blob: ffb96c1c8fd1dbad1f58f9e89a9fe5b060f17b29 [file] [log] [blame]
// Copyright 2016 The Vanadium Authors. All rights reserved.
// Use of this source code is governed by a BSD-style
// license that can be found in the LICENSE file.
package internal
import (
"encoding/hex"
"encoding/json"
"hash/fnv"
"io/ioutil"
"path"
"sort"
"strconv"
"strings"
"time"
"google.golang.org/cloud/bigtable"
"v.io/v23"
"v.io/v23/context"
"v.io/v23/naming"
"v.io/v23/security/access"
vdltime "v.io/v23/vdlroot/time"
"v.io/v23/verror"
"v.io/x/ref/lib/timekeeper"
)
var gcGracePeriod = time.Minute
// SetGcGracePeriod sets the grace period for garbage collecting newly created
// nodes. Nodes are not eligible for garbage collection until after this time
// has passed. This function exists only for testing purposes.
func SetGcGracePeriod(p time.Duration) {
gcGracePeriod = p
}
type mtNode struct {
bt *BigTable
name string
sticky bool
creationTime bigtable.Timestamp
permissions access.Permissions
version string
creator string
mountFlags mtFlags
servers []naming.MountedServer
expiredServers []string
children []string
}
type mtFlags struct {
MT bool `json:"mt,omitempty"`
Leaf bool `json:"leaf,omitempty"`
}
func longTimeout(ctx *context.T) (*context.T, func()) {
return context.WithTimeout(v23.GetBackgroundContext(ctx), time.Hour)
}
func getNode(ctx *context.T, bt *BigTable, name string) (*mtNode, error) {
row, err := bt.readRow(ctx, rowKey(name), bigtable.RowFilter(bigtable.LatestNFilter(1)))
if err != nil {
return nil, err
}
return nodeFromRow(ctx, bt, row, clock), nil
}
func rowKey(name string) string {
// The row key is a hash of the node name followed by the node name
// itself.
// This spreads the rows evenly across the tablet servers to avoid
// traffic imbalance.
name = naming.Clean("/" + name)
h := fnv.New32()
h.Write([]byte(name))
return hex.EncodeToString(h.Sum(nil)) + name
}
func nodeFromRow(ctx *context.T, bt *BigTable, row bigtable.Row, clock timekeeper.TimeKeeper) *mtNode {
const offset = 9 // 32-bit value in hex + '/'
name := row.Key()
if len(name) < offset {
return nil
}
n := &mtNode{
bt: bt,
name: name[offset:],
}
for _, i := range row[metadataFamily] {
col := strings.TrimPrefix(i.Column, metadataFamily+":")
switch col {
case stickyColumn:
n.sticky = true
case versionColumn:
n.version = string(i.Value)
case permissionsColumn:
if err := json.Unmarshal(i.Value, &n.permissions); err != nil {
ctx.Errorf("Failed to decode permissions for %s", name)
return nil
}
case creatorColumn:
n.creationTime = i.Timestamp
n.creator = string(i.Value)
}
}
n.servers = make([]naming.MountedServer, 0, len(row[serversFamily]))
for _, i := range row[serversFamily] {
deadline := i.Timestamp.Time()
server := i.Column[2:]
if deadline.Before(clock.Now()) {
n.expiredServers = append(n.expiredServers, server)
continue
}
if err := json.Unmarshal(i.Value, &n.mountFlags); err != nil {
ctx.Errorf("Failed to decode mount flags for %s", name)
return nil
}
n.servers = append(n.servers, naming.MountedServer{
Server: server,
Deadline: vdltime.Deadline{deadline},
})
}
n.children = make([]string, 0, len(row[childrenFamily]))
for _, i := range row[childrenFamily] {
child := strings.TrimPrefix(i.Column, childrenFamily+":")
n.children = append(n.children, child)
}
return n
}
func (n *mtNode) createChild(ctx *context.T, child string, perms access.Permissions, creator string, limit int64) (*mtNode, error) {
ts := n.bt.now()
mut := bigtable.NewMutation()
mut.Set(childrenFamily, child, ts, []byte{1})
if err := n.mutate(ctx, mut, false); err != nil {
return nil, err
}
// If the current process dies right here, it will leave the parent with
// a reference to a child row that doesn't exist. This means that the
// parent will never be seen as "empty" and will not be garbage
// collected. This will be corrected when:
// - the child is created again, or
// - the parent is forcibly deleted with Delete().
childName := naming.Join(n.name, child)
longCtx, cancel := longTimeout(ctx)
defer cancel()
if err := n.bt.createRow(longCtx, childName, perms, creator, ts, limit); err != nil {
return nil, err
}
n, err := getNode(ctx, n.bt, childName)
if err != nil {
return nil, err
}
if n == nil {
return nil, verror.New(errConcurrentAccess, ctx, childName)
}
return n, nil
}
func (n *mtNode) mount(ctx *context.T, server string, deadline time.Time, flags naming.MountFlag, limit int64) error {
delta := int64(1)
mut := bigtable.NewMutation()
for _, s := range n.servers {
// Mount replaces an already mounted server with the same name,
// or all servers if the Replace flag is set.
if s.Server != server && flags&naming.Replace == 0 {
continue
}
delta--
mut.DeleteCellsInColumn(serversFamily, s.Server)
}
if err := incrementCreatorServerCount(ctx, n.bt, n.creator, delta, limit); err != nil {
return err
}
defer func() {
if err := incrementCreatorServerCount(ctx, n.bt, n.creator, -delta, 0); err != nil {
ctx.Errorf("incrementCreatorServerCount failed: %v", err)
}
}()
f := mtFlags{
MT: flags&naming.MT != 0,
Leaf: flags&naming.Leaf != 0,
}
jsonValue, err := json.Marshal(f)
if err != nil {
return err
}
mut.Set(serversFamily, server, n.bt.time(deadline), jsonValue)
if err := n.mutate(ctx, mut, false); err != nil {
return err
}
delta = 0
return nil
}
func (n *mtNode) unmount(ctx *context.T, server string) error {
delta := int64(0)
mut := bigtable.NewMutation()
for _, s := range n.servers {
// Unmount removes the specified server, or all servers if
// server == "".
if server != "" && s.Server != server {
continue
}
delta--
mut.DeleteCellsInColumn(serversFamily, s.Server)
}
if delta == 0 {
return nil
}
if err := n.mutate(ctx, mut, false); err != nil {
return err
}
if n, err := getNode(ctx, n.bt, n.name); err == nil {
n.gc(ctx)
}
return incrementCreatorServerCount(ctx, n.bt, n.creator, delta, 0)
}
func (n *mtNode) gc(ctx *context.T) (deletedSomething bool, err error) {
for n != nil && n.name != "" {
if len(n.expiredServers) > 0 {
mut := bigtable.NewMutation()
for _, s := range n.expiredServers {
mut.DeleteCellsInColumn(serversFamily, s)
}
if err = n.mutate(ctx, mut, false); err != nil {
break
}
delta := -int64(len(n.expiredServers))
if err = incrementCreatorServerCount(ctx, n.bt, n.creator, delta, 0); err != nil {
// TODO(rthellend): Since counters are stored in different rows,
// there is no way to update them atomically, e.g. if the server
// dies here, or if incrementCreatorServerCount returns an error,
// the server counter will be off.
// The same thing could happen everywhere the counters are updated.
// If/when we start using these counters for quota enforcement, we
// should also come up with a way to make sure the counters aren't
// too far off.
break
}
deletedSomething = true
break
}
if n.sticky || len(n.children) > 0 || len(n.servers) > 0 {
break
}
if time.Since(n.creationTime.Time()) < gcGracePeriod {
break
}
if err = n.delete(ctx, false); err != nil {
break
}
ctx.VI(2).Infof("Deleted empty node %q", n.name)
deletedSomething = true
parent := path.Dir(n.name)
if parent == "." {
break
}
if n, err = getNode(ctx, n.bt, parent); err != nil {
break
}
}
return
}
func (n *mtNode) deleteAndGC(ctx *context.T, deleteSubtree bool) error {
if err := n.delete(ctx, deleteSubtree); err != nil {
return err
}
parentName, _ := path.Split(n.name)
if parent, err := getNode(ctx, n.bt, parentName); err == nil {
parent.gc(ctx)
}
return nil
}
func (n *mtNode) delete(ctx *context.T, deleteSubtree bool) error {
if !deleteSubtree && len(n.children) > 0 {
return verror.New(errNotEmpty, ctx, n.name)
}
// TODO(rthellend): This naive approach could be very expensive in
// terms of memory. A smarter, but slower, approach would be to walk
// the tree without holding on to all the node data.
for _, c := range n.children {
cn, err := getNode(ctx, n.bt, naming.Join(n.name, c))
if err != nil {
return err
}
if cn == nil {
// Node 'n' has a reference to a child that doesn't
// exist. It could be that it is being created or
// deleted concurrently. To be sure, we have to create
// it before deleting it.
if cn, err = n.createChild(ctx, c, n.permissions, "", 0); err != nil {
return err
}
}
if err := cn.delete(ctx, true); err != nil {
return err
}
}
mut := bigtable.NewMutation()
mut.DeleteRow()
if err := n.mutate(ctx, mut, true); err != nil {
return err
}
// If the current process dies right here, it will leave the parent with
// a reference to a child row that no longer exists. This means that the
// parent will never be seen as "empty" and will not be garbage
// collected. This will be corrected when:
// - the child is re-created, or
// - the parent is forcibly deleted with Delete().
// Delete from parent node.
parent, child := path.Split(n.name)
mut = bigtable.NewMutation()
// DeleteTimestampRange deletes the cells whose timestamp is in the
// half open range [start,end). We need to delete the cell with
// timestamp n.creationTime (and any older ones).
mut.DeleteTimestampRange(childrenFamily, child, 0, n.bt.timeNext(n.creationTime))
longCtx, cancel := longTimeout(ctx)
defer cancel()
if err := n.bt.apply(longCtx, rowKey(parent), mut); err != nil {
return err
}
// Adjust the server count for the node creator. Since delete() can be
// called directly without gc(), we need to include the expired servers
// in the counter adjustment.
delta := -int64(len(n.servers) + len(n.expiredServers))
if err := incrementCreatorServerCount(ctx, n.bt, n.creator, delta, 0); err != nil {
return err
}
return incrementCreatorNodeCount(ctx, n.bt, n.creator, -1, 0)
}
func (n *mtNode) setPermissions(ctx *context.T, perms access.Permissions) error {
jsonPerms, err := json.Marshal(perms)
if err != nil {
return err
}
mut := bigtable.NewMutation()
mut.Set(metadataFamily, permissionsColumn, bigtable.ServerTime, jsonPerms)
mut.Set(metadataFamily, stickyColumn, bigtable.ServerTime, []byte{1})
if err := n.mutate(ctx, mut, false); err != nil {
return err
}
return nil
}
func (n *mtNode) mutate(ctx *context.T, mut *bigtable.Mutation, delete bool) error {
if !delete {
v, err := strconv.ParseUint(n.version, 10, 64)
if err != nil {
return err
}
newVersion := strconv.FormatUint(v+1, 10)
mut.Set(metadataFamily, versionColumn, bigtable.ServerTime, []byte(newVersion))
}
// The mutation will succeed iff the row already exists with the
// expected version.
filter := bigtable.ChainFilters(
bigtable.FamilyFilter(metadataFamily),
bigtable.ColumnFilter(versionColumn),
bigtable.LatestNFilter(1),
bigtable.ValueFilter(n.version),
)
condMut := bigtable.NewCondMutation(filter, mut, nil)
var success bool
if err := n.bt.apply(ctx, rowKey(n.name), condMut, bigtable.GetCondMutationResult(&success)); err != nil {
return err
}
if !success {
return verror.New(errConcurrentAccess, ctx, n.name)
}
return nil
}
func createNodesFromFile(ctx *context.T, bt *BigTable, fileName string) error {
var nodes map[string]access.Permissions
data, err := ioutil.ReadFile(fileName)
if err != nil {
return err
}
if err := json.Unmarshal(data, &nodes); err != nil {
return err
}
// This loop adds backward compatibility with the older template format,
// e.g. "a/b/%%" { "Admin": { "In": [ "%%" ] } }
// With the new format, this is equivalent to:
// "a/b" { "%%/Admin": { "In": [ "%%" ] } }
for node, perms := range nodes {
if strings.HasSuffix(node, "/%%") {
delete(nodes, node)
node = strings.TrimSuffix(node, "/%%")
p := nodes[node]
if p == nil {
p = make(access.Permissions)
}
for tag := range perms {
p["%%/"+tag] = perms[tag]
}
nodes[node] = p
}
}
// Create the nodes in alphanumeric order so that children are
// created after their parents.
sortedNodes := []string{}
for node := range nodes {
sortedNodes = append(sortedNodes, node)
}
sort.Strings(sortedNodes)
ts := bt.now()
for _, node := range sortedNodes {
perms := nodes[node]
if node == "" {
if err := bt.createRow(ctx, "", perms, "", ts, 0); err != nil {
return err
}
continue
}
parentName := ""
for _, e := range strings.Split(node, "/") {
n, err := getNode(ctx, bt, naming.Join(parentName, e))
if err != nil {
return err
}
if n == nil {
parent, err := getNode(ctx, bt, parentName)
if err != nil {
return err
}
if n, err = parent.createChild(ctx, e, parent.permissions, "", 0); err != nil {
return err
}
}
if n.name == node {
// setPermissions also makes the node sticky.
if err := n.setPermissions(ctx, perms); err != nil {
return err
}
}
parentName = n.name
}
}
return nil
}