• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 /*
2  *
3  * Copyright 2017 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 
19 #include <grpc/support/port_platform.h>
20 
21 #include "src/core/lib/iomgr/port.h"
22 
23 #include <grpc/support/log.h>
24 
25 /* This polling engine is only relevant on linux kernels supporting epoll() */
26 #ifdef GRPC_LINUX_EPOLL_CREATE1
27 
28 #include "src/core/lib/iomgr/ev_epollex_linux.h"
29 
30 #include <assert.h>
31 #include <errno.h>
32 #include <limits.h>
33 #include <poll.h>
34 #include <pthread.h>
35 #include <string.h>
36 #include <sys/epoll.h>
37 #include <sys/socket.h>
38 #include <sys/syscall.h>
39 #include <unistd.h>
40 
41 #include <grpc/support/alloc.h>
42 #include <grpc/support/string_util.h>
43 
44 #include "src/core/lib/debug/stats.h"
45 #include "src/core/lib/gpr/spinlock.h"
46 #include "src/core/lib/gpr/tls.h"
47 #include "src/core/lib/gpr/useful.h"
48 #include "src/core/lib/gprpp/manual_constructor.h"
49 #include "src/core/lib/gprpp/mutex_lock.h"
50 #include "src/core/lib/iomgr/block_annotate.h"
51 #include "src/core/lib/iomgr/iomgr_internal.h"
52 #include "src/core/lib/iomgr/is_epollexclusive_available.h"
53 #include "src/core/lib/iomgr/lockfree_event.h"
54 #include "src/core/lib/iomgr/sys_epoll_wrapper.h"
55 #include "src/core/lib/iomgr/timer.h"
56 #include "src/core/lib/iomgr/wakeup_fd_posix.h"
57 #include "src/core/lib/profiling/timers.h"
58 
59 // debug aid: create workers on the heap (allows asan to spot
60 // use-after-destruction)
61 //#define GRPC_EPOLLEX_CREATE_WORKERS_ON_HEAP 1
62 
63 #define MAX_EPOLL_EVENTS 100
64 // TODO(juanlishen): We use a greater-than-one value here as a workaround fix to
65 // a keepalive ping timeout issue. We may want to revert https://github
66 // .com/grpc/grpc/pull/14943 once we figure out the root cause.
67 #define MAX_EPOLL_EVENTS_HANDLED_EACH_POLL_CALL 16
68 #define MAX_FDS_IN_CACHE 32
69 
70 grpc_core::DebugOnlyTraceFlag grpc_trace_pollable_refcount(false,
71                                                            "pollable_refcount");
72 
73 /*******************************************************************************
74  * pollable Declarations
75  */
76 
77 typedef enum { PO_MULTI, PO_FD, PO_EMPTY } pollable_type;
78 
79 typedef struct pollable pollable;
80 
81 typedef struct cached_fd {
82   // Set to the grpc_fd's salt value. See 'salt' variable' in grpc_fd for more
83   // details
84   intptr_t salt;
85 
86   // The underlying fd
87   int fd;
88 
89   // A recency time counter that helps to determine the LRU fd in the cache
90   uint64_t last_used;
91 } cached_fd;
92 
93 /// A pollable is something that can be polled: it has an epoll set to poll on,
94 /// and a wakeup fd for kicks
95 /// There are three broad types:
96 ///  - PO_EMPTY - the empty pollable, used before file descriptors are added to
97 ///               a pollset
98 ///  - PO_FD - a pollable containing only one FD - used to optimize single-fd
99 ///            pollsets (which are common with synchronous api usage)
100 ///  - PO_MULTI - a pollable containing many fds
101 struct pollable {
102   pollable_type type;  // immutable
103   gpr_refcount refs;
104 
105   int epfd;
106   grpc_wakeup_fd wakeup;
107 
108   // The following are relevant only for type PO_FD
109   grpc_fd* owner_fd;       // Set to the owner_fd if the type is PO_FD
110   gpr_mu owner_orphan_mu;  // Synchronizes access to owner_orphaned field
111   bool owner_orphaned;     // Is the owner fd orphaned
112 
113   grpc_pollset_set* pollset_set;
114   pollable* next;
115   pollable* prev;
116 
117   gpr_mu mu;
118   grpc_pollset_worker* root_worker;
119 
120   int event_cursor;
121   int event_count;
122   struct epoll_event events[MAX_EPOLL_EVENTS];
123 
124   // We may be calling pollable_add_fd() on the same (pollable, fd) multiple
125   // times. To prevent pollable_add_fd() from making multiple sys calls to
126   // epoll_ctl() to add the fd, we maintain a cache of what fds are already
127   // present in the underlying epoll-set.
128   //
129   // Since this is not a correctness issue, we do not need to maintain all the
130   // fds in the cache. Hence we just use an LRU cache of size 'MAX_FDS_IN_CACHE'
131   //
132   // NOTE: An ideal implementation of this should do the following:
133   //  1) Add fds to the cache in pollable_add_fd() function (i.e whenever the fd
134   //     is added to the pollable's epoll set)
135   //  2) Remove the fd from the cache whenever the fd is removed from the
136   //     underlying epoll set (i.e whenever fd_orphan() is called).
137   //
138   // Implementing (2) above (i.e removing fds from cache on fd_orphan) adds a
139   // lot of complexity since an fd can be present in multiple pollables. So our
140   // implementation ONLY DOES (1) and NOT (2).
141   //
142   // The cache_fd.salt variable helps here to maintain correctness (it serves as
143   // an epoch that differentiates one grpc_fd from the other even though both of
144   // them may have the same fd number)
145   //
146   // The following implements LRU-eviction cache of fds in this pollable
147   cached_fd fd_cache[MAX_FDS_IN_CACHE];
148   int fd_cache_size;
149   uint64_t fd_cache_counter;  // Recency timer tick counter
150 };
151 
pollable_type_string(pollable_type t)152 static const char* pollable_type_string(pollable_type t) {
153   switch (t) {
154     case PO_MULTI:
155       return "pollset";
156     case PO_FD:
157       return "fd";
158     case PO_EMPTY:
159       return "empty";
160   }
161   return "<invalid>";
162 }
163 
pollable_desc(pollable * p)164 static char* pollable_desc(pollable* p) {
165   char* out;
166   gpr_asprintf(&out, "type=%s epfd=%d wakeup=%d", pollable_type_string(p->type),
167                p->epfd, p->wakeup.read_fd);
168   return out;
169 }
170 
171 /// Shared empty pollable - used by pollset to poll on until the first fd is
172 /// added
173 static pollable* g_empty_pollable;
174 
175 static grpc_error* pollable_create(pollable_type type, pollable** p);
176 #ifdef NDEBUG
177 static pollable* pollable_ref(pollable* p);
178 static void pollable_unref(pollable* p);
179 #define POLLABLE_REF(p, r) pollable_ref(p)
180 #define POLLABLE_UNREF(p, r) pollable_unref(p)
181 #else
182 static pollable* pollable_ref(pollable* p, int line, const char* reason);
183 static void pollable_unref(pollable* p, int line, const char* reason);
184 #define POLLABLE_REF(p, r) pollable_ref((p), __LINE__, (r))
185 #define POLLABLE_UNREF(p, r) pollable_unref((p), __LINE__, (r))
186 #endif
187 
188 /*******************************************************************************
189  * Fd Declarations
190  */
191 
192 // Monotonically increasing Epoch counter that is assinged to each grpc_fd. See
193 // the description of 'salt' variable in 'grpc_fd' for more details
194 // TODO: (sreek/kpayson) gpr_atm is intptr_t which may not be wide-enough on
195 // 32-bit systems. Change this to int_64 - atleast on 32-bit systems
196 static gpr_atm g_fd_salt;
197 
198 struct grpc_fd {
199   int fd;
200 
201   // Since fd numbers can be reused (after old fds are closed), this serves as
202   // an epoch that uniquely identifies this fd (i.e the pair (salt, fd) is
203   // unique (until the salt counter (i.e g_fd_salt) overflows)
204   intptr_t salt;
205 
206   // refst format:
207   //     bit 0    : 1=Active / 0=Orphaned
208   //     bits 1-n : refcount
209   //  Ref/Unref by two to avoid altering the orphaned bit
210   gpr_atm refst;
211 
212   gpr_mu orphan_mu;
213 
214   gpr_mu pollable_mu;
215   pollable* pollable_obj;
216 
217   grpc_core::ManualConstructor<grpc_core::LockfreeEvent> read_closure;
218   grpc_core::ManualConstructor<grpc_core::LockfreeEvent> write_closure;
219   grpc_core::ManualConstructor<grpc_core::LockfreeEvent> error_closure;
220 
221   struct grpc_fd* freelist_next;
222   grpc_closure* on_done_closure;
223 
224   grpc_iomgr_object iomgr_object;
225 
226   // Do we need to track EPOLLERR events separately?
227   bool track_err;
228 };
229 
230 static void fd_global_init(void);
231 static void fd_global_shutdown(void);
232 
233 /*******************************************************************************
234  * Pollset Declarations
235  */
236 
237 typedef struct {
238   grpc_pollset_worker* next;
239   grpc_pollset_worker* prev;
240 } pwlink;
241 
242 typedef enum { PWLINK_POLLABLE = 0, PWLINK_POLLSET, PWLINK_COUNT } pwlinks;
243 
244 struct grpc_pollset_worker {
245   bool kicked;
246   bool initialized_cv;
247 #ifndef NDEBUG
248   // debug aid: which thread started this worker
249   pid_t originator;
250 #endif
251   gpr_cv cv;
252   grpc_pollset* pollset;
253   pollable* pollable_obj;
254 
255   pwlink links[PWLINK_COUNT];
256 };
257 
258 struct grpc_pollset {
259   gpr_mu mu;
260   gpr_atm worker_count;
261   pollable* active_pollable;
262   bool kicked_without_poller;
263   grpc_closure* shutdown_closure;
264   bool already_shutdown;
265   grpc_pollset_worker* root_worker;
266   int containing_pollset_set_count;
267 };
268 
269 /*******************************************************************************
270  * Pollset-set Declarations
271  */
272 
273 struct grpc_pollset_set {
274   gpr_refcount refs;
275   gpr_mu mu;
276   grpc_pollset_set* parent;
277 
278   size_t pollset_count;
279   size_t pollset_capacity;
280   grpc_pollset** pollsets;
281 
282   size_t fd_count;
283   size_t fd_capacity;
284   grpc_fd** fds;
285 };
286 
287 /*******************************************************************************
288  * Common helpers
289  */
290 
append_error(grpc_error ** composite,grpc_error * error,const char * desc)291 static bool append_error(grpc_error** composite, grpc_error* error,
292                          const char* desc) {
293   if (error == GRPC_ERROR_NONE) return true;
294   if (*composite == GRPC_ERROR_NONE) {
295     *composite = GRPC_ERROR_CREATE_FROM_COPIED_STRING(desc);
296   }
297   *composite = grpc_error_add_child(*composite, error);
298   return false;
299 }
300 
301 /*******************************************************************************
302  * Fd Definitions
303  */
304 
305 /* We need to keep a freelist not because of any concerns of malloc performance
306  * but instead so that implementations with multiple threads in (for example)
307  * epoll_wait deal with the race between pollset removal and incoming poll
308  * notifications.
309  *
310  * The problem is that the poller ultimately holds a reference to this
311  * object, so it is very difficult to know when is safe to free it, at least
312  * without some expensive synchronization.
313  *
314  * If we keep the object freelisted, in the worst case losing this race just
315  * becomes a spurious read notification on a reused fd.
316  */
317 
318 static grpc_fd* fd_freelist = nullptr;
319 static gpr_mu fd_freelist_mu;
320 
321 #ifndef NDEBUG
322 #define REF_BY(fd, n, reason) ref_by(fd, n, reason, __FILE__, __LINE__)
323 #define UNREF_BY(fd, n, reason) unref_by(fd, n, reason, __FILE__, __LINE__)
ref_by(grpc_fd * fd,int n,const char * reason,const char * file,int line)324 static void ref_by(grpc_fd* fd, int n, const char* reason, const char* file,
325                    int line) {
326   if (grpc_trace_fd_refcount.enabled()) {
327     gpr_log(GPR_DEBUG,
328             "FD %d %p   ref %d %" PRIdPTR " -> %" PRIdPTR " [%s; %s:%d]",
329             fd->fd, fd, n, gpr_atm_no_barrier_load(&fd->refst),
330             gpr_atm_no_barrier_load(&fd->refst) + n, reason, file, line);
331   }
332 #else
333 #define REF_BY(fd, n, reason) ref_by(fd, n)
334 #define UNREF_BY(fd, n, reason) unref_by(fd, n)
335 static void ref_by(grpc_fd* fd, int n) {
336 #endif
337   GPR_ASSERT(gpr_atm_no_barrier_fetch_add(&fd->refst, n) > 0);
338 }
339 
340 #ifndef NDEBUG
341 #define INVALIDATE_FD(fd) invalidate_fd(fd)
342 /* Since an fd is never really destroyed (i.e gpr_free() is not called), it is
343  * hard to cases where fd fields are accessed even after calling fd_destroy().
344  * The following invalidates fd fields to make catching such errors easier */
345 static void invalidate_fd(grpc_fd* fd) {
346   fd->fd = -1;
347   fd->salt = -1;
348   gpr_atm_no_barrier_store(&fd->refst, -1);
349   memset(&fd->orphan_mu, -1, sizeof(fd->orphan_mu));
350   memset(&fd->pollable_mu, -1, sizeof(fd->pollable_mu));
351   fd->pollable_obj = nullptr;
352   fd->on_done_closure = nullptr;
353   memset(&fd->iomgr_object, -1, sizeof(fd->iomgr_object));
354   fd->track_err = false;
355 }
356 #else
357 #define INVALIDATE_FD(fd)
358 #endif
359 
360 /* Uninitialize and add to the freelist */
361 static void fd_destroy(void* arg, grpc_error* error) {
362   grpc_fd* fd = static_cast<grpc_fd*>(arg);
363   grpc_iomgr_unregister_object(&fd->iomgr_object);
364   POLLABLE_UNREF(fd->pollable_obj, "fd_pollable");
365   gpr_mu_destroy(&fd->pollable_mu);
366   gpr_mu_destroy(&fd->orphan_mu);
367 
368   fd->read_closure->DestroyEvent();
369   fd->write_closure->DestroyEvent();
370   fd->error_closure->DestroyEvent();
371 
372   INVALIDATE_FD(fd);
373 
374   /* Add the fd to the freelist */
375   gpr_mu_lock(&fd_freelist_mu);
376   fd->freelist_next = fd_freelist;
377   fd_freelist = fd;
378   gpr_mu_unlock(&fd_freelist_mu);
379 }
380 
381 #ifndef NDEBUG
382 static void unref_by(grpc_fd* fd, int n, const char* reason, const char* file,
383                      int line) {
384   if (grpc_trace_fd_refcount.enabled()) {
385     gpr_log(GPR_DEBUG,
386             "FD %d %p unref %d %" PRIdPTR " -> %" PRIdPTR " [%s; %s:%d]",
387             fd->fd, fd, n, gpr_atm_no_barrier_load(&fd->refst),
388             gpr_atm_no_barrier_load(&fd->refst) - n, reason, file, line);
389   }
390 #else
391 static void unref_by(grpc_fd* fd, int n) {
392 #endif
393   gpr_atm old = gpr_atm_full_fetch_add(&fd->refst, -n);
394   if (old == n) {
395     GRPC_CLOSURE_SCHED(
396         GRPC_CLOSURE_CREATE(fd_destroy, fd, grpc_schedule_on_exec_ctx),
397         GRPC_ERROR_NONE);
398   } else {
399     GPR_ASSERT(old > n);
400   }
401 }
402 
403 static void fd_global_init(void) { gpr_mu_init(&fd_freelist_mu); }
404 
405 static void fd_global_shutdown(void) {
406   gpr_mu_lock(&fd_freelist_mu);
407   gpr_mu_unlock(&fd_freelist_mu);
408   while (fd_freelist != nullptr) {
409     grpc_fd* fd = fd_freelist;
410     fd_freelist = fd_freelist->freelist_next;
411     gpr_free(fd);
412   }
413   gpr_mu_destroy(&fd_freelist_mu);
414 }
415 
416 static grpc_fd* fd_create(int fd, const char* name, bool track_err) {
417   grpc_fd* new_fd = nullptr;
418 
419   gpr_mu_lock(&fd_freelist_mu);
420   if (fd_freelist != nullptr) {
421     new_fd = fd_freelist;
422     fd_freelist = fd_freelist->freelist_next;
423   }
424   gpr_mu_unlock(&fd_freelist_mu);
425 
426   if (new_fd == nullptr) {
427     new_fd = static_cast<grpc_fd*>(gpr_malloc(sizeof(grpc_fd)));
428     new_fd->read_closure.Init();
429     new_fd->write_closure.Init();
430     new_fd->error_closure.Init();
431   }
432 
433   new_fd->fd = fd;
434   new_fd->salt = gpr_atm_no_barrier_fetch_add(&g_fd_salt, 1);
435   gpr_atm_rel_store(&new_fd->refst, (gpr_atm)1);
436   gpr_mu_init(&new_fd->orphan_mu);
437   gpr_mu_init(&new_fd->pollable_mu);
438   new_fd->pollable_obj = nullptr;
439   new_fd->read_closure->InitEvent();
440   new_fd->write_closure->InitEvent();
441   new_fd->error_closure->InitEvent();
442   new_fd->freelist_next = nullptr;
443   new_fd->on_done_closure = nullptr;
444 
445   char* fd_name;
446   gpr_asprintf(&fd_name, "%s fd=%d", name, fd);
447   grpc_iomgr_register_object(&new_fd->iomgr_object, fd_name);
448 #ifndef NDEBUG
449   if (grpc_trace_fd_refcount.enabled()) {
450     gpr_log(GPR_DEBUG, "FD %d %p create %s", fd, new_fd, fd_name);
451   }
452 #endif
453   gpr_free(fd_name);
454 
455   new_fd->track_err = track_err;
456   return new_fd;
457 }
458 
459 static int fd_wrapped_fd(grpc_fd* fd) {
460   int ret_fd = fd->fd;
461   return (gpr_atm_acq_load(&fd->refst) & 1) ? ret_fd : -1;
462 }
463 
464 static void fd_orphan(grpc_fd* fd, grpc_closure* on_done, int* release_fd,
465                       const char* reason) {
466   bool is_fd_closed = false;
467 
468   gpr_mu_lock(&fd->orphan_mu);
469 
470   // Get the fd->pollable_obj and set the owner_orphaned on that pollable to
471   // true so that the pollable will no longer access its owner_fd field.
472   gpr_mu_lock(&fd->pollable_mu);
473   pollable* pollable_obj = fd->pollable_obj;
474   gpr_mu_unlock(&fd->pollable_mu);
475 
476   if (pollable_obj) {
477     gpr_mu_lock(&pollable_obj->owner_orphan_mu);
478     pollable_obj->owner_orphaned = true;
479   }
480 
481   fd->on_done_closure = on_done;
482 
483   /* If release_fd is not NULL, we should be relinquishing control of the file
484      descriptor fd->fd (but we still own the grpc_fd structure). */
485   if (release_fd != nullptr) {
486     *release_fd = fd->fd;
487   } else {
488     close(fd->fd);
489     is_fd_closed = true;
490   }
491 
492   // TODO(sreek): handle fd removal (where is_fd_closed=false)
493   if (!is_fd_closed) {
494     GRPC_FD_TRACE("epoll_fd %p (%d) was orphaned but not closed.", fd, fd->fd);
495   }
496 
497   /* Remove the active status but keep referenced. We want this grpc_fd struct
498      to be alive (and not added to freelist) until the end of this function */
499   REF_BY(fd, 1, reason);
500 
501   GRPC_CLOSURE_SCHED(fd->on_done_closure, GRPC_ERROR_NONE);
502 
503   if (pollable_obj) {
504     gpr_mu_unlock(&pollable_obj->owner_orphan_mu);
505   }
506 
507   gpr_mu_unlock(&fd->orphan_mu);
508 
509   UNREF_BY(fd, 2, reason); /* Drop the reference */
510 }
511 
512 static bool fd_is_shutdown(grpc_fd* fd) {
513   return fd->read_closure->IsShutdown();
514 }
515 
516 /* Might be called multiple times */
517 static void fd_shutdown(grpc_fd* fd, grpc_error* why) {
518   if (fd->read_closure->SetShutdown(GRPC_ERROR_REF(why))) {
519     if (shutdown(fd->fd, SHUT_RDWR)) {
520       if (errno != ENOTCONN) {
521         gpr_log(GPR_ERROR, "Error shutting down fd %d. errno: %d",
522                 grpc_fd_wrapped_fd(fd), errno);
523       }
524     }
525     fd->write_closure->SetShutdown(GRPC_ERROR_REF(why));
526     fd->error_closure->SetShutdown(GRPC_ERROR_REF(why));
527   }
528   GRPC_ERROR_UNREF(why);
529 }
530 
531 static void fd_notify_on_read(grpc_fd* fd, grpc_closure* closure) {
532   fd->read_closure->NotifyOn(closure);
533 }
534 
535 static void fd_notify_on_write(grpc_fd* fd, grpc_closure* closure) {
536   fd->write_closure->NotifyOn(closure);
537 }
538 
539 static void fd_notify_on_error(grpc_fd* fd, grpc_closure* closure) {
540   fd->error_closure->NotifyOn(closure);
541 }
542 
543 /*******************************************************************************
544  * Pollable Definitions
545  */
546 
547 static grpc_error* pollable_create(pollable_type type, pollable** p) {
548   *p = nullptr;
549 
550   int epfd = epoll_create1(EPOLL_CLOEXEC);
551   if (epfd == -1) {
552     return GRPC_OS_ERROR(errno, "epoll_create1");
553   }
554   GRPC_FD_TRACE("Pollable_create: created epfd: %d (type: %d)", epfd, type);
555   *p = static_cast<pollable*>(gpr_malloc(sizeof(**p)));
556   grpc_error* err = grpc_wakeup_fd_init(&(*p)->wakeup);
557   if (err != GRPC_ERROR_NONE) {
558     GRPC_FD_TRACE(
559         "Pollable_create: closed epfd: %d (type: %d). wakeupfd_init error",
560         epfd, type);
561     close(epfd);
562     gpr_free(*p);
563     *p = nullptr;
564     return err;
565   }
566   struct epoll_event ev;
567   ev.events = static_cast<uint32_t>(EPOLLIN | EPOLLET);
568   ev.data.ptr = (void*)(1 | (intptr_t) & (*p)->wakeup);
569   if (epoll_ctl(epfd, EPOLL_CTL_ADD, (*p)->wakeup.read_fd, &ev) != 0) {
570     err = GRPC_OS_ERROR(errno, "epoll_ctl");
571     GRPC_FD_TRACE(
572         "Pollable_create: closed epfd: %d (type: %d). epoll_ctl error", epfd,
573         type);
574     close(epfd);
575     grpc_wakeup_fd_destroy(&(*p)->wakeup);
576     gpr_free(*p);
577     *p = nullptr;
578     return err;
579   }
580 
581   (*p)->type = type;
582   gpr_ref_init(&(*p)->refs, 1);
583   gpr_mu_init(&(*p)->mu);
584   (*p)->epfd = epfd;
585   (*p)->owner_fd = nullptr;
586   gpr_mu_init(&(*p)->owner_orphan_mu);
587   (*p)->owner_orphaned = false;
588   (*p)->pollset_set = nullptr;
589   (*p)->next = (*p)->prev = *p;
590   (*p)->root_worker = nullptr;
591   (*p)->event_cursor = 0;
592   (*p)->event_count = 0;
593   (*p)->fd_cache_size = 0;
594   (*p)->fd_cache_counter = 0;
595   return GRPC_ERROR_NONE;
596 }
597 
598 #ifdef NDEBUG
599 static pollable* pollable_ref(pollable* p) {
600 #else
601 static pollable* pollable_ref(pollable* p, int line, const char* reason) {
602   if (grpc_trace_pollable_refcount.enabled()) {
603     int r = static_cast<int> gpr_atm_no_barrier_load(&p->refs.count);
604     gpr_log(__FILE__, line, GPR_LOG_SEVERITY_DEBUG,
605             "POLLABLE:%p   ref %d->%d %s", p, r, r + 1, reason);
606   }
607 #endif
608   gpr_ref(&p->refs);
609   return p;
610 }
611 
612 #ifdef NDEBUG
613 static void pollable_unref(pollable* p) {
614 #else
615 static void pollable_unref(pollable* p, int line, const char* reason) {
616   if (p == nullptr) return;
617   if (grpc_trace_pollable_refcount.enabled()) {
618     int r = static_cast<int> gpr_atm_no_barrier_load(&p->refs.count);
619     gpr_log(__FILE__, line, GPR_LOG_SEVERITY_DEBUG,
620             "POLLABLE:%p unref %d->%d %s", p, r, r - 1, reason);
621   }
622 #endif
623   if (p != nullptr && gpr_unref(&p->refs)) {
624     GRPC_FD_TRACE("pollable_unref: Closing epfd: %d", p->epfd);
625     close(p->epfd);
626     grpc_wakeup_fd_destroy(&p->wakeup);
627     gpr_mu_destroy(&p->owner_orphan_mu);
628     gpr_free(p);
629   }
630 }
631 
632 static grpc_error* pollable_add_fd(pollable* p, grpc_fd* fd) {
633   grpc_error* error = GRPC_ERROR_NONE;
634   static const char* err_desc = "pollable_add_fd";
635   const int epfd = p->epfd;
636   gpr_mu_lock(&p->mu);
637   p->fd_cache_counter++;
638 
639   // Handle the case of overflow for our cache counter by
640   // reseting the recency-counter on all cache objects
641   if (p->fd_cache_counter == 0) {
642     for (int i = 0; i < p->fd_cache_size; i++) {
643       p->fd_cache[i].last_used = 0;
644     }
645   }
646 
647   int lru_idx = 0;
648   for (int i = 0; i < p->fd_cache_size; i++) {
649     if (p->fd_cache[i].fd == fd->fd && p->fd_cache[i].salt == fd->salt) {
650       GRPC_STATS_INC_POLLSET_FD_CACHE_HITS();
651       p->fd_cache[i].last_used = p->fd_cache_counter;
652       gpr_mu_unlock(&p->mu);
653       return GRPC_ERROR_NONE;
654     } else if (p->fd_cache[i].last_used < p->fd_cache[lru_idx].last_used) {
655       lru_idx = i;
656     }
657   }
658 
659   // Add to cache
660   if (p->fd_cache_size < MAX_FDS_IN_CACHE) {
661     lru_idx = p->fd_cache_size;
662     p->fd_cache_size++;
663   }
664   p->fd_cache[lru_idx].fd = fd->fd;
665   p->fd_cache[lru_idx].salt = fd->salt;
666   p->fd_cache[lru_idx].last_used = p->fd_cache_counter;
667   gpr_mu_unlock(&p->mu);
668 
669   if (grpc_polling_trace.enabled()) {
670     gpr_log(GPR_INFO, "add fd %p (%d) to pollable %p", fd, fd->fd, p);
671   }
672 
673   struct epoll_event ev_fd;
674   ev_fd.events =
675       static_cast<uint32_t>(EPOLLET | EPOLLIN | EPOLLOUT | EPOLLEXCLUSIVE);
676   /* Use the second least significant bit of ev_fd.data.ptr to store track_err
677    * to avoid synchronization issues when accessing it after receiving an event.
678    * Accessing fd would be a data race there because the fd might have been
679    * returned to the free list at that point. */
680   ev_fd.data.ptr = reinterpret_cast<void*>(reinterpret_cast<intptr_t>(fd) |
681                                            (fd->track_err ? 2 : 0));
682   GRPC_STATS_INC_SYSCALL_EPOLL_CTL();
683   if (epoll_ctl(epfd, EPOLL_CTL_ADD, fd->fd, &ev_fd) != 0) {
684     switch (errno) {
685       case EEXIST:
686         break;
687       default:
688         append_error(&error, GRPC_OS_ERROR(errno, "epoll_ctl"), err_desc);
689     }
690   }
691 
692   return error;
693 }
694 
695 /*******************************************************************************
696  * Pollset Definitions
697  */
698 
699 GPR_TLS_DECL(g_current_thread_pollset);
700 GPR_TLS_DECL(g_current_thread_worker);
701 
702 /* Global state management */
703 static grpc_error* pollset_global_init(void) {
704   gpr_tls_init(&g_current_thread_pollset);
705   gpr_tls_init(&g_current_thread_worker);
706   return pollable_create(PO_EMPTY, &g_empty_pollable);
707 }
708 
709 static void pollset_global_shutdown(void) {
710   POLLABLE_UNREF(g_empty_pollable, "g_empty_pollable");
711   gpr_tls_destroy(&g_current_thread_pollset);
712   gpr_tls_destroy(&g_current_thread_worker);
713 }
714 
715 /* pollset->mu must be held while calling this function */
716 static void pollset_maybe_finish_shutdown(grpc_pollset* pollset) {
717   if (grpc_polling_trace.enabled()) {
718     gpr_log(GPR_INFO,
719             "PS:%p (pollable:%p) maybe_finish_shutdown sc=%p (target:!NULL) "
720             "rw=%p (target:NULL) cpsc=%d (target:0)",
721             pollset, pollset->active_pollable, pollset->shutdown_closure,
722             pollset->root_worker, pollset->containing_pollset_set_count);
723   }
724   if (pollset->shutdown_closure != nullptr && pollset->root_worker == nullptr &&
725       pollset->containing_pollset_set_count == 0) {
726     GPR_TIMER_MARK("pollset_finish_shutdown", 0);
727     GRPC_CLOSURE_SCHED(pollset->shutdown_closure, GRPC_ERROR_NONE);
728     pollset->shutdown_closure = nullptr;
729     pollset->already_shutdown = true;
730   }
731 }
732 
733 /* pollset->mu must be held before calling this function,
734  * pollset->active_pollable->mu & specific_worker->pollable_obj->mu must not be
735  * held */
736 static grpc_error* kick_one_worker(grpc_pollset_worker* specific_worker) {
737   GPR_TIMER_SCOPE("kick_one_worker", 0);
738   pollable* p = specific_worker->pollable_obj;
739   grpc_core::MutexLock lock(&p->mu);
740   GPR_ASSERT(specific_worker != nullptr);
741   if (specific_worker->kicked) {
742     if (grpc_polling_trace.enabled()) {
743       gpr_log(GPR_INFO, "PS:%p kicked_specific_but_already_kicked", p);
744     }
745     GRPC_STATS_INC_POLLSET_KICKED_AGAIN();
746     return GRPC_ERROR_NONE;
747   }
748   if (gpr_tls_get(&g_current_thread_worker) == (intptr_t)specific_worker) {
749     if (grpc_polling_trace.enabled()) {
750       gpr_log(GPR_INFO, "PS:%p kicked_specific_but_awake", p);
751     }
752     GRPC_STATS_INC_POLLSET_KICK_OWN_THREAD();
753     specific_worker->kicked = true;
754     return GRPC_ERROR_NONE;
755   }
756   if (specific_worker == p->root_worker) {
757     GRPC_STATS_INC_POLLSET_KICK_WAKEUP_FD();
758     if (grpc_polling_trace.enabled()) {
759       gpr_log(GPR_INFO, "PS:%p kicked_specific_via_wakeup_fd", p);
760     }
761     specific_worker->kicked = true;
762     grpc_error* error = grpc_wakeup_fd_wakeup(&p->wakeup);
763     return error;
764   }
765   if (specific_worker->initialized_cv) {
766     GRPC_STATS_INC_POLLSET_KICK_WAKEUP_CV();
767     if (grpc_polling_trace.enabled()) {
768       gpr_log(GPR_INFO, "PS:%p kicked_specific_via_cv", p);
769     }
770     specific_worker->kicked = true;
771     gpr_cv_signal(&specific_worker->cv);
772     return GRPC_ERROR_NONE;
773   }
774   // we can get here during end_worker after removing specific_worker from the
775   // pollable list but before removing it from the pollset list
776   return GRPC_ERROR_NONE;
777 }
778 
779 static grpc_error* pollset_kick(grpc_pollset* pollset,
780                                 grpc_pollset_worker* specific_worker) {
781   GPR_TIMER_SCOPE("pollset_kick", 0);
782   GRPC_STATS_INC_POLLSET_KICK();
783   if (grpc_polling_trace.enabled()) {
784     gpr_log(GPR_INFO,
785             "PS:%p kick %p tls_pollset=%p tls_worker=%p pollset.root_worker=%p",
786             pollset, specific_worker,
787             (void*)gpr_tls_get(&g_current_thread_pollset),
788             (void*)gpr_tls_get(&g_current_thread_worker), pollset->root_worker);
789   }
790   if (specific_worker == nullptr) {
791     if (gpr_tls_get(&g_current_thread_pollset) != (intptr_t)pollset) {
792       if (pollset->root_worker == nullptr) {
793         if (grpc_polling_trace.enabled()) {
794           gpr_log(GPR_INFO, "PS:%p kicked_any_without_poller", pollset);
795         }
796         GRPC_STATS_INC_POLLSET_KICKED_WITHOUT_POLLER();
797         pollset->kicked_without_poller = true;
798         return GRPC_ERROR_NONE;
799       } else {
800         // We've been asked to kick a poller, but we haven't been told which one
801         // ... any will do
802         // We look at the pollset worker list because:
803         // 1. the pollable list may include workers from other pollers, so we'd
804         //    need to do an O(N) search
805         // 2. we'd additionally need to take the pollable lock, which we've so
806         //    far avoided
807         // Now, we would prefer to wake a poller in cv_wait, and not in
808         // epoll_wait (since the latter would imply the need to do an additional
809         // wakeup)
810         // We know that if a worker is at the root of a pollable, it's (likely)
811         // also the root of a pollset, and we know that if a worker is NOT at
812         // the root of a pollset, it's (likely) not at the root of a pollable,
813         // so we take our chances and choose the SECOND worker enqueued against
814         // the pollset as a worker that's likely to be in cv_wait
815         return kick_one_worker(
816             pollset->root_worker->links[PWLINK_POLLSET].next);
817       }
818     } else {
819       if (grpc_polling_trace.enabled()) {
820         gpr_log(GPR_INFO, "PS:%p kicked_any_but_awake", pollset);
821       }
822       GRPC_STATS_INC_POLLSET_KICK_OWN_THREAD();
823       return GRPC_ERROR_NONE;
824     }
825   } else {
826     return kick_one_worker(specific_worker);
827   }
828 }
829 
830 static grpc_error* pollset_kick_all(grpc_pollset* pollset) {
831   GPR_TIMER_SCOPE("pollset_kick_all", 0);
832   grpc_error* error = GRPC_ERROR_NONE;
833   const char* err_desc = "pollset_kick_all";
834   grpc_pollset_worker* w = pollset->root_worker;
835   if (w != nullptr) {
836     do {
837       GRPC_STATS_INC_POLLSET_KICK();
838       append_error(&error, kick_one_worker(w), err_desc);
839       w = w->links[PWLINK_POLLSET].next;
840     } while (w != pollset->root_worker);
841   }
842   return error;
843 }
844 
845 static void pollset_init(grpc_pollset* pollset, gpr_mu** mu) {
846   gpr_mu_init(&pollset->mu);
847   gpr_atm_no_barrier_store(&pollset->worker_count, 0);
848   pollset->active_pollable = POLLABLE_REF(g_empty_pollable, "pollset");
849   pollset->kicked_without_poller = false;
850   pollset->shutdown_closure = nullptr;
851   pollset->already_shutdown = false;
852   pollset->root_worker = nullptr;
853   pollset->containing_pollset_set_count = 0;
854   *mu = &pollset->mu;
855 }
856 
857 static int poll_deadline_to_millis_timeout(grpc_millis millis) {
858   if (millis == GRPC_MILLIS_INF_FUTURE) return -1;
859   grpc_millis delta = millis - grpc_core::ExecCtx::Get()->Now();
860   if (delta > INT_MAX)
861     return INT_MAX;
862   else if (delta < 0)
863     return 0;
864   else
865     return static_cast<int>(delta);
866 }
867 
868 static void fd_become_readable(grpc_fd* fd) { fd->read_closure->SetReady(); }
869 
870 static void fd_become_writable(grpc_fd* fd) { fd->write_closure->SetReady(); }
871 
872 static void fd_has_errors(grpc_fd* fd) { fd->error_closure->SetReady(); }
873 
874 /* Get the pollable_obj attached to this fd. If none is attached, create a new
875  * pollable object (of type PO_FD), attach it to the fd and return it
876  *
877  * Note that if a pollable object is already attached to the fd, it may be of
878  * either PO_FD or PO_MULTI type */
879 static grpc_error* get_fd_pollable(grpc_fd* fd, pollable** p) {
880   gpr_mu_lock(&fd->pollable_mu);
881   grpc_error* error = GRPC_ERROR_NONE;
882   static const char* err_desc = "get_fd_pollable";
883   if (fd->pollable_obj == nullptr) {
884     if (append_error(&error, pollable_create(PO_FD, &fd->pollable_obj),
885                      err_desc)) {
886       fd->pollable_obj->owner_fd = fd;
887       if (!append_error(&error, pollable_add_fd(fd->pollable_obj, fd),
888                         err_desc)) {
889         POLLABLE_UNREF(fd->pollable_obj, "fd_pollable");
890         fd->pollable_obj = nullptr;
891       }
892     }
893   }
894   if (error == GRPC_ERROR_NONE) {
895     GPR_ASSERT(fd->pollable_obj != nullptr);
896     *p = POLLABLE_REF(fd->pollable_obj, "pollset");
897   } else {
898     GPR_ASSERT(fd->pollable_obj == nullptr);
899     *p = nullptr;
900   }
901   gpr_mu_unlock(&fd->pollable_mu);
902   return error;
903 }
904 
905 /* pollset->po.mu lock must be held by the caller before calling this */
906 static void pollset_shutdown(grpc_pollset* pollset, grpc_closure* closure) {
907   GPR_TIMER_SCOPE("pollset_shutdown", 0);
908   GPR_ASSERT(pollset->shutdown_closure == nullptr);
909   pollset->shutdown_closure = closure;
910   GRPC_LOG_IF_ERROR("pollset_shutdown", pollset_kick_all(pollset));
911   pollset_maybe_finish_shutdown(pollset);
912 }
913 
914 static grpc_error* pollable_process_events(grpc_pollset* pollset,
915                                            pollable* pollable_obj, bool drain) {
916   GPR_TIMER_SCOPE("pollable_process_events", 0);
917   static const char* err_desc = "pollset_process_events";
918   // Use a simple heuristic to determine how many fd events to process
919   // per loop iteration.  (events/workers)
920   int handle_count = 1;
921   int worker_count = gpr_atm_no_barrier_load(&pollset->worker_count);
922   GPR_ASSERT(worker_count > 0);
923   handle_count =
924       (pollable_obj->event_count - pollable_obj->event_cursor) / worker_count;
925   if (handle_count == 0) {
926     handle_count = 1;
927   } else if (handle_count > MAX_EPOLL_EVENTS_HANDLED_EACH_POLL_CALL) {
928     handle_count = MAX_EPOLL_EVENTS_HANDLED_EACH_POLL_CALL;
929   }
930   grpc_error* error = GRPC_ERROR_NONE;
931   for (int i = 0; (drain || i < handle_count) &&
932                   pollable_obj->event_cursor != pollable_obj->event_count;
933        i++) {
934     int n = pollable_obj->event_cursor++;
935     struct epoll_event* ev = &pollable_obj->events[n];
936     void* data_ptr = ev->data.ptr;
937     if (1 & (intptr_t)data_ptr) {
938       if (grpc_polling_trace.enabled()) {
939         gpr_log(GPR_INFO, "PS:%p got pollset_wakeup %p", pollset, data_ptr);
940       }
941       append_error(&error,
942                    grpc_wakeup_fd_consume_wakeup(
943                        (grpc_wakeup_fd*)((~static_cast<intptr_t>(1)) &
944                                          (intptr_t)data_ptr)),
945                    err_desc);
946     } else {
947       grpc_fd* fd =
948           reinterpret_cast<grpc_fd*>(reinterpret_cast<intptr_t>(data_ptr) & ~2);
949       bool track_err = reinterpret_cast<intptr_t>(data_ptr) & 2;
950       bool cancel = (ev->events & EPOLLHUP) != 0;
951       bool error = (ev->events & EPOLLERR) != 0;
952       bool read_ev = (ev->events & (EPOLLIN | EPOLLPRI)) != 0;
953       bool write_ev = (ev->events & EPOLLOUT) != 0;
954       bool err_fallback = error && !track_err;
955 
956       if (grpc_polling_trace.enabled()) {
957         gpr_log(GPR_INFO,
958                 "PS:%p got fd %p: cancel=%d read=%d "
959                 "write=%d",
960                 pollset, fd, cancel, read_ev, write_ev);
961       }
962       if (error && !err_fallback) {
963         fd_has_errors(fd);
964       }
965       if (read_ev || cancel || err_fallback) {
966         fd_become_readable(fd);
967       }
968       if (write_ev || cancel || err_fallback) {
969         fd_become_writable(fd);
970       }
971     }
972   }
973 
974   return error;
975 }
976 
977 /* pollset_shutdown is guaranteed to be called before pollset_destroy. */
978 static void pollset_destroy(grpc_pollset* pollset) {
979   POLLABLE_UNREF(pollset->active_pollable, "pollset");
980   pollset->active_pollable = nullptr;
981   gpr_mu_destroy(&pollset->mu);
982 }
983 
984 static grpc_error* pollable_epoll(pollable* p, grpc_millis deadline) {
985   GPR_TIMER_SCOPE("pollable_epoll", 0);
986   int timeout = poll_deadline_to_millis_timeout(deadline);
987 
988   if (grpc_polling_trace.enabled()) {
989     char* desc = pollable_desc(p);
990     gpr_log(GPR_INFO, "POLLABLE:%p[%s] poll for %dms", p, desc, timeout);
991     gpr_free(desc);
992   }
993 
994   if (timeout != 0) {
995     GRPC_SCHEDULING_START_BLOCKING_REGION;
996   }
997   int r;
998   do {
999     GRPC_STATS_INC_SYSCALL_POLL();
1000     r = epoll_wait(p->epfd, p->events, MAX_EPOLL_EVENTS, timeout);
1001   } while (r < 0 && errno == EINTR);
1002   if (timeout != 0) {
1003     GRPC_SCHEDULING_END_BLOCKING_REGION;
1004   }
1005 
1006   if (r < 0) return GRPC_OS_ERROR(errno, "epoll_wait");
1007 
1008   if (grpc_polling_trace.enabled()) {
1009     gpr_log(GPR_INFO, "POLLABLE:%p got %d events", p, r);
1010   }
1011 
1012   p->event_cursor = 0;
1013   p->event_count = r;
1014 
1015   return GRPC_ERROR_NONE;
1016 }
1017 
1018 /* Return true if first in list */
1019 static bool worker_insert(grpc_pollset_worker** root_worker,
1020                           grpc_pollset_worker* worker, pwlinks link) {
1021   if (*root_worker == nullptr) {
1022     *root_worker = worker;
1023     worker->links[link].next = worker->links[link].prev = worker;
1024     return true;
1025   } else {
1026     worker->links[link].next = *root_worker;
1027     worker->links[link].prev = worker->links[link].next->links[link].prev;
1028     worker->links[link].next->links[link].prev = worker;
1029     worker->links[link].prev->links[link].next = worker;
1030     return false;
1031   }
1032 }
1033 
1034 /* returns the new root IFF the root changed */
1035 typedef enum { WRR_NEW_ROOT, WRR_EMPTIED, WRR_REMOVED } worker_remove_result;
1036 
1037 static worker_remove_result worker_remove(grpc_pollset_worker** root_worker,
1038                                           grpc_pollset_worker* worker,
1039                                           pwlinks link) {
1040   if (worker == *root_worker) {
1041     if (worker == worker->links[link].next) {
1042       *root_worker = nullptr;
1043       return WRR_EMPTIED;
1044     } else {
1045       *root_worker = worker->links[link].next;
1046       worker->links[link].prev->links[link].next = worker->links[link].next;
1047       worker->links[link].next->links[link].prev = worker->links[link].prev;
1048       return WRR_NEW_ROOT;
1049     }
1050   } else {
1051     worker->links[link].prev->links[link].next = worker->links[link].next;
1052     worker->links[link].next->links[link].prev = worker->links[link].prev;
1053     return WRR_REMOVED;
1054   }
1055 }
1056 
1057 /* Return true if this thread should poll */
1058 static bool begin_worker(grpc_pollset* pollset, grpc_pollset_worker* worker,
1059                          grpc_pollset_worker** worker_hdl,
1060                          grpc_millis deadline) {
1061   GPR_TIMER_SCOPE("begin_worker", 0);
1062   bool do_poll =
1063       (pollset->shutdown_closure == nullptr && !pollset->already_shutdown);
1064   gpr_atm_no_barrier_fetch_add(&pollset->worker_count, 1);
1065   if (worker_hdl != nullptr) *worker_hdl = worker;
1066   worker->initialized_cv = false;
1067   worker->kicked = false;
1068   worker->pollset = pollset;
1069   worker->pollable_obj =
1070       POLLABLE_REF(pollset->active_pollable, "pollset_worker");
1071   worker_insert(&pollset->root_worker, worker, PWLINK_POLLSET);
1072   gpr_mu_lock(&worker->pollable_obj->mu);
1073   if (!worker_insert(&worker->pollable_obj->root_worker, worker,
1074                      PWLINK_POLLABLE)) {
1075     worker->initialized_cv = true;
1076     gpr_cv_init(&worker->cv);
1077     gpr_mu_unlock(&pollset->mu);
1078     if (grpc_polling_trace.enabled() &&
1079         worker->pollable_obj->root_worker != worker) {
1080       gpr_log(GPR_INFO, "PS:%p wait %p w=%p for %dms", pollset,
1081               worker->pollable_obj, worker,
1082               poll_deadline_to_millis_timeout(deadline));
1083     }
1084     while (do_poll && worker->pollable_obj->root_worker != worker) {
1085       if (gpr_cv_wait(&worker->cv, &worker->pollable_obj->mu,
1086                       grpc_millis_to_timespec(deadline, GPR_CLOCK_REALTIME))) {
1087         if (grpc_polling_trace.enabled()) {
1088           gpr_log(GPR_INFO, "PS:%p timeout_wait %p w=%p", pollset,
1089                   worker->pollable_obj, worker);
1090         }
1091         do_poll = false;
1092       } else if (worker->kicked) {
1093         if (grpc_polling_trace.enabled()) {
1094           gpr_log(GPR_INFO, "PS:%p wakeup %p w=%p", pollset,
1095                   worker->pollable_obj, worker);
1096         }
1097         do_poll = false;
1098       } else if (grpc_polling_trace.enabled() &&
1099                  worker->pollable_obj->root_worker != worker) {
1100         gpr_log(GPR_INFO, "PS:%p spurious_wakeup %p w=%p", pollset,
1101                 worker->pollable_obj, worker);
1102       }
1103     }
1104     grpc_core::ExecCtx::Get()->InvalidateNow();
1105   } else {
1106     gpr_mu_unlock(&pollset->mu);
1107   }
1108   gpr_mu_unlock(&worker->pollable_obj->mu);
1109 
1110   return do_poll;
1111 }
1112 
1113 static void end_worker(grpc_pollset* pollset, grpc_pollset_worker* worker,
1114                        grpc_pollset_worker** worker_hdl) {
1115   GPR_TIMER_SCOPE("end_worker", 0);
1116   gpr_mu_lock(&pollset->mu);
1117   gpr_mu_lock(&worker->pollable_obj->mu);
1118   switch (worker_remove(&worker->pollable_obj->root_worker, worker,
1119                         PWLINK_POLLABLE)) {
1120     case WRR_NEW_ROOT: {
1121       // wakeup new poller
1122       grpc_pollset_worker* new_root = worker->pollable_obj->root_worker;
1123       GPR_ASSERT(new_root->initialized_cv);
1124       gpr_cv_signal(&new_root->cv);
1125       break;
1126     }
1127     case WRR_EMPTIED:
1128       if (pollset->active_pollable != worker->pollable_obj) {
1129         // pollable no longer being polled: flush events
1130         pollable_process_events(pollset, worker->pollable_obj, true);
1131       }
1132       break;
1133     case WRR_REMOVED:
1134       break;
1135   }
1136   gpr_mu_unlock(&worker->pollable_obj->mu);
1137   POLLABLE_UNREF(worker->pollable_obj, "pollset_worker");
1138   if (worker_remove(&pollset->root_worker, worker, PWLINK_POLLSET) ==
1139       WRR_EMPTIED) {
1140     pollset_maybe_finish_shutdown(pollset);
1141   }
1142   if (worker->initialized_cv) {
1143     gpr_cv_destroy(&worker->cv);
1144   }
1145   gpr_atm_no_barrier_fetch_add(&pollset->worker_count, -1);
1146 }
1147 
1148 #ifndef NDEBUG
1149 static long gettid(void) { return syscall(__NR_gettid); }
1150 #endif
1151 
1152 /* pollset->mu lock must be held by the caller before calling this.
1153    The function pollset_work() may temporarily release the lock (pollset->po.mu)
1154    during the course of its execution but it will always re-acquire the lock and
1155    ensure that it is held by the time the function returns */
1156 static grpc_error* pollset_work(grpc_pollset* pollset,
1157                                 grpc_pollset_worker** worker_hdl,
1158                                 grpc_millis deadline) {
1159   GPR_TIMER_SCOPE("pollset_work", 0);
1160 #ifdef GRPC_EPOLLEX_CREATE_WORKERS_ON_HEAP
1161   grpc_pollset_worker* worker =
1162       (grpc_pollset_worker*)gpr_malloc(sizeof(*worker));
1163 #define WORKER_PTR (worker)
1164 #else
1165   grpc_pollset_worker worker;
1166 #define WORKER_PTR (&worker)
1167 #endif
1168 #ifndef NDEBUG
1169   WORKER_PTR->originator = gettid();
1170 #endif
1171   if (grpc_polling_trace.enabled()) {
1172     gpr_log(GPR_INFO,
1173             "PS:%p work hdl=%p worker=%p now=%" PRId64 " deadline=%" PRId64
1174             " kwp=%d pollable=%p",
1175             pollset, worker_hdl, WORKER_PTR, grpc_core::ExecCtx::Get()->Now(),
1176             deadline, pollset->kicked_without_poller, pollset->active_pollable);
1177   }
1178   static const char* err_desc = "pollset_work";
1179   grpc_error* error = GRPC_ERROR_NONE;
1180   if (pollset->kicked_without_poller) {
1181     pollset->kicked_without_poller = false;
1182   } else {
1183     if (begin_worker(pollset, WORKER_PTR, worker_hdl, deadline)) {
1184       gpr_tls_set(&g_current_thread_pollset, (intptr_t)pollset);
1185       gpr_tls_set(&g_current_thread_worker, (intptr_t)WORKER_PTR);
1186       if (WORKER_PTR->pollable_obj->event_cursor ==
1187           WORKER_PTR->pollable_obj->event_count) {
1188         append_error(&error, pollable_epoll(WORKER_PTR->pollable_obj, deadline),
1189                      err_desc);
1190       }
1191       append_error(
1192           &error,
1193           pollable_process_events(pollset, WORKER_PTR->pollable_obj, false),
1194           err_desc);
1195       grpc_core::ExecCtx::Get()->Flush();
1196       gpr_tls_set(&g_current_thread_pollset, 0);
1197       gpr_tls_set(&g_current_thread_worker, 0);
1198     }
1199     end_worker(pollset, WORKER_PTR, worker_hdl);
1200   }
1201 #ifdef GRPC_EPOLLEX_CREATE_WORKERS_ON_HEAP
1202   gpr_free(worker);
1203 #endif
1204 #undef WORKER_PTR
1205   return error;
1206 }
1207 
1208 static grpc_error* pollset_transition_pollable_from_empty_to_fd_locked(
1209     grpc_pollset* pollset, grpc_fd* fd) {
1210   static const char* err_desc = "pollset_transition_pollable_from_empty_to_fd";
1211   grpc_error* error = GRPC_ERROR_NONE;
1212   if (grpc_polling_trace.enabled()) {
1213     gpr_log(GPR_INFO,
1214             "PS:%p add fd %p (%d); transition pollable from empty to fd",
1215             pollset, fd, fd->fd);
1216   }
1217   append_error(&error, pollset_kick_all(pollset), err_desc);
1218   POLLABLE_UNREF(pollset->active_pollable, "pollset");
1219   append_error(&error, get_fd_pollable(fd, &pollset->active_pollable),
1220                err_desc);
1221   return error;
1222 }
1223 
1224 static grpc_error* pollset_transition_pollable_from_fd_to_multi_locked(
1225     grpc_pollset* pollset, grpc_fd* and_add_fd) {
1226   static const char* err_desc = "pollset_transition_pollable_from_fd_to_multi";
1227   grpc_error* error = GRPC_ERROR_NONE;
1228   if (grpc_polling_trace.enabled()) {
1229     gpr_log(
1230         GPR_INFO,
1231         "PS:%p add fd %p (%d); transition pollable from fd %p to multipoller",
1232         pollset, and_add_fd, and_add_fd ? and_add_fd->fd : -1,
1233         pollset->active_pollable->owner_fd);
1234   }
1235   append_error(&error, pollset_kick_all(pollset), err_desc);
1236   grpc_fd* initial_fd = pollset->active_pollable->owner_fd;
1237   POLLABLE_UNREF(pollset->active_pollable, "pollset");
1238   pollset->active_pollable = nullptr;
1239   if (append_error(&error, pollable_create(PO_MULTI, &pollset->active_pollable),
1240                    err_desc)) {
1241     append_error(&error, pollable_add_fd(pollset->active_pollable, initial_fd),
1242                  err_desc);
1243     if (and_add_fd != nullptr) {
1244       append_error(&error,
1245                    pollable_add_fd(pollset->active_pollable, and_add_fd),
1246                    err_desc);
1247     }
1248   }
1249   return error;
1250 }
1251 
1252 /* expects pollsets locked, flag whether fd is locked or not */
1253 static grpc_error* pollset_add_fd_locked(grpc_pollset* pollset, grpc_fd* fd) {
1254   grpc_error* error = GRPC_ERROR_NONE;
1255   pollable* po_at_start =
1256       POLLABLE_REF(pollset->active_pollable, "pollset_add_fd");
1257   switch (pollset->active_pollable->type) {
1258     case PO_EMPTY:
1259       /* empty pollable --> single fd pollable */
1260       error = pollset_transition_pollable_from_empty_to_fd_locked(pollset, fd);
1261       break;
1262     case PO_FD:
1263       gpr_mu_lock(&po_at_start->owner_orphan_mu);
1264       if (po_at_start->owner_orphaned) {
1265         error =
1266             pollset_transition_pollable_from_empty_to_fd_locked(pollset, fd);
1267       } else {
1268         /* fd --> multipoller */
1269         error =
1270             pollset_transition_pollable_from_fd_to_multi_locked(pollset, fd);
1271       }
1272       gpr_mu_unlock(&po_at_start->owner_orphan_mu);
1273       break;
1274     case PO_MULTI:
1275       error = pollable_add_fd(pollset->active_pollable, fd);
1276       break;
1277   }
1278   if (error != GRPC_ERROR_NONE) {
1279     POLLABLE_UNREF(pollset->active_pollable, "pollset");
1280     pollset->active_pollable = po_at_start;
1281   } else {
1282     POLLABLE_UNREF(po_at_start, "pollset_add_fd");
1283   }
1284   return error;
1285 }
1286 
1287 static grpc_error* pollset_as_multipollable_locked(grpc_pollset* pollset,
1288                                                    pollable** pollable_obj) {
1289   grpc_error* error = GRPC_ERROR_NONE;
1290   pollable* po_at_start =
1291       POLLABLE_REF(pollset->active_pollable, "pollset_as_multipollable");
1292   switch (pollset->active_pollable->type) {
1293     case PO_EMPTY:
1294       POLLABLE_UNREF(pollset->active_pollable, "pollset");
1295       error = pollable_create(PO_MULTI, &pollset->active_pollable);
1296       /* Any workers currently polling on this pollset must now be woked up so
1297        * that they can pick up the new active_pollable */
1298       if (grpc_polling_trace.enabled()) {
1299         gpr_log(GPR_INFO,
1300                 "PS:%p active pollable transition from empty to multi",
1301                 pollset);
1302       }
1303       static const char* err_desc =
1304           "pollset_as_multipollable_locked: empty -> multi";
1305       append_error(&error, pollset_kick_all(pollset), err_desc);
1306       break;
1307     case PO_FD:
1308       gpr_mu_lock(&po_at_start->owner_orphan_mu);
1309       if (po_at_start->owner_orphaned) {
1310         // Unlock before Unref'ing the pollable
1311         gpr_mu_unlock(&po_at_start->owner_orphan_mu);
1312         POLLABLE_UNREF(pollset->active_pollable, "pollset");
1313         error = pollable_create(PO_MULTI, &pollset->active_pollable);
1314       } else {
1315         error = pollset_transition_pollable_from_fd_to_multi_locked(pollset,
1316                                                                     nullptr);
1317         gpr_mu_unlock(&po_at_start->owner_orphan_mu);
1318       }
1319       break;
1320     case PO_MULTI:
1321       break;
1322   }
1323   if (error != GRPC_ERROR_NONE) {
1324     POLLABLE_UNREF(pollset->active_pollable, "pollset");
1325     pollset->active_pollable = po_at_start;
1326     *pollable_obj = nullptr;
1327   } else {
1328     *pollable_obj = POLLABLE_REF(pollset->active_pollable, "pollset_set");
1329     POLLABLE_UNREF(po_at_start, "pollset_as_multipollable");
1330   }
1331   return error;
1332 }
1333 
1334 static void pollset_add_fd(grpc_pollset* pollset, grpc_fd* fd) {
1335   GPR_TIMER_SCOPE("pollset_add_fd", 0);
1336   gpr_mu_lock(&pollset->mu);
1337   grpc_error* error = pollset_add_fd_locked(pollset, fd);
1338   gpr_mu_unlock(&pollset->mu);
1339   GRPC_LOG_IF_ERROR("pollset_add_fd", error);
1340 }
1341 
1342 /*******************************************************************************
1343  * Pollset-set Definitions
1344  */
1345 
1346 static grpc_pollset_set* pss_lock_adam(grpc_pollset_set* pss) {
1347   gpr_mu_lock(&pss->mu);
1348   while (pss->parent != nullptr) {
1349     gpr_mu_unlock(&pss->mu);
1350     pss = pss->parent;
1351     gpr_mu_lock(&pss->mu);
1352   }
1353   return pss;
1354 }
1355 
1356 static grpc_pollset_set* pollset_set_create(void) {
1357   grpc_pollset_set* pss =
1358       static_cast<grpc_pollset_set*>(gpr_zalloc(sizeof(*pss)));
1359   gpr_mu_init(&pss->mu);
1360   gpr_ref_init(&pss->refs, 1);
1361   return pss;
1362 }
1363 
1364 static void pollset_set_unref(grpc_pollset_set* pss) {
1365   if (pss == nullptr) return;
1366   if (!gpr_unref(&pss->refs)) return;
1367   pollset_set_unref(pss->parent);
1368   gpr_mu_destroy(&pss->mu);
1369   for (size_t i = 0; i < pss->pollset_count; i++) {
1370     gpr_mu_lock(&pss->pollsets[i]->mu);
1371     if (0 == --pss->pollsets[i]->containing_pollset_set_count) {
1372       pollset_maybe_finish_shutdown(pss->pollsets[i]);
1373     }
1374     gpr_mu_unlock(&pss->pollsets[i]->mu);
1375   }
1376   for (size_t i = 0; i < pss->fd_count; i++) {
1377     UNREF_BY(pss->fds[i], 2, "pollset_set");
1378   }
1379   gpr_free(pss->pollsets);
1380   gpr_free(pss->fds);
1381   gpr_free(pss);
1382 }
1383 
1384 static void pollset_set_add_fd(grpc_pollset_set* pss, grpc_fd* fd) {
1385   GPR_TIMER_SCOPE("pollset_set_add_fd", 0);
1386   if (grpc_polling_trace.enabled()) {
1387     gpr_log(GPR_INFO, "PSS:%p: add fd %p (%d)", pss, fd, fd->fd);
1388   }
1389   grpc_error* error = GRPC_ERROR_NONE;
1390   static const char* err_desc = "pollset_set_add_fd";
1391   pss = pss_lock_adam(pss);
1392   for (size_t i = 0; i < pss->pollset_count; i++) {
1393     append_error(&error, pollable_add_fd(pss->pollsets[i]->active_pollable, fd),
1394                  err_desc);
1395   }
1396   if (pss->fd_count == pss->fd_capacity) {
1397     pss->fd_capacity = GPR_MAX(pss->fd_capacity * 2, 8);
1398     pss->fds = static_cast<grpc_fd**>(
1399         gpr_realloc(pss->fds, pss->fd_capacity * sizeof(*pss->fds)));
1400   }
1401   REF_BY(fd, 2, "pollset_set");
1402   pss->fds[pss->fd_count++] = fd;
1403   gpr_mu_unlock(&pss->mu);
1404 
1405   GRPC_LOG_IF_ERROR(err_desc, error);
1406 }
1407 
1408 static void pollset_set_del_fd(grpc_pollset_set* pss, grpc_fd* fd) {
1409   GPR_TIMER_SCOPE("pollset_set_del_fd", 0);
1410   if (grpc_polling_trace.enabled()) {
1411     gpr_log(GPR_INFO, "PSS:%p: del fd %p", pss, fd);
1412   }
1413   pss = pss_lock_adam(pss);
1414   size_t i;
1415   for (i = 0; i < pss->fd_count; i++) {
1416     if (pss->fds[i] == fd) {
1417       UNREF_BY(fd, 2, "pollset_set");
1418       break;
1419     }
1420   }
1421   GPR_ASSERT(i != pss->fd_count);
1422   for (; i < pss->fd_count - 1; i++) {
1423     pss->fds[i] = pss->fds[i + 1];
1424   }
1425   pss->fd_count--;
1426   gpr_mu_unlock(&pss->mu);
1427 }
1428 
1429 static void pollset_set_del_pollset(grpc_pollset_set* pss, grpc_pollset* ps) {
1430   GPR_TIMER_SCOPE("pollset_set_del_pollset", 0);
1431   if (grpc_polling_trace.enabled()) {
1432     gpr_log(GPR_INFO, "PSS:%p: del pollset %p", pss, ps);
1433   }
1434   pss = pss_lock_adam(pss);
1435   size_t i;
1436   for (i = 0; i < pss->pollset_count; i++) {
1437     if (pss->pollsets[i] == ps) {
1438       break;
1439     }
1440   }
1441   GPR_ASSERT(i != pss->pollset_count);
1442   for (; i < pss->pollset_count - 1; i++) {
1443     pss->pollsets[i] = pss->pollsets[i + 1];
1444   }
1445   pss->pollset_count--;
1446   gpr_mu_unlock(&pss->mu);
1447   gpr_mu_lock(&ps->mu);
1448   if (0 == --ps->containing_pollset_set_count) {
1449     pollset_maybe_finish_shutdown(ps);
1450   }
1451   gpr_mu_unlock(&ps->mu);
1452 }
1453 
1454 // add all fds to pollables, and output a new array of unorphaned out_fds
1455 // assumes pollsets are multipollable
1456 static grpc_error* add_fds_to_pollsets(grpc_fd** fds, size_t fd_count,
1457                                        grpc_pollset** pollsets,
1458                                        size_t pollset_count,
1459                                        const char* err_desc, grpc_fd** out_fds,
1460                                        size_t* out_fd_count) {
1461   GPR_TIMER_SCOPE("add_fds_to_pollsets", 0);
1462   grpc_error* error = GRPC_ERROR_NONE;
1463   for (size_t i = 0; i < fd_count; i++) {
1464     gpr_mu_lock(&fds[i]->orphan_mu);
1465     if ((gpr_atm_no_barrier_load(&fds[i]->refst) & 1) == 0) {
1466       gpr_mu_unlock(&fds[i]->orphan_mu);
1467       UNREF_BY(fds[i], 2, "pollset_set");
1468     } else {
1469       for (size_t j = 0; j < pollset_count; j++) {
1470         append_error(&error,
1471                      pollable_add_fd(pollsets[j]->active_pollable, fds[i]),
1472                      err_desc);
1473       }
1474       gpr_mu_unlock(&fds[i]->orphan_mu);
1475       out_fds[(*out_fd_count)++] = fds[i];
1476     }
1477   }
1478   return error;
1479 }
1480 
1481 static void pollset_set_add_pollset(grpc_pollset_set* pss, grpc_pollset* ps) {
1482   GPR_TIMER_SCOPE("pollset_set_add_pollset", 0);
1483   if (grpc_polling_trace.enabled()) {
1484     gpr_log(GPR_INFO, "PSS:%p: add pollset %p", pss, ps);
1485   }
1486   grpc_error* error = GRPC_ERROR_NONE;
1487   static const char* err_desc = "pollset_set_add_pollset";
1488   pollable* pollable_obj = nullptr;
1489   gpr_mu_lock(&ps->mu);
1490   if (!GRPC_LOG_IF_ERROR(err_desc,
1491                          pollset_as_multipollable_locked(ps, &pollable_obj))) {
1492     GPR_ASSERT(pollable_obj == nullptr);
1493     gpr_mu_unlock(&ps->mu);
1494     return;
1495   }
1496   ps->containing_pollset_set_count++;
1497   gpr_mu_unlock(&ps->mu);
1498   pss = pss_lock_adam(pss);
1499   size_t initial_fd_count = pss->fd_count;
1500   pss->fd_count = 0;
1501   append_error(&error,
1502                add_fds_to_pollsets(pss->fds, initial_fd_count, &ps, 1, err_desc,
1503                                    pss->fds, &pss->fd_count),
1504                err_desc);
1505   if (pss->pollset_count == pss->pollset_capacity) {
1506     pss->pollset_capacity = GPR_MAX(pss->pollset_capacity * 2, 8);
1507     pss->pollsets = static_cast<grpc_pollset**>(gpr_realloc(
1508         pss->pollsets, pss->pollset_capacity * sizeof(*pss->pollsets)));
1509   }
1510   pss->pollsets[pss->pollset_count++] = ps;
1511   gpr_mu_unlock(&pss->mu);
1512   POLLABLE_UNREF(pollable_obj, "pollset_set");
1513 
1514   GRPC_LOG_IF_ERROR(err_desc, error);
1515 }
1516 
1517 static void pollset_set_add_pollset_set(grpc_pollset_set* a,
1518                                         grpc_pollset_set* b) {
1519   GPR_TIMER_SCOPE("pollset_set_add_pollset_set", 0);
1520   if (grpc_polling_trace.enabled()) {
1521     gpr_log(GPR_INFO, "PSS: merge (%p, %p)", a, b);
1522   }
1523   grpc_error* error = GRPC_ERROR_NONE;
1524   static const char* err_desc = "pollset_set_add_fd";
1525   for (;;) {
1526     if (a == b) {
1527       // pollset ancestors are the same: nothing to do
1528       return;
1529     }
1530     if (a > b) {
1531       GPR_SWAP(grpc_pollset_set*, a, b);
1532     }
1533     gpr_mu* a_mu = &a->mu;
1534     gpr_mu* b_mu = &b->mu;
1535     gpr_mu_lock(a_mu);
1536     gpr_mu_lock(b_mu);
1537     if (a->parent != nullptr) {
1538       a = a->parent;
1539     } else if (b->parent != nullptr) {
1540       b = b->parent;
1541     } else {
1542       break;  // exit loop, both pollsets locked
1543     }
1544     gpr_mu_unlock(a_mu);
1545     gpr_mu_unlock(b_mu);
1546   }
1547   // try to do the least copying possible
1548   // TODO(sreek): there's probably a better heuristic here
1549   const size_t a_size = a->fd_count + a->pollset_count;
1550   const size_t b_size = b->fd_count + b->pollset_count;
1551   if (b_size > a_size) {
1552     GPR_SWAP(grpc_pollset_set*, a, b);
1553   }
1554   if (grpc_polling_trace.enabled()) {
1555     gpr_log(GPR_INFO, "PSS: parent %p to %p", b, a);
1556   }
1557   gpr_ref(&a->refs);
1558   b->parent = a;
1559   if (a->fd_capacity < a->fd_count + b->fd_count) {
1560     a->fd_capacity = GPR_MAX(2 * a->fd_capacity, a->fd_count + b->fd_count);
1561     a->fds = static_cast<grpc_fd**>(
1562         gpr_realloc(a->fds, a->fd_capacity * sizeof(*a->fds)));
1563   }
1564   size_t initial_a_fd_count = a->fd_count;
1565   a->fd_count = 0;
1566   append_error(
1567       &error,
1568       add_fds_to_pollsets(a->fds, initial_a_fd_count, b->pollsets,
1569                           b->pollset_count, "merge_a2b", a->fds, &a->fd_count),
1570       err_desc);
1571   append_error(
1572       &error,
1573       add_fds_to_pollsets(b->fds, b->fd_count, a->pollsets, a->pollset_count,
1574                           "merge_b2a", a->fds, &a->fd_count),
1575       err_desc);
1576   if (a->pollset_capacity < a->pollset_count + b->pollset_count) {
1577     a->pollset_capacity =
1578         GPR_MAX(2 * a->pollset_capacity, a->pollset_count + b->pollset_count);
1579     a->pollsets = static_cast<grpc_pollset**>(
1580         gpr_realloc(a->pollsets, a->pollset_capacity * sizeof(*a->pollsets)));
1581   }
1582   if (b->pollset_count > 0) {
1583     memcpy(a->pollsets + a->pollset_count, b->pollsets,
1584            b->pollset_count * sizeof(*b->pollsets));
1585   }
1586   a->pollset_count += b->pollset_count;
1587   gpr_free(b->fds);
1588   gpr_free(b->pollsets);
1589   b->fds = nullptr;
1590   b->pollsets = nullptr;
1591   b->fd_count = b->fd_capacity = b->pollset_count = b->pollset_capacity = 0;
1592   gpr_mu_unlock(&a->mu);
1593   gpr_mu_unlock(&b->mu);
1594 }
1595 
1596 static void pollset_set_del_pollset_set(grpc_pollset_set* bag,
1597                                         grpc_pollset_set* item) {}
1598 
1599 /*******************************************************************************
1600  * Event engine binding
1601  */
1602 
1603 static void shutdown_engine(void) {
1604   fd_global_shutdown();
1605   pollset_global_shutdown();
1606 }
1607 
1608 static const grpc_event_engine_vtable vtable = {
1609     sizeof(grpc_pollset),
1610     true,
1611 
1612     fd_create,
1613     fd_wrapped_fd,
1614     fd_orphan,
1615     fd_shutdown,
1616     fd_notify_on_read,
1617     fd_notify_on_write,
1618     fd_notify_on_error,
1619     fd_become_readable,
1620     fd_become_writable,
1621     fd_has_errors,
1622     fd_is_shutdown,
1623 
1624     pollset_init,
1625     pollset_shutdown,
1626     pollset_destroy,
1627     pollset_work,
1628     pollset_kick,
1629     pollset_add_fd,
1630 
1631     pollset_set_create,
1632     pollset_set_unref,  // destroy ==> unref 1 public ref
1633     pollset_set_add_pollset,
1634     pollset_set_del_pollset,
1635     pollset_set_add_pollset_set,
1636     pollset_set_del_pollset_set,
1637     pollset_set_add_fd,
1638     pollset_set_del_fd,
1639 
1640     shutdown_engine,
1641 };
1642 
1643 const grpc_event_engine_vtable* grpc_init_epollex_linux(
1644     bool explicitly_requested) {
1645   if (!grpc_has_wakeup_fd()) {
1646     gpr_log(GPR_ERROR, "Skipping epollex because of no wakeup fd.");
1647     return nullptr;
1648   }
1649 
1650   if (!grpc_is_epollexclusive_available()) {
1651     gpr_log(GPR_INFO, "Skipping epollex because it is not supported.");
1652     return nullptr;
1653   }
1654 
1655   fd_global_init();
1656 
1657   if (!GRPC_LOG_IF_ERROR("pollset_global_init", pollset_global_init())) {
1658     pollset_global_shutdown();
1659     fd_global_shutdown();
1660     return nullptr;
1661   }
1662 
1663   return &vtable;
1664 }
1665 
1666 #else /* defined(GRPC_LINUX_EPOLL_CREATE1) */
1667 #if defined(GRPC_POSIX_SOCKET_EV_EPOLLEX)
1668 #include "src/core/lib/iomgr/ev_epollex_linux.h"
1669 /* If GRPC_LINUX_EPOLL_CREATE1 is not defined, it means
1670    epoll_create1 is not available. Return NULL */
1671 const grpc_event_engine_vtable* grpc_init_epollex_linux(
1672     bool explicitly_requested) {
1673   return nullptr;
1674 }
1675 #endif /* defined(GRPC_POSIX_SOCKET_EV_EPOLLEX) */
1676 
1677 #endif /* !defined(GRPC_LINUX_EPOLL_CREATE1) */
1678