|
| 1 | +#include "postgres.h" |
| 2 | + |
| 3 | +#include "access/htup_details.h" |
| 4 | +#include "catalog/pg_type.h" |
| 5 | +#include "funcapi.h" |
| 6 | +#include "miscadmin.h" |
| 7 | +#include "postmaster/bgworker.h" |
| 8 | +#include "storage/ipc.h" |
| 9 | +#include "storage/procarray.h" |
| 10 | +#include "storage/procsignal.h" |
| 11 | +#include "storage/s_lock.h" |
| 12 | +#include "storage/shm_mq.h" |
| 13 | +#include "storage/shm_toc.h" |
| 14 | +#include "storage/spin.h" |
| 15 | +#include "storage/wait.h" |
| 16 | +#include "utils/guc.h" |
| 17 | +#include "utils/memutils.h" |
| 18 | +#include "utils/resowner.h" |
| 19 | +#include "portability/instr_time.h" |
| 20 | + |
| 21 | +#include "pg_stat_wait.h" |
| 22 | + |
| 23 | +CollectorShmqHeader *hdr; |
| 24 | + |
| 25 | +static void *pgsw; |
| 26 | +shm_toc *toc; |
| 27 | +shm_mq *mq; |
| 28 | +static volatile sig_atomic_t shutdown_requested = false; |
| 29 | + |
| 30 | +static void handle_sigterm(SIGNAL_ARGS); |
| 31 | +static void collector_main(Datum main_arg); |
| 32 | + |
| 33 | +/* |
| 34 | + * Estimate shared memory space needed. |
| 35 | + */ |
| 36 | +Size |
| 37 | +CollectorShmemSize(void) |
| 38 | +{ |
| 39 | + shm_toc_estimator e; |
| 40 | + Size size; |
| 41 | + |
| 42 | + shm_toc_initialize_estimator(&e); |
| 43 | + shm_toc_estimate_chunk(&e, sizeof(CollectorShmqHeader)); |
| 44 | + shm_toc_estimate_chunk(&e, (Size) COLLECTOR_QUEUE_SIZE); |
| 45 | + shm_toc_estimate_keys(&e, 2); |
| 46 | + size = shm_toc_estimate(&e); |
| 47 | + |
| 48 | + return size; |
| 49 | +} |
| 50 | + |
| 51 | +void |
| 52 | +AllocateCollectorMem(void) |
| 53 | +{ |
| 54 | + bool found; |
| 55 | + Size segsize= CollectorShmemSize(); |
| 56 | + |
| 57 | + pgsw = ShmemInitStruct("pg_stat_wait", segsize, &found); |
| 58 | + |
| 59 | + if (!found) |
| 60 | + { |
| 61 | + void *mq_mem; |
| 62 | + |
| 63 | + toc = shm_toc_create(PG_STAT_WAIT_MAGIC, pgsw, segsize); |
| 64 | + hdr = shm_toc_allocate(toc, sizeof(CollectorShmqHeader)); |
| 65 | + shm_toc_insert(toc, 0, hdr); |
| 66 | + |
| 67 | + mq_mem = shm_toc_allocate(toc, COLLECTOR_QUEUE_SIZE); |
| 68 | + shm_toc_insert(toc, 1, mq_mem); |
| 69 | + |
| 70 | + DefineCustomIntVariable("pg_stat_wait.history_size", |
| 71 | + "Sets size of waits history.", NULL, |
| 72 | + &hdr->historySize, 5000, 100, INT_MAX, |
| 73 | + PGC_SUSET, 0, NULL, NULL, NULL); |
| 74 | + |
| 75 | + DefineCustomIntVariable("pg_stat_wait.history_period", |
| 76 | + "Sets period of waits history sampling.", NULL, |
| 77 | + &hdr->historyPeriod, 10, 1, INT_MAX, |
| 78 | + PGC_SUSET, 0, NULL, NULL, NULL); |
| 79 | + |
| 80 | + DefineCustomBoolVariable("pg_stat_wait.history_skip_latch", |
| 81 | + "Skip latch events in waits history", NULL, |
| 82 | + &hdr->historySkipLatch, false, PGC_SUSET, 0, NULL, NULL, NULL); |
| 83 | + } |
| 84 | + else |
| 85 | + { |
| 86 | + toc = shm_toc_attach(PG_STAT_WAIT_MAGIC, pgsw); |
| 87 | + hdr = shm_toc_lookup(toc, 0); |
| 88 | + } |
| 89 | +} |
| 90 | + |
| 91 | +void |
| 92 | +RegisterWaitsCollector(void) |
| 93 | +{ |
| 94 | + BackgroundWorker worker; |
| 95 | + |
| 96 | + /* set up common data for all our workers */ |
| 97 | + worker.bgw_flags = BGWORKER_SHMEM_ACCESS | |
| 98 | + BGWORKER_BACKEND_DATABASE_CONNECTION; |
| 99 | + worker.bgw_start_time = BgWorkerStart_RecoveryFinished; |
| 100 | + worker.bgw_restart_time = BGW_NEVER_RESTART; |
| 101 | + worker.bgw_main = collector_main; |
| 102 | + worker.bgw_notify_pid = 0; |
| 103 | + snprintf(worker.bgw_name, BGW_MAXLEN, "pg_stat_wait collector"); |
| 104 | + worker.bgw_main_arg = (Datum)0; |
| 105 | + RegisterBackgroundWorker(&worker); |
| 106 | +} |
| 107 | + |
| 108 | +void |
| 109 | +AllocHistory(History *observations, int count) |
| 110 | +{ |
| 111 | + observations->items = (HistoryItem *) palloc0(sizeof(HistoryItem) * count); |
| 112 | + observations->index = 0; |
| 113 | + observations->count = count; |
| 114 | + observations->wraparound = false; |
| 115 | +} |
| 116 | + |
| 117 | +/* Read current wait information from proc, if readCurrent is true, |
| 118 | + * then it reads from currently going wait, and can be inconsistent |
| 119 | + */ |
| 120 | +int |
| 121 | +GetCurrentWaitsState(PGPROC *proc, HistoryItem *item, int idx) |
| 122 | +{ |
| 123 | + instr_time currentTime; |
| 124 | + ProcWait *wait; |
| 125 | + |
| 126 | + if (idx == -1) |
| 127 | + return 0; |
| 128 | + |
| 129 | + INSTR_TIME_SET_CURRENT(currentTime); |
| 130 | + wait = &proc->waits.waitsBuf[idx]; |
| 131 | + item->backendPid = proc->pid; |
| 132 | + item->classId = (int)wait->classId; |
| 133 | + if (item->classId == 0) |
| 134 | + return 0; |
| 135 | + |
| 136 | + item->eventId = (int)wait->eventId; |
| 137 | + |
| 138 | + INSTR_TIME_SUBTRACT(currentTime, wait->startTime); |
| 139 | + item->waitTime = INSTR_TIME_GET_MICROSEC(currentTime); |
| 140 | + memcpy(item->params, wait->params, sizeof(item->params)); |
| 141 | + return 1; |
| 142 | +} |
| 143 | + |
| 144 | +static void |
| 145 | +handle_sigterm(SIGNAL_ARGS) |
| 146 | +{ |
| 147 | + int save_errno = errno; |
| 148 | + shutdown_requested = true; |
| 149 | + if (MyProc) |
| 150 | + SetLatch(&MyProc->procLatch); |
| 151 | + errno = save_errno; |
| 152 | +} |
| 153 | + |
| 154 | +/* Circulation in history */ |
| 155 | +static HistoryItem * |
| 156 | +get_next_observation(History *observations) |
| 157 | +{ |
| 158 | + HistoryItem *result; |
| 159 | + |
| 160 | + result = &observations->items[observations->index]; |
| 161 | + observations->index++; |
| 162 | + if (observations->index >= observations->count) |
| 163 | + { |
| 164 | + observations->index = 0; |
| 165 | + observations->wraparound = true; |
| 166 | + } |
| 167 | + return result; |
| 168 | +} |
| 169 | + |
| 170 | +/* Gets current waits from backends */ |
| 171 | +static void |
| 172 | +write_waits_history(History *observations, TimestampTz current_ts) |
| 173 | +{ |
| 174 | + int i; |
| 175 | + |
| 176 | + LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 177 | + for (i = 0; i < ProcGlobal->allProcCount; ++i) |
| 178 | + { |
| 179 | + HistoryItem item, *observation; |
| 180 | + PGPROC *proc = &ProcGlobal->allProcs[i]; |
| 181 | + int stateOk = GetCurrentWaitsState(proc, &item, proc->waits.readIdx); |
| 182 | + |
| 183 | + /* mark waits as read */ |
| 184 | + proc->waits.readIdx = -1; |
| 185 | + |
| 186 | + if (stateOk) |
| 187 | + { |
| 188 | + if (hdr->historySkipLatch && item.classId == WAIT_LATCH) |
| 189 | + continue; |
| 190 | + |
| 191 | + item.ts = current_ts; |
| 192 | + observation = get_next_observation(observations); |
| 193 | + *observation = item; |
| 194 | + } |
| 195 | + } |
| 196 | + LWLockRelease(ProcArrayLock); |
| 197 | +} |
| 198 | + |
| 199 | +static void |
| 200 | +send_history(History *observations, shm_mq_handle *mqh) |
| 201 | +{ |
| 202 | + int count, i; |
| 203 | + |
| 204 | + if (observations->wraparound) |
| 205 | + count = observations->count; |
| 206 | + else |
| 207 | + count = observations->index; |
| 208 | + |
| 209 | + shm_mq_send(mqh, sizeof(count), &count, false); |
| 210 | + for (i = 0; i < count; i++) |
| 211 | + shm_mq_send(mqh, sizeof(HistoryItem), &observations->items[i], false); |
| 212 | +} |
| 213 | + |
| 214 | +static void |
| 215 | +collector_main(Datum main_arg) |
| 216 | +{ |
| 217 | + shm_mq *mq; |
| 218 | + shm_mq_handle *mqh; |
| 219 | + History observations; |
| 220 | + MemoryContext old_context, collector_context; |
| 221 | + |
| 222 | + /* |
| 223 | + * Establish signal handlers. |
| 224 | + * |
| 225 | + * We want CHECK_FOR_INTERRUPTS() to kill off this worker process just as |
| 226 | + * it would a normal user backend. To make that happen, we establish a |
| 227 | + * signal handler that is a stripped-down version of die(). We don't have |
| 228 | + * any equivalent of the backend's command-read loop, where interrupts can |
| 229 | + * be processed immediately, so make sure ImmediateInterruptOK is turned |
| 230 | + * off. |
| 231 | + */ |
| 232 | + pqsignal(SIGTERM, handle_sigterm); |
| 233 | + BackgroundWorkerUnblockSignals(); |
| 234 | + |
| 235 | + hdr->latch = &MyProc->procLatch; |
| 236 | + CurrentResourceOwner = ResourceOwnerCreate(NULL, "pg_stat_wait collector"); |
| 237 | + collector_context = AllocSetContextCreate(TopMemoryContext, |
| 238 | + "pg_stat_wait context", |
| 239 | + ALLOCSET_DEFAULT_MINSIZE, |
| 240 | + ALLOCSET_DEFAULT_INITSIZE, |
| 241 | + ALLOCSET_DEFAULT_MAXSIZE); |
| 242 | + old_context = MemoryContextSwitchTo(collector_context); |
| 243 | + AllocHistory(&observations, hdr->historySize); |
| 244 | + MemoryContextSwitchTo(old_context); |
| 245 | + |
| 246 | + while (1) |
| 247 | + { |
| 248 | + int rc; |
| 249 | + TimestampTz current_ts; |
| 250 | + |
| 251 | + ResetLatch(&MyProc->procLatch); |
| 252 | + current_ts = GetCurrentTimestamp(); |
| 253 | + write_waits_history(&observations, current_ts); |
| 254 | + |
| 255 | + if (shutdown_requested) |
| 256 | + break; |
| 257 | + |
| 258 | + rc = WaitLatch(&MyProc->procLatch, |
| 259 | + WL_LATCH_SET | WL_TIMEOUT | WL_POSTMASTER_DEATH, |
| 260 | + hdr->historyPeriod); |
| 261 | + |
| 262 | + if (rc & WL_POSTMASTER_DEATH) |
| 263 | + exit(1); |
| 264 | + |
| 265 | + if (hdr->request == HISTORY_REQUEST) |
| 266 | + { |
| 267 | + hdr->request = NO_REQUEST; |
| 268 | + |
| 269 | + mq = (shm_mq *)shm_toc_lookup(toc, 1); |
| 270 | + shm_mq_set_sender(mq, MyProc); |
| 271 | + mqh = shm_mq_attach(mq, NULL, NULL); |
| 272 | + shm_mq_wait_for_attach(mqh); |
| 273 | + |
| 274 | + if (shm_mq_get_receiver(mq) != NULL) |
| 275 | + send_history(&observations, mqh); |
| 276 | + |
| 277 | + shm_mq_detach(mq); |
| 278 | + } |
| 279 | + } |
| 280 | + |
| 281 | + MemoryContextReset(collector_context); |
| 282 | + |
| 283 | + /* |
| 284 | + * We're done. Explicitly detach the shared memory segment so that we |
| 285 | + * don't get a resource leak warning at commit time. This will fire any |
| 286 | + * on_dsm_detach callbacks we've registered, as well. Once that's done, |
| 287 | + * we can go ahead and exit. |
| 288 | + */ |
| 289 | + proc_exit(0); |
| 290 | +} |
0 commit comments