Append(src []byte) { - p := m.GrowNoZero(len(src)) - if p == nil { - panic(fmt.Sprintf("Can't grow %d bytes", len(src))) +func (m *OutMessage) Append(src ...[]byte) { + if m.Sglist == nil { + // First element of Sglist is pre-filled with a pointer to the header + // to allow sending it with a single writev() call without copying the + // slice again + m.Sglist = append(m.Sglist, m.OutHeaderBytes()) } - - sh := (*reflect.SliceHeader)(unsafe.Pointer(&src)) - jacobsa_fuse_memmove(p, unsafe.Pointer(sh.Data), uintptr(sh.Len)) - + m.Sglist = append(m.Sglist, src...) return } // AppendString is like Append, but accepts string input. func (m *OutMessage) AppendString(src string) { - p := m.GrowNoZero(len(src)) - if p == nil { - panic(fmt.Sprintf("Can't grow %d bytes", len(src))) - } - - sh := (*reflect.StringHeader)(unsafe.Pointer(&src)) - jacobsa_fuse_memmove(p, unsafe.Pointer(sh.Data), uintptr(sh.Len)) - + m.Append([]byte(src)) return } // Len returns the current size of the message, including the leading header. func (m *OutMessage) Len() int { - return OutMessageHeaderSize + m.payloadOffset + if m.Sglist == nil { + return OutMessageHeaderSize + } + // First element of Sglist is the header, so we don't need to count it here + r := 0 + for _, b := range m.Sglist { + r += len(b) + } + return r } -// Bytes returns a reference to the current contents of the buffer, including -// the leading header. -func (m *OutMessage) Bytes() []byte { - l := m.Len() +// OutHeaderBytes returns a byte slice containing the current header. +func (m *OutMessage) OutHeaderBytes() []byte { + l := OutMessageHeaderSize sh := reflect.SliceHeader{ Data: uintptr(unsafe.Pointer(&m.header)), Len: l, Cap: l, } - return *(*[]byte)(unsafe.Pointer(&sh)) } diff --git a/internal/buffer/out_message_test.go b/internal/buffer/out_message_test.go index 6292794..1762b12 100644 --- a/internal/buffer/out_message_test.go +++ b/internal/buffer/out_message_test.go @@ -107,9 +107,12 @@ func TestOutMessageAppend(t *testing.T) { t.Errorf("om.Len() = %d, want %d", got, want) } - b := om.Bytes() + b := []byte(nil) + for i := 0; i < len(om.Sglist); i++ { + b = append(b, om.Sglist[i]...) + } if got, want := len(b), wantLen; got != want { - t.Fatalf("len(om.Bytes()) = %d, want %d", got, want) + t.Fatalf("len(om.OutHeaderBytes()) = %d, want %d", got, want) } want := append( @@ -137,9 +140,12 @@ func TestOutMessageAppendString(t *testing.T) { t.Errorf("om.Len() = %d, want %d", got, want) } - b := om.Bytes() + b := []byte(nil) + for i := 0; i < len(om.Sglist); i++ { + b = append(b, om.Sglist[i]...) + } if got, want := len(b), wantLen; got != want { - t.Fatalf("len(om.Bytes()) = %d, want %d", got, want) + t.Fatalf("len(om.OutHeaderBytes()) = %d, want %d", got, want) } want := append( @@ -168,9 +174,12 @@ func TestOutMessageShrinkTo(t *testing.T) { t.Errorf("om.Len() = %d, want %d", got, want) } - b := om.Bytes() + b := []byte(nil) + for i := 0; i < len(om.Sglist); i++ { + b = append(b, om.Sglist[i]...) + } if got, want := len(b), wantLen; got != want { - t.Fatalf("len(om.Bytes()) = %d, want %d", got, want) + t.Fatalf("len(om.OutHeaderBytes()) = %d, want %d", got, want) } want := append( @@ -201,7 +210,7 @@ func TestOutMessageHeader(t *testing.T) { *h = want // Check that the result is as expected. - b := om.Bytes() + b := om.OutHeaderBytes() if len(b) != int(unsafe.Sizeof(want)) { t.Fatalf("unexpected length %d; want %d", len(b), unsafe.Sizeof(want)) } @@ -225,9 +234,7 @@ func TestOutMessageReset(t *testing.T) { } // Ensure a non-zero payload length. - if p := om.GrowNoZero(128); p == nil { - t.Fatal("GrowNoZero failed") - } + p := om.Grow(128) // Reset. om.Reset() @@ -259,10 +266,7 @@ func TestOutMessageGrow(t *testing.T) { // Set up garbage where the payload will soon be. const payloadSize = 1234 { - p := om.GrowNoZero(payloadSize) - if p == nil { - t.Fatal("GrowNoZero failed") - } + p := om.Grow(payloadSize) err := fillWithGarbage(p, payloadSize) if err != nil { @@ -283,7 +287,10 @@ func TestOutMessageGrow(t *testing.T) { t.Errorf("om.Len() = %d, want %d", got, want) } - b := om.Bytes() + b := []byte(nil) + for i := 0; i < len(om.Sglist); i++ { + b = append(b, om.Sglist[i]...) + } if got, want := len(b), wantLen; got != want { t.Fatalf("len(om.Len()) = %d, want %d", got, want) } @@ -304,7 +311,7 @@ func BenchmarkOutMessageReset(b *testing.B) { om.Reset() } - b.SetBytes(int64(unsafe.Offsetof(om.payload))) + b.SetBytes(int64(om.Len())) }) // Many megabytes worth of buffers, which should defeat the CPU cache. @@ -321,7 +328,7 @@ func BenchmarkOutMessageReset(b *testing.B) { oms[i%numMessages].Reset() } - b.SetBytes(int64(unsafe.Offsetof(oms[0].payload))) + b.SetBytes(int64(oms[0].Len())) }) } diff --git a/mount_config.go b/mount_config.go index d23224f..5d827ab 100644 --- a/mount_config.go +++ b/mount_config.go @@ -156,6 +156,11 @@ type MountConfig struct { // actually utilise any form of qualifiable UNIX permissions. DisableDefaultPermissions bool + // Use VectoredReadOp instead of ReadFileOp. + // Vectored read allows file systems to reduce memory copying overhead if + // the data is already in memory when they return it to FUSE. + UseVectoredRead bool + // OS X only. // // The name of the mounted volume, as displayed in the Finder. If empty, a diff --git a/writev.go b/writev.go new file mode 100644 index 0000000..05c9358 --- /dev/null +++ b/writev.go @@ -0,0 +1,29 @@ +package fuse + +import ( + "syscall" + "unsafe" +) + +func writev(fd int, packet [][]byte) (n int, err error) { + iovecs := make([]syscall.Iovec, 0, len(packet)) + for _, v := range packet { + if len(v) == 0 { + continue + } + vec := syscall.Iovec{ + Base: &v[0], + } + vec.SetLen(len(v)) + iovecs = append(iovecs, vec) + } + n1, _, e1 := syscall.Syscall( + syscall.SYS_WRITEV, + uintptr(fd), uintptr(unsafe.Pointer(&iovecs[0])), uintptr(len(iovecs)), + ) + n = int(n1) + if e1 != 0 { + err = syscall.Errno(e1) + } + return +}