| 1 |
/*------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* s_lock.c |
| 4 |
* 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) 1996-2025, 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 "storage/s_lock.h" |
| 55 |
#include "utils/wait_event.h" |
| 56 |
|
| 57 |
#define MIN_SPINS_PER_DELAY 10 |
| 58 |
#define MAX_SPINS_PER_DELAY 1000 |
| 59 |
#define NUM_DELAYS 1000 |
| 60 |
#define MIN_DELAY_USEC 1000L |
| 61 |
#define MAX_DELAY_USEC 1000000L |
| 62 |
|
| 63 |
#ifdef S_LOCK_TEST |
| 64 |
/* |
| 65 |
* These are needed by pgstat_report_wait_start in the standalone compile of |
| 66 |
* s_lock_test. |
| 67 |
*/ |
| 68 |
static uint32 local_my_wait_event_info; |
| 69 |
uint32 *my_wait_event_info = &local_my_wait_event_info; |
| 70 |
#endif |
| 71 |
|
| 72 |
static int spins_per_delay = DEFAULT_SPINS_PER_DELAY; |
| 73 |
|
| 74 |
|
| 75 |
/* |
| 76 |
* s_lock_stuck() - complain about a stuck spinlock |
| 77 |
*/ |
| 78 |
static void |
| 79 |
s_lock_stuck(const char *file, int line, const char *func) |
| 80 |
{ |
| 81 |
if (!func) |
| 82 |
func = "(unknown)"; |
| 83 |
#if defined(S_LOCK_TEST) |
| 84 |
fprintf(stderr, |
| 85 |
"\nStuck spinlock detected at %s, %s:%d.\n", |
| 86 |
func, file, line); |
| 87 |
exit(1); |
| 88 |
#else |
| 89 |
elog(PANIC, "stuck spinlock detected at %s, %s:%d", |
| 90 |
func, file, line); |
| 91 |
#endif |
| 92 |
} |
| 93 |
|
| 94 |
/* |
| 95 |
* s_lock(lock) - platform-independent portion of waiting for a spinlock. |
| 96 |
*/ |
| 97 |
int |
| 98 |
s_lock(volatile slock_t *lock, const char *file, int line, const char *func) |
| 99 |
{ |
| 100 |
SpinDelayStatus delayStatus; |
| 101 |
|
| 102 |
init_spin_delay(&delayStatus, file, line, func); |
| 103 |
|
| 104 |
while (TAS_SPIN(lock)) |
| 105 |
{ |
| 106 |
perform_spin_delay(&delayStatus); |
| 107 |
} |
| 108 |
|
| 109 |
finish_spin_delay(&delayStatus); |
| 110 |
|
| 111 |
return delayStatus.delays; |
| 112 |
} |
| 113 |
|
| 114 |
#ifdef USE_DEFAULT_S_UNLOCK |
| 115 |
void |
| 116 |
s_unlock(volatile slock_t *lock) |
| 117 |
{ |
| 118 |
*lock = 0; |
| 119 |
} |
| 120 |
#endif |
| 121 |
|
| 122 |
/* |
| 123 |
* Wait while spinning on a contended spinlock. |
| 124 |
*/ |
| 125 |
void |
| 126 |
perform_spin_delay(SpinDelayStatus *status) |
| 127 |
{ |
| 128 |
/* CPU-specific delay each time through the loop */ |
| 129 |
SPIN_DELAY(); |
| 130 |
|
| 131 |
/* Block the process every spins_per_delay tries */ |
| 132 |
if (++(status->spins) >= spins_per_delay) |
| 133 |
{ |
| 134 |
if (++(status->delays) > NUM_DELAYS) |
| 135 |
s_lock_stuck(status->file, status->line, status->func); |
| 136 |
|
| 137 |
if (status->cur_delay == 0) /* first time to delay? */ |
| 138 |
status->cur_delay = MIN_DELAY_USEC; |
| 139 |
|
| 140 |
/* |
| 141 |
* Once we start sleeping, the overhead of reporting a wait event is |
| 142 |
* justified. Actively spinning easily stands out in profilers, but |
| 143 |
* sleeping with an exponential backoff is harder to spot... |
| 144 |
* |
| 145 |
* We might want to report something more granular at some point, but |
| 146 |
* this is better than nothing. |
| 147 |
*/ |
| 148 |
pgstat_report_wait_start(WAIT_EVENT_SPIN_DELAY); |
| 149 |
pg_usleep(status->cur_delay); |
| 150 |
pgstat_report_wait_end(); |
| 151 |
|
| 152 |
#if defined(S_LOCK_TEST) |
| 153 |
fprintf(stdout, "*"); |
| 154 |
fflush(stdout); |
| 155 |
#endif |
| 156 |
|
| 157 |
/* increase delay by a random fraction between 1X and 2X */ |
| 158 |
status->cur_delay += (int) (status->cur_delay * |
| 159 |
pg_prng_double(&pg_global_prng_state) + 0.5); |
| 160 |
/* wrap back to minimum delay when max is exceeded */ |
| 161 |
if (status->cur_delay > MAX_DELAY_USEC) |
| 162 |
status->cur_delay = MIN_DELAY_USEC; |
| 163 |
|
| 164 |
status->spins = 0; |
| 165 |
} |
| 166 |
} |
| 167 |
|
| 168 |
/* |
| 169 |
* After acquiring a spinlock, update estimates about how long to loop. |
| 170 |
* |
| 171 |
* If we were able to acquire the lock without delaying, it's a good |
| 172 |
* indication we are in a multiprocessor. If we had to delay, it's a sign |
| 173 |
* (but not a sure thing) that we are in a uniprocessor. Hence, we |
| 174 |
* decrement spins_per_delay slowly when we had to delay, and increase it |
| 175 |
* rapidly when we didn't. It's expected that spins_per_delay will |
| 176 |
* converge to the minimum value on a uniprocessor and to the maximum |
| 177 |
* value on a multiprocessor. |
| 178 |
* |
| 179 |
* Note: spins_per_delay is local within our current process. We want to |
| 180 |
* average these observations across multiple backends, since it's |
| 181 |
* relatively rare for this function to even get entered, and so a single |
| 182 |
* backend might not live long enough to converge on a good value. That |
| 183 |
* is handled by the two routines below. |
| 184 |
*/ |
| 185 |
void |
| 186 |
finish_spin_delay(SpinDelayStatus *status) |
| 187 |
{ |
| 188 |
if (status->cur_delay == 0) |
| 189 |
{ |
| 190 |
/* we never had to delay */ |
| 191 |
if (spins_per_delay < MAX_SPINS_PER_DELAY) |
| 192 |
spins_per_delay = Min(spins_per_delay + 100, MAX_SPINS_PER_DELAY); |
| 193 |
} |
| 194 |
else |
| 195 |
{ |
| 196 |
if (spins_per_delay > MIN_SPINS_PER_DELAY) |
| 197 |
spins_per_delay = Max(spins_per_delay - 1, MIN_SPINS_PER_DELAY); |
| 198 |
} |
| 199 |
} |
| 200 |
|
| 201 |
/* |
| 202 |
* Set local copy of spins_per_delay during backend startup. |
| 203 |
* |
| 204 |
* NB: this has to be pretty fast as it is called while holding a spinlock |
| 205 |
*/ |
| 206 |
void |
| 207 |
set_spins_per_delay(int shared_spins_per_delay) |
| 208 |
{ |
| 209 |
spins_per_delay = shared_spins_per_delay; |
| 210 |
} |
| 211 |
|
| 212 |
/* |
| 213 |
* Update shared estimate of spins_per_delay during backend exit. |
| 214 |
* |
| 215 |
* NB: this has to be pretty fast as it is called while holding a spinlock |
| 216 |
*/ |
| 217 |
int |
| 218 |
update_spins_per_delay(int shared_spins_per_delay) |
| 219 |
{ |
| 220 |
/* |
| 221 |
* We use an exponential moving average with a relatively slow adaption |
| 222 |
* rate, so that noise in any one backend's result won't affect the shared |
| 223 |
* value too much. As long as both inputs are within the allowed range, |
| 224 |
* the result must be too, so we need not worry about clamping the result. |
| 225 |
* |
| 226 |
* We deliberately truncate rather than rounding; this is so that single |
| 227 |
* adjustments inside a backend can affect the shared estimate (see the |
| 228 |
* asymmetric adjustment rules above). |
| 229 |
*/ |
| 230 |
return (shared_spins_per_delay * 15 + spins_per_delay) / 16; |
| 231 |
} |
| 232 |
|
| 233 |
|
| 234 |
/*****************************************************************************/ |
| 235 |
#if defined(S_LOCK_TEST) |
| 236 |
|
| 237 |
/* |
| 238 |
* test program for verifying a port's spinlock support. |
| 239 |
*/ |
| 240 |
|
| 241 |
struct test_lock_struct |
| 242 |
{ |
| 243 |
char pad1; |
| 244 |
slock_t lock; |
| 245 |
char pad2; |
| 246 |
}; |
| 247 |
|
| 248 |
volatile struct test_lock_struct test_lock; |
| 249 |
|
| 250 |
int |
| 251 |
main() |
| 252 |
{ |
| 253 |
pg_prng_seed(&pg_global_prng_state, (uint64) time(NULL)); |
| 254 |
|
| 255 |
test_lock.pad1 = test_lock.pad2 = 0x44; |
| 256 |
|
| 257 |
S_INIT_LOCK(&test_lock.lock); |
| 258 |
|
| 259 |
if (test_lock.pad1 != 0x44 || test_lock.pad2 != 0x44) |
| 260 |
{ |
| 261 |
printf("S_LOCK_TEST: failed, declared datatype is wrong size\n"); |
| 262 |
return 1; |
| 263 |
} |
| 264 |
|
| 265 |
if (!S_LOCK_FREE(&test_lock.lock)) |
| 266 |
{ |
| 267 |
printf("S_LOCK_TEST: failed, lock not initialized\n"); |
| 268 |
return 1; |
| 269 |
} |
| 270 |
|
| 271 |
S_LOCK(&test_lock.lock); |
| 272 |
|
| 273 |
if (test_lock.pad1 != 0x44 || test_lock.pad2 != 0x44) |
| 274 |
{ |
| 275 |
printf("S_LOCK_TEST: failed, declared datatype is wrong size\n"); |
| 276 |
return 1; |
| 277 |
} |
| 278 |
|
| 279 |
if (S_LOCK_FREE(&test_lock.lock)) |
| 280 |
{ |
| 281 |
printf("S_LOCK_TEST: failed, lock not locked\n"); |
| 282 |
return 1; |
| 283 |
} |
| 284 |
|
| 285 |
S_UNLOCK(&test_lock.lock); |
| 286 |
|
| 287 |
if (test_lock.pad1 != 0x44 || test_lock.pad2 != 0x44) |
| 288 |
{ |
| 289 |
printf("S_LOCK_TEST: failed, declared datatype is wrong size\n"); |
| 290 |
return 1; |
| 291 |
} |
| 292 |
|
| 293 |
if (!S_LOCK_FREE(&test_lock.lock)) |
| 294 |
{ |
| 295 |
printf("S_LOCK_TEST: failed, lock not unlocked\n"); |
| 296 |
return 1; |
| 297 |
} |
| 298 |
|
| 299 |
S_LOCK(&test_lock.lock); |
| 300 |
|
| 301 |
if (test_lock.pad1 != 0x44 || test_lock.pad2 != 0x44) |
| 302 |
{ |
| 303 |
printf("S_LOCK_TEST: failed, declared datatype is wrong size\n"); |
| 304 |
return 1; |
| 305 |
} |
| 306 |
|
| 307 |
if (S_LOCK_FREE(&test_lock.lock)) |
| 308 |
{ |
| 309 |
printf("S_LOCK_TEST: failed, lock not re-locked\n"); |
| 310 |
return 1; |
| 311 |
} |
| 312 |
|
| 313 |
printf("S_LOCK_TEST: this will print %d stars and then\n", NUM_DELAYS); |
| 314 |
printf(" exit with a 'stuck spinlock' message\n"); |
| 315 |
printf(" if S_LOCK() and TAS() are working.\n"); |
| 316 |
fflush(stdout); |
| 317 |
|
| 318 |
s_lock(&test_lock.lock, __FILE__, __LINE__, __func__); |
| 319 |
|
| 320 |
printf("S_LOCK_TEST: failed, lock not locked\n"); |
| 321 |
return 1; |
| 322 |
} |
| 323 |
|
| 324 |
#endif /* S_LOCK_TEST */ |