fix: lock maintenance should honor quorum (#9138)

The staleness of a lock should be determined by
the quorum number of entries returning stale,
this allows for situations when locks are held
when nodes are down - we don't accidentally
clear locks unintentionally when they are valid
and correct.

Also lock maintenance should be run by all servers,
not one server, stale locks need to be run outside
the requirement for holding distributed locks.

Thanks @klauspost for reproducing this issue
This commit is contained in:
Harshavardhana 2020-03-15 11:55:52 -07:00 committed by GitHub
parent 10fd53d6bb
commit c9212819af
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23

View file

@ -225,8 +225,6 @@ func getLongLivedLocks(interval time.Duration) map[Endpoint][]nameLockRequesterI
return nlripMap
}
var lockMaintenanceTimeout = newDynamicTimeout(60*time.Second, time.Second)
// lockMaintenance loops over locks that have been active for some time and checks back
// with the original server whether it is still alive or not
//
@ -235,28 +233,19 @@ var lockMaintenanceTimeout = newDynamicTimeout(60*time.Second, time.Second)
// - some network error (and server is up normally)
//
// We will ignore the error, and we will retry later to get a resolve on this lock
func lockMaintenance(ctx context.Context, interval time.Duration, objAPI ObjectLayer) error {
// Lock to avoid concurrent lock maintenance loops
maintenanceLock := objAPI.NewNSLock(ctx, "system", "lock-maintenance-ops")
if err := maintenanceLock.GetLock(lockMaintenanceTimeout); err != nil {
return err
}
defer maintenanceLock.Unlock()
func lockMaintenance(ctx context.Context, interval time.Duration) error {
// Validate if long lived locks are indeed clean.
// Get list of long lived locks to check for staleness.
for lendpoint, nlrips := range getLongLivedLocks(interval) {
nlripsMap := make(map[string]int, len(nlrips))
for _, nlrip := range nlrips {
// Locks are only held on first zone, make sure that
// we only look for ownership of locks from endpoints
// on first zone.
for _, endpoint := range globalEndpoints[0].Endpoints {
if endpoint.String() == lendpoint.String() {
continue
}
c := newLockAPI(endpoint)
if !c.IsOnline() {
nlripsMap[nlrip.name]++
continue
}
@ -266,25 +255,37 @@ func lockMaintenance(ctx context.Context, interval time.Duration, objAPI ObjectL
UID: nlrip.lri.UID,
Resources: []string{nlrip.name},
})
if err != nil {
nlripsMap[nlrip.name]++
c.Close()
continue
}
// For successful response, verify if lock was indeed active or stale.
if expired {
// The lock is no longer active at server that originated
// the lock, attempt to remove the lock.
globalLockServers[lendpoint].mutex.Lock()
// Purge the stale entry if it exists.
globalLockServers[lendpoint].removeEntryIfExists(nlrip)
globalLockServers[lendpoint].mutex.Unlock()
if !expired {
nlripsMap[nlrip.name]++
}
// Close the connection regardless of the call response.
c.Close()
}
// Read locks we assume quorum for be N/2 success
quorum := globalXLSetDriveCount / 2
if nlrip.lri.Writer {
// For write locks we need N/2+1 success
quorum = globalXLSetDriveCount/2 + 1
}
// less than the quorum, we have locks expired.
if nlripsMap[nlrip.name] < quorum {
// The lock is no longer active at server that originated
// the lock, attempt to remove the lock.
globalLockServers[lendpoint].mutex.Lock()
// Purge the stale entry if it exists.
globalLockServers[lendpoint].removeEntryIfExists(nlrip)
globalLockServers[lendpoint].mutex.Unlock()
}
}
}
@ -293,12 +294,13 @@ func lockMaintenance(ctx context.Context, interval time.Duration, objAPI ObjectL
// Start lock maintenance from all lock servers.
func startLockMaintenance() {
var objAPI ObjectLayer
var ctx = context.Background()
// Wait until the object API is ready
// no need to start the lock maintenance
// if ObjectAPI is not initialized.
for {
objAPI = newObjectLayerWithoutSafeModeFn()
objAPI := newObjectLayerWithoutSafeModeFn()
if objAPI == nil {
time.Sleep(time.Second)
continue
@ -322,7 +324,7 @@ func startLockMaintenance() {
// "synchronous checks" between servers
duration := time.Duration(r.Float64() * float64(lockMaintenanceInterval))
time.Sleep(duration)
if err := lockMaintenance(ctx, lockValidityCheckInterval, objAPI); err != nil {
if err := lockMaintenance(ctx, lockValidityCheckInterval); err != nil {
// Sleep right after an error.
duration := time.Duration(r.Float64() * float64(lockMaintenanceInterval))
time.Sleep(duration)