Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * condition_variable.c
4 : * Implementation of condition variables. Condition variables provide
5 : * a way for one process to wait until a specific condition occurs,
6 : * without needing to know the specific identity of the process for
7 : * which they are waiting. Waits for condition variables can be
8 : * interrupted, unlike LWLock waits. Condition variables are safe
9 : * to use within dynamic shared memory segments.
10 : *
11 : * Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group
12 : * Portions Copyright (c) 1994, Regents of the University of California
13 : *
14 : * src/backend/storage/lmgr/condition_variable.c
15 : *
16 : *-------------------------------------------------------------------------
17 : */
18 :
19 : #include "postgres.h"
20 :
21 : #include "miscadmin.h"
22 : #include "portability/instr_time.h"
23 : #include "storage/condition_variable.h"
24 : #include "storage/proc.h"
25 : #include "storage/proclist.h"
26 : #include "storage/spin.h"
27 :
28 : /* Initially, we are not prepared to sleep on any condition variable. */
29 : static ConditionVariable *cv_sleep_target = NULL;
30 :
31 : /*
32 : * Initialize a condition variable.
33 : */
34 : void
35 34069400 : ConditionVariableInit(ConditionVariable *cv)
36 : {
37 34069400 : SpinLockInit(&cv->mutex);
38 34069400 : proclist_init(&cv->wakeup);
39 34069400 : }
40 :
41 : /*
42 : * Prepare to wait on a given condition variable.
43 : *
44 : * This can optionally be called before entering a test/sleep loop.
45 : * Doing so is more efficient if we'll need to sleep at least once.
46 : * However, if the first test of the exit condition is likely to succeed,
47 : * it's more efficient to omit the ConditionVariablePrepareToSleep call.
48 : * See comments in ConditionVariableSleep for more detail.
49 : *
50 : * Caution: "before entering the loop" means you *must* test the exit
51 : * condition between calling ConditionVariablePrepareToSleep and calling
52 : * ConditionVariableSleep. If that is inconvenient, omit calling
53 : * ConditionVariablePrepareToSleep.
54 : */
55 : void
56 1328340 : ConditionVariablePrepareToSleep(ConditionVariable *cv)
57 : {
58 1328340 : int pgprocno = MyProcNumber;
59 :
60 : /*
61 : * If some other sleep is already prepared, cancel it; this is necessary
62 : * because we have just one static variable tracking the prepared sleep,
63 : * and also only one cvWaitLink in our PGPROC. It's okay to do this
64 : * because whenever control does return to the other test-and-sleep loop,
65 : * its ConditionVariableSleep call will just re-establish that sleep as
66 : * the prepared one.
67 : */
68 1328340 : if (cv_sleep_target != NULL)
69 3192 : ConditionVariableCancelSleep();
70 :
71 : /* Record the condition variable on which we will sleep. */
72 1328340 : cv_sleep_target = cv;
73 :
74 : /* Add myself to the wait queue. */
75 1328340 : SpinLockAcquire(&cv->mutex);
76 1328340 : proclist_push_tail(&cv->wakeup, pgprocno, cvWaitLink);
77 1328340 : SpinLockRelease(&cv->mutex);
78 1328340 : }
79 :
80 : /*
81 : * Wait for the given condition variable to be signaled.
82 : *
83 : * This should be called in a predicate loop that tests for a specific exit
84 : * condition and otherwise sleeps, like so:
85 : *
86 : * ConditionVariablePrepareToSleep(cv); // optional
87 : * while (condition for which we are waiting is not true)
88 : * ConditionVariableSleep(cv, wait_event_info);
89 : * ConditionVariableCancelSleep();
90 : *
91 : * wait_event_info should be a value from one of the WaitEventXXX enums
92 : * defined in pgstat.h. This controls the contents of pg_stat_activity's
93 : * wait_event_type and wait_event columns while waiting.
94 : */
95 : void
96 1900962 : ConditionVariableSleep(ConditionVariable *cv, uint32 wait_event_info)
97 : {
98 1900962 : (void) ConditionVariableTimedSleep(cv, -1 /* no timeout */ ,
99 : wait_event_info);
100 1900960 : }
101 :
102 : /*
103 : * Wait for a condition variable to be signaled or a timeout to be reached.
104 : *
105 : * The "timeout" is given in milliseconds.
106 : *
107 : * Returns true when timeout expires, otherwise returns false.
108 : *
109 : * See ConditionVariableSleep() for general usage.
110 : */
111 : bool
112 1901756 : ConditionVariableTimedSleep(ConditionVariable *cv, long timeout,
113 : uint32 wait_event_info)
114 : {
115 1901756 : long cur_timeout = -1;
116 : instr_time start_time;
117 : instr_time cur_time;
118 : int wait_events;
119 :
120 : /*
121 : * If the caller didn't prepare to sleep explicitly, then do so now and
122 : * return immediately. The caller's predicate loop should immediately
123 : * call again if its exit condition is not yet met. This will result in
124 : * the exit condition being tested twice before we first sleep. The extra
125 : * test can be prevented by calling ConditionVariablePrepareToSleep(cv)
126 : * first. Whether it's worth doing that depends on whether you expect the
127 : * exit condition to be met initially, in which case skipping the prepare
128 : * is recommended because it avoids manipulations of the wait list, or not
129 : * met initially, in which case preparing first is better because it
130 : * avoids one extra test of the exit condition.
131 : *
132 : * If we are currently prepared to sleep on some other CV, we just cancel
133 : * that and prepare this one; see ConditionVariablePrepareToSleep.
134 : */
135 1901756 : if (cv_sleep_target != cv)
136 : {
137 704894 : ConditionVariablePrepareToSleep(cv);
138 704894 : return false;
139 : }
140 :
141 : /*
142 : * Record the current time so that we can calculate the remaining timeout
143 : * if we are woken up spuriously.
144 : */
145 1196862 : if (timeout >= 0)
146 : {
147 398 : INSTR_TIME_SET_CURRENT(start_time);
148 : Assert(timeout >= 0 && timeout <= INT_MAX);
149 398 : cur_timeout = timeout;
150 398 : wait_events = WL_LATCH_SET | WL_TIMEOUT | WL_EXIT_ON_PM_DEATH;
151 : }
152 : else
153 1196464 : wait_events = WL_LATCH_SET | WL_EXIT_ON_PM_DEATH;
154 :
155 : while (true)
156 25154 : {
157 1222016 : bool done = false;
158 :
159 : /*
160 : * Wait for latch to be set. (If we're awakened for some other
161 : * reason, the code below will cope anyway.)
162 : */
163 1222016 : (void) WaitLatch(MyLatch, wait_events, cur_timeout, wait_event_info);
164 :
165 : /* Reset latch before examining the state of the wait list. */
166 1222016 : ResetLatch(MyLatch);
167 :
168 : /*
169 : * If this process has been taken out of the wait list, then we know
170 : * that it has been signaled by ConditionVariableSignal (or
171 : * ConditionVariableBroadcast), so we should return to the caller. But
172 : * that doesn't guarantee that the exit condition is met, only that we
173 : * ought to check it. So we must put the process back into the wait
174 : * list, to ensure we don't miss any additional wakeup occurring while
175 : * the caller checks its exit condition. We can take ourselves out of
176 : * the wait list only when the caller calls
177 : * ConditionVariableCancelSleep.
178 : *
179 : * If we're still in the wait list, then the latch must have been set
180 : * by something other than ConditionVariableSignal; though we don't
181 : * guarantee not to return spuriously, we'll avoid this obvious case.
182 : */
183 1222016 : SpinLockAcquire(&cv->mutex);
184 1222016 : if (!proclist_contains(&cv->wakeup, MyProcNumber, cvWaitLink))
185 : {
186 1196700 : done = true;
187 1196700 : proclist_push_tail(&cv->wakeup, MyProcNumber, cvWaitLink);
188 : }
189 1222016 : SpinLockRelease(&cv->mutex);
190 :
191 : /*
192 : * Check for interrupts, and return spuriously if that caused the
193 : * current sleep target to change (meaning that interrupt handler code
194 : * waited for a different condition variable).
195 : */
196 1222016 : CHECK_FOR_INTERRUPTS();
197 1222014 : if (cv != cv_sleep_target)
198 156 : done = true;
199 :
200 : /* We were signaled, so return */
201 1222014 : if (done)
202 1196854 : return false;
203 :
204 : /* If we're not done, update cur_timeout for next iteration */
205 25160 : if (timeout >= 0)
206 : {
207 88 : INSTR_TIME_SET_CURRENT(cur_time);
208 88 : INSTR_TIME_SUBTRACT(cur_time, start_time);
209 88 : cur_timeout = timeout - (long) INSTR_TIME_GET_MILLISEC(cur_time);
210 :
211 : /* Have we crossed the timeout threshold? */
212 88 : if (cur_timeout <= 0)
213 6 : return true;
214 : }
215 : }
216 : }
217 :
218 : /*
219 : * Cancel any pending sleep operation.
220 : *
221 : * We just need to remove ourselves from the wait queue of any condition
222 : * variable for which we have previously prepared a sleep.
223 : *
224 : * Do nothing if nothing is pending; this allows this function to be called
225 : * during transaction abort to clean up any unfinished CV sleep.
226 : *
227 : * Return true if we've been signaled.
228 : */
229 : bool
230 10445818 : ConditionVariableCancelSleep(void)
231 : {
232 10445818 : ConditionVariable *cv = cv_sleep_target;
233 10445818 : bool signaled = false;
234 :
235 10445818 : if (cv == NULL)
236 9117478 : return false;
237 :
238 1328340 : SpinLockAcquire(&cv->mutex);
239 1328340 : if (proclist_contains(&cv->wakeup, MyProcNumber, cvWaitLink))
240 1280682 : proclist_delete(&cv->wakeup, MyProcNumber, cvWaitLink);
241 : else
242 47658 : signaled = true;
243 1328340 : SpinLockRelease(&cv->mutex);
244 :
245 1328340 : cv_sleep_target = NULL;
246 :
247 1328340 : return signaled;
248 : }
249 :
250 : /*
251 : * Wake up the oldest process sleeping on the CV, if there is any.
252 : *
253 : * Note: it's difficult to tell whether this has any real effect: we know
254 : * whether we took an entry off the list, but the entry might only be a
255 : * sentinel. Hence, think twice before proposing that this should return
256 : * a flag telling whether it woke somebody.
257 : */
258 : void
259 1652 : ConditionVariableSignal(ConditionVariable *cv)
260 : {
261 1652 : PGPROC *proc = NULL;
262 :
263 : /* Remove the first process from the wakeup queue (if any). */
264 1652 : SpinLockAcquire(&cv->mutex);
265 1652 : if (!proclist_is_empty(&cv->wakeup))
266 112 : proc = proclist_pop_head_node(&cv->wakeup, cvWaitLink);
267 1652 : SpinLockRelease(&cv->mutex);
268 :
269 : /* If we found someone sleeping, set their latch to wake them up. */
270 1652 : if (proc != NULL)
271 112 : SetLatch(&proc->procLatch);
272 1652 : }
273 :
274 : /*
275 : * Wake up all processes sleeping on the given CV.
276 : *
277 : * This guarantees to wake all processes that were sleeping on the CV
278 : * at time of call, but processes that add themselves to the list mid-call
279 : * will typically not get awakened.
280 : */
281 : void
282 17083496 : ConditionVariableBroadcast(ConditionVariable *cv)
283 : {
284 17083496 : int pgprocno = MyProcNumber;
285 17083496 : PGPROC *proc = NULL;
286 17083496 : bool have_sentinel = false;
287 :
288 : /*
289 : * In some use-cases, it is common for awakened processes to immediately
290 : * re-queue themselves. If we just naively try to reduce the wakeup list
291 : * to empty, we'll get into a potentially-indefinite loop against such a
292 : * process. The semantics we really want are just to be sure that we have
293 : * wakened all processes that were in the list at entry. We can use our
294 : * own cvWaitLink as a sentinel to detect when we've finished.
295 : *
296 : * A seeming flaw in this approach is that someone else might signal the
297 : * CV and in doing so remove our sentinel entry. But that's fine: since
298 : * CV waiters are always added and removed in order, that must mean that
299 : * every previous waiter has been wakened, so we're done. We'll get an
300 : * extra "set" on our latch from the someone else's signal, which is
301 : * slightly inefficient but harmless.
302 : *
303 : * We can't insert our cvWaitLink as a sentinel if it's already in use in
304 : * some other proclist. While that's not expected to be true for typical
305 : * uses of this function, we can deal with it by simply canceling any
306 : * prepared CV sleep. The next call to ConditionVariableSleep will take
307 : * care of re-establishing the lost state.
308 : */
309 17083496 : if (cv_sleep_target != NULL)
310 176 : ConditionVariableCancelSleep();
311 :
312 : /*
313 : * Inspect the state of the queue. If it's empty, we have nothing to do.
314 : * If there's exactly one entry, we need only remove and signal that
315 : * entry. Otherwise, remove the first entry and insert our sentinel.
316 : */
317 17083496 : SpinLockAcquire(&cv->mutex);
318 : /* While we're here, let's assert we're not in the list. */
319 : Assert(!proclist_contains(&cv->wakeup, pgprocno, cvWaitLink));
320 :
321 17083496 : if (!proclist_is_empty(&cv->wakeup))
322 : {
323 1095300 : proc = proclist_pop_head_node(&cv->wakeup, cvWaitLink);
324 1095300 : if (!proclist_is_empty(&cv->wakeup))
325 : {
326 98764 : proclist_push_tail(&cv->wakeup, pgprocno, cvWaitLink);
327 98764 : have_sentinel = true;
328 : }
329 : }
330 17083496 : SpinLockRelease(&cv->mutex);
331 :
332 : /* Awaken first waiter, if there was one. */
333 17083496 : if (proc != NULL)
334 1095300 : SetLatch(&proc->procLatch);
335 :
336 17296140 : while (have_sentinel)
337 : {
338 : /*
339 : * Each time through the loop, remove the first wakeup list entry, and
340 : * signal it unless it's our sentinel. Repeat as long as the sentinel
341 : * remains in the list.
342 : *
343 : * Notice that if someone else removes our sentinel, we will waken one
344 : * additional process before exiting. That's intentional, because if
345 : * someone else signals the CV, they may be intending to waken some
346 : * third process that added itself to the list after we added the
347 : * sentinel. Better to give a spurious wakeup (which should be
348 : * harmless beyond wasting some cycles) than to lose a wakeup.
349 : */
350 212644 : proc = NULL;
351 212644 : SpinLockAcquire(&cv->mutex);
352 212644 : if (!proclist_is_empty(&cv->wakeup))
353 197490 : proc = proclist_pop_head_node(&cv->wakeup, cvWaitLink);
354 212644 : have_sentinel = proclist_contains(&cv->wakeup, pgprocno, cvWaitLink);
355 212644 : SpinLockRelease(&cv->mutex);
356 :
357 212644 : if (proc != NULL && proc != MyProc)
358 117826 : SetLatch(&proc->procLatch);
359 : }
360 17083496 : }
|
