Files
seaweedfs/weed/filer/posixlock/manager.go
T
Chris Lu f8caaa4464 mount,filer: re-assert POSIX locks via keepalive (ownership migration + restart) (#9668)
* mount: renew POSIX lock leases via keepalive

The mount tracks the inode keys it holds locks on and a background loop
renews its session lease (KEEP_ALIVE) with each key's owner filer every
5s, within the filer's 15s TTL. A live mount is never reaped; a dead one
stops renewing and owners reclaim its locks. Tracking is a superset:
holds are added on grant and dropped only on owner release, so a still
held lock is never under-renewed.

* mount,filer: re-assert held POSIX locks via keepalive

The owner filer holds POSIX advisory locks as in-memory soft state, so a key's
owner change (ring rebalance) or an owner restart lost or stranded them: the new
or restarted owner was blind to existing holders and would double-grant.

Make the keepalive carry the mount's held lock ranges per key. The mount mirrors
its own granted locks (posixOwn), and each tick re-asserts them to the key's
current owner, which rebuilds that session's locks from the assertion — self
-healing after a takeover or restart. The owner arbitrates re-asserted locks
against other sessions so it never double-grants; a lock that lost a migration
race is reported, not forced. A bare keepalive (no ranges) still just renews.
2026-05-25 01:02:45 -07:00

260 lines
6.7 KiB
Go

package posixlock
import (
"sync"
"time"
)
// Manager is the owner filer's in-memory authority for POSIX advisory locks
// across inodes. Lock state lives here, not in replicated metadata: it is
// transient coordination, so keeping it out of the meta-log avoids churn and
// does not pollute what subscribers see (the distributed lock manager holds its
// locks the same way). A `Set` per inode key, plus a session index so a dead
// mount's locks are reaped in O(locks held) rather than by scanning every inode.
//
// key is an opaque inode identity supplied by the caller — the file's path, or
// "hl:"+hex(HardLinkId) for a hardlinked inode — so all names of one inode share
// a Set. The Manager is safe for concurrent use.
type Manager struct {
mu sync.Mutex
byKey map[string]*Set // inode key -> held locks
bySid map[uint64]map[string]bool // session -> keys it currently holds locks on
lastSeen map[uint64]time.Time // session -> last keepalive; only renewing sessions are leased
}
func NewManager() *Manager {
return &Manager{
byKey: make(map[string]*Set),
bySid: make(map[uint64]map[string]bool),
lastSeen: make(map[uint64]time.Time),
}
}
// Renew records a keepalive from a session, placing it under lease management.
// Only sessions that have renewed are subject to ReapExpired, so a session that
// never sends keepalives (e.g. before the mount keepalive exists) is never reaped.
func (m *Manager) Renew(sid uint64) {
m.mu.Lock()
defer m.mu.Unlock()
m.lastSeen[sid] = time.Now()
}
// ReapExpired releases the locks of every leased session whose last keepalive is
// older than ttl — a dead or partitioned mount. Sessions that never renewed are
// left untouched. Returns the reaped session ids.
func (m *Manager) ReapExpired(ttl time.Duration) []uint64 {
m.mu.Lock()
defer m.mu.Unlock()
cutoff := time.Now().Add(-ttl)
var reaped []uint64
for sid, seen := range m.lastSeen {
if seen.After(cutoff) {
continue
}
for key := range m.bySid[sid] {
s := m.byKey[key]
if s == nil {
continue
}
s.ReleaseSession(sid)
if s.Empty() {
delete(m.byKey, key)
}
}
delete(m.bySid, sid)
delete(m.lastSeen, sid)
reaped = append(reaped, sid)
}
return reaped
}
// TryLock grants lk on key, or returns the conflicting lock and false. The set
// is created on first use and dropped again when it empties.
func (m *Manager) TryLock(key string, lk Range) (Range, bool) {
m.mu.Lock()
defer m.mu.Unlock()
s, ok := m.byKey[key]
if !ok {
s = &Set{}
}
if c, granted := s.Acquire(lk); !granted {
return c, false
}
if !ok {
m.byKey[key] = s
}
m.index(lk.Sid, key)
return Range{}, true
}
// Track records a lock the server already granted, without arbitration. A mount
// uses it to mirror its own held locks so it can re-assert them to the inode's
// current owner filer after an ownership change or owner restart.
func (m *Manager) Track(key string, lk Range) {
m.mu.Lock()
defer m.mu.Unlock()
s, ok := m.byKey[key]
if !ok {
s = &Set{}
m.byKey[key] = s
}
s.Grant(lk)
m.index(lk.Sid, key)
}
// Snapshot returns a copy of the held locks per key. A mount calls it to drive
// re-assertion keepalives; the filer never does.
func (m *Manager) Snapshot() map[string][]Range {
m.mu.Lock()
defer m.mu.Unlock()
out := make(map[string][]Range, len(m.byKey))
for key, s := range m.byKey {
out[key] = append([]Range(nil), s.locks...)
}
return out
}
// Reassert rebuilds session sid's locks on key from the client's authoritative
// list, renewing the lease. It replaces sid's existing locks on the key, then
// re-acquires each asserted lock — arbitrating against other sessions so it never
// double-grants. Locks that lost to another session in a migration window are
// returned as conflicts. The owner filer calls this on a re-assertion keepalive.
func (m *Manager) Reassert(key string, sid uint64, locks []Range) (conflicts []Range) {
m.mu.Lock()
defer m.mu.Unlock()
m.lastSeen[sid] = time.Now()
s := m.byKey[key]
if s == nil {
if len(locks) == 0 {
return nil
}
s = &Set{}
m.byKey[key] = s
}
s.ReleaseSession(sid)
for _, lk := range locks {
lk.Sid = sid
if c, granted := s.Acquire(lk); !granted {
conflicts = append(conflicts, c)
}
}
if setHasSession(s, sid) {
m.index(sid, key)
} else {
m.deindex(sid, key)
}
if s.Empty() {
delete(m.byKey, key)
}
return conflicts
}
// Unlock releases lk's owner's locks within its namespace over lk's range.
func (m *Manager) Unlock(key string, lk Range) {
m.mu.Lock()
defer m.mu.Unlock()
s := m.byKey[key]
if s == nil {
return
}
s.Release(lk)
m.afterRelease(key, s, lk.Sid)
}
// GetLk reports the lock that would block proposed on key, if any.
func (m *Manager) GetLk(key string, proposed Range) (Range, bool) {
m.mu.Lock()
defer m.mu.Unlock()
s := m.byKey[key]
if s == nil {
return Range{}, false
}
return s.Conflict(proposed)
}
// ReleasePosixOwner drops (sid, owner)'s fcntl locks on key — the flush-time path.
func (m *Manager) ReleasePosixOwner(key string, sid, owner uint64) {
m.mu.Lock()
defer m.mu.Unlock()
s := m.byKey[key]
if s == nil {
return
}
s.ReleasePosixOwner(sid, owner)
m.afterRelease(key, s, sid)
}
// ReleaseFlockOwner drops (sid, owner)'s flock locks on key — the release-time path.
func (m *Manager) ReleaseFlockOwner(key string, sid, owner uint64) {
m.mu.Lock()
defer m.mu.Unlock()
s := m.byKey[key]
if s == nil {
return
}
s.ReleaseFlockOwner(sid, owner)
m.afterRelease(key, s, sid)
}
// ReleaseSession drops every lock held by a session across all inodes it touched,
// reaping a mount whose lease expired. O(locks held by the session).
func (m *Manager) ReleaseSession(sid uint64) {
m.mu.Lock()
defer m.mu.Unlock()
for key := range m.bySid[sid] {
s := m.byKey[key]
if s == nil {
continue
}
s.ReleaseSession(sid)
if s.Empty() {
delete(m.byKey, key)
}
}
delete(m.bySid, sid)
}
// afterRelease prunes the session index when sid no longer holds any lock on key,
// and drops the set when it empties. Only sid's presence can have changed, since
// a release only removes sid's locks.
func (m *Manager) afterRelease(key string, s *Set, sid uint64) {
if s.Empty() {
delete(m.byKey, key)
m.deindex(sid, key)
return
}
if !setHasSession(s, sid) {
m.deindex(sid, key)
}
}
func (m *Manager) index(sid uint64, key string) {
keys := m.bySid[sid]
if keys == nil {
keys = make(map[string]bool)
m.bySid[sid] = keys
}
keys[key] = true
}
func (m *Manager) deindex(sid uint64, key string) {
keys := m.bySid[sid]
if keys == nil {
return
}
delete(keys, key)
if len(keys) == 0 {
delete(m.bySid, sid)
}
}
func setHasSession(s *Set, sid uint64) bool {
for _, l := range s.Locks() {
if l.Sid == sid {
return true
}
}
return false
}