Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * s_lock.c
4 : * Hardware-dependent implementation of spinlocks.
5 : *
6 : * When waiting for a contended spinlock we loop tightly for awhile, then
7 : * delay using pg_usleep() and try again. Preferably, "awhile" should be a
8 : * small multiple of the maximum time we expect a spinlock to be held. 100
9 : * iterations seems about right as an initial guess. However, on a
10 : * uniprocessor the loop is a waste of cycles, while in a multi-CPU scenario
11 : * it's usually better to spin a bit longer than to call the kernel, so we try
12 : * to adapt the spin loop count depending on whether we seem to be in a
13 : * uniprocessor or multiprocessor.
14 : *
15 : * Note: you might think MIN_SPINS_PER_DELAY should be just 1, but you'd
16 : * be wrong; there are platforms where that can result in a "stuck
17 : * spinlock" failure. This has been seen particularly on Alphas; it seems
18 : * that the first TAS after returning from kernel space will always fail
19 : * on that hardware.
20 : *
21 : * Once we do decide to block, we use randomly increasing pg_usleep()
22 : * delays. The first delay is 1 msec, then the delay randomly increases to
23 : * about one second, after which we reset to 1 msec and start again. The
24 : * idea here is that in the presence of heavy contention we need to
25 : * increase the delay, else the spinlock holder may never get to run and
26 : * release the lock. (Consider situation where spinlock holder has been
27 : * nice'd down in priority by the scheduler --- it will not get scheduled
28 : * until all would-be acquirers are sleeping, so if we always use a 1-msec
29 : * sleep, there is a real possibility of starvation.) But we can't just
30 : * clamp the delay to an upper bound, else it would take a long time to
31 : * make a reasonable number of tries.
32 : *
33 : * We time out and declare error after NUM_DELAYS delays (thus, exactly
34 : * that many tries). With the given settings, this will usually take 2 or
35 : * so minutes. It seems better to fix the total number of tries (and thus
36 : * the probability of unintended failure) than to fix the total time
37 : * spent.
38 : *
39 : * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
40 : * Portions Copyright (c) 1994, Regents of the University of California
41 : *
42 : *
43 : * IDENTIFICATION
44 : * src/backend/storage/lmgr/s_lock.c
45 : *
46 : *-------------------------------------------------------------------------
47 : */
48 : #include "postgres.h"
49 :
50 : #include <time.h>
51 : #include <unistd.h>
52 :
53 : #include "common/pg_prng.h"
54 : #include "port/atomics.h"
55 : #include "storage/s_lock.h"
56 : #include "utils/wait_event.h"
57 :
58 : #define MIN_SPINS_PER_DELAY 10
59 : #define MAX_SPINS_PER_DELAY 1000
60 : #define NUM_DELAYS 1000
61 : #define MIN_DELAY_USEC 1000L
62 : #define MAX_DELAY_USEC 1000000L
63 :
64 : #ifdef S_LOCK_TEST
65 : /*
66 : * These are needed by pgstat_report_wait_start in the standalone compile of
67 : * s_lock_test.
68 : */
69 : static uint32 local_my_wait_event_info;
70 : uint32 *my_wait_event_info = &local_my_wait_event_info;
71 : #endif
72 :
73 : static int spins_per_delay = DEFAULT_SPINS_PER_DELAY;
74 :
75 :
76 : /*
77 : * s_lock_stuck() - complain about a stuck spinlock
78 : */
79 : static void
80 0 : s_lock_stuck(const char *file, int line, const char *func)
81 : {
82 0 : if (!func)
83 0 : func = "(unknown)";
84 : #if defined(S_LOCK_TEST)
85 : fprintf(stderr,
86 : "\nStuck spinlock detected at %s, %s:%d.\n",
87 : func, file, line);
88 : exit(1);
89 : #else
90 0 : elog(PANIC, "stuck spinlock detected at %s, %s:%d",
91 : func, file, line);
92 : #endif
93 : }
94 :
95 : /*
96 : * s_lock(lock) - platform-independent portion of waiting for a spinlock.
97 : */
98 : int
99 12604 : s_lock(volatile slock_t *lock, const char *file, int line, const char *func)
100 : {
101 : SpinDelayStatus delayStatus;
102 :
103 12604 : init_spin_delay(&delayStatus, file, line, func);
104 :
105 199504 : while (TAS_SPIN(lock))
106 : {
107 186900 : perform_spin_delay(&delayStatus);
108 : }
109 :
110 12604 : finish_spin_delay(&delayStatus);
111 :
112 12604 : return delayStatus.delays;
113 : }
114 :
115 : #ifdef USE_DEFAULT_S_UNLOCK
116 : void
117 : s_unlock(volatile slock_t *lock)
118 : {
119 : #ifdef TAS_ACTIVE_WORD
120 : /* HP's PA-RISC */
121 : *TAS_ACTIVE_WORD(lock) = -1;
122 : #else
123 : *lock = 0;
124 : #endif
125 : }
126 : #endif
127 :
128 : /*
129 : * Wait while spinning on a contended spinlock.
130 : */
131 : void
132 204286 : perform_spin_delay(SpinDelayStatus *status)
133 : {
134 : /* CPU-specific delay each time through the loop */
135 204286 : SPIN_DELAY();
136 :
137 : /* Block the process every spins_per_delay tries */
138 204286 : if (++(status->spins) >= spins_per_delay)
139 : {
140 198 : if (++(status->delays) > NUM_DELAYS)
141 0 : s_lock_stuck(status->file, status->line, status->func);
142 :
143 198 : if (status->cur_delay == 0) /* first time to delay? */
144 90 : status->cur_delay = MIN_DELAY_USEC;
145 :
146 : /*
147 : * Once we start sleeping, the overhead of reporting a wait event is
148 : * justified. Actively spinning easily stands out in profilers, but
149 : * sleeping with an exponential backoff is harder to spot...
150 : *
151 : * We might want to report something more granular at some point, but
152 : * this is better than nothing.
153 : */
154 198 : pgstat_report_wait_start(WAIT_EVENT_SPIN_DELAY);
155 198 : pg_usleep(status->cur_delay);
156 198 : pgstat_report_wait_end();
157 :
158 : #if defined(S_LOCK_TEST)
159 : fprintf(stdout, "*");
160 : fflush(stdout);
161 : #endif
162 :
163 : /* increase delay by a random fraction between 1X and 2X */
164 396 : status->cur_delay += (int) (status->cur_delay *
165 198 : pg_prng_double(&pg_global_prng_state) + 0.5);
166 : /* wrap back to minimum delay when max is exceeded */
167 198 : if (status->cur_delay > MAX_DELAY_USEC)
168 0 : status->cur_delay = MIN_DELAY_USEC;
169 :
170 198 : status->spins = 0;
171 : }
172 204286 : }
173 :
174 : /*
175 : * After acquiring a spinlock, update estimates about how long to loop.
176 : *
177 : * If we were able to acquire the lock without delaying, it's a good
178 : * indication we are in a multiprocessor. If we had to delay, it's a sign
179 : * (but not a sure thing) that we are in a uniprocessor. Hence, we
180 : * decrement spins_per_delay slowly when we had to delay, and increase it
181 : * rapidly when we didn't. It's expected that spins_per_delay will
182 : * converge to the minimum value on a uniprocessor and to the maximum
183 : * value on a multiprocessor.
184 : *
185 : * Note: spins_per_delay is local within our current process. We want to
186 : * average these observations across multiple backends, since it's
187 : * relatively rare for this function to even get entered, and so a single
188 : * backend might not live long enough to converge on a good value. That
189 : * is handled by the two routines below.
190 : */
191 : void
192 47369574 : finish_spin_delay(SpinDelayStatus *status)
193 : {
194 47369574 : if (status->cur_delay == 0)
195 : {
196 : /* we never had to delay */
197 47369484 : if (spins_per_delay < MAX_SPINS_PER_DELAY)
198 91994 : spins_per_delay = Min(spins_per_delay + 100, MAX_SPINS_PER_DELAY);
199 : }
200 : else
201 : {
202 90 : if (spins_per_delay > MIN_SPINS_PER_DELAY)
203 90 : spins_per_delay = Max(spins_per_delay - 1, MIN_SPINS_PER_DELAY);
204 : }
205 47369574 : }
206 :
207 : /*
208 : * Set local copy of spins_per_delay during backend startup.
209 : *
210 : * NB: this has to be pretty fast as it is called while holding a spinlock
211 : */
212 : void
213 29864 : set_spins_per_delay(int shared_spins_per_delay)
214 : {
215 29864 : spins_per_delay = shared_spins_per_delay;
216 29864 : }
217 :
218 : /*
219 : * Update shared estimate of spins_per_delay during backend exit.
220 : *
221 : * NB: this has to be pretty fast as it is called while holding a spinlock
222 : */
223 : int
224 29860 : update_spins_per_delay(int shared_spins_per_delay)
225 : {
226 : /*
227 : * We use an exponential moving average with a relatively slow adaption
228 : * rate, so that noise in any one backend's result won't affect the shared
229 : * value too much. As long as both inputs are within the allowed range,
230 : * the result must be too, so we need not worry about clamping the result.
231 : *
232 : * We deliberately truncate rather than rounding; this is so that single
233 : * adjustments inside a backend can affect the shared estimate (see the
234 : * asymmetric adjustment rules above).
235 : */
236 29860 : return (shared_spins_per_delay * 15 + spins_per_delay) / 16;
237 : }
238 :
239 :
240 : /*****************************************************************************/
241 : #if defined(S_LOCK_TEST)
242 :
243 : /*
244 : * test program for verifying a port's spinlock support.
245 : */
246 :
247 : struct test_lock_struct
248 : {
249 : char pad1;
250 : slock_t lock;
251 : char pad2;
252 : };
253 :
254 : volatile struct test_lock_struct test_lock;
255 :
256 : int
257 : main()
258 : {
259 : pg_prng_seed(&pg_global_prng_state, (uint64) time(NULL));
260 :
261 : test_lock.pad1 = test_lock.pad2 = 0x44;
262 :
263 : S_INIT_LOCK(&test_lock.lock);
264 :
265 : if (test_lock.pad1 != 0x44 || test_lock.pad2 != 0x44)
266 : {
267 : printf("S_LOCK_TEST: failed, declared datatype is wrong size\n");
268 : return 1;
269 : }
270 :
271 : if (!S_LOCK_FREE(&test_lock.lock))
272 : {
273 : printf("S_LOCK_TEST: failed, lock not initialized\n");
274 : return 1;
275 : }
276 :
277 : S_LOCK(&test_lock.lock);
278 :
279 : if (test_lock.pad1 != 0x44 || test_lock.pad2 != 0x44)
280 : {
281 : printf("S_LOCK_TEST: failed, declared datatype is wrong size\n");
282 : return 1;
283 : }
284 :
285 : if (S_LOCK_FREE(&test_lock.lock))
286 : {
287 : printf("S_LOCK_TEST: failed, lock not locked\n");
288 : return 1;
289 : }
290 :
291 : S_UNLOCK(&test_lock.lock);
292 :
293 : if (test_lock.pad1 != 0x44 || test_lock.pad2 != 0x44)
294 : {
295 : printf("S_LOCK_TEST: failed, declared datatype is wrong size\n");
296 : return 1;
297 : }
298 :
299 : if (!S_LOCK_FREE(&test_lock.lock))
300 : {
301 : printf("S_LOCK_TEST: failed, lock not unlocked\n");
302 : return 1;
303 : }
304 :
305 : S_LOCK(&test_lock.lock);
306 :
307 : if (test_lock.pad1 != 0x44 || test_lock.pad2 != 0x44)
308 : {
309 : printf("S_LOCK_TEST: failed, declared datatype is wrong size\n");
310 : return 1;
311 : }
312 :
313 : if (S_LOCK_FREE(&test_lock.lock))
314 : {
315 : printf("S_LOCK_TEST: failed, lock not re-locked\n");
316 : return 1;
317 : }
318 :
319 : printf("S_LOCK_TEST: this will print %d stars and then\n", NUM_DELAYS);
320 : printf(" exit with a 'stuck spinlock' message\n");
321 : printf(" if S_LOCK() and TAS() are working.\n");
322 : fflush(stdout);
323 :
324 : s_lock(&test_lock.lock, __FILE__, __LINE__, __func__);
325 :
326 : printf("S_LOCK_TEST: failed, lock not locked\n");
327 : return 1;
328 : }
329 :
330 : #endif /* S_LOCK_TEST */
|