| 1 |
/*-------------------------------------------------------------------- |
| 2 |
* bgworker.c |
| 3 |
* POSTGRES pluggable background workers implementation |
| 4 |
* |
| 5 |
* Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group |
| 6 |
* |
| 7 |
* IDENTIFICATION |
| 8 |
* src/backend/postmaster/bgworker.c |
| 9 |
* |
| 10 |
*------------------------------------------------------------------------- |
| 11 |
*/ |
| 12 |
|
| 13 |
#include "postgres.h" |
| 14 |
|
| 15 |
#include "access/parallel.h" |
| 16 |
#include "libpq/pqsignal.h" |
| 17 |
#include "miscadmin.h" |
| 18 |
#include "pgstat.h" |
| 19 |
#include "port/atomics.h" |
| 20 |
#include "postmaster/bgworker_internals.h" |
| 21 |
#include "postmaster/postmaster.h" |
| 22 |
#include "replication/logicallauncher.h" |
| 23 |
#include "replication/logicalworker.h" |
| 24 |
#include "storage/ipc.h" |
| 25 |
#include "storage/latch.h" |
| 26 |
#include "storage/lwlock.h" |
| 27 |
#include "storage/pmsignal.h" |
| 28 |
#include "storage/proc.h" |
| 29 |
#include "storage/procsignal.h" |
| 30 |
#include "storage/shmem.h" |
| 31 |
#include "tcop/tcopprot.h" |
| 32 |
#include "utils/ascii.h" |
| 33 |
#include "utils/memutils.h" |
| 34 |
#include "utils/ps_status.h" |
| 35 |
#include "utils/timeout.h" |
| 36 |
|
| 37 |
/* |
| 38 |
* The postmaster's list of registered background workers, in private memory. |
| 39 |
*/ |
| 40 |
dlist_head BackgroundWorkerList = DLIST_STATIC_INIT(BackgroundWorkerList); |
| 41 |
|
| 42 |
/* |
| 43 |
* BackgroundWorkerSlots exist in shared memory and can be accessed (via |
| 44 |
* the BackgroundWorkerArray) by both the postmaster and by regular backends. |
| 45 |
* However, the postmaster cannot take locks, even spinlocks, because this |
| 46 |
* might allow it to crash or become wedged if shared memory gets corrupted. |
| 47 |
* Such an outcome is intolerable. Therefore, we need a lockless protocol |
| 48 |
* for coordinating access to this data. |
| 49 |
* |
| 50 |
* The 'in_use' flag is used to hand off responsibility for the slot between |
| 51 |
* the postmaster and the rest of the system. When 'in_use' is false, |
| 52 |
* the postmaster will ignore the slot entirely, except for the 'in_use' flag |
| 53 |
* itself, which it may read. In this state, regular backends may modify the |
| 54 |
* slot. Once a backend sets 'in_use' to true, the slot becomes the |
| 55 |
* responsibility of the postmaster. Regular backends may no longer modify it, |
| 56 |
* but the postmaster may examine it. Thus, a backend initializing a slot |
| 57 |
* must fully initialize the slot - and insert a write memory barrier - before |
| 58 |
* marking it as in use. |
| 59 |
* |
| 60 |
* As an exception, however, even when the slot is in use, regular backends |
| 61 |
* may set the 'terminate' flag for a slot, telling the postmaster not |
| 62 |
* to restart it. Once the background worker is no longer running, the slot |
| 63 |
* will be released for reuse. |
| 64 |
* |
| 65 |
* In addition to coordinating with the postmaster, backends modifying this |
| 66 |
* data structure must coordinate with each other. Since they can take locks, |
| 67 |
* this is straightforward: any backend wishing to manipulate a slot must |
| 68 |
* take BackgroundWorkerLock in exclusive mode. Backends wishing to read |
| 69 |
* data that might get concurrently modified by other backends should take |
| 70 |
* this lock in shared mode. No matter what, backends reading this data |
| 71 |
* structure must be able to tolerate concurrent modifications by the |
| 72 |
* postmaster. |
| 73 |
*/ |
| 74 |
typedef struct BackgroundWorkerSlot |
| 75 |
{ |
| 76 |
bool in_use; |
| 77 |
bool terminate; |
| 78 |
pid_t pid; /* InvalidPid = not started yet; 0 = dead */ |
| 79 |
uint64 generation; /* incremented when slot is recycled */ |
| 80 |
BackgroundWorker worker; |
| 81 |
} BackgroundWorkerSlot; |
| 82 |
|
| 83 |
/* |
| 84 |
* In order to limit the total number of parallel workers (according to |
| 85 |
* max_parallel_workers GUC), we maintain the number of active parallel |
| 86 |
* workers. Since the postmaster cannot take locks, two variables are used for |
| 87 |
* this purpose: the number of registered parallel workers (modified by the |
| 88 |
* backends, protected by BackgroundWorkerLock) and the number of terminated |
| 89 |
* parallel workers (modified only by the postmaster, lockless). The active |
| 90 |
* number of parallel workers is the number of registered workers minus the |
| 91 |
* terminated ones. These counters can of course overflow, but it's not |
| 92 |
* important here since the subtraction will still give the right number. |
| 93 |
*/ |
| 94 |
typedef struct BackgroundWorkerArray |
| 95 |
{ |
| 96 |
int total_slots; |
| 97 |
uint32 parallel_register_count; |
| 98 |
uint32 parallel_terminate_count; |
| 99 |
BackgroundWorkerSlot slot[FLEXIBLE_ARRAY_MEMBER]; |
| 100 |
} BackgroundWorkerArray; |
| 101 |
|
| 102 |
struct BackgroundWorkerHandle |
| 103 |
{ |
| 104 |
int slot; |
| 105 |
uint64 generation; |
| 106 |
}; |
| 107 |
|
| 108 |
static BackgroundWorkerArray *BackgroundWorkerData; |
| 109 |
|
| 110 |
/* |
| 111 |
* List of internal background worker entry points. We need this for |
| 112 |
* reasons explained in LookupBackgroundWorkerFunction(), below. |
| 113 |
*/ |
| 114 |
static const struct |
| 115 |
{ |
| 116 |
const char *fn_name; |
| 117 |
bgworker_main_type fn_addr; |
| 118 |
} InternalBGWorkers[] = |
| 119 |
|
| 120 |
{ |
| 121 |
{ |
| 122 |
"ParallelWorkerMain", ParallelWorkerMain |
| 123 |
}, |
| 124 |
{ |
| 125 |
"ApplyLauncherMain", ApplyLauncherMain |
| 126 |
}, |
| 127 |
{ |
| 128 |
"ApplyWorkerMain", ApplyWorkerMain |
| 129 |
}, |
| 130 |
{ |
| 131 |
"ParallelApplyWorkerMain", ParallelApplyWorkerMain |
| 132 |
}, |
| 133 |
{ |
| 134 |
"TablesyncWorkerMain", TablesyncWorkerMain |
| 135 |
} |
| 136 |
}; |
| 137 |
|
| 138 |
/* Private functions. */ |
| 139 |
static bgworker_main_type LookupBackgroundWorkerFunction(const char *libraryname, const char *funcname); |
| 140 |
|
| 141 |
|
| 142 |
/* |
| 143 |
* Calculate shared memory needed. |
| 144 |
*/ |
| 145 |
Size |
| 146 |
BackgroundWorkerShmemSize(void) |
| 147 |
{ |
| 148 |
Size size; |
| 149 |
|
| 150 |
/* Array of workers is variably sized. */ |
| 151 |
size = offsetof(BackgroundWorkerArray, slot); |
| 152 |
size = add_size(size, mul_size(max_worker_processes, |
| 153 |
sizeof(BackgroundWorkerSlot))); |
| 154 |
|
| 155 |
return size; |
| 156 |
} |
| 157 |
|
| 158 |
/* |
| 159 |
* Initialize shared memory. |
| 160 |
*/ |
| 161 |
void |
| 162 |
BackgroundWorkerShmemInit(void) |
| 163 |
{ |
| 164 |
bool found; |
| 165 |
|
| 166 |
BackgroundWorkerData = ShmemInitStruct("Background Worker Data", |
| 167 |
BackgroundWorkerShmemSize(), |
| 168 |
&found); |
| 169 |
if (!IsUnderPostmaster) |
| 170 |
{ |
| 171 |
dlist_iter iter; |
| 172 |
int slotno = 0; |
| 173 |
|
| 174 |
BackgroundWorkerData->total_slots = max_worker_processes; |
| 175 |
BackgroundWorkerData->parallel_register_count = 0; |
| 176 |
BackgroundWorkerData->parallel_terminate_count = 0; |
| 177 |
|
| 178 |
/* |
| 179 |
* Copy contents of worker list into shared memory. Record the shared |
| 180 |
* memory slot assigned to each worker. This ensures a 1-to-1 |
| 181 |
* correspondence between the postmaster's private list and the array |
| 182 |
* in shared memory. |
| 183 |
*/ |
| 184 |
dlist_foreach(iter, &BackgroundWorkerList) |
| 185 |
{ |
| 186 |
BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno]; |
| 187 |
RegisteredBgWorker *rw; |
| 188 |
|
| 189 |
rw = dlist_container(RegisteredBgWorker, rw_lnode, iter.cur); |
| 190 |
Assert(slotno < max_worker_processes); |
| 191 |
slot->in_use = true; |
| 192 |
slot->terminate = false; |
| 193 |
slot->pid = InvalidPid; |
| 194 |
slot->generation = 0; |
| 195 |
rw->rw_shmem_slot = slotno; |
| 196 |
rw->rw_worker.bgw_notify_pid = 0; /* might be reinit after crash */ |
| 197 |
memcpy(&slot->worker, &rw->rw_worker, sizeof(BackgroundWorker)); |
| 198 |
++slotno; |
| 199 |
} |
| 200 |
|
| 201 |
/* |
| 202 |
* Mark any remaining slots as not in use. |
| 203 |
*/ |
| 204 |
while (slotno < max_worker_processes) |
| 205 |
{ |
| 206 |
BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno]; |
| 207 |
|
| 208 |
slot->in_use = false; |
| 209 |
++slotno; |
| 210 |
} |
| 211 |
} |
| 212 |
else |
| 213 |
Assert(found); |
| 214 |
} |
| 215 |
|
| 216 |
/* |
| 217 |
* Search the postmaster's backend-private list of RegisteredBgWorker objects |
| 218 |
* for the one that maps to the given slot number. |
| 219 |
*/ |
| 220 |
static RegisteredBgWorker * |
| 221 |
FindRegisteredWorkerBySlotNumber(int slotno) |
| 222 |
{ |
| 223 |
dlist_iter iter; |
| 224 |
|
| 225 |
dlist_foreach(iter, &BackgroundWorkerList) |
| 226 |
{ |
| 227 |
RegisteredBgWorker *rw; |
| 228 |
|
| 229 |
rw = dlist_container(RegisteredBgWorker, rw_lnode, iter.cur); |
| 230 |
if (rw->rw_shmem_slot == slotno) |
| 231 |
return rw; |
| 232 |
} |
| 233 |
|
| 234 |
return NULL; |
| 235 |
} |
| 236 |
|
| 237 |
/* |
| 238 |
* Notice changes to shared memory made by other backends. |
| 239 |
* Accept new worker requests only if allow_new_workers is true. |
| 240 |
* |
| 241 |
* This code runs in the postmaster, so we must be very careful not to assume |
| 242 |
* that shared memory contents are sane. Otherwise, a rogue backend could |
| 243 |
* take out the postmaster. |
| 244 |
*/ |
| 245 |
void |
| 246 |
BackgroundWorkerStateChange(bool allow_new_workers) |
| 247 |
{ |
| 248 |
int slotno; |
| 249 |
|
| 250 |
/* |
| 251 |
* The total number of slots stored in shared memory should match our |
| 252 |
* notion of max_worker_processes. If it does not, something is very |
| 253 |
* wrong. Further down, we always refer to this value as |
| 254 |
* max_worker_processes, in case shared memory gets corrupted while we're |
| 255 |
* looping. |
| 256 |
*/ |
| 257 |
if (max_worker_processes != BackgroundWorkerData->total_slots) |
| 258 |
{ |
| 259 |
ereport(LOG, |
| 260 |
(errmsg("inconsistent background worker state (\"max_worker_processes\"=%d, total slots=%d)", |
| 261 |
max_worker_processes, |
| 262 |
BackgroundWorkerData->total_slots))); |
| 263 |
return; |
| 264 |
} |
| 265 |
|
| 266 |
/* |
| 267 |
* Iterate through slots, looking for newly-registered workers or workers |
| 268 |
* who must die. |
| 269 |
*/ |
| 270 |
for (slotno = 0; slotno < max_worker_processes; ++slotno) |
| 271 |
{ |
| 272 |
BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno]; |
| 273 |
RegisteredBgWorker *rw; |
| 274 |
|
| 275 |
if (!slot->in_use) |
| 276 |
continue; |
| 277 |
|
| 278 |
/* |
| 279 |
* Make sure we don't see the in_use flag before the updated slot |
| 280 |
* contents. |
| 281 |
*/ |
| 282 |
pg_read_barrier(); |
| 283 |
|
| 284 |
/* See whether we already know about this worker. */ |
| 285 |
rw = FindRegisteredWorkerBySlotNumber(slotno); |
| 286 |
if (rw != NULL) |
| 287 |
{ |
| 288 |
/* |
| 289 |
* In general, the worker data can't change after it's initially |
| 290 |
* registered. However, someone can set the terminate flag. |
| 291 |
*/ |
| 292 |
if (slot->terminate && !rw->rw_terminate) |
| 293 |
{ |
| 294 |
rw->rw_terminate = true; |
| 295 |
if (rw->rw_pid != 0) |
| 296 |
kill(rw->rw_pid, SIGTERM); |
| 297 |
else |
| 298 |
{ |
| 299 |
/* Report never-started, now-terminated worker as dead. */ |
| 300 |
ReportBackgroundWorkerPID(rw); |
| 301 |
} |
| 302 |
} |
| 303 |
continue; |
| 304 |
} |
| 305 |
|
| 306 |
/* |
| 307 |
* If we aren't allowing new workers, then immediately mark it for |
| 308 |
* termination; the next stanza will take care of cleaning it up. |
| 309 |
* Doing this ensures that any process waiting for the worker will get |
| 310 |
* awoken, even though the worker will never be allowed to run. |
| 311 |
*/ |
| 312 |
if (!allow_new_workers) |
| 313 |
slot->terminate = true; |
| 314 |
|
| 315 |
/* |
| 316 |
* If the worker is marked for termination, we don't need to add it to |
| 317 |
* the registered workers list; we can just free the slot. However, if |
| 318 |
* bgw_notify_pid is set, the process that registered the worker may |
| 319 |
* need to know that we've processed the terminate request, so be sure |
| 320 |
* to signal it. |
| 321 |
*/ |
| 322 |
if (slot->terminate) |
| 323 |
{ |
| 324 |
int notify_pid; |
| 325 |
|
| 326 |
/* |
| 327 |
* We need a memory barrier here to make sure that the load of |
| 328 |
* bgw_notify_pid and the update of parallel_terminate_count |
| 329 |
* complete before the store to in_use. |
| 330 |
*/ |
| 331 |
notify_pid = slot->worker.bgw_notify_pid; |
| 332 |
if ((slot->worker.bgw_flags & BGWORKER_CLASS_PARALLEL) != 0) |
| 333 |
BackgroundWorkerData->parallel_terminate_count++; |
| 334 |
slot->pid = 0; |
| 335 |
|
| 336 |
pg_memory_barrier(); |
| 337 |
slot->in_use = false; |
| 338 |
|
| 339 |
if (notify_pid != 0) |
| 340 |
kill(notify_pid, SIGUSR1); |
| 341 |
|
| 342 |
continue; |
| 343 |
} |
| 344 |
|
| 345 |
/* |
| 346 |
* Copy the registration data into the registered workers list. |
| 347 |
*/ |
| 348 |
rw = MemoryContextAllocExtended(PostmasterContext, |
| 349 |
sizeof(RegisteredBgWorker), |
| 350 |
MCXT_ALLOC_NO_OOM | MCXT_ALLOC_ZERO); |
| 351 |
if (rw == NULL) |
| 352 |
{ |
| 353 |
ereport(LOG, |
| 354 |
(errcode(ERRCODE_OUT_OF_MEMORY), |
| 355 |
errmsg("out of memory"))); |
| 356 |
return; |
| 357 |
} |
| 358 |
|
| 359 |
/* |
| 360 |
* Copy strings in a paranoid way. If shared memory is corrupted, the |
| 361 |
* source data might not even be NUL-terminated. |
| 362 |
*/ |
| 363 |
ascii_safe_strlcpy(rw->rw_worker.bgw_name, |
| 364 |
slot->worker.bgw_name, BGW_MAXLEN); |
| 365 |
ascii_safe_strlcpy(rw->rw_worker.bgw_type, |
| 366 |
slot->worker.bgw_type, BGW_MAXLEN); |
| 367 |
ascii_safe_strlcpy(rw->rw_worker.bgw_library_name, |
| 368 |
slot->worker.bgw_library_name, MAXPGPATH); |
| 369 |
ascii_safe_strlcpy(rw->rw_worker.bgw_function_name, |
| 370 |
slot->worker.bgw_function_name, BGW_MAXLEN); |
| 371 |
|
| 372 |
/* |
| 373 |
* Copy various fixed-size fields. |
| 374 |
* |
| 375 |
* flags, start_time, and restart_time are examined by the postmaster, |
| 376 |
* but nothing too bad will happen if they are corrupted. The |
| 377 |
* remaining fields will only be examined by the child process. It |
| 378 |
* might crash, but we won't. |
| 379 |
*/ |
| 380 |
rw->rw_worker.bgw_flags = slot->worker.bgw_flags; |
| 381 |
rw->rw_worker.bgw_start_time = slot->worker.bgw_start_time; |
| 382 |
rw->rw_worker.bgw_restart_time = slot->worker.bgw_restart_time; |
| 383 |
rw->rw_worker.bgw_main_arg = slot->worker.bgw_main_arg; |
| 384 |
memcpy(rw->rw_worker.bgw_extra, slot->worker.bgw_extra, BGW_EXTRALEN); |
| 385 |
|
| 386 |
/* |
| 387 |
* Copy the PID to be notified about state changes, but only if the |
| 388 |
* postmaster knows about a backend with that PID. It isn't an error |
| 389 |
* if the postmaster doesn't know about the PID, because the backend |
| 390 |
* that requested the worker could have died (or been killed) just |
| 391 |
* after doing so. Nonetheless, at least until we get some experience |
| 392 |
* with how this plays out in the wild, log a message at a relative |
| 393 |
* high debug level. |
| 394 |
*/ |
| 395 |
rw->rw_worker.bgw_notify_pid = slot->worker.bgw_notify_pid; |
| 396 |
if (!PostmasterMarkPIDForWorkerNotify(rw->rw_worker.bgw_notify_pid)) |
| 397 |
{ |
| 398 |
elog(DEBUG1, "worker notification PID %d is not valid", |
| 399 |
(int) rw->rw_worker.bgw_notify_pid); |
| 400 |
rw->rw_worker.bgw_notify_pid = 0; |
| 401 |
} |
| 402 |
|
| 403 |
/* Initialize postmaster bookkeeping. */ |
| 404 |
rw->rw_pid = 0; |
| 405 |
rw->rw_crashed_at = 0; |
| 406 |
rw->rw_shmem_slot = slotno; |
| 407 |
rw->rw_terminate = false; |
| 408 |
|
| 409 |
/* Log it! */ |
| 410 |
ereport(DEBUG1, |
| 411 |
(errmsg_internal("registering background worker \"%s\"", |
| 412 |
rw->rw_worker.bgw_name))); |
| 413 |
|
| 414 |
dlist_push_head(&BackgroundWorkerList, &rw->rw_lnode); |
| 415 |
} |
| 416 |
} |
| 417 |
|
| 418 |
/* |
| 419 |
* Forget about a background worker that's no longer needed. |
| 420 |
* |
| 421 |
* NOTE: The entry is unlinked from BackgroundWorkerList. If the caller is |
| 422 |
* iterating through it, better use a mutable iterator! |
| 423 |
* |
| 424 |
* Caller is responsible for notifying bgw_notify_pid, if appropriate. |
| 425 |
* |
| 426 |
* This function must be invoked only in the postmaster. |
| 427 |
*/ |
| 428 |
void |
| 429 |
ForgetBackgroundWorker(RegisteredBgWorker *rw) |
| 430 |
{ |
| 431 |
BackgroundWorkerSlot *slot; |
| 432 |
|
| 433 |
Assert(rw->rw_shmem_slot < max_worker_processes); |
| 434 |
slot = &BackgroundWorkerData->slot[rw->rw_shmem_slot]; |
| 435 |
Assert(slot->in_use); |
| 436 |
|
| 437 |
/* |
| 438 |
* We need a memory barrier here to make sure that the update of |
| 439 |
* parallel_terminate_count completes before the store to in_use. |
| 440 |
*/ |
| 441 |
if ((rw->rw_worker.bgw_flags & BGWORKER_CLASS_PARALLEL) != 0) |
| 442 |
BackgroundWorkerData->parallel_terminate_count++; |
| 443 |
|
| 444 |
pg_memory_barrier(); |
| 445 |
slot->in_use = false; |
| 446 |
|
| 447 |
ereport(DEBUG1, |
| 448 |
(errmsg_internal("unregistering background worker \"%s\"", |
| 449 |
rw->rw_worker.bgw_name))); |
| 450 |
|
| 451 |
dlist_delete(&rw->rw_lnode); |
| 452 |
pfree(rw); |
| 453 |
} |
| 454 |
|
| 455 |
/* |
| 456 |
* Report the PID of a newly-launched background worker in shared memory. |
| 457 |
* |
| 458 |
* This function should only be called from the postmaster. |
| 459 |
*/ |
| 460 |
void |
| 461 |
ReportBackgroundWorkerPID(RegisteredBgWorker *rw) |
| 462 |
{ |
| 463 |
BackgroundWorkerSlot *slot; |
| 464 |
|
| 465 |
Assert(rw->rw_shmem_slot < max_worker_processes); |
| 466 |
slot = &BackgroundWorkerData->slot[rw->rw_shmem_slot]; |
| 467 |
slot->pid = rw->rw_pid; |
| 468 |
|
| 469 |
if (rw->rw_worker.bgw_notify_pid != 0) |
| 470 |
kill(rw->rw_worker.bgw_notify_pid, SIGUSR1); |
| 471 |
} |
| 472 |
|
| 473 |
/* |
| 474 |
* Report that the PID of a background worker is now zero because a |
| 475 |
* previously-running background worker has exited. |
| 476 |
* |
| 477 |
* NOTE: The entry may be unlinked from BackgroundWorkerList. If the caller |
| 478 |
* is iterating through it, better use a mutable iterator! |
| 479 |
* |
| 480 |
* This function should only be called from the postmaster. |
| 481 |
*/ |
| 482 |
void |
| 483 |
ReportBackgroundWorkerExit(RegisteredBgWorker *rw) |
| 484 |
{ |
| 485 |
BackgroundWorkerSlot *slot; |
| 486 |
int notify_pid; |
| 487 |
|
| 488 |
Assert(rw->rw_shmem_slot < max_worker_processes); |
| 489 |
slot = &BackgroundWorkerData->slot[rw->rw_shmem_slot]; |
| 490 |
slot->pid = rw->rw_pid; |
| 491 |
notify_pid = rw->rw_worker.bgw_notify_pid; |
| 492 |
|
| 493 |
/* |
| 494 |
* If this worker is slated for deregistration, do that before notifying |
| 495 |
* the process which started it. Otherwise, if that process tries to |
| 496 |
* reuse the slot immediately, it might not be available yet. In theory |
| 497 |
* that could happen anyway if the process checks slot->pid at just the |
| 498 |
* wrong moment, but this makes the window narrower. |
| 499 |
*/ |
| 500 |
if (rw->rw_terminate || |
| 501 |
rw->rw_worker.bgw_restart_time == BGW_NEVER_RESTART) |
| 502 |
ForgetBackgroundWorker(rw); |
| 503 |
|
| 504 |
if (notify_pid != 0) |
| 505 |
kill(notify_pid, SIGUSR1); |
| 506 |
} |
| 507 |
|
| 508 |
/* |
| 509 |
* Cancel SIGUSR1 notifications for a PID belonging to an exiting backend. |
| 510 |
* |
| 511 |
* This function should only be called from the postmaster. |
| 512 |
*/ |
| 513 |
void |
| 514 |
BackgroundWorkerStopNotifications(pid_t pid) |
| 515 |
{ |
| 516 |
dlist_iter iter; |
| 517 |
|
| 518 |
dlist_foreach(iter, &BackgroundWorkerList) |
| 519 |
{ |
| 520 |
RegisteredBgWorker *rw; |
| 521 |
|
| 522 |
rw = dlist_container(RegisteredBgWorker, rw_lnode, iter.cur); |
| 523 |
if (rw->rw_worker.bgw_notify_pid == pid) |
| 524 |
rw->rw_worker.bgw_notify_pid = 0; |
| 525 |
} |
| 526 |
} |
| 527 |
|
| 528 |
/* |
| 529 |
* Cancel any not-yet-started worker requests that have waiting processes. |
| 530 |
* |
| 531 |
* This is called during a normal ("smart" or "fast") database shutdown. |
| 532 |
* After this point, no new background workers will be started, so anything |
| 533 |
* that might be waiting for them needs to be kicked off its wait. We do |
| 534 |
* that by canceling the bgworker registration entirely, which is perhaps |
| 535 |
* overkill, but since we're shutting down it does not matter whether the |
| 536 |
* registration record sticks around. |
| 537 |
* |
| 538 |
* This function should only be called from the postmaster. |
| 539 |
*/ |
| 540 |
void |
| 541 |
ForgetUnstartedBackgroundWorkers(void) |
| 542 |
{ |
| 543 |
dlist_mutable_iter iter; |
| 544 |
|
| 545 |
dlist_foreach_modify(iter, &BackgroundWorkerList) |
| 546 |
{ |
| 547 |
RegisteredBgWorker *rw; |
| 548 |
BackgroundWorkerSlot *slot; |
| 549 |
|
| 550 |
rw = dlist_container(RegisteredBgWorker, rw_lnode, iter.cur); |
| 551 |
Assert(rw->rw_shmem_slot < max_worker_processes); |
| 552 |
slot = &BackgroundWorkerData->slot[rw->rw_shmem_slot]; |
| 553 |
|
| 554 |
/* If it's not yet started, and there's someone waiting ... */ |
| 555 |
if (slot->pid == InvalidPid && |
| 556 |
rw->rw_worker.bgw_notify_pid != 0) |
| 557 |
{ |
| 558 |
/* ... then zap it, and notify the waiter */ |
| 559 |
int notify_pid = rw->rw_worker.bgw_notify_pid; |
| 560 |
|
| 561 |
ForgetBackgroundWorker(rw); |
| 562 |
if (notify_pid != 0) |
| 563 |
kill(notify_pid, SIGUSR1); |
| 564 |
} |
| 565 |
} |
| 566 |
} |
| 567 |
|
| 568 |
/* |
| 569 |
* Reset background worker crash state. |
| 570 |
* |
| 571 |
* We assume that, after a crash-and-restart cycle, background workers without |
| 572 |
* the never-restart flag should be restarted immediately, instead of waiting |
| 573 |
* for bgw_restart_time to elapse. On the other hand, workers with that flag |
| 574 |
* should be forgotten immediately, since we won't ever restart them. |
| 575 |
* |
| 576 |
* This function should only be called from the postmaster. |
| 577 |
*/ |
| 578 |
void |
| 579 |
ResetBackgroundWorkerCrashTimes(void) |
| 580 |
{ |
| 581 |
dlist_mutable_iter iter; |
| 582 |
|
| 583 |
dlist_foreach_modify(iter, &BackgroundWorkerList) |
| 584 |
{ |
| 585 |
RegisteredBgWorker *rw; |
| 586 |
|
| 587 |
rw = dlist_container(RegisteredBgWorker, rw_lnode, iter.cur); |
| 588 |
|
| 589 |
if (rw->rw_worker.bgw_restart_time == BGW_NEVER_RESTART) |
| 590 |
{ |
| 591 |
/* |
| 592 |
* Workers marked BGW_NEVER_RESTART shouldn't get relaunched after |
| 593 |
* the crash, so forget about them. (If we wait until after the |
| 594 |
* crash to forget about them, and they are parallel workers, |
| 595 |
* parallel_terminate_count will get incremented after we've |
| 596 |
* already zeroed parallel_register_count, which would be bad.) |
| 597 |
*/ |
| 598 |
ForgetBackgroundWorker(rw); |
| 599 |
} |
| 600 |
else |
| 601 |
{ |
| 602 |
/* |
| 603 |
* The accounting which we do via parallel_register_count and |
| 604 |
* parallel_terminate_count would get messed up if a worker marked |
| 605 |
* parallel could survive a crash and restart cycle. All such |
| 606 |
* workers should be marked BGW_NEVER_RESTART, and thus control |
| 607 |
* should never reach this branch. |
| 608 |
*/ |
| 609 |
Assert((rw->rw_worker.bgw_flags & BGWORKER_CLASS_PARALLEL) == 0); |
| 610 |
|
| 611 |
/* |
| 612 |
* Allow this worker to be restarted immediately after we finish |
| 613 |
* resetting. |
| 614 |
*/ |
| 615 |
rw->rw_crashed_at = 0; |
| 616 |
rw->rw_pid = 0; |
| 617 |
|
| 618 |
/* |
| 619 |
* If there was anyone waiting for it, they're history. |
| 620 |
*/ |
| 621 |
rw->rw_worker.bgw_notify_pid = 0; |
| 622 |
} |
| 623 |
} |
| 624 |
} |
| 625 |
|
| 626 |
/* |
| 627 |
* Complain about the BackgroundWorker definition using error level elevel. |
| 628 |
* Return true if it looks ok, false if not (unless elevel >= ERROR, in |
| 629 |
* which case we won't return at all in the not-OK case). |
| 630 |
*/ |
| 631 |
static bool |
| 632 |
SanityCheckBackgroundWorker(BackgroundWorker *worker, int elevel) |
| 633 |
{ |
| 634 |
/* sanity check for flags */ |
| 635 |
|
| 636 |
/* |
| 637 |
* We used to support workers not connected to shared memory, but don't |
| 638 |
* anymore. Thus this is a required flag now. We're not removing the flag |
| 639 |
* for compatibility reasons and because the flag still provides some |
| 640 |
* signal when reading code. |
| 641 |
*/ |
| 642 |
if (!(worker->bgw_flags & BGWORKER_SHMEM_ACCESS)) |
| 643 |
{ |
| 644 |
ereport(elevel, |
| 645 |
(errcode(ERRCODE_INVALID_PARAMETER_VALUE), |
| 646 |
errmsg("background worker \"%s\": background workers without shared memory access are not supported", |
| 647 |
worker->bgw_name))); |
| 648 |
return false; |
| 649 |
} |
| 650 |
|
| 651 |
if (worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION) |
| 652 |
{ |
| 653 |
if (worker->bgw_start_time == BgWorkerStart_PostmasterStart) |
| 654 |
{ |
| 655 |
ereport(elevel, |
| 656 |
(errcode(ERRCODE_INVALID_PARAMETER_VALUE), |
| 657 |
errmsg("background worker \"%s\": cannot request database access if starting at postmaster start", |
| 658 |
worker->bgw_name))); |
| 659 |
return false; |
| 660 |
} |
| 661 |
|
| 662 |
/* XXX other checks? */ |
| 663 |
} |
| 664 |
|
| 665 |
if ((worker->bgw_restart_time < 0 && |
| 666 |
worker->bgw_restart_time != BGW_NEVER_RESTART) || |
| 667 |
(worker->bgw_restart_time > USECS_PER_DAY / 1000)) |
| 668 |
{ |
| 669 |
ereport(elevel, |
| 670 |
(errcode(ERRCODE_INVALID_PARAMETER_VALUE), |
| 671 |
errmsg("background worker \"%s\": invalid restart interval", |
| 672 |
worker->bgw_name))); |
| 673 |
return false; |
| 674 |
} |
| 675 |
|
| 676 |
/* |
| 677 |
* Parallel workers may not be configured for restart, because the |
| 678 |
* parallel_register_count/parallel_terminate_count accounting can't |
| 679 |
* handle parallel workers lasting through a crash-and-restart cycle. |
| 680 |
*/ |
| 681 |
if (worker->bgw_restart_time != BGW_NEVER_RESTART && |
| 682 |
(worker->bgw_flags & BGWORKER_CLASS_PARALLEL) != 0) |
| 683 |
{ |
| 684 |
ereport(elevel, |
| 685 |
(errcode(ERRCODE_INVALID_PARAMETER_VALUE), |
| 686 |
errmsg("background worker \"%s\": parallel workers may not be configured for restart", |
| 687 |
worker->bgw_name))); |
| 688 |
return false; |
| 689 |
} |
| 690 |
|
| 691 |
/* |
| 692 |
* If bgw_type is not filled in, use bgw_name. |
| 693 |
*/ |
| 694 |
if (strcmp(worker->bgw_type, "") == 0) |
| 695 |
strcpy(worker->bgw_type, worker->bgw_name); |
| 696 |
|
| 697 |
return true; |
| 698 |
} |
| 699 |
|
| 700 |
/* |
| 701 |
* Standard SIGTERM handler for background workers |
| 702 |
*/ |
| 703 |
static void |
| 704 |
bgworker_die(SIGNAL_ARGS) |
| 705 |
{ |
| 706 |
sigprocmask(SIG_SETMASK, &BlockSig, NULL); |
| 707 |
|
| 708 |
ereport(FATAL, |
| 709 |
(errcode(ERRCODE_ADMIN_SHUTDOWN), |
| 710 |
errmsg("terminating background worker \"%s\" due to administrator command", |
| 711 |
MyBgworkerEntry->bgw_type))); |
| 712 |
} |
| 713 |
|
| 714 |
/* |
| 715 |
* Main entry point for background worker processes. |
| 716 |
*/ |
| 717 |
void |
| 718 |
BackgroundWorkerMain(const void *startup_data, size_t startup_data_len) |
| 719 |
{ |
| 720 |
sigjmp_buf local_sigjmp_buf; |
| 721 |
BackgroundWorker *worker; |
| 722 |
bgworker_main_type entrypt; |
| 723 |
|
| 724 |
if (startup_data == NULL) |
| 725 |
elog(FATAL, "unable to find bgworker entry"); |
| 726 |
Assert(startup_data_len == sizeof(BackgroundWorker)); |
| 727 |
worker = MemoryContextAlloc(TopMemoryContext, sizeof(BackgroundWorker)); |
| 728 |
memcpy(worker, startup_data, sizeof(BackgroundWorker)); |
| 729 |
|
| 730 |
/* |
| 731 |
* Now that we're done reading the startup data, release postmaster's |
| 732 |
* working memory context. |
| 733 |
*/ |
| 734 |
if (PostmasterContext) |
| 735 |
{ |
| 736 |
MemoryContextDelete(PostmasterContext); |
| 737 |
PostmasterContext = NULL; |
| 738 |
} |
| 739 |
|
| 740 |
MyBgworkerEntry = worker; |
| 741 |
MyBackendType = B_BG_WORKER; |
| 742 |
init_ps_display(worker->bgw_name); |
| 743 |
|
| 744 |
Assert(GetProcessingMode() == InitProcessing); |
| 745 |
|
| 746 |
/* Apply PostAuthDelay */ |
| 747 |
if (PostAuthDelay > 0) |
| 748 |
pg_usleep(PostAuthDelay * 1000000L); |
| 749 |
|
| 750 |
/* |
| 751 |
* Set up signal handlers. |
| 752 |
*/ |
| 753 |
if (worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION) |
| 754 |
{ |
| 755 |
/* |
| 756 |
* SIGINT is used to signal canceling the current action |
| 757 |
*/ |
| 758 |
pqsignal(SIGINT, StatementCancelHandler); |
| 759 |
pqsignal(SIGUSR1, procsignal_sigusr1_handler); |
| 760 |
pqsignal(SIGFPE, FloatExceptionHandler); |
| 761 |
|
| 762 |
/* XXX Any other handlers needed here? */ |
| 763 |
} |
| 764 |
else |
| 765 |
{ |
| 766 |
pqsignal(SIGINT, SIG_IGN); |
| 767 |
pqsignal(SIGUSR1, SIG_IGN); |
| 768 |
pqsignal(SIGFPE, SIG_IGN); |
| 769 |
} |
| 770 |
pqsignal(SIGTERM, bgworker_die); |
| 771 |
/* SIGQUIT handler was already set up by InitPostmasterChild */ |
| 772 |
pqsignal(SIGHUP, SIG_IGN); |
| 773 |
|
| 774 |
InitializeTimeouts(); /* establishes SIGALRM handler */ |
| 775 |
|
| 776 |
pqsignal(SIGPIPE, SIG_IGN); |
| 777 |
pqsignal(SIGUSR2, SIG_IGN); |
| 778 |
pqsignal(SIGCHLD, SIG_DFL); |
| 779 |
|
| 780 |
/* |
| 781 |
* If an exception is encountered, processing resumes here. |
| 782 |
* |
| 783 |
* We just need to clean up, report the error, and go away. |
| 784 |
*/ |
| 785 |
if (sigsetjmp(local_sigjmp_buf, 1) != 0) |
| 786 |
{ |
| 787 |
/* Since not using PG_TRY, must reset error stack by hand */ |
| 788 |
error_context_stack = NULL; |
| 789 |
|
| 790 |
/* Prevent interrupts while cleaning up */ |
| 791 |
HOLD_INTERRUPTS(); |
| 792 |
|
| 793 |
/* |
| 794 |
* sigsetjmp will have blocked all signals, but we may need to accept |
| 795 |
* signals while communicating with our parallel leader. Once we've |
| 796 |
* done HOLD_INTERRUPTS() it should be safe to unblock signals. |
| 797 |
*/ |
| 798 |
BackgroundWorkerUnblockSignals(); |
| 799 |
|
| 800 |
/* Report the error to the parallel leader and the server log */ |
| 801 |
EmitErrorReport(); |
| 802 |
|
| 803 |
/* |
| 804 |
* Do we need more cleanup here? For shmem-connected bgworkers, we |
| 805 |
* will call InitProcess below, which will install ProcKill as exit |
| 806 |
* callback. That will take care of releasing locks, etc. |
| 807 |
*/ |
| 808 |
|
| 809 |
/* and go away */ |
| 810 |
proc_exit(1); |
| 811 |
} |
| 812 |
|
| 813 |
/* We can now handle ereport(ERROR) */ |
| 814 |
PG_exception_stack = &local_sigjmp_buf; |
| 815 |
|
| 816 |
/* |
| 817 |
* Create a per-backend PGPROC struct in shared memory. We must do this |
| 818 |
* before we can use LWLocks or access any shared memory. |
| 819 |
*/ |
| 820 |
InitProcess(); |
| 821 |
|
| 822 |
/* |
| 823 |
* Early initialization. |
| 824 |
*/ |
| 825 |
BaseInit(); |
| 826 |
|
| 827 |
/* |
| 828 |
* Look up the entry point function, loading its library if necessary. |
| 829 |
*/ |
| 830 |
entrypt = LookupBackgroundWorkerFunction(worker->bgw_library_name, |
| 831 |
worker->bgw_function_name); |
| 832 |
|
| 833 |
/* |
| 834 |
* Note that in normal processes, we would call InitPostgres here. For a |
| 835 |
* worker, however, we don't know what database to connect to, yet; so we |
| 836 |
* need to wait until the user code does it via |
| 837 |
* BackgroundWorkerInitializeConnection(). |
| 838 |
*/ |
| 839 |
|
| 840 |
/* |
| 841 |
* Now invoke the user-defined worker code |
| 842 |
*/ |
| 843 |
entrypt(worker->bgw_main_arg); |
| 844 |
|
| 845 |
/* ... and if it returns, we're done */ |
| 846 |
proc_exit(0); |
| 847 |
} |
| 848 |
|
| 849 |
/* |
| 850 |
* Connect background worker to a database. |
| 851 |
*/ |
| 852 |
void |
| 853 |
BackgroundWorkerInitializeConnection(const char *dbname, const char *username, uint32 flags) |
| 854 |
{ |
| 855 |
BackgroundWorker *worker = MyBgworkerEntry; |
| 856 |
bits32 init_flags = 0; /* never honor session_preload_libraries */ |
| 857 |
|
| 858 |
/* ignore datallowconn and ACL_CONNECT? */ |
| 859 |
if (flags & BGWORKER_BYPASS_ALLOWCONN) |
| 860 |
init_flags |= INIT_PG_OVERRIDE_ALLOW_CONNS; |
| 861 |
/* ignore rolcanlogin? */ |
| 862 |
if (flags & BGWORKER_BYPASS_ROLELOGINCHECK) |
| 863 |
init_flags |= INIT_PG_OVERRIDE_ROLE_LOGIN; |
| 864 |
|
| 865 |
/* XXX is this the right errcode? */ |
| 866 |
if (!(worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION)) |
| 867 |
ereport(FATAL, |
| 868 |
(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED), |
| 869 |
errmsg("database connection requirement not indicated during registration"))); |
| 870 |
|
| 871 |
InitPostgres(dbname, InvalidOid, /* database to connect to */ |
| 872 |
username, InvalidOid, /* role to connect as */ |
| 873 |
init_flags, |
| 874 |
NULL); /* no out_dbname */ |
| 875 |
|
| 876 |
/* it had better not gotten out of "init" mode yet */ |
| 877 |
if (!IsInitProcessingMode()) |
| 878 |
ereport(ERROR, |
| 879 |
(errmsg("invalid processing mode in background worker"))); |
| 880 |
SetProcessingMode(NormalProcessing); |
| 881 |
} |
| 882 |
|
| 883 |
/* |
| 884 |
* Connect background worker to a database using OIDs. |
| 885 |
*/ |
| 886 |
void |
| 887 |
BackgroundWorkerInitializeConnectionByOid(Oid dboid, Oid useroid, uint32 flags) |
| 888 |
{ |
| 889 |
BackgroundWorker *worker = MyBgworkerEntry; |
| 890 |
bits32 init_flags = 0; /* never honor session_preload_libraries */ |
| 891 |
|
| 892 |
/* ignore datallowconn and ACL_CONNECT? */ |
| 893 |
if (flags & BGWORKER_BYPASS_ALLOWCONN) |
| 894 |
init_flags |= INIT_PG_OVERRIDE_ALLOW_CONNS; |
| 895 |
/* ignore rolcanlogin? */ |
| 896 |
if (flags & BGWORKER_BYPASS_ROLELOGINCHECK) |
| 897 |
init_flags |= INIT_PG_OVERRIDE_ROLE_LOGIN; |
| 898 |
|
| 899 |
/* XXX is this the right errcode? */ |
| 900 |
if (!(worker->bgw_flags & BGWORKER_BACKEND_DATABASE_CONNECTION)) |
| 901 |
ereport(FATAL, |
| 902 |
(errcode(ERRCODE_PROGRAM_LIMIT_EXCEEDED), |
| 903 |
errmsg("database connection requirement not indicated during registration"))); |
| 904 |
|
| 905 |
InitPostgres(NULL, dboid, /* database to connect to */ |
| 906 |
NULL, useroid, /* role to connect as */ |
| 907 |
init_flags, |
| 908 |
NULL); /* no out_dbname */ |
| 909 |
|
| 910 |
/* it had better not gotten out of "init" mode yet */ |
| 911 |
if (!IsInitProcessingMode()) |
| 912 |
ereport(ERROR, |
| 913 |
(errmsg("invalid processing mode in background worker"))); |
| 914 |
SetProcessingMode(NormalProcessing); |
| 915 |
} |
| 916 |
|
| 917 |
/* |
| 918 |
* Block/unblock signals in a background worker |
| 919 |
*/ |
| 920 |
void |
| 921 |
BackgroundWorkerBlockSignals(void) |
| 922 |
{ |
| 923 |
sigprocmask(SIG_SETMASK, &BlockSig, NULL); |
| 924 |
} |
| 925 |
|
| 926 |
void |
| 927 |
BackgroundWorkerUnblockSignals(void) |
| 928 |
{ |
| 929 |
sigprocmask(SIG_SETMASK, &UnBlockSig, NULL); |
| 930 |
} |
| 931 |
|
| 932 |
/* |
| 933 |
* Register a new static background worker. |
| 934 |
* |
| 935 |
* This can only be called directly from postmaster or in the _PG_init |
| 936 |
* function of a module library that's loaded by shared_preload_libraries; |
| 937 |
* otherwise it will have no effect. |
| 938 |
*/ |
| 939 |
void |
| 940 |
RegisterBackgroundWorker(BackgroundWorker *worker) |
| 941 |
{ |
| 942 |
RegisteredBgWorker *rw; |
| 943 |
static int numworkers = 0; |
| 944 |
|
| 945 |
/* |
| 946 |
* Static background workers can only be registered in the postmaster |
| 947 |
* process. |
| 948 |
*/ |
| 949 |
if (IsUnderPostmaster || !IsPostmasterEnvironment) |
| 950 |
{ |
| 951 |
/* |
| 952 |
* In EXEC_BACKEND or single-user mode, we process |
| 953 |
* shared_preload_libraries in backend processes too. We cannot |
| 954 |
* register static background workers at that stage, but many |
| 955 |
* libraries' _PG_init() functions don't distinguish whether they're |
| 956 |
* being loaded in the postmaster or in a backend, they just check |
| 957 |
* process_shared_preload_libraries_in_progress. It's a bit sloppy, |
| 958 |
* but for historical reasons we tolerate it. In EXEC_BACKEND mode, |
| 959 |
* the background workers should already have been registered when the |
| 960 |
* library was loaded in postmaster. |
| 961 |
*/ |
| 962 |
if (process_shared_preload_libraries_in_progress) |
| 963 |
return; |
| 964 |
ereport(LOG, |
| 965 |
(errcode(ERRCODE_FEATURE_NOT_SUPPORTED), |
| 966 |
errmsg("background worker \"%s\": must be registered in \"shared_preload_libraries\"", |
| 967 |
worker->bgw_name))); |
| 968 |
return; |
| 969 |
} |
| 970 |
|
| 971 |
/* |
| 972 |
* Cannot register static background workers after calling |
| 973 |
* BackgroundWorkerShmemInit(). |
| 974 |
*/ |
| 975 |
if (BackgroundWorkerData != NULL) |
| 976 |
elog(ERROR, "cannot register background worker \"%s\" after shmem init", |
| 977 |
worker->bgw_name); |
| 978 |
|
| 979 |
ereport(DEBUG1, |
| 980 |
(errmsg_internal("registering background worker \"%s\"", worker->bgw_name))); |
| 981 |
|
| 982 |
if (!SanityCheckBackgroundWorker(worker, LOG)) |
| 983 |
return; |
| 984 |
|
| 985 |
if (worker->bgw_notify_pid != 0) |
| 986 |
{ |
| 987 |
ereport(LOG, |
| 988 |
(errcode(ERRCODE_FEATURE_NOT_SUPPORTED), |
| 989 |
errmsg("background worker \"%s\": only dynamic background workers can request notification", |
| 990 |
worker->bgw_name))); |
| 991 |
return; |
| 992 |
} |
| 993 |
|
| 994 |
/* |
| 995 |
* Enforce maximum number of workers. Note this is overly restrictive: we |
| 996 |
* could allow more non-shmem-connected workers, because these don't count |
| 997 |
* towards the MAX_BACKENDS limit elsewhere. For now, it doesn't seem |
| 998 |
* important to relax this restriction. |
| 999 |
*/ |
| 1000 |
if (++numworkers > max_worker_processes) |
| 1001 |
{ |
| 1002 |
ereport(LOG, |
| 1003 |
(errcode(ERRCODE_CONFIGURATION_LIMIT_EXCEEDED), |
| 1004 |
errmsg("too many background workers"), |
| 1005 |
errdetail_plural("Up to %d background worker can be registered with the current settings.", |
| 1006 |
"Up to %d background workers can be registered with the current settings.", |
| 1007 |
max_worker_processes, |
| 1008 |
max_worker_processes), |
| 1009 |
errhint("Consider increasing the configuration parameter \"%s\".", "max_worker_processes"))); |
| 1010 |
return; |
| 1011 |
} |
| 1012 |
|
| 1013 |
/* |
| 1014 |
* Copy the registration data into the registered workers list. |
| 1015 |
*/ |
| 1016 |
rw = MemoryContextAllocExtended(PostmasterContext, |
| 1017 |
sizeof(RegisteredBgWorker), |
| 1018 |
MCXT_ALLOC_NO_OOM); |
| 1019 |
if (rw == NULL) |
| 1020 |
{ |
| 1021 |
ereport(LOG, |
| 1022 |
(errcode(ERRCODE_OUT_OF_MEMORY), |
| 1023 |
errmsg("out of memory"))); |
| 1024 |
return; |
| 1025 |
} |
| 1026 |
|
| 1027 |
rw->rw_worker = *worker; |
| 1028 |
rw->rw_pid = 0; |
| 1029 |
rw->rw_crashed_at = 0; |
| 1030 |
rw->rw_terminate = false; |
| 1031 |
|
| 1032 |
dlist_push_head(&BackgroundWorkerList, &rw->rw_lnode); |
| 1033 |
} |
| 1034 |
|
| 1035 |
/* |
| 1036 |
* Register a new background worker from a regular backend. |
| 1037 |
* |
| 1038 |
* Returns true on success and false on failure. Failure typically indicates |
| 1039 |
* that no background worker slots are currently available. |
| 1040 |
* |
| 1041 |
* If handle != NULL, we'll set *handle to a pointer that can subsequently |
| 1042 |
* be used as an argument to GetBackgroundWorkerPid(). The caller can |
| 1043 |
* free this pointer using pfree(), if desired. |
| 1044 |
*/ |
| 1045 |
bool |
| 1046 |
RegisterDynamicBackgroundWorker(BackgroundWorker *worker, |
| 1047 |
BackgroundWorkerHandle **handle) |
| 1048 |
{ |
| 1049 |
int slotno; |
| 1050 |
bool success = false; |
| 1051 |
bool parallel; |
| 1052 |
uint64 generation = 0; |
| 1053 |
|
| 1054 |
/* |
| 1055 |
* We can't register dynamic background workers from the postmaster. If |
| 1056 |
* this is a standalone backend, we're the only process and can't start |
| 1057 |
* any more. In a multi-process environment, it might be theoretically |
| 1058 |
* possible, but we don't currently support it due to locking |
| 1059 |
* considerations; see comments on the BackgroundWorkerSlot data |
| 1060 |
* structure. |
| 1061 |
*/ |
| 1062 |
if (!IsUnderPostmaster) |
| 1063 |
return false; |
| 1064 |
|
| 1065 |
if (!SanityCheckBackgroundWorker(worker, ERROR)) |
| 1066 |
return false; |
| 1067 |
|
| 1068 |
parallel = (worker->bgw_flags & BGWORKER_CLASS_PARALLEL) != 0; |
| 1069 |
|
| 1070 |
LWLockAcquire(BackgroundWorkerLock, LW_EXCLUSIVE); |
| 1071 |
|
| 1072 |
/* |
| 1073 |
* If this is a parallel worker, check whether there are already too many |
| 1074 |
* parallel workers; if so, don't register another one. Our view of |
| 1075 |
* parallel_terminate_count may be slightly stale, but that doesn't really |
| 1076 |
* matter: we would have gotten the same result if we'd arrived here |
| 1077 |
* slightly earlier anyway. There's no help for it, either, since the |
| 1078 |
* postmaster must not take locks; a memory barrier wouldn't guarantee |
| 1079 |
* anything useful. |
| 1080 |
*/ |
| 1081 |
if (parallel && (BackgroundWorkerData->parallel_register_count - |
| 1082 |
BackgroundWorkerData->parallel_terminate_count) >= |
| 1083 |
max_parallel_workers) |
| 1084 |
{ |
| 1085 |
Assert(BackgroundWorkerData->parallel_register_count - |
| 1086 |
BackgroundWorkerData->parallel_terminate_count <= |
| 1087 |
MAX_PARALLEL_WORKER_LIMIT); |
| 1088 |
LWLockRelease(BackgroundWorkerLock); |
| 1089 |
return false; |
| 1090 |
} |
| 1091 |
|
| 1092 |
/* |
| 1093 |
* Look for an unused slot. If we find one, grab it. |
| 1094 |
*/ |
| 1095 |
for (slotno = 0; slotno < BackgroundWorkerData->total_slots; ++slotno) |
| 1096 |
{ |
| 1097 |
BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno]; |
| 1098 |
|
| 1099 |
if (!slot->in_use) |
| 1100 |
{ |
| 1101 |
memcpy(&slot->worker, worker, sizeof(BackgroundWorker)); |
| 1102 |
slot->pid = InvalidPid; /* indicates not started yet */ |
| 1103 |
slot->generation++; |
| 1104 |
slot->terminate = false; |
| 1105 |
generation = slot->generation; |
| 1106 |
if (parallel) |
| 1107 |
BackgroundWorkerData->parallel_register_count++; |
| 1108 |
|
| 1109 |
/* |
| 1110 |
* Make sure postmaster doesn't see the slot as in use before it |
| 1111 |
* sees the new contents. |
| 1112 |
*/ |
| 1113 |
pg_write_barrier(); |
| 1114 |
|
| 1115 |
slot->in_use = true; |
| 1116 |
success = true; |
| 1117 |
break; |
| 1118 |
} |
| 1119 |
} |
| 1120 |
|
| 1121 |
LWLockRelease(BackgroundWorkerLock); |
| 1122 |
|
| 1123 |
/* If we found a slot, tell the postmaster to notice the change. */ |
| 1124 |
if (success) |
| 1125 |
SendPostmasterSignal(PMSIGNAL_BACKGROUND_WORKER_CHANGE); |
| 1126 |
|
| 1127 |
/* |
| 1128 |
* If we found a slot and the user has provided a handle, initialize it. |
| 1129 |
*/ |
| 1130 |
if (success && handle) |
| 1131 |
{ |
| 1132 |
*handle = palloc(sizeof(BackgroundWorkerHandle)); |
| 1133 |
(*handle)->slot = slotno; |
| 1134 |
(*handle)->generation = generation; |
| 1135 |
} |
| 1136 |
|
| 1137 |
return success; |
| 1138 |
} |
| 1139 |
|
| 1140 |
/* |
| 1141 |
* Get the PID of a dynamically-registered background worker. |
| 1142 |
* |
| 1143 |
* If the worker is determined to be running, the return value will be |
| 1144 |
* BGWH_STARTED and *pidp will get the PID of the worker process. If the |
| 1145 |
* postmaster has not yet attempted to start the worker, the return value will |
| 1146 |
* be BGWH_NOT_YET_STARTED. Otherwise, the return value is BGWH_STOPPED. |
| 1147 |
* |
| 1148 |
* BGWH_STOPPED can indicate either that the worker is temporarily stopped |
| 1149 |
* (because it is configured for automatic restart and exited non-zero), |
| 1150 |
* or that the worker is permanently stopped (because it exited with exit |
| 1151 |
* code 0, or was not configured for automatic restart), or even that the |
| 1152 |
* worker was unregistered without ever starting (either because startup |
| 1153 |
* failed and the worker is not configured for automatic restart, or because |
| 1154 |
* TerminateBackgroundWorker was used before the worker was successfully |
| 1155 |
* started). |
| 1156 |
*/ |
| 1157 |
BgwHandleStatus |
| 1158 |
GetBackgroundWorkerPid(BackgroundWorkerHandle *handle, pid_t *pidp) |
| 1159 |
{ |
| 1160 |
BackgroundWorkerSlot *slot; |
| 1161 |
pid_t pid; |
| 1162 |
|
| 1163 |
Assert(handle->slot < max_worker_processes); |
| 1164 |
slot = &BackgroundWorkerData->slot[handle->slot]; |
| 1165 |
|
| 1166 |
/* |
| 1167 |
* We could probably arrange to synchronize access to data using memory |
| 1168 |
* barriers only, but for now, let's just keep it simple and grab the |
| 1169 |
* lock. It seems unlikely that there will be enough traffic here to |
| 1170 |
* result in meaningful contention. |
| 1171 |
*/ |
| 1172 |
LWLockAcquire(BackgroundWorkerLock, LW_SHARED); |
| 1173 |
|
| 1174 |
/* |
| 1175 |
* The generation number can't be concurrently changed while we hold the |
| 1176 |
* lock. The pid, which is updated by the postmaster, can change at any |
| 1177 |
* time, but we assume such changes are atomic. So the value we read |
| 1178 |
* won't be garbage, but it might be out of date by the time the caller |
| 1179 |
* examines it (but that's unavoidable anyway). |
| 1180 |
* |
| 1181 |
* The in_use flag could be in the process of changing from true to false, |
| 1182 |
* but if it is already false then it can't change further. |
| 1183 |
*/ |
| 1184 |
if (handle->generation != slot->generation || !slot->in_use) |
| 1185 |
pid = 0; |
| 1186 |
else |
| 1187 |
pid = slot->pid; |
| 1188 |
|
| 1189 |
/* All done. */ |
| 1190 |
LWLockRelease(BackgroundWorkerLock); |
| 1191 |
|
| 1192 |
if (pid == 0) |
| 1193 |
return BGWH_STOPPED; |
| 1194 |
else if (pid == InvalidPid) |
| 1195 |
return BGWH_NOT_YET_STARTED; |
| 1196 |
*pidp = pid; |
| 1197 |
return BGWH_STARTED; |
| 1198 |
} |
| 1199 |
|
| 1200 |
/* |
| 1201 |
* Wait for a background worker to start up. |
| 1202 |
* |
| 1203 |
* This is like GetBackgroundWorkerPid(), except that if the worker has not |
| 1204 |
* yet started, we wait for it to do so; thus, BGWH_NOT_YET_STARTED is never |
| 1205 |
* returned. However, if the postmaster has died, we give up and return |
| 1206 |
* BGWH_POSTMASTER_DIED, since it that case we know that startup will not |
| 1207 |
* take place. |
| 1208 |
* |
| 1209 |
* The caller *must* have set our PID as the worker's bgw_notify_pid, |
| 1210 |
* else we will not be awoken promptly when the worker's state changes. |
| 1211 |
*/ |
| 1212 |
BgwHandleStatus |
| 1213 |
WaitForBackgroundWorkerStartup(BackgroundWorkerHandle *handle, pid_t *pidp) |
| 1214 |
{ |
| 1215 |
BgwHandleStatus status; |
| 1216 |
int rc; |
| 1217 |
|
| 1218 |
for (;;) |
| 1219 |
{ |
| 1220 |
pid_t pid; |
| 1221 |
|
| 1222 |
CHECK_FOR_INTERRUPTS(); |
| 1223 |
|
| 1224 |
status = GetBackgroundWorkerPid(handle, &pid); |
| 1225 |
if (status == BGWH_STARTED) |
| 1226 |
*pidp = pid; |
| 1227 |
if (status != BGWH_NOT_YET_STARTED) |
| 1228 |
break; |
| 1229 |
|
| 1230 |
rc = WaitLatch(MyLatch, |
| 1231 |
WL_LATCH_SET | WL_POSTMASTER_DEATH, 0, |
| 1232 |
WAIT_EVENT_BGWORKER_STARTUP); |
| 1233 |
|
| 1234 |
if (rc & WL_POSTMASTER_DEATH) |
| 1235 |
{ |
| 1236 |
status = BGWH_POSTMASTER_DIED; |
| 1237 |
break; |
| 1238 |
} |
| 1239 |
|
| 1240 |
ResetLatch(MyLatch); |
| 1241 |
} |
| 1242 |
|
| 1243 |
return status; |
| 1244 |
} |
| 1245 |
|
| 1246 |
/* |
| 1247 |
* Wait for a background worker to stop. |
| 1248 |
* |
| 1249 |
* If the worker hasn't yet started, or is running, we wait for it to stop |
| 1250 |
* and then return BGWH_STOPPED. However, if the postmaster has died, we give |
| 1251 |
* up and return BGWH_POSTMASTER_DIED, because it's the postmaster that |
| 1252 |
* notifies us when a worker's state changes. |
| 1253 |
* |
| 1254 |
* The caller *must* have set our PID as the worker's bgw_notify_pid, |
| 1255 |
* else we will not be awoken promptly when the worker's state changes. |
| 1256 |
*/ |
| 1257 |
BgwHandleStatus |
| 1258 |
WaitForBackgroundWorkerShutdown(BackgroundWorkerHandle *handle) |
| 1259 |
{ |
| 1260 |
BgwHandleStatus status; |
| 1261 |
int rc; |
| 1262 |
|
| 1263 |
for (;;) |
| 1264 |
{ |
| 1265 |
pid_t pid; |
| 1266 |
|
| 1267 |
CHECK_FOR_INTERRUPTS(); |
| 1268 |
|
| 1269 |
status = GetBackgroundWorkerPid(handle, &pid); |
| 1270 |
if (status == BGWH_STOPPED) |
| 1271 |
break; |
| 1272 |
|
| 1273 |
rc = WaitLatch(MyLatch, |
| 1274 |
WL_LATCH_SET | WL_POSTMASTER_DEATH, 0, |
| 1275 |
WAIT_EVENT_BGWORKER_SHUTDOWN); |
| 1276 |
|
| 1277 |
if (rc & WL_POSTMASTER_DEATH) |
| 1278 |
{ |
| 1279 |
status = BGWH_POSTMASTER_DIED; |
| 1280 |
break; |
| 1281 |
} |
| 1282 |
|
| 1283 |
ResetLatch(MyLatch); |
| 1284 |
} |
| 1285 |
|
| 1286 |
return status; |
| 1287 |
} |
| 1288 |
|
| 1289 |
/* |
| 1290 |
* Instruct the postmaster to terminate a background worker. |
| 1291 |
* |
| 1292 |
* Note that it's safe to do this without regard to whether the worker is |
| 1293 |
* still running, or even if the worker may already have exited and been |
| 1294 |
* unregistered. |
| 1295 |
*/ |
| 1296 |
void |
| 1297 |
TerminateBackgroundWorker(BackgroundWorkerHandle *handle) |
| 1298 |
{ |
| 1299 |
BackgroundWorkerSlot *slot; |
| 1300 |
bool signal_postmaster = false; |
| 1301 |
|
| 1302 |
Assert(handle->slot < max_worker_processes); |
| 1303 |
slot = &BackgroundWorkerData->slot[handle->slot]; |
| 1304 |
|
| 1305 |
/* Set terminate flag in shared memory, unless slot has been reused. */ |
| 1306 |
LWLockAcquire(BackgroundWorkerLock, LW_EXCLUSIVE); |
| 1307 |
if (handle->generation == slot->generation) |
| 1308 |
{ |
| 1309 |
slot->terminate = true; |
| 1310 |
signal_postmaster = true; |
| 1311 |
} |
| 1312 |
LWLockRelease(BackgroundWorkerLock); |
| 1313 |
|
| 1314 |
/* Make sure the postmaster notices the change to shared memory. */ |
| 1315 |
if (signal_postmaster) |
| 1316 |
SendPostmasterSignal(PMSIGNAL_BACKGROUND_WORKER_CHANGE); |
| 1317 |
} |
| 1318 |
|
| 1319 |
/* |
| 1320 |
* Look up (and possibly load) a bgworker entry point function. |
| 1321 |
* |
| 1322 |
* For functions contained in the core code, we use library name "postgres" |
| 1323 |
* and consult the InternalBGWorkers array. External functions are |
| 1324 |
* looked up, and loaded if necessary, using load_external_function(). |
| 1325 |
* |
| 1326 |
* The point of this is to pass function names as strings across process |
| 1327 |
* boundaries. We can't pass actual function addresses because of the |
| 1328 |
* possibility that the function has been loaded at a different address |
| 1329 |
* in a different process. This is obviously a hazard for functions in |
| 1330 |
* loadable libraries, but it can happen even for functions in the core code |
| 1331 |
* on platforms using EXEC_BACKEND (e.g., Windows). |
| 1332 |
* |
| 1333 |
* At some point it might be worthwhile to get rid of InternalBGWorkers[] |
| 1334 |
* in favor of applying load_external_function() for core functions too; |
| 1335 |
* but that raises portability issues that are not worth addressing now. |
| 1336 |
*/ |
| 1337 |
static bgworker_main_type |
| 1338 |
LookupBackgroundWorkerFunction(const char *libraryname, const char *funcname) |
| 1339 |
{ |
| 1340 |
/* |
| 1341 |
* If the function is to be loaded from postgres itself, search the |
| 1342 |
* InternalBGWorkers array. |
| 1343 |
*/ |
| 1344 |
if (strcmp(libraryname, "postgres") == 0) |
| 1345 |
{ |
| 1346 |
int i; |
| 1347 |
|
| 1348 |
for (i = 0; i < lengthof(InternalBGWorkers); i++) |
| 1349 |
{ |
| 1350 |
if (strcmp(InternalBGWorkers[i].fn_name, funcname) == 0) |
| 1351 |
return InternalBGWorkers[i].fn_addr; |
| 1352 |
} |
| 1353 |
|
| 1354 |
/* We can only reach this by programming error. */ |
| 1355 |
elog(ERROR, "internal function \"%s\" not found", funcname); |
| 1356 |
} |
| 1357 |
|
| 1358 |
/* Otherwise load from external library. */ |
| 1359 |
return (bgworker_main_type) |
| 1360 |
load_external_function(libraryname, funcname, true, NULL); |
| 1361 |
} |
| 1362 |
|
| 1363 |
/* |
| 1364 |
* Given a PID, get the bgw_type of the background worker. Returns NULL if |
| 1365 |
* not a valid background worker. |
| 1366 |
* |
| 1367 |
* The return value is in static memory belonging to this function, so it has |
| 1368 |
* to be used before calling this function again. This is so that the caller |
| 1369 |
* doesn't have to worry about the background worker locking protocol. |
| 1370 |
*/ |
| 1371 |
const char * |
| 1372 |
GetBackgroundWorkerTypeByPid(pid_t pid) |
| 1373 |
{ |
| 1374 |
int slotno; |
| 1375 |
bool found = false; |
| 1376 |
static char result[BGW_MAXLEN]; |
| 1377 |
|
| 1378 |
LWLockAcquire(BackgroundWorkerLock, LW_SHARED); |
| 1379 |
|
| 1380 |
for (slotno = 0; slotno < BackgroundWorkerData->total_slots; slotno++) |
| 1381 |
{ |
| 1382 |
BackgroundWorkerSlot *slot = &BackgroundWorkerData->slot[slotno]; |
| 1383 |
|
| 1384 |
if (slot->pid > 0 && slot->pid == pid) |
| 1385 |
{ |
| 1386 |
strcpy(result, slot->worker.bgw_type); |
| 1387 |
found = true; |
| 1388 |
break; |
| 1389 |
} |
| 1390 |
} |
| 1391 |
|
| 1392 |
LWLockRelease(BackgroundWorkerLock); |
| 1393 |
|
| 1394 |
if (!found) |
| 1395 |
return NULL; |
| 1396 |
|
| 1397 |
return result; |
| 1398 |
} |