• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  *
3  * Copyright 2015-2016 gRPC authors.
4  *
5  * Licensed under the Apache License, Version 2.0 (the "License");
6  * you may not use this file except in compliance with the License.
7  * You may obtain a copy of the License at
8  *
9  *     http://www.apache.org/licenses/LICENSE-2.0
10  *
11  * Unless required by applicable law or agreed to in writing, software
12  * distributed under the License is distributed on an "AS IS" BASIS,
13  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
14  * See the License for the specific language governing permissions and
15  * limitations under the License.
16  *
17  */
18 #include <grpc/support/port_platform.h>
19 
20 #include "src/core/lib/surface/completion_queue.h"
21 
22 #include <inttypes.h>
23 #include <stdio.h>
24 #include <string.h>
25 
26 #include <vector>
27 
28 #include "absl/strings/str_format.h"
29 #include "absl/strings/str_join.h"
30 
31 #include <grpc/support/alloc.h>
32 #include <grpc/support/atm.h>
33 #include <grpc/support/log.h>
34 #include <grpc/support/string_util.h>
35 #include <grpc/support/time.h>
36 
37 #include "src/core/lib/debug/stats.h"
38 #include "src/core/lib/gpr/spinlock.h"
39 #include "src/core/lib/gpr/string.h"
40 #include "src/core/lib/gpr/tls.h"
41 #include "src/core/lib/gprpp/atomic.h"
42 #include "src/core/lib/iomgr/executor.h"
43 #include "src/core/lib/iomgr/pollset.h"
44 #include "src/core/lib/iomgr/timer.h"
45 #include "src/core/lib/profiling/timers.h"
46 #include "src/core/lib/surface/api_trace.h"
47 #include "src/core/lib/surface/call.h"
48 #include "src/core/lib/surface/event_string.h"
49 
50 grpc_core::TraceFlag grpc_trace_operation_failures(false, "op_failure");
51 grpc_core::DebugOnlyTraceFlag grpc_trace_pending_tags(false, "pending_tags");
52 grpc_core::DebugOnlyTraceFlag grpc_trace_cq_refcount(false, "cq_refcount");
53 
54 namespace {
55 
56 // Specifies a cq thread local cache.
57 // The first event that occurs on a thread
58 // with a cq cache will go into that cache, and
59 // will only be returned on the thread that initialized the cache.
60 // NOTE: Only one event will ever be cached.
61 GPR_TLS_DECL(g_cached_event);
62 GPR_TLS_DECL(g_cached_cq);
63 
64 struct plucker {
65   grpc_pollset_worker** worker;
66   void* tag;
67 };
68 struct cq_poller_vtable {
69   bool can_get_pollset;
70   bool can_listen;
71   size_t (*size)(void);
72   void (*init)(grpc_pollset* pollset, gpr_mu** mu);
73   grpc_error* (*kick)(grpc_pollset* pollset,
74                       grpc_pollset_worker* specific_worker);
75   grpc_error* (*work)(grpc_pollset* pollset, grpc_pollset_worker** worker,
76                       grpc_millis deadline);
77   void (*shutdown)(grpc_pollset* pollset, grpc_closure* closure);
78   void (*destroy)(grpc_pollset* pollset);
79 };
80 typedef struct non_polling_worker {
81   gpr_cv cv;
82   bool kicked;
83   struct non_polling_worker* next;
84   struct non_polling_worker* prev;
85 } non_polling_worker;
86 
87 struct non_polling_poller {
88   gpr_mu mu;
89   bool kicked_without_poller;
90   non_polling_worker* root;
91   grpc_closure* shutdown;
92 };
non_polling_poller_size(void)93 size_t non_polling_poller_size(void) { return sizeof(non_polling_poller); }
94 
non_polling_poller_init(grpc_pollset * pollset,gpr_mu ** mu)95 void non_polling_poller_init(grpc_pollset* pollset, gpr_mu** mu) {
96   non_polling_poller* npp = reinterpret_cast<non_polling_poller*>(pollset);
97   gpr_mu_init(&npp->mu);
98   *mu = &npp->mu;
99 }
100 
non_polling_poller_destroy(grpc_pollset * pollset)101 void non_polling_poller_destroy(grpc_pollset* pollset) {
102   non_polling_poller* npp = reinterpret_cast<non_polling_poller*>(pollset);
103   gpr_mu_destroy(&npp->mu);
104 }
105 
non_polling_poller_work(grpc_pollset * pollset,grpc_pollset_worker ** worker,grpc_millis deadline)106 grpc_error* non_polling_poller_work(grpc_pollset* pollset,
107                                     grpc_pollset_worker** worker,
108                                     grpc_millis deadline) {
109   non_polling_poller* npp = reinterpret_cast<non_polling_poller*>(pollset);
110   if (npp->shutdown) return GRPC_ERROR_NONE;
111   if (npp->kicked_without_poller) {
112     npp->kicked_without_poller = false;
113     return GRPC_ERROR_NONE;
114   }
115   non_polling_worker w;
116   gpr_cv_init(&w.cv);
117   if (worker != nullptr) *worker = reinterpret_cast<grpc_pollset_worker*>(&w);
118   if (npp->root == nullptr) {
119     npp->root = w.next = w.prev = &w;
120   } else {
121     w.next = npp->root;
122     w.prev = w.next->prev;
123     w.next->prev = w.prev->next = &w;
124   }
125   w.kicked = false;
126   gpr_timespec deadline_ts =
127       grpc_millis_to_timespec(deadline, GPR_CLOCK_MONOTONIC);
128   while (!npp->shutdown && !w.kicked &&
129          !gpr_cv_wait(&w.cv, &npp->mu, deadline_ts)) {
130   }
131   grpc_core::ExecCtx::Get()->InvalidateNow();
132   if (&w == npp->root) {
133     npp->root = w.next;
134     if (&w == npp->root) {
135       if (npp->shutdown) {
136         grpc_core::ExecCtx::Run(DEBUG_LOCATION, npp->shutdown, GRPC_ERROR_NONE);
137       }
138       npp->root = nullptr;
139     }
140   }
141   w.next->prev = w.prev;
142   w.prev->next = w.next;
143   gpr_cv_destroy(&w.cv);
144   if (worker != nullptr) *worker = nullptr;
145   return GRPC_ERROR_NONE;
146 }
147 
non_polling_poller_kick(grpc_pollset * pollset,grpc_pollset_worker * specific_worker)148 grpc_error* non_polling_poller_kick(grpc_pollset* pollset,
149                                     grpc_pollset_worker* specific_worker) {
150   non_polling_poller* p = reinterpret_cast<non_polling_poller*>(pollset);
151   if (specific_worker == nullptr) {
152     specific_worker = reinterpret_cast<grpc_pollset_worker*>(p->root);
153   }
154   if (specific_worker != nullptr) {
155     non_polling_worker* w =
156         reinterpret_cast<non_polling_worker*>(specific_worker);
157     if (!w->kicked) {
158       w->kicked = true;
159       gpr_cv_signal(&w->cv);
160     }
161   } else {
162     p->kicked_without_poller = true;
163   }
164   return GRPC_ERROR_NONE;
165 }
166 
non_polling_poller_shutdown(grpc_pollset * pollset,grpc_closure * closure)167 void non_polling_poller_shutdown(grpc_pollset* pollset, grpc_closure* closure) {
168   non_polling_poller* p = reinterpret_cast<non_polling_poller*>(pollset);
169   GPR_ASSERT(closure != nullptr);
170   p->shutdown = closure;
171   if (p->root == nullptr) {
172     grpc_core::ExecCtx::Run(DEBUG_LOCATION, closure, GRPC_ERROR_NONE);
173   } else {
174     non_polling_worker* w = p->root;
175     do {
176       gpr_cv_signal(&w->cv);
177       w = w->next;
178     } while (w != p->root);
179   }
180 }
181 
182 const cq_poller_vtable g_poller_vtable_by_poller_type[] = {
183     /* GRPC_CQ_DEFAULT_POLLING */
184     {true, true, grpc_pollset_size, grpc_pollset_init, grpc_pollset_kick,
185      grpc_pollset_work, grpc_pollset_shutdown, grpc_pollset_destroy},
186     /* GRPC_CQ_NON_LISTENING */
187     {true, false, grpc_pollset_size, grpc_pollset_init, grpc_pollset_kick,
188      grpc_pollset_work, grpc_pollset_shutdown, grpc_pollset_destroy},
189     /* GRPC_CQ_NON_POLLING */
190     {false, false, non_polling_poller_size, non_polling_poller_init,
191      non_polling_poller_kick, non_polling_poller_work,
192      non_polling_poller_shutdown, non_polling_poller_destroy},
193 };
194 
195 }  // namespace
196 
197 struct cq_vtable {
198   grpc_cq_completion_type cq_completion_type;
199   size_t data_size;
200   void (*init)(void* data,
201                grpc_experimental_completion_queue_functor* shutdown_callback);
202   void (*shutdown)(grpc_completion_queue* cq);
203   void (*destroy)(void* data);
204   bool (*begin_op)(grpc_completion_queue* cq, void* tag);
205   void (*end_op)(grpc_completion_queue* cq, void* tag, grpc_error* error,
206                  void (*done)(void* done_arg, grpc_cq_completion* storage),
207                  void* done_arg, grpc_cq_completion* storage, bool internal);
208   grpc_event (*next)(grpc_completion_queue* cq, gpr_timespec deadline,
209                      void* reserved);
210   grpc_event (*pluck)(grpc_completion_queue* cq, void* tag,
211                       gpr_timespec deadline, void* reserved);
212 };
213 
214 namespace {
215 
216 /* Queue that holds the cq_completion_events. Internally uses
217  * MultiProducerSingleConsumerQueue (a lockfree multiproducer single consumer
218  * queue). It uses a queue_lock to support multiple consumers.
219  * Only used in completion queues whose completion_type is GRPC_CQ_NEXT */
220 class CqEventQueue {
221  public:
222   CqEventQueue() = default;
223   ~CqEventQueue() = default;
224 
225   /* Note: The counter is not incremented/decremented atomically with push/pop.
226    * The count is only eventually consistent */
num_items() const227   intptr_t num_items() const {
228     return num_queue_items_.Load(grpc_core::MemoryOrder::RELAXED);
229   }
230 
231   bool Push(grpc_cq_completion* c);
232   grpc_cq_completion* Pop();
233 
234  private:
235   /* Spinlock to serialize consumers i.e pop() operations */
236   gpr_spinlock queue_lock_ = GPR_SPINLOCK_INITIALIZER;
237 
238   grpc_core::MultiProducerSingleConsumerQueue queue_;
239 
240   /* A lazy counter of number of items in the queue. This is NOT atomically
241      incremented/decremented along with push/pop operations and hence is only
242      eventually consistent */
243   grpc_core::Atomic<intptr_t> num_queue_items_{0};
244 };
245 
246 struct cq_next_data {
~cq_next_data__anon158538fc0211::cq_next_data247   ~cq_next_data() {
248     GPR_ASSERT(queue.num_items() == 0);
249 #ifndef NDEBUG
250     if (pending_events.Load(grpc_core::MemoryOrder::ACQUIRE) != 0) {
251       gpr_log(GPR_ERROR, "Destroying CQ without draining it fully.");
252     }
253 #endif
254   }
255 
256   /** Completed events for completion-queues of type GRPC_CQ_NEXT */
257   CqEventQueue queue;
258 
259   /** Counter of how many things have ever been queued on this completion queue
260       useful for avoiding locks to check the queue */
261   grpc_core::Atomic<intptr_t> things_queued_ever{0};
262 
263   /** Number of outstanding events (+1 if not shut down)
264       Initial count is dropped by grpc_completion_queue_shutdown */
265   grpc_core::Atomic<intptr_t> pending_events{1};
266 
267   /** 0 initially. 1 once we initiated shutdown */
268   bool shutdown_called = false;
269 };
270 
271 struct cq_pluck_data {
cq_pluck_data__anon158538fc0211::cq_pluck_data272   cq_pluck_data() {
273     completed_tail = &completed_head;
274     completed_head.next = reinterpret_cast<uintptr_t>(completed_tail);
275   }
276 
~cq_pluck_data__anon158538fc0211::cq_pluck_data277   ~cq_pluck_data() {
278     GPR_ASSERT(completed_head.next ==
279                reinterpret_cast<uintptr_t>(&completed_head));
280 #ifndef NDEBUG
281     if (pending_events.Load(grpc_core::MemoryOrder::ACQUIRE) != 0) {
282       gpr_log(GPR_ERROR, "Destroying CQ without draining it fully.");
283     }
284 #endif
285   }
286 
287   /** Completed events for completion-queues of type GRPC_CQ_PLUCK */
288   grpc_cq_completion completed_head;
289   grpc_cq_completion* completed_tail;
290 
291   /** Number of pending events (+1 if we're not shutdown).
292       Initial count is dropped by grpc_completion_queue_shutdown. */
293   grpc_core::Atomic<intptr_t> pending_events{1};
294 
295   /** Counter of how many things have ever been queued on this completion queue
296       useful for avoiding locks to check the queue */
297   grpc_core::Atomic<intptr_t> things_queued_ever{0};
298 
299   /** 0 initially. 1 once we completed shutting */
300   /* TODO: (sreek) This is not needed since (shutdown == 1) if and only if
301    * (pending_events == 0). So consider removing this in future and use
302    * pending_events */
303   grpc_core::Atomic<bool> shutdown{false};
304 
305   /** 0 initially. 1 once we initiated shutdown */
306   bool shutdown_called = false;
307 
308   int num_pluckers = 0;
309   plucker pluckers[GRPC_MAX_COMPLETION_QUEUE_PLUCKERS];
310 };
311 
312 struct cq_callback_data {
cq_callback_data__anon158538fc0211::cq_callback_data313   explicit cq_callback_data(
314       grpc_experimental_completion_queue_functor* shutdown_callback)
315       : shutdown_callback(shutdown_callback) {}
316 
~cq_callback_data__anon158538fc0211::cq_callback_data317   ~cq_callback_data() {
318 #ifndef NDEBUG
319     if (pending_events.Load(grpc_core::MemoryOrder::ACQUIRE) != 0) {
320       gpr_log(GPR_ERROR, "Destroying CQ without draining it fully.");
321     }
322 #endif
323   }
324 
325   /** No actual completed events queue, unlike other types */
326 
327   /** Number of pending events (+1 if we're not shutdown).
328       Initial count is dropped by grpc_completion_queue_shutdown. */
329   grpc_core::Atomic<intptr_t> pending_events{1};
330 
331   /** 0 initially. 1 once we initiated shutdown */
332   bool shutdown_called = false;
333 
334   /** A callback that gets invoked when the CQ completes shutdown */
335   grpc_experimental_completion_queue_functor* shutdown_callback;
336 };
337 
338 }  // namespace
339 
340 /* Completion queue structure */
341 struct grpc_completion_queue {
342   /** Once owning_refs drops to zero, we will destroy the cq */
343   grpc_core::RefCount owning_refs;
344 
345   gpr_mu* mu;
346 
347   const cq_vtable* vtable;
348   const cq_poller_vtable* poller_vtable;
349 
350 #ifndef NDEBUG
351   void** outstanding_tags;
352   size_t outstanding_tag_count;
353   size_t outstanding_tag_capacity;
354 #endif
355 
356   grpc_closure pollset_shutdown_done;
357   int num_polls;
358 };
359 
360 /* Forward declarations */
361 static void cq_finish_shutdown_next(grpc_completion_queue* cq);
362 static void cq_finish_shutdown_pluck(grpc_completion_queue* cq);
363 static void cq_finish_shutdown_callback(grpc_completion_queue* cq);
364 static void cq_shutdown_next(grpc_completion_queue* cq);
365 static void cq_shutdown_pluck(grpc_completion_queue* cq);
366 static void cq_shutdown_callback(grpc_completion_queue* cq);
367 
368 static bool cq_begin_op_for_next(grpc_completion_queue* cq, void* tag);
369 static bool cq_begin_op_for_pluck(grpc_completion_queue* cq, void* tag);
370 static bool cq_begin_op_for_callback(grpc_completion_queue* cq, void* tag);
371 
372 // A cq_end_op function is called when an operation on a given CQ with
373 // a given tag has completed. The storage argument is a reference to the
374 // space reserved for this completion as it is placed into the corresponding
375 // queue. The done argument is a callback that will be invoked when it is
376 // safe to free up that storage. The storage MUST NOT be freed until the
377 // done callback is invoked.
378 static void cq_end_op_for_next(
379     grpc_completion_queue* cq, void* tag, grpc_error* error,
380     void (*done)(void* done_arg, grpc_cq_completion* storage), void* done_arg,
381     grpc_cq_completion* storage, bool internal);
382 
383 static void cq_end_op_for_pluck(
384     grpc_completion_queue* cq, void* tag, grpc_error* error,
385     void (*done)(void* done_arg, grpc_cq_completion* storage), void* done_arg,
386     grpc_cq_completion* storage, bool internal);
387 
388 static void cq_end_op_for_callback(
389     grpc_completion_queue* cq, void* tag, grpc_error* error,
390     void (*done)(void* done_arg, grpc_cq_completion* storage), void* done_arg,
391     grpc_cq_completion* storage, bool internal);
392 
393 static grpc_event cq_next(grpc_completion_queue* cq, gpr_timespec deadline,
394                           void* reserved);
395 
396 static grpc_event cq_pluck(grpc_completion_queue* cq, void* tag,
397                            gpr_timespec deadline, void* reserved);
398 
399 // Note that cq_init_next and cq_init_pluck do not use the shutdown_callback
400 static void cq_init_next(
401     void* data, grpc_experimental_completion_queue_functor* shutdown_callback);
402 static void cq_init_pluck(
403     void* data, grpc_experimental_completion_queue_functor* shutdown_callback);
404 static void cq_init_callback(
405     void* data, grpc_experimental_completion_queue_functor* shutdown_callback);
406 static void cq_destroy_next(void* data);
407 static void cq_destroy_pluck(void* data);
408 static void cq_destroy_callback(void* data);
409 
410 /* Completion queue vtables based on the completion-type */
411 static const cq_vtable g_cq_vtable[] = {
412     /* GRPC_CQ_NEXT */
413     {GRPC_CQ_NEXT, sizeof(cq_next_data), cq_init_next, cq_shutdown_next,
414      cq_destroy_next, cq_begin_op_for_next, cq_end_op_for_next, cq_next,
415      nullptr},
416     /* GRPC_CQ_PLUCK */
417     {GRPC_CQ_PLUCK, sizeof(cq_pluck_data), cq_init_pluck, cq_shutdown_pluck,
418      cq_destroy_pluck, cq_begin_op_for_pluck, cq_end_op_for_pluck, nullptr,
419      cq_pluck},
420     /* GRPC_CQ_CALLBACK */
421     {GRPC_CQ_CALLBACK, sizeof(cq_callback_data), cq_init_callback,
422      cq_shutdown_callback, cq_destroy_callback, cq_begin_op_for_callback,
423      cq_end_op_for_callback, nullptr, nullptr},
424 };
425 
426 #define DATA_FROM_CQ(cq) ((void*)((cq) + 1))
427 #define POLLSET_FROM_CQ(cq) \
428   ((grpc_pollset*)((cq)->vtable->data_size + (char*)DATA_FROM_CQ(cq)))
429 
430 grpc_core::TraceFlag grpc_cq_pluck_trace(false, "queue_pluck");
431 
432 #define GRPC_SURFACE_TRACE_RETURNED_EVENT(cq, event)     \
433   do {                                                   \
434     if (GRPC_TRACE_FLAG_ENABLED(grpc_api_trace) &&       \
435         (GRPC_TRACE_FLAG_ENABLED(grpc_cq_pluck_trace) || \
436          (event)->type != GRPC_QUEUE_TIMEOUT)) {         \
437       gpr_log(GPR_INFO, "RETURN_EVENT[%p]: %s", cq,      \
438               grpc_event_string(event).c_str());         \
439     }                                                    \
440   } while (0)
441 
442 static void on_pollset_shutdown_done(void* arg, grpc_error* error);
443 
grpc_cq_global_init()444 void grpc_cq_global_init() {
445   gpr_tls_init(&g_cached_event);
446   gpr_tls_init(&g_cached_cq);
447 }
448 
grpc_completion_queue_thread_local_cache_init(grpc_completion_queue * cq)449 void grpc_completion_queue_thread_local_cache_init(grpc_completion_queue* cq) {
450   if (reinterpret_cast<grpc_completion_queue*>(gpr_tls_get(&g_cached_cq)) ==
451       nullptr) {
452     gpr_tls_set(&g_cached_event, (intptr_t)0);
453     gpr_tls_set(&g_cached_cq, (intptr_t)cq);
454   }
455 }
456 
grpc_completion_queue_thread_local_cache_flush(grpc_completion_queue * cq,void ** tag,int * ok)457 int grpc_completion_queue_thread_local_cache_flush(grpc_completion_queue* cq,
458                                                    void** tag, int* ok) {
459   grpc_cq_completion* storage =
460       reinterpret_cast<grpc_cq_completion*>(gpr_tls_get(&g_cached_event));
461   int ret = 0;
462   if (storage != nullptr && reinterpret_cast<grpc_completion_queue*>(
463                                 gpr_tls_get(&g_cached_cq)) == cq) {
464     *tag = storage->tag;
465     grpc_core::ExecCtx exec_ctx;
466     *ok = (storage->next & static_cast<uintptr_t>(1)) == 1;
467     storage->done(storage->done_arg, storage);
468     ret = 1;
469     cq_next_data* cqd = static_cast<cq_next_data*> DATA_FROM_CQ(cq);
470     if (cqd->pending_events.FetchSub(1, grpc_core::MemoryOrder::ACQ_REL) == 1) {
471       GRPC_CQ_INTERNAL_REF(cq, "shutting_down");
472       gpr_mu_lock(cq->mu);
473       cq_finish_shutdown_next(cq);
474       gpr_mu_unlock(cq->mu);
475       GRPC_CQ_INTERNAL_UNREF(cq, "shutting_down");
476     }
477   }
478   gpr_tls_set(&g_cached_event, (intptr_t)0);
479   gpr_tls_set(&g_cached_cq, (intptr_t)0);
480 
481   return ret;
482 }
483 
Push(grpc_cq_completion * c)484 bool CqEventQueue::Push(grpc_cq_completion* c) {
485   queue_.Push(
486       reinterpret_cast<grpc_core::MultiProducerSingleConsumerQueue::Node*>(c));
487   return num_queue_items_.FetchAdd(1, grpc_core::MemoryOrder::RELAXED) == 0;
488 }
489 
Pop()490 grpc_cq_completion* CqEventQueue::Pop() {
491   grpc_cq_completion* c = nullptr;
492 
493   if (gpr_spinlock_trylock(&queue_lock_)) {
494     GRPC_STATS_INC_CQ_EV_QUEUE_TRYLOCK_SUCCESSES();
495 
496     bool is_empty = false;
497     c = reinterpret_cast<grpc_cq_completion*>(queue_.PopAndCheckEnd(&is_empty));
498     gpr_spinlock_unlock(&queue_lock_);
499 
500     if (c == nullptr && !is_empty) {
501       GRPC_STATS_INC_CQ_EV_QUEUE_TRANSIENT_POP_FAILURES();
502     }
503   } else {
504     GRPC_STATS_INC_CQ_EV_QUEUE_TRYLOCK_FAILURES();
505   }
506 
507   if (c) {
508     num_queue_items_.FetchSub(1, grpc_core::MemoryOrder::RELAXED);
509   }
510 
511   return c;
512 }
513 
grpc_completion_queue_create_internal(grpc_cq_completion_type completion_type,grpc_cq_polling_type polling_type,grpc_experimental_completion_queue_functor * shutdown_callback)514 grpc_completion_queue* grpc_completion_queue_create_internal(
515     grpc_cq_completion_type completion_type, grpc_cq_polling_type polling_type,
516     grpc_experimental_completion_queue_functor* shutdown_callback) {
517   GPR_TIMER_SCOPE("grpc_completion_queue_create_internal", 0);
518 
519   grpc_completion_queue* cq;
520 
521   GRPC_API_TRACE(
522       "grpc_completion_queue_create_internal(completion_type=%d, "
523       "polling_type=%d)",
524       2, (completion_type, polling_type));
525 
526   const cq_vtable* vtable = &g_cq_vtable[completion_type];
527   const cq_poller_vtable* poller_vtable =
528       &g_poller_vtable_by_poller_type[polling_type];
529 
530   grpc_core::ExecCtx exec_ctx;
531   GRPC_STATS_INC_CQS_CREATED();
532 
533   cq = static_cast<grpc_completion_queue*>(
534       gpr_zalloc(sizeof(grpc_completion_queue) + vtable->data_size +
535                  poller_vtable->size()));
536 
537   cq->vtable = vtable;
538   cq->poller_vtable = poller_vtable;
539 
540   /* One for destroy(), one for pollset_shutdown */
541   new (&cq->owning_refs) grpc_core::RefCount(2);
542 
543   poller_vtable->init(POLLSET_FROM_CQ(cq), &cq->mu);
544   vtable->init(DATA_FROM_CQ(cq), shutdown_callback);
545 
546   GRPC_CLOSURE_INIT(&cq->pollset_shutdown_done, on_pollset_shutdown_done, cq,
547                     grpc_schedule_on_exec_ctx);
548   return cq;
549 }
550 
cq_init_next(void * data,grpc_experimental_completion_queue_functor *)551 static void cq_init_next(
552     void* data,
553     grpc_experimental_completion_queue_functor* /*shutdown_callback*/) {
554   new (data) cq_next_data();
555 }
556 
cq_destroy_next(void * data)557 static void cq_destroy_next(void* data) {
558   cq_next_data* cqd = static_cast<cq_next_data*>(data);
559   cqd->~cq_next_data();
560 }
561 
cq_init_pluck(void * data,grpc_experimental_completion_queue_functor *)562 static void cq_init_pluck(
563     void* data,
564     grpc_experimental_completion_queue_functor* /*shutdown_callback*/) {
565   new (data) cq_pluck_data();
566 }
567 
cq_destroy_pluck(void * data)568 static void cq_destroy_pluck(void* data) {
569   cq_pluck_data* cqd = static_cast<cq_pluck_data*>(data);
570   cqd->~cq_pluck_data();
571 }
572 
cq_init_callback(void * data,grpc_experimental_completion_queue_functor * shutdown_callback)573 static void cq_init_callback(
574     void* data, grpc_experimental_completion_queue_functor* shutdown_callback) {
575   new (data) cq_callback_data(shutdown_callback);
576 }
577 
cq_destroy_callback(void * data)578 static void cq_destroy_callback(void* data) {
579   cq_callback_data* cqd = static_cast<cq_callback_data*>(data);
580   cqd->~cq_callback_data();
581 }
582 
grpc_get_cq_completion_type(grpc_completion_queue * cq)583 grpc_cq_completion_type grpc_get_cq_completion_type(grpc_completion_queue* cq) {
584   return cq->vtable->cq_completion_type;
585 }
586 
grpc_get_cq_poll_num(grpc_completion_queue * cq)587 int grpc_get_cq_poll_num(grpc_completion_queue* cq) {
588   int cur_num_polls;
589   gpr_mu_lock(cq->mu);
590   cur_num_polls = cq->num_polls;
591   gpr_mu_unlock(cq->mu);
592   return cur_num_polls;
593 }
594 
595 #ifndef NDEBUG
grpc_cq_internal_ref(grpc_completion_queue * cq,const char * reason,const char * file,int line)596 void grpc_cq_internal_ref(grpc_completion_queue* cq, const char* reason,
597                           const char* file, int line) {
598   grpc_core::DebugLocation debug_location(file, line);
599 #else
600 void grpc_cq_internal_ref(grpc_completion_queue* cq) {
601   grpc_core::DebugLocation debug_location;
602   const char* reason = nullptr;
603 #endif
604   cq->owning_refs.Ref(debug_location, reason);
605 }
606 
607 static void on_pollset_shutdown_done(void* arg, grpc_error* /*error*/) {
608   grpc_completion_queue* cq = static_cast<grpc_completion_queue*>(arg);
609   GRPC_CQ_INTERNAL_UNREF(cq, "pollset_destroy");
610 }
611 
612 #ifndef NDEBUG
613 void grpc_cq_internal_unref(grpc_completion_queue* cq, const char* reason,
614                             const char* file, int line) {
615   grpc_core::DebugLocation debug_location(file, line);
616 #else
617 void grpc_cq_internal_unref(grpc_completion_queue* cq) {
618   grpc_core::DebugLocation debug_location;
619   const char* reason = nullptr;
620 #endif
621   if (GPR_UNLIKELY(cq->owning_refs.Unref(debug_location, reason))) {
622     cq->vtable->destroy(DATA_FROM_CQ(cq));
623     cq->poller_vtable->destroy(POLLSET_FROM_CQ(cq));
624 #ifndef NDEBUG
625     gpr_free(cq->outstanding_tags);
626 #endif
627     gpr_free(cq);
628   }
629 }
630 
631 #ifndef NDEBUG
632 static void cq_check_tag(grpc_completion_queue* cq, void* tag, bool lock_cq) {
633   int found = 0;
634   if (lock_cq) {
635     gpr_mu_lock(cq->mu);
636   }
637 
638   for (int i = 0; i < static_cast<int>(cq->outstanding_tag_count); i++) {
639     if (cq->outstanding_tags[i] == tag) {
640       cq->outstanding_tag_count--;
641       GPR_SWAP(void*, cq->outstanding_tags[i],
642                cq->outstanding_tags[cq->outstanding_tag_count]);
643       found = 1;
644       break;
645     }
646   }
647 
648   if (lock_cq) {
649     gpr_mu_unlock(cq->mu);
650   }
651 
652   GPR_ASSERT(found);
653 }
654 #else
655 static void cq_check_tag(grpc_completion_queue* /*cq*/, void* /*tag*/,
656                          bool /*lock_cq*/) {}
657 #endif
658 
659 static bool cq_begin_op_for_next(grpc_completion_queue* cq, void* /*tag*/) {
660   cq_next_data* cqd = static_cast<cq_next_data*> DATA_FROM_CQ(cq);
661   return cqd->pending_events.IncrementIfNonzero();
662 }
663 
664 static bool cq_begin_op_for_pluck(grpc_completion_queue* cq, void* /*tag*/) {
665   cq_pluck_data* cqd = static_cast<cq_pluck_data*> DATA_FROM_CQ(cq);
666   return cqd->pending_events.IncrementIfNonzero();
667 }
668 
669 static bool cq_begin_op_for_callback(grpc_completion_queue* cq, void* /*tag*/) {
670   cq_callback_data* cqd = static_cast<cq_callback_data*> DATA_FROM_CQ(cq);
671   return cqd->pending_events.IncrementIfNonzero();
672 }
673 
674 bool grpc_cq_begin_op(grpc_completion_queue* cq, void* tag) {
675 #ifndef NDEBUG
676   gpr_mu_lock(cq->mu);
677   if (cq->outstanding_tag_count == cq->outstanding_tag_capacity) {
678     cq->outstanding_tag_capacity = GPR_MAX(4, 2 * cq->outstanding_tag_capacity);
679     cq->outstanding_tags = static_cast<void**>(gpr_realloc(
680         cq->outstanding_tags,
681         sizeof(*cq->outstanding_tags) * cq->outstanding_tag_capacity));
682   }
683   cq->outstanding_tags[cq->outstanding_tag_count++] = tag;
684   gpr_mu_unlock(cq->mu);
685 #endif
686   return cq->vtable->begin_op(cq, tag);
687 }
688 
689 /* Queue a GRPC_OP_COMPLETED operation to a completion queue (with a
690  * completion
691  * type of GRPC_CQ_NEXT) */
692 static void cq_end_op_for_next(
693     grpc_completion_queue* cq, void* tag, grpc_error* error,
694     void (*done)(void* done_arg, grpc_cq_completion* storage), void* done_arg,
695     grpc_cq_completion* storage, bool /*internal*/) {
696   GPR_TIMER_SCOPE("cq_end_op_for_next", 0);
697 
698   if (GRPC_TRACE_FLAG_ENABLED(grpc_api_trace) ||
699       (GRPC_TRACE_FLAG_ENABLED(grpc_trace_operation_failures) &&
700        error != GRPC_ERROR_NONE)) {
701     const char* errmsg = grpc_error_string(error);
702     GRPC_API_TRACE(
703         "cq_end_op_for_next(cq=%p, tag=%p, error=%s, "
704         "done=%p, done_arg=%p, storage=%p)",
705         6, (cq, tag, errmsg, done, done_arg, storage));
706     if (GRPC_TRACE_FLAG_ENABLED(grpc_trace_operation_failures) &&
707         error != GRPC_ERROR_NONE) {
708       gpr_log(GPR_ERROR, "Operation failed: tag=%p, error=%s", tag, errmsg);
709     }
710   }
711   cq_next_data* cqd = static_cast<cq_next_data*> DATA_FROM_CQ(cq);
712   int is_success = (error == GRPC_ERROR_NONE);
713 
714   storage->tag = tag;
715   storage->done = done;
716   storage->done_arg = done_arg;
717   storage->next = static_cast<uintptr_t>(is_success);
718 
719   cq_check_tag(cq, tag, true); /* Used in debug builds only */
720 
721   if (reinterpret_cast<grpc_completion_queue*>(gpr_tls_get(&g_cached_cq)) ==
722           cq &&
723       reinterpret_cast<grpc_cq_completion*>(gpr_tls_get(&g_cached_event)) ==
724           nullptr) {
725     gpr_tls_set(&g_cached_event, (intptr_t)storage);
726   } else {
727     /* Add the completion to the queue */
728     bool is_first = cqd->queue.Push(storage);
729     cqd->things_queued_ever.FetchAdd(1, grpc_core::MemoryOrder::RELAXED);
730     /* Since we do not hold the cq lock here, it is important to do an 'acquire'
731        load here (instead of a 'no_barrier' load) to match with the release
732        store
733        (done via pending_events.FetchSub(1, ACQ_REL)) in cq_shutdown_next
734        */
735     if (cqd->pending_events.Load(grpc_core::MemoryOrder::ACQUIRE) != 1) {
736       /* Only kick if this is the first item queued */
737       if (is_first) {
738         gpr_mu_lock(cq->mu);
739         grpc_error* kick_error =
740             cq->poller_vtable->kick(POLLSET_FROM_CQ(cq), nullptr);
741         gpr_mu_unlock(cq->mu);
742 
743         if (kick_error != GRPC_ERROR_NONE) {
744           const char* msg = grpc_error_string(kick_error);
745           gpr_log(GPR_ERROR, "Kick failed: %s", msg);
746           GRPC_ERROR_UNREF(kick_error);
747         }
748       }
749       if (cqd->pending_events.FetchSub(1, grpc_core::MemoryOrder::ACQ_REL) ==
750           1) {
751         GRPC_CQ_INTERNAL_REF(cq, "shutting_down");
752         gpr_mu_lock(cq->mu);
753         cq_finish_shutdown_next(cq);
754         gpr_mu_unlock(cq->mu);
755         GRPC_CQ_INTERNAL_UNREF(cq, "shutting_down");
756       }
757     } else {
758       GRPC_CQ_INTERNAL_REF(cq, "shutting_down");
759       cqd->pending_events.Store(0, grpc_core::MemoryOrder::RELEASE);
760       gpr_mu_lock(cq->mu);
761       cq_finish_shutdown_next(cq);
762       gpr_mu_unlock(cq->mu);
763       GRPC_CQ_INTERNAL_UNREF(cq, "shutting_down");
764     }
765   }
766 
767   GRPC_ERROR_UNREF(error);
768 }
769 
770 /* Queue a GRPC_OP_COMPLETED operation to a completion queue (with a
771  * completion
772  * type of GRPC_CQ_PLUCK) */
773 static void cq_end_op_for_pluck(
774     grpc_completion_queue* cq, void* tag, grpc_error* error,
775     void (*done)(void* done_arg, grpc_cq_completion* storage), void* done_arg,
776     grpc_cq_completion* storage, bool /*internal*/) {
777   GPR_TIMER_SCOPE("cq_end_op_for_pluck", 0);
778 
779   cq_pluck_data* cqd = static_cast<cq_pluck_data*> DATA_FROM_CQ(cq);
780   int is_success = (error == GRPC_ERROR_NONE);
781 
782   if (GRPC_TRACE_FLAG_ENABLED(grpc_api_trace) ||
783       (GRPC_TRACE_FLAG_ENABLED(grpc_trace_operation_failures) &&
784        error != GRPC_ERROR_NONE)) {
785     const char* errmsg = grpc_error_string(error);
786     GRPC_API_TRACE(
787         "cq_end_op_for_pluck(cq=%p, tag=%p, error=%s, "
788         "done=%p, done_arg=%p, storage=%p)",
789         6, (cq, tag, errmsg, done, done_arg, storage));
790     if (GRPC_TRACE_FLAG_ENABLED(grpc_trace_operation_failures) &&
791         error != GRPC_ERROR_NONE) {
792       gpr_log(GPR_ERROR, "Operation failed: tag=%p, error=%s", tag, errmsg);
793     }
794   }
795 
796   storage->tag = tag;
797   storage->done = done;
798   storage->done_arg = done_arg;
799   storage->next = reinterpret_cast<uintptr_t>(&cqd->completed_head) |
800                   static_cast<uintptr_t>(is_success);
801 
802   gpr_mu_lock(cq->mu);
803   cq_check_tag(cq, tag, false); /* Used in debug builds only */
804 
805   /* Add to the list of completions */
806   cqd->things_queued_ever.FetchAdd(1, grpc_core::MemoryOrder::RELAXED);
807   cqd->completed_tail->next =
808       reinterpret_cast<uintptr_t>(storage) | (1u & cqd->completed_tail->next);
809   cqd->completed_tail = storage;
810 
811   if (cqd->pending_events.FetchSub(1, grpc_core::MemoryOrder::ACQ_REL) == 1) {
812     cq_finish_shutdown_pluck(cq);
813     gpr_mu_unlock(cq->mu);
814   } else {
815     grpc_pollset_worker* pluck_worker = nullptr;
816     for (int i = 0; i < cqd->num_pluckers; i++) {
817       if (cqd->pluckers[i].tag == tag) {
818         pluck_worker = *cqd->pluckers[i].worker;
819         break;
820       }
821     }
822 
823     grpc_error* kick_error =
824         cq->poller_vtable->kick(POLLSET_FROM_CQ(cq), pluck_worker);
825 
826     gpr_mu_unlock(cq->mu);
827 
828     if (kick_error != GRPC_ERROR_NONE) {
829       const char* msg = grpc_error_string(kick_error);
830       gpr_log(GPR_ERROR, "Kick failed: %s", msg);
831 
832       GRPC_ERROR_UNREF(kick_error);
833     }
834   }
835 
836   GRPC_ERROR_UNREF(error);
837 }
838 
839 static void functor_callback(void* arg, grpc_error* error) {
840   auto* functor = static_cast<grpc_experimental_completion_queue_functor*>(arg);
841   functor->functor_run(functor, error == GRPC_ERROR_NONE);
842 }
843 
844 /* Complete an event on a completion queue of type GRPC_CQ_CALLBACK */
845 static void cq_end_op_for_callback(
846     grpc_completion_queue* cq, void* tag, grpc_error* error,
847     void (*done)(void* done_arg, grpc_cq_completion* storage), void* done_arg,
848     grpc_cq_completion* storage, bool internal) {
849   GPR_TIMER_SCOPE("cq_end_op_for_callback", 0);
850 
851   cq_callback_data* cqd = static_cast<cq_callback_data*> DATA_FROM_CQ(cq);
852 
853   if (GRPC_TRACE_FLAG_ENABLED(grpc_api_trace) ||
854       (GRPC_TRACE_FLAG_ENABLED(grpc_trace_operation_failures) &&
855        error != GRPC_ERROR_NONE)) {
856     const char* errmsg = grpc_error_string(error);
857     GRPC_API_TRACE(
858         "cq_end_op_for_callback(cq=%p, tag=%p, error=%s, "
859         "done=%p, done_arg=%p, storage=%p)",
860         6, (cq, tag, errmsg, done, done_arg, storage));
861     if (GRPC_TRACE_FLAG_ENABLED(grpc_trace_operation_failures) &&
862         error != GRPC_ERROR_NONE) {
863       gpr_log(GPR_ERROR, "Operation failed: tag=%p, error=%s", tag, errmsg);
864     }
865   }
866 
867   // The callback-based CQ isn't really a queue at all and thus has no need
868   // for reserved storage. Invoke the done callback right away to release it.
869   done(done_arg, storage);
870 
871   cq_check_tag(cq, tag, true); /* Used in debug builds only */
872 
873   if (cqd->pending_events.FetchSub(1, grpc_core::MemoryOrder::ACQ_REL) == 1) {
874     cq_finish_shutdown_callback(cq);
875   }
876 
877   // If possible, schedule the callback onto an existing thread-local
878   // ApplicationCallbackExecCtx, which is a work queue. This is possible for:
879   // 1. The callback is internally-generated and there is an ACEC available
880   // 2. The callback is marked inlineable and there is an ACEC available
881   // 3. We are already running in a background poller thread (which always has
882   //    an ACEC available at the base of the stack).
883   auto* functor = static_cast<grpc_experimental_completion_queue_functor*>(tag);
884   if (((internal || functor->inlineable) &&
885        grpc_core::ApplicationCallbackExecCtx::Available()) ||
886       grpc_iomgr_is_any_background_poller_thread()) {
887     grpc_core::ApplicationCallbackExecCtx::Enqueue(functor,
888                                                    (error == GRPC_ERROR_NONE));
889     GRPC_ERROR_UNREF(error);
890     return;
891   }
892 
893   // Schedule the callback on a closure if not internal or triggered
894   // from a background poller thread.
895   grpc_core::Executor::Run(
896       GRPC_CLOSURE_CREATE(functor_callback, functor, nullptr), error);
897 }
898 
899 void grpc_cq_end_op(grpc_completion_queue* cq, void* tag, grpc_error* error,
900                     void (*done)(void* done_arg, grpc_cq_completion* storage),
901                     void* done_arg, grpc_cq_completion* storage,
902                     bool internal) {
903   cq->vtable->end_op(cq, tag, error, done, done_arg, storage, internal);
904 }
905 
906 struct cq_is_finished_arg {
907   gpr_atm last_seen_things_queued_ever;
908   grpc_completion_queue* cq;
909   grpc_millis deadline;
910   grpc_cq_completion* stolen_completion;
911   void* tag; /* for pluck */
912   bool first_loop;
913 };
914 class ExecCtxNext : public grpc_core::ExecCtx {
915  public:
916   explicit ExecCtxNext(void* arg)
917       : ExecCtx(0), check_ready_to_finish_arg_(arg) {}
918 
919   bool CheckReadyToFinish() override {
920     cq_is_finished_arg* a =
921         static_cast<cq_is_finished_arg*>(check_ready_to_finish_arg_);
922     grpc_completion_queue* cq = a->cq;
923     cq_next_data* cqd = static_cast<cq_next_data*> DATA_FROM_CQ(cq);
924     GPR_ASSERT(a->stolen_completion == nullptr);
925 
926     intptr_t current_last_seen_things_queued_ever =
927         cqd->things_queued_ever.Load(grpc_core::MemoryOrder::RELAXED);
928 
929     if (current_last_seen_things_queued_ever !=
930         a->last_seen_things_queued_ever) {
931       a->last_seen_things_queued_ever =
932           cqd->things_queued_ever.Load(grpc_core::MemoryOrder::RELAXED);
933 
934       /* Pop a cq_completion from the queue. Returns NULL if the queue is empty
935        * might return NULL in some cases even if the queue is not empty; but
936        * that
937        * is ok and doesn't affect correctness. Might effect the tail latencies a
938        * bit) */
939       a->stolen_completion = cqd->queue.Pop();
940       if (a->stolen_completion != nullptr) {
941         return true;
942       }
943     }
944     return !a->first_loop && a->deadline < grpc_core::ExecCtx::Get()->Now();
945   }
946 
947  private:
948   void* check_ready_to_finish_arg_;
949 };
950 
951 #ifndef NDEBUG
952 static void dump_pending_tags(grpc_completion_queue* cq) {
953   if (!GRPC_TRACE_FLAG_ENABLED(grpc_trace_pending_tags)) return;
954   std::vector<std::string> parts;
955   parts.push_back("PENDING TAGS:");
956   gpr_mu_lock(cq->mu);
957   for (size_t i = 0; i < cq->outstanding_tag_count; i++) {
958     parts.push_back(absl::StrFormat(" %p", cq->outstanding_tags[i]));
959   }
960   gpr_mu_unlock(cq->mu);
961   gpr_log(GPR_DEBUG, "%s", absl::StrJoin(parts, "").c_str());
962 }
963 #else
964 static void dump_pending_tags(grpc_completion_queue* /*cq*/) {}
965 #endif
966 
967 static grpc_event cq_next(grpc_completion_queue* cq, gpr_timespec deadline,
968                           void* reserved) {
969   GPR_TIMER_SCOPE("grpc_completion_queue_next", 0);
970 
971   grpc_event ret;
972   cq_next_data* cqd = static_cast<cq_next_data*> DATA_FROM_CQ(cq);
973 
974   GRPC_API_TRACE(
975       "grpc_completion_queue_next("
976       "cq=%p, "
977       "deadline=gpr_timespec { tv_sec: %" PRId64
978       ", tv_nsec: %d, clock_type: %d }, "
979       "reserved=%p)",
980       5,
981       (cq, deadline.tv_sec, deadline.tv_nsec, (int)deadline.clock_type,
982        reserved));
983   GPR_ASSERT(!reserved);
984 
985   dump_pending_tags(cq);
986 
987   GRPC_CQ_INTERNAL_REF(cq, "next");
988 
989   grpc_millis deadline_millis = grpc_timespec_to_millis_round_up(deadline);
990   cq_is_finished_arg is_finished_arg = {
991       cqd->things_queued_ever.Load(grpc_core::MemoryOrder::RELAXED),
992       cq,
993       deadline_millis,
994       nullptr,
995       nullptr,
996       true};
997   ExecCtxNext exec_ctx(&is_finished_arg);
998   for (;;) {
999     grpc_millis iteration_deadline = deadline_millis;
1000 
1001     if (is_finished_arg.stolen_completion != nullptr) {
1002       grpc_cq_completion* c = is_finished_arg.stolen_completion;
1003       is_finished_arg.stolen_completion = nullptr;
1004       ret.type = GRPC_OP_COMPLETE;
1005       ret.success = c->next & 1u;
1006       ret.tag = c->tag;
1007       c->done(c->done_arg, c);
1008       break;
1009     }
1010 
1011     grpc_cq_completion* c = cqd->queue.Pop();
1012 
1013     if (c != nullptr) {
1014       ret.type = GRPC_OP_COMPLETE;
1015       ret.success = c->next & 1u;
1016       ret.tag = c->tag;
1017       c->done(c->done_arg, c);
1018       break;
1019     } else {
1020       /* If c == NULL it means either the queue is empty OR in an transient
1021          inconsistent state. If it is the latter, we shold do a 0-timeout poll
1022          so that the thread comes back quickly from poll to make a second
1023          attempt at popping. Not doing this can potentially deadlock this
1024          thread forever (if the deadline is infinity) */
1025       if (cqd->queue.num_items() > 0) {
1026         iteration_deadline = 0;
1027       }
1028     }
1029 
1030     if (cqd->pending_events.Load(grpc_core::MemoryOrder::ACQUIRE) == 0) {
1031       /* Before returning, check if the queue has any items left over (since
1032          MultiProducerSingleConsumerQueue::Pop() can sometimes return NULL
1033          even if the queue is not empty. If so, keep retrying but do not
1034          return GRPC_QUEUE_SHUTDOWN */
1035       if (cqd->queue.num_items() > 0) {
1036         /* Go to the beginning of the loop. No point doing a poll because
1037            (cq->shutdown == true) is only possible when there is no pending
1038            work (i.e cq->pending_events == 0) and any outstanding completion
1039            events should have already been queued on this cq */
1040         continue;
1041       }
1042 
1043       ret.type = GRPC_QUEUE_SHUTDOWN;
1044       ret.success = 0;
1045       break;
1046     }
1047 
1048     if (!is_finished_arg.first_loop &&
1049         grpc_core::ExecCtx::Get()->Now() >= deadline_millis) {
1050       ret.type = GRPC_QUEUE_TIMEOUT;
1051       ret.success = 0;
1052       dump_pending_tags(cq);
1053       break;
1054     }
1055 
1056     /* The main polling work happens in grpc_pollset_work */
1057     gpr_mu_lock(cq->mu);
1058     cq->num_polls++;
1059     grpc_error* err = cq->poller_vtable->work(POLLSET_FROM_CQ(cq), nullptr,
1060                                               iteration_deadline);
1061     gpr_mu_unlock(cq->mu);
1062 
1063     if (err != GRPC_ERROR_NONE) {
1064       const char* msg = grpc_error_string(err);
1065       gpr_log(GPR_ERROR, "Completion queue next failed: %s", msg);
1066 
1067       GRPC_ERROR_UNREF(err);
1068       ret.type = GRPC_QUEUE_TIMEOUT;
1069       ret.success = 0;
1070       dump_pending_tags(cq);
1071       break;
1072     }
1073     is_finished_arg.first_loop = false;
1074   }
1075 
1076   if (cqd->queue.num_items() > 0 &&
1077       cqd->pending_events.Load(grpc_core::MemoryOrder::ACQUIRE) > 0) {
1078     gpr_mu_lock(cq->mu);
1079     cq->poller_vtable->kick(POLLSET_FROM_CQ(cq), nullptr);
1080     gpr_mu_unlock(cq->mu);
1081   }
1082 
1083   GRPC_SURFACE_TRACE_RETURNED_EVENT(cq, &ret);
1084   GRPC_CQ_INTERNAL_UNREF(cq, "next");
1085 
1086   GPR_ASSERT(is_finished_arg.stolen_completion == nullptr);
1087 
1088   return ret;
1089 }
1090 
1091 /* Finishes the completion queue shutdown. This means that there are no more
1092    completion events / tags expected from the completion queue
1093    - Must be called under completion queue lock
1094    - Must be called only once in completion queue's lifetime
1095    - grpc_completion_queue_shutdown() MUST have been called before calling
1096    this function */
1097 static void cq_finish_shutdown_next(grpc_completion_queue* cq) {
1098   cq_next_data* cqd = static_cast<cq_next_data*> DATA_FROM_CQ(cq);
1099 
1100   GPR_ASSERT(cqd->shutdown_called);
1101   GPR_ASSERT(cqd->pending_events.Load(grpc_core::MemoryOrder::RELAXED) == 0);
1102 
1103   cq->poller_vtable->shutdown(POLLSET_FROM_CQ(cq), &cq->pollset_shutdown_done);
1104 }
1105 
1106 static void cq_shutdown_next(grpc_completion_queue* cq) {
1107   cq_next_data* cqd = static_cast<cq_next_data*> DATA_FROM_CQ(cq);
1108 
1109   /* Need an extra ref for cq here because:
1110    * We call cq_finish_shutdown_next() below, that would call pollset shutdown.
1111    * Pollset shutdown decrements the cq ref count which can potentially destroy
1112    * the cq (if that happens to be the last ref).
1113    * Creating an extra ref here prevents the cq from getting destroyed while
1114    * this function is still active */
1115   GRPC_CQ_INTERNAL_REF(cq, "shutting_down");
1116   gpr_mu_lock(cq->mu);
1117   if (cqd->shutdown_called) {
1118     gpr_mu_unlock(cq->mu);
1119     GRPC_CQ_INTERNAL_UNREF(cq, "shutting_down");
1120     return;
1121   }
1122   cqd->shutdown_called = true;
1123   /* Doing acq/release FetchSub here to match with
1124    * cq_begin_op_for_next and cq_end_op_for_next functions which read/write
1125    * on this counter without necessarily holding a lock on cq */
1126   if (cqd->pending_events.FetchSub(1, grpc_core::MemoryOrder::ACQ_REL) == 1) {
1127     cq_finish_shutdown_next(cq);
1128   }
1129   gpr_mu_unlock(cq->mu);
1130   GRPC_CQ_INTERNAL_UNREF(cq, "shutting_down");
1131 }
1132 
1133 grpc_event grpc_completion_queue_next(grpc_completion_queue* cq,
1134                                       gpr_timespec deadline, void* reserved) {
1135   return cq->vtable->next(cq, deadline, reserved);
1136 }
1137 
1138 static int add_plucker(grpc_completion_queue* cq, void* tag,
1139                        grpc_pollset_worker** worker) {
1140   cq_pluck_data* cqd = static_cast<cq_pluck_data*> DATA_FROM_CQ(cq);
1141   if (cqd->num_pluckers == GRPC_MAX_COMPLETION_QUEUE_PLUCKERS) {
1142     return 0;
1143   }
1144   cqd->pluckers[cqd->num_pluckers].tag = tag;
1145   cqd->pluckers[cqd->num_pluckers].worker = worker;
1146   cqd->num_pluckers++;
1147   return 1;
1148 }
1149 
1150 static void del_plucker(grpc_completion_queue* cq, void* tag,
1151                         grpc_pollset_worker** worker) {
1152   cq_pluck_data* cqd = static_cast<cq_pluck_data*> DATA_FROM_CQ(cq);
1153   for (int i = 0; i < cqd->num_pluckers; i++) {
1154     if (cqd->pluckers[i].tag == tag && cqd->pluckers[i].worker == worker) {
1155       cqd->num_pluckers--;
1156       GPR_SWAP(plucker, cqd->pluckers[i], cqd->pluckers[cqd->num_pluckers]);
1157       return;
1158     }
1159   }
1160   GPR_UNREACHABLE_CODE(return );
1161 }
1162 
1163 class ExecCtxPluck : public grpc_core::ExecCtx {
1164  public:
1165   explicit ExecCtxPluck(void* arg)
1166       : ExecCtx(0), check_ready_to_finish_arg_(arg) {}
1167 
1168   bool CheckReadyToFinish() override {
1169     cq_is_finished_arg* a =
1170         static_cast<cq_is_finished_arg*>(check_ready_to_finish_arg_);
1171     grpc_completion_queue* cq = a->cq;
1172     cq_pluck_data* cqd = static_cast<cq_pluck_data*> DATA_FROM_CQ(cq);
1173 
1174     GPR_ASSERT(a->stolen_completion == nullptr);
1175     gpr_atm current_last_seen_things_queued_ever =
1176         cqd->things_queued_ever.Load(grpc_core::MemoryOrder::RELAXED);
1177     if (current_last_seen_things_queued_ever !=
1178         a->last_seen_things_queued_ever) {
1179       gpr_mu_lock(cq->mu);
1180       a->last_seen_things_queued_ever =
1181           cqd->things_queued_ever.Load(grpc_core::MemoryOrder::RELAXED);
1182       grpc_cq_completion* c;
1183       grpc_cq_completion* prev = &cqd->completed_head;
1184       while ((c = reinterpret_cast<grpc_cq_completion*>(
1185                   prev->next & ~static_cast<uintptr_t>(1))) !=
1186              &cqd->completed_head) {
1187         if (c->tag == a->tag) {
1188           prev->next = (prev->next & static_cast<uintptr_t>(1)) |
1189                        (c->next & ~static_cast<uintptr_t>(1));
1190           if (c == cqd->completed_tail) {
1191             cqd->completed_tail = prev;
1192           }
1193           gpr_mu_unlock(cq->mu);
1194           a->stolen_completion = c;
1195           return true;
1196         }
1197         prev = c;
1198       }
1199       gpr_mu_unlock(cq->mu);
1200     }
1201     return !a->first_loop && a->deadline < grpc_core::ExecCtx::Get()->Now();
1202   }
1203 
1204  private:
1205   void* check_ready_to_finish_arg_;
1206 };
1207 
1208 static grpc_event cq_pluck(grpc_completion_queue* cq, void* tag,
1209                            gpr_timespec deadline, void* reserved) {
1210   GPR_TIMER_SCOPE("grpc_completion_queue_pluck", 0);
1211 
1212   grpc_event ret;
1213   grpc_cq_completion* c;
1214   grpc_cq_completion* prev;
1215   grpc_pollset_worker* worker = nullptr;
1216   cq_pluck_data* cqd = static_cast<cq_pluck_data*> DATA_FROM_CQ(cq);
1217 
1218   if (GRPC_TRACE_FLAG_ENABLED(grpc_cq_pluck_trace)) {
1219     GRPC_API_TRACE(
1220         "grpc_completion_queue_pluck("
1221         "cq=%p, tag=%p, "
1222         "deadline=gpr_timespec { tv_sec: %" PRId64
1223         ", tv_nsec: %d, clock_type: %d }, "
1224         "reserved=%p)",
1225         6,
1226         (cq, tag, deadline.tv_sec, deadline.tv_nsec, (int)deadline.clock_type,
1227          reserved));
1228   }
1229   GPR_ASSERT(!reserved);
1230 
1231   dump_pending_tags(cq);
1232 
1233   GRPC_CQ_INTERNAL_REF(cq, "pluck");
1234   gpr_mu_lock(cq->mu);
1235   grpc_millis deadline_millis = grpc_timespec_to_millis_round_up(deadline);
1236   cq_is_finished_arg is_finished_arg = {
1237       cqd->things_queued_ever.Load(grpc_core::MemoryOrder::RELAXED),
1238       cq,
1239       deadline_millis,
1240       nullptr,
1241       tag,
1242       true};
1243   ExecCtxPluck exec_ctx(&is_finished_arg);
1244   for (;;) {
1245     if (is_finished_arg.stolen_completion != nullptr) {
1246       gpr_mu_unlock(cq->mu);
1247       c = is_finished_arg.stolen_completion;
1248       is_finished_arg.stolen_completion = nullptr;
1249       ret.type = GRPC_OP_COMPLETE;
1250       ret.success = c->next & 1u;
1251       ret.tag = c->tag;
1252       c->done(c->done_arg, c);
1253       break;
1254     }
1255     prev = &cqd->completed_head;
1256     while ((c = reinterpret_cast<grpc_cq_completion*>(
1257                 prev->next & ~static_cast<uintptr_t>(1))) !=
1258            &cqd->completed_head) {
1259       if (c->tag == tag) {
1260         prev->next = (prev->next & static_cast<uintptr_t>(1)) |
1261                      (c->next & ~static_cast<uintptr_t>(1));
1262         if (c == cqd->completed_tail) {
1263           cqd->completed_tail = prev;
1264         }
1265         gpr_mu_unlock(cq->mu);
1266         ret.type = GRPC_OP_COMPLETE;
1267         ret.success = c->next & 1u;
1268         ret.tag = c->tag;
1269         c->done(c->done_arg, c);
1270         goto done;
1271       }
1272       prev = c;
1273     }
1274     if (cqd->shutdown.Load(grpc_core::MemoryOrder::RELAXED)) {
1275       gpr_mu_unlock(cq->mu);
1276       ret.type = GRPC_QUEUE_SHUTDOWN;
1277       ret.success = 0;
1278       break;
1279     }
1280     if (!add_plucker(cq, tag, &worker)) {
1281       gpr_log(GPR_DEBUG,
1282               "Too many outstanding grpc_completion_queue_pluck calls: maximum "
1283               "is %d",
1284               GRPC_MAX_COMPLETION_QUEUE_PLUCKERS);
1285       gpr_mu_unlock(cq->mu);
1286       /* TODO(ctiller): should we use a different result here */
1287       ret.type = GRPC_QUEUE_TIMEOUT;
1288       ret.success = 0;
1289       dump_pending_tags(cq);
1290       break;
1291     }
1292     if (!is_finished_arg.first_loop &&
1293         grpc_core::ExecCtx::Get()->Now() >= deadline_millis) {
1294       del_plucker(cq, tag, &worker);
1295       gpr_mu_unlock(cq->mu);
1296       ret.type = GRPC_QUEUE_TIMEOUT;
1297       ret.success = 0;
1298       dump_pending_tags(cq);
1299       break;
1300     }
1301     cq->num_polls++;
1302     grpc_error* err =
1303         cq->poller_vtable->work(POLLSET_FROM_CQ(cq), &worker, deadline_millis);
1304     if (err != GRPC_ERROR_NONE) {
1305       del_plucker(cq, tag, &worker);
1306       gpr_mu_unlock(cq->mu);
1307       const char* msg = grpc_error_string(err);
1308       gpr_log(GPR_ERROR, "Completion queue pluck failed: %s", msg);
1309 
1310       GRPC_ERROR_UNREF(err);
1311       ret.type = GRPC_QUEUE_TIMEOUT;
1312       ret.success = 0;
1313       dump_pending_tags(cq);
1314       break;
1315     }
1316     is_finished_arg.first_loop = false;
1317     del_plucker(cq, tag, &worker);
1318   }
1319 done:
1320   GRPC_SURFACE_TRACE_RETURNED_EVENT(cq, &ret);
1321   GRPC_CQ_INTERNAL_UNREF(cq, "pluck");
1322 
1323   GPR_ASSERT(is_finished_arg.stolen_completion == nullptr);
1324 
1325   return ret;
1326 }
1327 
1328 grpc_event grpc_completion_queue_pluck(grpc_completion_queue* cq, void* tag,
1329                                        gpr_timespec deadline, void* reserved) {
1330   return cq->vtable->pluck(cq, tag, deadline, reserved);
1331 }
1332 
1333 static void cq_finish_shutdown_pluck(grpc_completion_queue* cq) {
1334   cq_pluck_data* cqd = static_cast<cq_pluck_data*> DATA_FROM_CQ(cq);
1335 
1336   GPR_ASSERT(cqd->shutdown_called);
1337   GPR_ASSERT(!cqd->shutdown.Load(grpc_core::MemoryOrder::RELAXED));
1338   cqd->shutdown.Store(true, grpc_core::MemoryOrder::RELAXED);
1339 
1340   cq->poller_vtable->shutdown(POLLSET_FROM_CQ(cq), &cq->pollset_shutdown_done);
1341 }
1342 
1343 /* NOTE: This function is almost exactly identical to cq_shutdown_next() but
1344  * merging them is a bit tricky and probably not worth it */
1345 static void cq_shutdown_pluck(grpc_completion_queue* cq) {
1346   cq_pluck_data* cqd = static_cast<cq_pluck_data*> DATA_FROM_CQ(cq);
1347 
1348   /* Need an extra ref for cq here because:
1349    * We call cq_finish_shutdown_pluck() below, that would call pollset shutdown.
1350    * Pollset shutdown decrements the cq ref count which can potentially destroy
1351    * the cq (if that happens to be the last ref).
1352    * Creating an extra ref here prevents the cq from getting destroyed while
1353    * this function is still active */
1354   GRPC_CQ_INTERNAL_REF(cq, "shutting_down (pluck cq)");
1355   gpr_mu_lock(cq->mu);
1356   if (cqd->shutdown_called) {
1357     gpr_mu_unlock(cq->mu);
1358     GRPC_CQ_INTERNAL_UNREF(cq, "shutting_down (pluck cq)");
1359     return;
1360   }
1361   cqd->shutdown_called = true;
1362   if (cqd->pending_events.FetchSub(1, grpc_core::MemoryOrder::ACQ_REL) == 1) {
1363     cq_finish_shutdown_pluck(cq);
1364   }
1365   gpr_mu_unlock(cq->mu);
1366   GRPC_CQ_INTERNAL_UNREF(cq, "shutting_down (pluck cq)");
1367 }
1368 
1369 static void cq_finish_shutdown_callback(grpc_completion_queue* cq) {
1370   cq_callback_data* cqd = static_cast<cq_callback_data*> DATA_FROM_CQ(cq);
1371   auto* callback = cqd->shutdown_callback;
1372 
1373   GPR_ASSERT(cqd->shutdown_called);
1374 
1375   cq->poller_vtable->shutdown(POLLSET_FROM_CQ(cq), &cq->pollset_shutdown_done);
1376   if (grpc_iomgr_is_any_background_poller_thread()) {
1377     grpc_core::ApplicationCallbackExecCtx::Enqueue(callback, true);
1378     return;
1379   }
1380 
1381   // Schedule the callback on a closure if not internal or triggered
1382   // from a background poller thread.
1383   grpc_core::Executor::Run(
1384       GRPC_CLOSURE_CREATE(functor_callback, callback, nullptr),
1385       GRPC_ERROR_NONE);
1386 }
1387 
1388 static void cq_shutdown_callback(grpc_completion_queue* cq) {
1389   cq_callback_data* cqd = static_cast<cq_callback_data*> DATA_FROM_CQ(cq);
1390 
1391   /* Need an extra ref for cq here because:
1392    * We call cq_finish_shutdown_callback() below, which calls pollset shutdown.
1393    * Pollset shutdown decrements the cq ref count which can potentially destroy
1394    * the cq (if that happens to be the last ref).
1395    * Creating an extra ref here prevents the cq from getting destroyed while
1396    * this function is still active */
1397   GRPC_CQ_INTERNAL_REF(cq, "shutting_down (callback cq)");
1398   gpr_mu_lock(cq->mu);
1399   if (cqd->shutdown_called) {
1400     gpr_mu_unlock(cq->mu);
1401     GRPC_CQ_INTERNAL_UNREF(cq, "shutting_down (callback cq)");
1402     return;
1403   }
1404   cqd->shutdown_called = true;
1405   if (cqd->pending_events.FetchSub(1, grpc_core::MemoryOrder::ACQ_REL) == 1) {
1406     gpr_mu_unlock(cq->mu);
1407     cq_finish_shutdown_callback(cq);
1408   } else {
1409     gpr_mu_unlock(cq->mu);
1410   }
1411   GRPC_CQ_INTERNAL_UNREF(cq, "shutting_down (callback cq)");
1412 }
1413 
1414 /* Shutdown simply drops a ref that we reserved at creation time; if we drop
1415    to zero here, then enter shutdown mode and wake up any waiters */
1416 void grpc_completion_queue_shutdown(grpc_completion_queue* cq) {
1417   GPR_TIMER_SCOPE("grpc_completion_queue_shutdown", 0);
1418   grpc_core::ApplicationCallbackExecCtx callback_exec_ctx;
1419   grpc_core::ExecCtx exec_ctx;
1420   GRPC_API_TRACE("grpc_completion_queue_shutdown(cq=%p)", 1, (cq));
1421   cq->vtable->shutdown(cq);
1422 }
1423 
1424 void grpc_completion_queue_destroy(grpc_completion_queue* cq) {
1425   GPR_TIMER_SCOPE("grpc_completion_queue_destroy", 0);
1426   GRPC_API_TRACE("grpc_completion_queue_destroy(cq=%p)", 1, (cq));
1427   grpc_completion_queue_shutdown(cq);
1428 
1429   grpc_core::ExecCtx exec_ctx;
1430   GRPC_CQ_INTERNAL_UNREF(cq, "destroy");
1431 }
1432 
1433 grpc_pollset* grpc_cq_pollset(grpc_completion_queue* cq) {
1434   return cq->poller_vtable->can_get_pollset ? POLLSET_FROM_CQ(cq) : nullptr;
1435 }
1436 
1437 bool grpc_cq_can_listen(grpc_completion_queue* cq) {
1438   return cq->poller_vtable->can_listen;
1439 }
1440