| 1 | // SPDX-License-Identifier: GPL-3.0-or-later |
| 2 | |
| 3 | package functions |
| 4 | |
| 5 | import ( |
| 6 | "errors" |
| 7 | "sync" |
| 8 | ) |
| 9 | |
| 10 | var ( |
| 11 | errSchedulerStopping = errors.New("scheduler is stopping") |
| 12 | errSchedulerInvalid = errors.New("scheduler invalid request") |
| 13 | ) |
| 14 | |
| 15 | type scheduleLane struct { |
| 16 | ownerUID string |
| 17 | queue []*invocationRequest |
| 18 | } |
| 19 | |
| 20 | // keyScheduler serializes execution by schedule key while allowing concurrency |
| 21 | // across different keys. |
| 22 | type keyScheduler struct { |
| 23 | mux sync.Mutex |
| 24 | |
| 25 | cond *sync.Cond |
| 26 | |
| 27 | ready []*invocationRequest |
| 28 | lanes map[string]*scheduleLane |
| 29 | |
| 30 | maxPending int |
| 31 | pending int |
| 32 | accepting bool |
| 33 | stopping bool |
| 34 | |
| 35 | // enqueueWaiters counts goroutines currently blocked inside enqueue() |
| 36 | // waiting for space. Used by tests to synchronize deterministically |
| 37 | // instead of sleeping. |
| 38 | enqueueWaiters int |
| 39 | } |
| 40 | |
| 41 | func newKeyScheduler(maxPending int) *keyScheduler { |
| 42 | s := &keyScheduler{ |
| 43 | lanes: make(map[string]*scheduleLane), |
| 44 | maxPending: maxPending, |
| 45 | accepting: true, |
| 46 | } |
| 47 | s.cond = sync.NewCond(&s.mux) |
| 48 | return s |
| 49 | } |
| 50 | |
| 51 | func (s *keyScheduler) enqueue(req *invocationRequest) error { |
| 52 | if req == nil || req.fn == nil || req.fn.UID == "" || req.scheduleKey == "" { |
| 53 | return errSchedulerInvalid |
| 54 | } |
| 55 | |
| 56 | s.mux.Lock() |
| 57 | defer s.mux.Unlock() |
| 58 | |
| 59 | // Block until there is space, or the scheduler is stopped. We must not |
| 60 | // drop dyncfg commands silently: an awaited enable/disable that is dropped |
| 61 | // here would wedge jobmgr's wait gate by leaving it waiting for a |
| 62 | // completion that will never arrive. Back-pressure flows upstream: the |
| 63 | // manager run-loop stops draining stdin, and netdata's write blocks on |
| 64 | // the OS pipe. |
| 65 | for { |
| 66 | if s.stopping || !s.accepting { |
| 67 | return errSchedulerStopping |
| 68 | } |
| 69 | if s.maxPending <= 0 || s.pending < s.maxPending { |
| 70 | break |
| 71 | } |
| 72 | s.enqueueWaiters++ |
| 73 | s.cond.Wait() |
| 74 | s.enqueueWaiters-- |
| 75 | } |
| 76 | |
| 77 | lane := s.lanes[req.scheduleKey] |
| 78 | if lane == nil { |
| 79 | lane = &scheduleLane{} |
| 80 | s.lanes[req.scheduleKey] = lane |
| 81 | } |
| 82 | |
| 83 | s.pending++ |
| 84 | if lane.ownerUID == "" { |
| 85 | lane.ownerUID = req.fn.UID |
| 86 | s.ready = append(s.ready, req) |
| 87 | s.cond.Signal() |
| 88 | return nil |
| 89 | } |
| 90 | |
| 91 | lane.queue = append(lane.queue, req) |
| 92 | return nil |
| 93 | } |
| 94 | |
| 95 | func (s *keyScheduler) next() (*invocationRequest, bool) { |
| 96 | s.mux.Lock() |
| 97 | defer s.mux.Unlock() |
| 98 | |
| 99 | for len(s.ready) == 0 && !s.stopping { |
| 100 | if s.drainedLocked() { |
| 101 | return nil, false |
| 102 | } |
| 103 | s.cond.Wait() |
| 104 | } |
| 105 | |
| 106 | if len(s.ready) == 0 || s.stopping { |
| 107 | return nil, false |
| 108 | } |
| 109 | |
| 110 | req := s.ready[0] |
| 111 | s.ready = s.ready[1:] |
| 112 | if s.pending > 0 { |
| 113 | s.pending-- |
| 114 | } |
| 115 | // Wake any enqueue() waiters blocked on a full queue. |
| 116 | s.cond.Broadcast() |
| 117 | return req, true |
| 118 | } |
| 119 | |
| 120 | func (s *keyScheduler) cancelQueued(scheduleKey, uid string) bool { |
| 121 | if scheduleKey == "" || uid == "" { |
| 122 | return false |
| 123 | } |
| 124 | |
| 125 | s.mux.Lock() |
| 126 | defer s.mux.Unlock() |
| 127 | |
| 128 | lane := s.lanes[scheduleKey] |
| 129 | if lane == nil || len(lane.queue) == 0 { |
| 130 | return false |
| 131 | } |
| 132 | |
| 133 | for i, req := range lane.queue { |
| 134 | if req == nil || req.fn == nil || req.fn.UID != uid { |
| 135 | continue |
| 136 | } |
| 137 | |
| 138 | copy(lane.queue[i:], lane.queue[i+1:]) |
| 139 | lane.queue = lane.queue[:len(lane.queue)-1] |
| 140 | if s.pending > 0 { |
| 141 | s.pending-- |
| 142 | } |
| 143 | |
| 144 | if lane.ownerUID == "" && len(lane.queue) == 0 { |
| 145 | delete(s.lanes, scheduleKey) |
| 146 | } |
| 147 | // Wake any enqueue() waiters blocked on a full queue. |
| 148 | s.cond.Broadcast() |
| 149 | return true |
| 150 | } |
| 151 | return false |
| 152 | } |
| 153 | |
| 154 | func (s *keyScheduler) complete(scheduleKey, uid string) { |
| 155 | if scheduleKey == "" || uid == "" { |
| 156 | return |
| 157 | } |
| 158 | |
| 159 | s.mux.Lock() |
| 160 | defer s.mux.Unlock() |
| 161 | |
| 162 | lane := s.lanes[scheduleKey] |
| 163 | if lane == nil || lane.ownerUID != uid { |
| 164 | return |
| 165 | } |
| 166 | |
| 167 | if s.removeReadyLocked(uid) && s.pending > 0 { |
| 168 | s.pending-- |
| 169 | } |
| 170 | |
| 171 | if s.stopping { |
| 172 | delete(s.lanes, scheduleKey) |
| 173 | s.cond.Broadcast() |
| 174 | return |
| 175 | } |
| 176 | |
| 177 | for len(lane.queue) > 0 { |
| 178 | next := lane.queue[0] |
| 179 | lane.queue = lane.queue[1:] |
| 180 | if next == nil || next.fn == nil { |
| 181 | if s.pending > 0 { |
| 182 | s.pending-- |
| 183 | } |
| 184 | continue |
| 185 | } |
| 186 | |
| 187 | lane.ownerUID = next.fn.UID |
| 188 | s.ready = append(s.ready, next) |
| 189 | // Broadcast: wakes both next() consumers and enqueue() producers. |
| 190 | s.cond.Broadcast() |
| 191 | return |
| 192 | } |
| 193 | |
| 194 | lane.ownerUID = "" |
| 195 | if len(lane.queue) == 0 { |
| 196 | delete(s.lanes, scheduleKey) |
| 197 | } |
| 198 | s.cond.Broadcast() |
| 199 | } |
| 200 | |
| 201 | func (s *keyScheduler) stopAccepting() { |
| 202 | s.mux.Lock() |
| 203 | s.accepting = false |
| 204 | if s.drainedLocked() { |
| 205 | s.cond.Broadcast() |
| 206 | } |
| 207 | s.mux.Unlock() |
| 208 | } |
| 209 | |
| 210 | func (s *keyScheduler) stop() { |
| 211 | s.mux.Lock() |
| 212 | s.stopping = true |
| 213 | s.cond.Broadcast() |
| 214 | s.mux.Unlock() |
| 215 | } |
| 216 | |
| 217 | func (s *keyScheduler) drainedLocked() bool { |
| 218 | return !s.accepting && s.pending == 0 && len(s.ready) == 0 |
| 219 | } |
| 220 | |
| 221 | func (s *keyScheduler) removeReadyLocked(uid string) bool { |
| 222 | for i, req := range s.ready { |
| 223 | if req == nil || req.fn == nil || req.fn.UID != uid { |
| 224 | continue |
| 225 | } |
| 226 | |
| 227 | copy(s.ready[i:], s.ready[i+1:]) |
| 228 | s.ready = s.ready[:len(s.ready)-1] |
| 229 | return true |
| 230 | } |
| 231 | return false |
| 232 | } |
| 233 | |
| 234 | func (s *keyScheduler) pendingCount() int { |
| 235 | s.mux.Lock() |
| 236 | defer s.mux.Unlock() |
| 237 | return s.pending |
| 238 | } |
| 239 | |
| 240 | // enqueueWaiterCount reports how many goroutines are currently blocked |
| 241 | // inside enqueue() waiting for queue space. Intended for tests. |
| 242 | func (s *keyScheduler) enqueueWaiterCount() int { |
| 243 | s.mux.Lock() |
| 244 | defer s.mux.Unlock() |
| 245 | return s.enqueueWaiters |
| 246 | } |