p2p/dnsdisc: fix hot-spin when all trees are empty (#22313)

In the random sync algorithm used by the DNS node iterator, we first pick a random
tree and then perform one sync action on that tree. This happens in a loop until any
node is found. If no trees contain any nodes, the iterator will enter a hot loop spinning
at 100% CPU.

The fix is complicated. The iterator now checks if a meaningful sync action can
be performed on any tree. If there is nothing to do, it waits for the next root record
recheck time to arrive and then tries again.

Fixes #22306
This commit is contained in:
Felix Lange 2021-02-19 09:54:46 +01:00 committed by GitHub
parent 6ec1561044
commit d36276d85e
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
3 changed files with 140 additions and 21 deletions

@ -217,8 +217,11 @@ type randomIterator struct {
c *Client
mu sync.Mutex
trees map[string]*clientTree // all trees
lc linkCache // tracks tree dependencies
trees map[string]*clientTree // all trees
// buffers for syncableTrees
syncableList []*clientTree
disabledList []*clientTree
}
func (c *Client) newRandomIterator() *randomIterator {
@ -238,10 +241,10 @@ func (it *randomIterator) Node() *enode.Node {
// Close closes the iterator.
func (it *randomIterator) Close() {
it.cancelFn()
it.mu.Lock()
defer it.mu.Unlock()
it.cancelFn()
it.trees = nil
}
@ -264,7 +267,7 @@ func (it *randomIterator) addTree(url string) error {
// nextNode syncs random tree entries until it finds a node.
func (it *randomIterator) nextNode() *enode.Node {
for {
ct := it.nextTree()
ct := it.pickTree()
if ct == nil {
return nil
}
@ -282,26 +285,79 @@ func (it *randomIterator) nextNode() *enode.Node {
}
}
// nextTree returns a random tree.
func (it *randomIterator) nextTree() *clientTree {
// pickTree returns a random tree to sync from.
func (it *randomIterator) pickTree() *clientTree {
it.mu.Lock()
defer it.mu.Unlock()
// Rebuild the trees map if any links have changed.
if it.lc.changed {
it.rebuildTrees()
it.lc.changed = false
}
if len(it.trees) == 0 {
for {
canSync, trees := it.syncableTrees()
switch {
case canSync:
// Pick a random tree.
return trees[rand.Intn(len(trees))]
case len(trees) > 0:
// No sync action can be performed on any tree right now. The only meaningful
// thing to do is waiting for any root record to get updated.
if !it.waitForRootUpdates(trees) {
// Iterator was closed while waiting.
return nil
}
limit := rand.Intn(len(it.trees))
default:
// There are no trees left, the iterator was closed.
return nil
}
}
}
// syncableTrees finds trees on which any meaningful sync action can be performed.
func (it *randomIterator) syncableTrees() (canSync bool, trees []*clientTree) {
// Resize tree lists.
it.syncableList = it.syncableList[:0]
it.disabledList = it.disabledList[:0]
// Partition them into the two lists.
for _, ct := range it.trees {
if limit == 0 {
return ct
if ct.canSyncRandom() {
it.syncableList = append(it.syncableList, ct)
} else {
it.disabledList = append(it.disabledList, ct)
}
limit--
}
return nil
if len(it.syncableList) > 0 {
return true, it.syncableList
}
return false, it.disabledList
}
// waitForRootUpdates waits for the closest scheduled root check time on the given trees.
func (it *randomIterator) waitForRootUpdates(trees []*clientTree) bool {
var minTree *clientTree
var nextCheck mclock.AbsTime
for _, ct := range trees {
check := ct.nextScheduledRootCheck()
if minTree == nil || check < nextCheck {
minTree = ct
nextCheck = check
}
}
sleep := nextCheck.Sub(it.c.clock.Now())
it.c.cfg.Logger.Debug("DNS iterator waiting for root updates", "sleep", sleep, "tree", minTree.loc.domain)
timeout := it.c.clock.NewTimer(sleep)
defer timeout.Stop()
select {
case <-timeout.C():
return true
case <-it.ctx.Done():
return false // Iterator was closed.
}
}
// rebuildTrees rebuilds the 'trees' map.

@ -231,6 +231,53 @@ func TestIteratorRootRecheckOnFail(t *testing.T) {
checkIterator(t, it, nodes)
}
// This test checks that the iterator works correctly when the tree is initially empty.
func TestIteratorEmptyTree(t *testing.T) {
var (
clock = new(mclock.Simulated)
nodes = testNodes(nodesSeed1, 1)
resolver = newMapResolver()
c = NewClient(Config{
Resolver: resolver,
Logger: testlog.Logger(t, log.LvlTrace),
RecheckInterval: 20 * time.Minute,
RateLimit: 500,
})
)
c.clock = clock
tree1, url := makeTestTree("n", nil, nil)
tree2, _ := makeTestTree("n", nodes, nil)
resolver.add(tree1.ToTXT("n"))
// Start the iterator.
node := make(chan *enode.Node)
it, err := c.NewIterator(url)
if err != nil {
t.Fatal(err)
}
go func() {
it.Next()
node <- it.Node()
}()
// Wait for the client to get stuck in waitForRootUpdates.
clock.WaitForTimers(1)
// Now update the root.
resolver.add(tree2.ToTXT("n"))
// Wait for it to pick up the root change.
clock.Run(c.cfg.RecheckInterval)
select {
case n := <-node:
if n.ID() != nodes[0].ID() {
t.Fatalf("wrong node returned")
}
case <-time.After(5 * time.Second):
t.Fatal("it.Next() did not unblock within 5s of real time")
}
}
// updateSomeNodes applies ENR updates to some of the given nodes.
func updateSomeNodes(keySeed int64, nodes []*enode.Node) {
keys := testKeys(nodesSeed1, len(nodes))

@ -25,9 +25,9 @@ import (
"github.com/ethereum/go-ethereum/p2p/enode"
)
const (
rootRecheckFailCount = 5 // update root if this many leaf requests fail
)
// This is the number of consecutive leaf requests that may fail before
// we consider re-resolving the tree root.
const rootRecheckFailCount = 5
// clientTree is a full tree being synced.
type clientTree struct {
@ -89,13 +89,22 @@ func (ct *clientTree) syncRandom(ctx context.Context) (n *enode.Node, err error)
ct.gcLinks()
// Sync next random entry in ENR tree. Once every node has been visited, we simply
// start over. This is fine because entries are cached.
// start over. This is fine because entries are cached internally by the client LRU
// also by DNS resolvers.
if ct.enrs.done() {
ct.enrs = newSubtreeSync(ct.c, ct.loc, ct.root.eroot, false)
}
return ct.syncNextRandomENR(ctx)
}
// canSyncRandom checks if any meaningful action can be performed by syncRandom.
func (ct *clientTree) canSyncRandom() bool {
// Note: the check for non-zero leaf count is very important here.
// If we're done syncing all nodes, and no leaves were found, the tree
// is empty and we can't use it for sync.
return ct.rootUpdateDue() || !ct.links.done() || !ct.enrs.done() || ct.enrs.leaves != 0
}
// gcLinks removes outdated links from the global link cache. GC runs once
// when the link sync finishes.
func (ct *clientTree) gcLinks() {
@ -184,10 +193,14 @@ func (ct *clientTree) updateRoot(ctx context.Context) error {
// rootUpdateDue returns true when a root update is needed.
func (ct *clientTree) rootUpdateDue() bool {
tooManyFailures := ct.leafFailCount > rootRecheckFailCount
scheduledCheck := ct.c.clock.Now().Sub(ct.lastRootCheck) > ct.c.cfg.RecheckInterval
scheduledCheck := ct.c.clock.Now() >= ct.nextScheduledRootCheck()
return ct.root == nil || tooManyFailures || scheduledCheck
}
func (ct *clientTree) nextScheduledRootCheck() mclock.AbsTime {
return ct.lastRootCheck.Add(ct.c.cfg.RecheckInterval)
}
// slowdownRootUpdate applies a delay to root resolution if is tried
// too frequently. This avoids busy polling when the client is offline.
// Returns true if the timeout passed, false if sync was canceled.
@ -218,10 +231,11 @@ type subtreeSync struct {
root string
missing []string // missing tree node hashes
link bool // true if this sync is for the link tree
leaves int // counter of synced leaves
}
func newSubtreeSync(c *Client, loc *linkEntry, root string, link bool) *subtreeSync {
return &subtreeSync{c, loc, root, []string{root}, link}
return &subtreeSync{c, loc, root, []string{root}, link, 0}
}
func (ts *subtreeSync) done() bool {
@ -253,10 +267,12 @@ func (ts *subtreeSync) resolveNext(ctx context.Context, hash string) (entry, err
if ts.link {
return nil, errENRInLinkTree
}
ts.leaves++
case *linkEntry:
if !ts.link {
return nil, errLinkInENRTree
}
ts.leaves++
case *branchEntry:
ts.missing = append(ts.missing, e.children...)
}