// compareBiogoNode compares a biogo node and a range tree node to determine if both
// contain the same values in the same order. It recursively calls itself on
// both children if they exist.
func compareBiogoNode(db *client.DB, biogoNode *llrb.Node, key *proto.Key) error {
// Retrieve the node form the range tree.
rtNode := &proto.RangeTreeNode{}
if err := db.GetProto(keys.RangeTreeNodeKey(*key), rtNode); err != nil {
return err
}
bNode := &proto.RangeTreeNode{
Key: proto.Key(biogoNode.Elem.(Key)),
ParentKey: proto.KeyMin,
Black: bool(biogoNode.Color),
}
if biogoNode.Left != nil {
leftKey := proto.Key(biogoNode.Left.Elem.(Key))
bNode.LeftKey = &leftKey
}
if biogoNode.Right != nil {
rightKey := proto.Key(biogoNode.Right.Elem.(Key))
bNode.RightKey = &rightKey
}
if err := nodesEqual(*key, *bNode, *rtNode); err != nil {
return err
}
if rtNode.LeftKey != nil {
if err := compareBiogoNode(db, biogoNode.Left, rtNode.LeftKey); err != nil {
return err
}
}
if rtNode.RightKey != nil {
if err := compareBiogoNode(db, biogoNode.Right, rtNode.RightKey); err != nil {
return err
}
}
return nil
}
// getConfig retrieves the configuration for the specified key. If the
// key is empty, all configurations are returned. Otherwise, the
// leading "/" path delimiter is stripped and the configuration
// matching the remainder is retrieved. Note that this will retrieve
// the default config if "key" is equal to "/", and will list all
// configs if "key" is equal to "". The body result contains a listing
// of keys and retrieval of a config. The output format is determined
// by the request header.
func getConfig(db *client.DB, configPrefix proto.Key, config gogoproto.Message,
path string, r *http.Request) (body []byte, contentType string, err error) {
// Scan all configs if the key is empty.
if len(path) == 0 {
var rows []client.KeyValue
if rows, err = db.Scan(configPrefix, configPrefix.PrefixEnd(), maxGetResults); err != nil {
return
}
if len(rows) == maxGetResults {
log.Warningf("retrieved maximum number of results (%d); some may be missing", maxGetResults)
}
var prefixes []string
for _, row := range rows {
trimmed := bytes.TrimPrefix(row.Key, configPrefix)
prefixes = append(prefixes, url.QueryEscape(string(trimmed)))
}
// Encode the response.
body, contentType, err = util.MarshalResponse(r, prefixes, util.AllEncodings)
} else {
configkey := keys.MakeKey(configPrefix, proto.Key(path[1:]))
if err = db.GetProto(configkey, config); err != nil {
return
}
body, contentType, err = util.MarshalResponse(r, config, util.AllEncodings)
}
return
}
// compareBiogoTree walks both a biogo tree and the range tree to determine if both
// contain the same values in the same order.
func compareBiogoTree(db *client.DB, biogoTree *llrb.Tree) error {
rt := &proto.RangeTree{}
if err := db.GetProto(keys.RangeTreeRoot, rt); err != nil {
return err
}
return compareBiogoNode(db, biogoTree.Root, &rt.RootKey)
}
func countRangeReplicas(db *client.DB) (int, error) {
desc := &roachpb.RangeDescriptor{}
if err := db.GetProto(keys.RangeDescriptorKey(roachpb.KeyMin), desc); err != nil {
return 0, err
}
return len(desc.Replicas), nil
}
// allocateNodeID increments the node id generator key to allocate
// a new, unique node id.
func allocateNodeID(db *client.DB) (roachpb.NodeID, error) {
r, err := db.Inc(keys.NodeIDGenerator, 1)
if err != nil {
return 0, util.Errorf("unable to allocate node ID: %s", err)
}
return roachpb.NodeID(r.ValueInt()), nil
}
func (p *planner) releaseLeases(db client.DB) {
if p.leases != nil {
for _, lease := range p.leases {
if err := p.leaseMgr.Release(lease); err != nil {
log.Warning(err)
}
}
p.leases = nil
}
// TODO(pmattis): This is a hack. Remove when schema change operations work
// properly.
if p.modifiedSchemas != nil {
for _, d := range p.modifiedSchemas {
var lease *LeaseState
err := db.Txn(func(txn *client.Txn) error {
var err error
lease, err = p.leaseMgr.Acquire(txn, d.id, d.version)
return err
})
if err != nil {
log.Warning(err)
continue
}
if err := p.leaseMgr.Release(lease); err != nil {
log.Warning(err)
}
}
p.modifiedSchemas = nil
}
}
// allocateStoreIDs increments the store id generator key for the
// specified node to allocate "inc" new, unique store ids. The
// first ID in a contiguous range is returned on success.
func allocateStoreIDs(nodeID roachpb.NodeID, inc int64, db *client.DB) (roachpb.StoreID, error) {
r, err := db.Inc(keys.StoreIDGenerator, inc)
if err != nil {
return 0, util.Errorf("unable to allocate %d store IDs for node %d: %s", inc, nodeID, err)
}
return roachpb.StoreID(r.ValueInt() - inc + 1), nil
}
// getPermConfig fetches the permissions config for 'prefix'.
func getPermConfig(db *client.DB, prefix string) (*config.PermConfig, error) {
config := &config.PermConfig{}
if err := db.GetProto(keys.MakeKey(keys.ConfigPermissionPrefix, proto.Key(prefix)), config); err != nil {
return nil, err
}
return config, nil
}
func split(db *client.DB) {
for i := 1; i < 10; i++ {
fmt.Printf("split key: %v\n", i)
err := db.AdminSplit([]byte(fmt.Sprintf("%d", i)))
if err != nil {
panic(fmt.Sprintf("split fail. key: %v, err: %v\n", i, err))
}
}
}
// deleteConfig removes the config specified by key.
func deleteConfig(db *client.DB, configPrefix proto.Key, path string, r *http.Request) error {
if len(path) == 0 {
return util.Errorf("no path specified for config Delete")
}
if path == "/" {
return util.Errorf("the default configuration cannot be deleted")
}
configKey := keys.MakeKey(configPrefix, proto.Key(path[1:]))
return db.Del(configKey)
}
// loadTree loads the tree root and all of its nodes. It puts all of the nodes
// into a map.
func loadTree(t *testing.T, db *client.DB) (*roachpb.RangeTree, map[string]roachpb.RangeTreeNode) {
tree := new(roachpb.RangeTree)
if err := db.GetProto(keys.RangeTreeRoot, tree); err != nil {
t.Fatal(err)
}
nodes := make(map[string]roachpb.RangeTreeNode)
if tree.RootKey != nil {
loadNodes(t, db, tree.RootKey, nodes)
}
return tree, nodes
}
// loadTree loads the tree root and all of its nodes. It puts all of the nodes
// into a map.
func loadTree(t *testing.T, db *client.DB) (storage.RangeTree, map[string]storage.RangeTreeNode) {
var tree storage.RangeTree
if err := db.GetProto(keys.RangeTreeRoot, &tree); err != nil {
t.Fatal(err)
}
nodes := make(map[string]storage.RangeTreeNode)
if tree.RootKey != nil {
loadNodes(t, db, tree.RootKey, nodes)
}
return tree, nodes
}
// getRangeKeys returns the end keys of all ranges.
func getRangeKeys(db *client.DB) ([]roachpb.Key, error) {
rows, err := db.Scan(keys.Meta2Prefix, keys.MetaMax, 0)
if err != nil {
return nil, err
}
ret := make([]roachpb.Key, len(rows), len(rows))
for i := 0; i < len(rows); i++ {
ret[i] = bytes.TrimPrefix(rows[i].Key, keys.Meta2Prefix)
}
return ret, nil
}
// treesEqual compares the expectedTree and expectedNodes to the actual range
// tree stored in the db.
func treesEqual(db *client.DB, expected testRangeTree) error {
// Compare the tree roots.
actualTree := &proto.RangeTree{}
if err := db.GetProto(keys.RangeTreeRoot, actualTree); err != nil {
return err
}
if !reflect.DeepEqual(&expected.Tree, actualTree) {
return util.Errorf("Range tree root is not as expected - expected:%+v - actual:%+v", expected.Tree, actualTree)
}
return treeNodesEqual(db, expected, expected.Tree.RootKey)
}
func insert(r *rand.Rand, db *client.DB, wg *sync.WaitGroup) {
wg.Add(1)
for i := 0; i < 1000; i++ {
key := getKey(r)
value := getValue(1024 * 10)
err := db.Put(key, value)
if err != nil {
fmt.Printf("put fail. err: %v, key: %v\n", err, key)
}
}
wg.Done()
}
// loadNodes fetches a node and recursively all of its children.
func loadNodes(t *testing.T, db *client.DB, key roachpb.RKey, nodes map[string]roachpb.RangeTreeNode) {
node := new(roachpb.RangeTreeNode)
if err := db.GetProto(keys.RangeTreeNodeKey(key), node); err != nil {
t.Fatal(err)
}
nodes[node.Key.String()] = *node
if node.LeftKey != nil {
loadNodes(t, db, node.LeftKey, nodes)
}
if node.RightKey != nil {
loadNodes(t, db, node.RightKey, nodes)
}
}
// setDefaultRangeMaxBytes sets the range-max-bytes value for the default zone.
func setDefaultRangeMaxBytes(t *testing.T, db *client.DB, maxBytes int64) {
zone := &proto.ZoneConfig{}
if err := db.GetProto(keys.ConfigZonePrefix, zone); err != nil {
t.Fatal(err)
}
if zone.RangeMaxBytes == maxBytes {
return
}
zone.RangeMaxBytes = maxBytes
if err := db.Put(keys.ConfigZonePrefix, zone); err != nil {
t.Fatal(err)
}
}
func waitForInitialSplits(db *client.DB) error {
expectedRanges := ExpectedInitialRangeCount()
return util.RetryForDuration(initialSplitsTimeout, func() error {
// Scan all keys in the Meta2Prefix; we only need a count.
rows, err := db.Scan(keys.Meta2Prefix, keys.MetaMax, 0)
if err != nil {
return err
}
if a, e := len(rows), expectedRanges; a != e {
return util.Errorf("had %d ranges at startup, expected %d", a, e)
}
return nil
})
}
// purgeOldLeases refreshes the leases on a table. Unused leases older than
// minVersion will be released.
// If deleted is set, minVersion is ignored; no lease is acquired and all
// existing unused leases are released. The table is further marked for
// deletion, which will cause existing in-use leases to be eagerly released once
// they're not in use any more.
// If t has no active leases, nothing is done.
func (t *tableState) purgeOldLeases(
db *client.DB, deleted bool, minVersion sqlbase.DescriptorVersion, store LeaseStore,
) error {
t.mu.Lock()
empty := len(t.active.data) == 0
t.mu.Unlock()
if empty {
// We don't currently have a lease on this table, so no need to refresh
// anything.
return nil
}
// Acquire and release a lease on the table at a version >= minVersion.
var lease *LeaseState
err := db.Txn(func(txn *client.Txn) error {
var err error
if !deleted {
lease, err = t.acquire(txn, minVersion, store)
if err == errTableDeleted {
deleted = true
}
}
if err == nil || deleted {
t.mu.Lock()
defer t.mu.Unlock()
var toRelease []*LeaseState
if deleted {
t.deleted = true
// If the table has been deleted, all leases are stale.
toRelease = append([]*LeaseState(nil), t.active.data...)
} else {
// Otherwise, all but the lease we just took are stale.
toRelease = append([]*LeaseState(nil), t.active.data[:len(t.active.data)-1]...)
}
if err := t.releaseLeasesIfNotActive(toRelease, store); err != nil {
return err
}
return nil
}
return err
})
if err != nil {
return err
}
if lease == nil {
return nil
}
return t.release(lease, store)
}
// allocateNodeID increments the node id generator key to allocate
// a new, unique node id. It will retry indefinitely on retryable
// errors.
func allocateNodeID(db *client.DB) (proto.NodeID, error) {
var id proto.NodeID
err := retry.WithBackoff(allocRetryOptions, func() (retry.Status, error) {
r, err := db.Inc(keys.NodeIDGenerator, 1)
if err != nil {
status := retry.Break
if _, ok := err.(util.Retryable); ok {
status = retry.Continue
}
return status, util.Errorf("unable to allocate node ID: %s", err)
}
id = proto.NodeID(r.ValueInt())
return retry.Break, nil
})
return id, err
}
// allocateStoreIDs increments the store id generator key for the
// specified node to allocate "inc" new, unique store ids. The
// first ID in a contiguous range is returned on success. The call
// will retry indefinitely on retryable errors.
func allocateStoreIDs(nodeID proto.NodeID, inc int64, db *client.DB) (proto.StoreID, error) {
var id proto.StoreID
err := retry.WithBackoff(allocRetryOptions, func() (retry.Status, error) {
r, err := db.Inc(keys.StoreIDGenerator, inc)
if err != nil {
status := retry.Break
if _, ok := err.(util.Retryable); ok {
status = retry.Continue
}
return status, util.Errorf("unable to allocate %d store IDs for node %d: %s", inc, nodeID, err)
}
id = proto.StoreID(r.ValueInt() - inc + 1)
return retry.Break, nil
})
return id, err
}
// putConfig writes a config for the specified key prefix (which is
// treated as a key). The config is parsed from the input "body". The
// config is stored proto-encoded. The specified body must validly
// parse into a config struct and must pass a given validation check (if
// validate is not nil).
func putConfig(db *client.DB, configPrefix proto.Key, config gogoproto.Message,
path string, body []byte, r *http.Request,
validate func(gogoproto.Message) error) error {
if len(path) == 0 {
return util.Errorf("no path specified for Put")
}
if err := util.UnmarshalRequest(r, body, config, util.AllEncodings); err != nil {
return util.Errorf("config has invalid format: %+v: %s", config, err)
}
if validate != nil {
if err := validate(config); err != nil {
return err
}
}
key := keys.MakeKey(configPrefix, proto.Key(path[1:]))
return db.Put(key, config)
}
func (hv *historyVerifier) runCmds(cmds []*cmd, historyIdx int, db *client.DB, t *testing.T) (string, map[string]int64, error) {
var strs []string
env := map[string]int64{}
err := db.Txn(func(txn *client.Txn) error {
for _, c := range cmds {
c.historyIdx = historyIdx
c.env = env
c.init(nil)
fmtStr, err := c.execute(txn, t)
if err != nil {
return err
}
strs = append(strs, fmt.Sprintf(fmtStr, 0, 0))
}
return nil
})
return strings.Join(strs, " "), env, err
}
// startTestWriter creates a writer which initiates a sequence of
// transactions, each which writes up to 10 times to random keys with
// random values. If not nil, txnChannel is written to non-blockingly
// every time a new transaction starts.
func startTestWriter(db *client.DB, i int64, valBytes int32, wg *sync.WaitGroup, retries *int32,
txnChannel chan struct{}, done <-chan struct{}, t *testing.T) {
src := rand.New(rand.NewSource(i))
defer func() {
if wg != nil {
wg.Done()
}
}()
for j := 0; ; j++ {
select {
case <-done:
return
default:
first := true
err := db.Txn(func(txn *client.Txn) error {
if first && txnChannel != nil {
select {
case txnChannel <- struct{}{}:
default:
}
} else if !first && retries != nil {
atomic.AddInt32(retries, 1)
}
first = false
for j := 0; j <= int(src.Int31n(10)); j++ {
key := randutil.RandBytes(src, 10)
val := randutil.RandBytes(src, int(src.Int31n(valBytes)))
if err := txn.Put(key, val); err != nil {
log.Infof("experienced an error in routine %d: %s", i, err)
return err
}
}
return nil
})
if err != nil {
t.Error(err)
} else {
time.Sleep(1 * time.Millisecond)
}
}
}
}
func (hv *historyVerifier) runTxn(txnIdx int, priority int32,
isolation roachpb.IsolationType, cmds []*cmd, db *client.DB, t *testing.T) error {
var retry int
txnName := fmt.Sprintf("txn%d", txnIdx)
err := db.Txn(func(txn *client.Txn) error {
txn.SetDebugName(txnName, 0)
if isolation == roachpb.SNAPSHOT {
if err := txn.SetIsolation(roachpb.SNAPSHOT); err != nil {
return err
}
}
txn.InternalSetPriority(priority)
env := map[string]int64{}
// TODO(spencer): restarts must create additional histories. They
// look like: given the current partial history and a restart on
// txn txnIdx, re-enumerate a set of all histories containing the
// remaining commands from extant txns and all commands from this
// restarted txn.
// If this is attempt > 1, reset cmds so no waits.
if retry++; retry == 2 {
for _, c := range cmds {
c.done()
}
}
if log.V(2) {
log.Infof("%s, retry=%d", txnName, retry)
}
for i := range cmds {
cmds[i].env = env
if err := hv.runCmd(txn, txnIdx, retry, i, cmds, t); err != nil {
if log.V(1) {
log.Infof("%s encountered error: %s", cmds[i], err)
}
return err
}
}
return nil
})
hv.wg.Done()
return err
}
// GetTableDescriptor retrieves a table descriptor directly from the KV layer.
func GetTableDescriptor(kvDB *client.DB, database string, table string) *TableDescriptor {
dbNameKey := MakeNameMetadataKey(keys.RootNamespaceID, database)
gr, err := kvDB.Get(dbNameKey)
if err != nil {
panic(err)
}
if !gr.Exists() {
panic("database missing")
}
dbDescID := ID(gr.ValueInt())
tableNameKey := MakeNameMetadataKey(dbDescID, table)
gr, err = kvDB.Get(tableNameKey)
if err != nil {
panic(err)
}
if !gr.Exists() {
panic("table missing")
}
descKey := MakeDescMetadataKey(ID(gr.ValueInt()))
desc := &Descriptor{}
if err := kvDB.GetProto(descKey, desc); err != nil {
panic("proto missing")
}
return desc.GetTable()
}
// refreshLease tries to refresh the node's table lease.
func (m *LeaseManager) refreshLease(db *client.DB, id ID, minVersion DescriptorVersion) error {
// Only attempt to update a lease for a table that is already leased.
if t := m.findTableState(id, false); t == nil {
return nil
}
// Acquire and release a lease on the table at a version >= minVersion.
var lease *LeaseState
if pErr := db.Txn(func(txn *client.Txn) *roachpb.Error {
var pErr *roachpb.Error
// Acquire() can only acquire a lease at a version if it has
// already been acquired at that version, or that version
// is the latest version. If the latest version is > minVersion
// then the node acquires a lease at the latest version but
// Acquire() itself returns an error. This is okay, because
// we want to update the node lease.
lease, pErr = m.Acquire(txn, id, minVersion)
return pErr
}); pErr != nil {
return pErr.GoError()
}
return m.Release(lease)
}
// treeNodesEqual compares the expectedTree from the provided key to the actual
// nodes retrieved from the db. It recursively calls itself on both left and
// right children if they exist.
func treeNodesEqual(db *client.DB, expected testRangeTree, key proto.Key) error {
expectedNode, ok := expected.Nodes[string(key)]
if !ok {
return util.Errorf("Expected does not contain a node for %s", key)
}
actualNode := &proto.RangeTreeNode{}
if err := db.GetProto(keys.RangeTreeNodeKey(key), actualNode); err != nil {
return err
}
if err := nodesEqual(key, expectedNode, *actualNode); err != nil {
return err
}
if expectedNode.LeftKey != nil {
if err := treeNodesEqual(db, expected, *expectedNode.LeftKey); err != nil {
return err
}
}
if expectedNode.RightKey != nil {
if err := treeNodesEqual(db, expected, *expectedNode.RightKey); err != nil {
return err
}
}
return nil
}
func (hv *historyVerifier) runHistory(historyIdx int, priorities []int32,
isolations []roachpb.IsolationType, cmds []*cmd, db *client.DB, t *testing.T) error {
plannedStr := historyString(cmds)
if log.V(1) {
log.Infof("attempting iso=%v pri=%v history=%s", isolations, priorities, plannedStr)
}
hv.actual = []string{}
hv.wg.Add(len(priorities))
txnMap := map[int][]*cmd{}
var prev *cmd
for _, c := range cmds {
c.historyIdx = historyIdx
txnMap[c.txnIdx] = append(txnMap[c.txnIdx], c)
c.init(prev)
prev = c
}
for i, txnCmds := range txnMap {
go func(i int, txnCmds []*cmd) {
if err := hv.runTxn(i, priorities[i-1], isolations[i-1], txnCmds, db, t); err != nil {
t.Errorf("(%s): unexpected failure running %s: %v", cmds, cmds[i], err)
}
}(i, txnCmds)
}
hv.wg.Wait()
// Construct string for actual history.
actualStr := strings.Join(hv.actual, " ")
// Verify history.
var verifyStrs []string
verifyEnv := map[string]int64{}
for _, c := range hv.verifyCmds {
c.historyIdx = historyIdx
c.env = verifyEnv
c.init(nil)
pErr := db.Txn(func(txn *client.Txn) *roachpb.Error {
fmtStr, pErr := c.execute(txn, t)
if pErr != nil {
return pErr
}
cmdStr := fmt.Sprintf(fmtStr, 0, 0)
verifyStrs = append(verifyStrs, cmdStr)
return nil
})
if pErr != nil {
t.Errorf("failed on execution of verification cmd %s: %s", c, pErr)
return pErr.GoError()
}
}
err := hv.verify.checkFn(verifyEnv)
if err == nil {
if log.V(1) {
log.Infof("PASSED: iso=%v, pri=%v, history=%q", isolations, priorities, actualStr)
}
}
if hv.expSuccess && err != nil {
verifyStr := strings.Join(verifyStrs, " ")
t.Errorf("%d: iso=%v, pri=%v, history=%q: actual=%q, verify=%q: %s",
historyIdx, isolations, priorities, plannedStr, actualStr, verifyStr, err)
}
return err
}
// concurrentIncrements starts two Goroutines in parallel, both of which
// read the integers stored at the other's key and add it onto their own.
// It is checked that the outcome is serializable, i.e. exactly one of the
// two Goroutines (the later write) sees the previous write by the other.
func concurrentIncrements(db *client.DB, t *testing.T) {
// wgStart waits for all transactions to line up, wgEnd has the main
// function wait for them to finish.
var wgStart, wgEnd sync.WaitGroup
wgStart.Add(2 + 1)
wgEnd.Add(2)
for i := 0; i < 2; i++ {
go func(i int) {
// Read the other key, write key i.
readKey := []byte(fmt.Sprintf(testUser+"/value-%d", (i+1)%2))
writeKey := []byte(fmt.Sprintf(testUser+"/value-%d", i))
defer wgEnd.Done()
wgStart.Done()
// Wait until the other goroutines are running.
wgStart.Wait()
if pErr := db.Txn(func(txn *client.Txn) *roachpb.Error {
txn.SetDebugName(fmt.Sprintf("test-%d", i), 0)
// Retrieve the other key.
gr, pErr := txn.Get(readKey)
if pErr != nil {
return pErr
}
otherValue := int64(0)
if gr.Value != nil {
otherValue = gr.ValueInt()
}
_, pErr = txn.Inc(writeKey, 1+otherValue)
return pErr
}); pErr != nil {
t.Error(pErr)
}
}(i)
}
// Kick the goroutines loose.
wgStart.Done()
// Wait for the goroutines to finish.
wgEnd.Wait()
// Verify that both keys contain something and, more importantly, that
// one key actually contains the value of the first writer and not only
// its own.
total := int64(0)
results := []int64(nil)
for i := 0; i < 2; i++ {
readKey := []byte(fmt.Sprintf(testUser+"/value-%d", i))
gr, pErr := db.Get(readKey)
if pErr != nil {
t.Fatal(pErr)
}
if gr.Value == nil {
t.Fatalf("unexpected empty key: %s=%v", readKey, gr.Value)
}
total += gr.ValueInt()
results = append(results, gr.ValueInt())
}
// First writer should have 1, second one 2
if total != 3 {
t.Fatalf("got unserializable values %v", results)
}
}