| 1 |
/*------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* procarray.c |
| 4 |
* POSTGRES process array code. |
| 5 |
* |
| 6 |
* |
| 7 |
* This module maintains arrays of PGPROC substructures, as well as associated |
| 8 |
* arrays in ProcGlobal, for all active backends. Although there are several |
| 9 |
* uses for this, the principal one is as a means of determining the set of |
| 10 |
* currently running transactions. |
| 11 |
* |
| 12 |
* Because of various subtle race conditions it is critical that a backend |
| 13 |
* hold the correct locks while setting or clearing its xid (in |
| 14 |
* ProcGlobal->xids[]/MyProc->xid). See notes in |
| 15 |
* src/backend/access/transam/README. |
| 16 |
* |
| 17 |
* The process arrays now also include structures representing prepared |
| 18 |
* transactions. The xid and subxids fields of these are valid, as are the |
| 19 |
* myProcLocks lists. They can be distinguished from regular backend PGPROCs |
| 20 |
* at need by checking for pid == 0. |
| 21 |
* |
| 22 |
* During hot standby, we also keep a list of XIDs representing transactions |
| 23 |
* that are known to be running on the primary (or more precisely, were running |
| 24 |
* as of the current point in the WAL stream). This list is kept in the |
| 25 |
* KnownAssignedXids array, and is updated by watching the sequence of |
| 26 |
* arriving XIDs. This is necessary because if we leave those XIDs out of |
| 27 |
* snapshots taken for standby queries, then they will appear to be already |
| 28 |
* complete, leading to MVCC failures. Note that in hot standby, the PGPROC |
| 29 |
* array represents standby processes, which by definition are not running |
| 30 |
* transactions that have XIDs. |
| 31 |
* |
| 32 |
* It is perhaps possible for a backend on the primary to terminate without |
| 33 |
* writing an abort record for its transaction. While that shouldn't really |
| 34 |
* happen, it would tie up KnownAssignedXids indefinitely, so we protect |
| 35 |
* ourselves by pruning the array when a valid list of running XIDs arrives. |
| 36 |
* |
| 37 |
* Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group |
| 38 |
* Portions Copyright (c) 1994, Regents of the University of California |
| 39 |
* |
| 40 |
* |
| 41 |
* IDENTIFICATION |
| 42 |
* src/backend/storage/ipc/procarray.c |
| 43 |
* |
| 44 |
*------------------------------------------------------------------------- |
| 45 |
*/ |
| 46 |
#include "postgres.h" |
| 47 |
|
| 48 |
#include <signal.h> |
| 49 |
|
| 50 |
#include "access/subtrans.h" |
| 51 |
#include "access/transam.h" |
| 52 |
#include "access/twophase.h" |
| 53 |
#include "access/xact.h" |
| 54 |
#include "access/xlogutils.h" |
| 55 |
#include "catalog/catalog.h" |
| 56 |
#include "catalog/pg_authid.h" |
| 57 |
#include "commands/dbcommands.h" |
| 58 |
#include "miscadmin.h" |
| 59 |
#include "pgstat.h" |
| 60 |
#include "port/pg_lfind.h" |
| 61 |
#include "storage/proc.h" |
| 62 |
#include "storage/procarray.h" |
| 63 |
#include "utils/acl.h" |
| 64 |
#include "utils/builtins.h" |
| 65 |
#include "utils/rel.h" |
| 66 |
#include "utils/snapmgr.h" |
| 67 |
|
| 68 |
#define UINT32_ACCESS_ONCE(var) ((uint32)(*((volatile uint32 *)&(var)))) |
| 69 |
|
| 70 |
/* Our shared memory area */ |
| 71 |
typedef struct ProcArrayStruct |
| 72 |
{ |
| 73 |
int numProcs; /* number of valid procs entries */ |
| 74 |
int maxProcs; /* allocated size of procs array */ |
| 75 |
|
| 76 |
/* |
| 77 |
* Known assigned XIDs handling |
| 78 |
*/ |
| 79 |
int maxKnownAssignedXids; /* allocated size of array */ |
| 80 |
int numKnownAssignedXids; /* current # of valid entries */ |
| 81 |
int tailKnownAssignedXids; /* index of oldest valid element */ |
| 82 |
int headKnownAssignedXids; /* index of newest element, + 1 */ |
| 83 |
|
| 84 |
/* |
| 85 |
* Highest subxid that has been removed from KnownAssignedXids array to |
| 86 |
* prevent overflow; or InvalidTransactionId if none. We track this for |
| 87 |
* similar reasons to tracking overflowing cached subxids in PGPROC |
| 88 |
* entries. Must hold exclusive ProcArrayLock to change this, and shared |
| 89 |
* lock to read it. |
| 90 |
*/ |
| 91 |
TransactionId lastOverflowedXid; |
| 92 |
|
| 93 |
/* oldest xmin of any replication slot */ |
| 94 |
TransactionId replication_slot_xmin; |
| 95 |
/* oldest catalog xmin of any replication slot */ |
| 96 |
TransactionId replication_slot_catalog_xmin; |
| 97 |
|
| 98 |
/* indexes into allProcs[], has PROCARRAY_MAXPROCS entries */ |
| 99 |
int pgprocnos[FLEXIBLE_ARRAY_MEMBER]; |
| 100 |
} ProcArrayStruct; |
| 101 |
|
| 102 |
/* |
| 103 |
* State for the GlobalVisTest* family of functions. Those functions can |
| 104 |
* e.g. be used to decide if a deleted row can be removed without violating |
| 105 |
* MVCC semantics: If the deleted row's xmax is not considered to be running |
| 106 |
* by anyone, the row can be removed. |
| 107 |
* |
| 108 |
* To avoid slowing down GetSnapshotData(), we don't calculate a precise |
| 109 |
* cutoff XID while building a snapshot (looking at the frequently changing |
| 110 |
* xmins scales badly). Instead we compute two boundaries while building the |
| 111 |
* snapshot: |
| 112 |
* |
| 113 |
* 1) definitely_needed, indicating that rows deleted by XIDs >= |
| 114 |
* definitely_needed are definitely still visible. |
| 115 |
* |
| 116 |
* 2) maybe_needed, indicating that rows deleted by XIDs < maybe_needed can |
| 117 |
* definitely be removed |
| 118 |
* |
| 119 |
* When testing an XID that falls in between the two (i.e. XID >= maybe_needed |
| 120 |
* && XID < definitely_needed), the boundaries can be recomputed (using |
| 121 |
* ComputeXidHorizons()) to get a more accurate answer. This is cheaper than |
| 122 |
* maintaining an accurate value all the time. |
| 123 |
* |
| 124 |
* As it is not cheap to compute accurate boundaries, we limit the number of |
| 125 |
* times that happens in short succession. See GlobalVisTestShouldUpdate(). |
| 126 |
* |
| 127 |
* |
| 128 |
* There are three backend lifetime instances of this struct, optimized for |
| 129 |
* different types of relations. As e.g. a normal user defined table in one |
| 130 |
* database is inaccessible to backends connected to another database, a test |
| 131 |
* specific to a relation can be more aggressive than a test for a shared |
| 132 |
* relation. Currently we track four different states: |
| 133 |
* |
| 134 |
* 1) GlobalVisSharedRels, which only considers an XID's |
| 135 |
* effects visible-to-everyone if neither snapshots in any database, nor a |
| 136 |
* replication slot's xmin, nor a replication slot's catalog_xmin might |
| 137 |
* still consider XID as running. |
| 138 |
* |
| 139 |
* 2) GlobalVisCatalogRels, which only considers an XID's |
| 140 |
* effects visible-to-everyone if neither snapshots in the current |
| 141 |
* database, nor a replication slot's xmin, nor a replication slot's |
| 142 |
* catalog_xmin might still consider XID as running. |
| 143 |
* |
| 144 |
* I.e. the difference to GlobalVisSharedRels is that |
| 145 |
* snapshot in other databases are ignored. |
| 146 |
* |
| 147 |
* 3) GlobalVisDataRels, which only considers an XID's |
| 148 |
* effects visible-to-everyone if neither snapshots in the current |
| 149 |
* database, nor a replication slot's xmin consider XID as running. |
| 150 |
* |
| 151 |
* I.e. the difference to GlobalVisCatalogRels is that |
| 152 |
* replication slot's catalog_xmin is not taken into account. |
| 153 |
* |
| 154 |
* 4) GlobalVisTempRels, which only considers the current session, as temp |
| 155 |
* tables are not visible to other sessions. |
| 156 |
* |
| 157 |
* GlobalVisTestFor(relation) returns the appropriate state |
| 158 |
* for the relation. |
| 159 |
* |
| 160 |
* The boundaries are FullTransactionIds instead of TransactionIds to avoid |
| 161 |
* wraparound dangers. There e.g. would otherwise exist no procarray state to |
| 162 |
* prevent maybe_needed to become old enough after the GetSnapshotData() |
| 163 |
* call. |
| 164 |
* |
| 165 |
* The typedef is in the header. |
| 166 |
*/ |
| 167 |
struct GlobalVisState |
| 168 |
{ |
| 169 |
/* XIDs >= are considered running by some backend */ |
| 170 |
FullTransactionId definitely_needed; |
| 171 |
|
| 172 |
/* XIDs < are not considered to be running by any backend */ |
| 173 |
FullTransactionId maybe_needed; |
| 174 |
}; |
| 175 |
|
| 176 |
/* |
| 177 |
* Result of ComputeXidHorizons(). |
| 178 |
*/ |
| 179 |
typedef struct ComputeXidHorizonsResult |
| 180 |
{ |
| 181 |
/* |
| 182 |
* The value of TransamVariables->latestCompletedXid when |
| 183 |
* ComputeXidHorizons() held ProcArrayLock. |
| 184 |
*/ |
| 185 |
FullTransactionId latest_completed; |
| 186 |
|
| 187 |
/* |
| 188 |
* The same for procArray->replication_slot_xmin and |
| 189 |
* procArray->replication_slot_catalog_xmin. |
| 190 |
*/ |
| 191 |
TransactionId slot_xmin; |
| 192 |
TransactionId slot_catalog_xmin; |
| 193 |
|
| 194 |
/* |
| 195 |
* Oldest xid that any backend might still consider running. This needs to |
| 196 |
* include processes running VACUUM, in contrast to the normal visibility |
| 197 |
* cutoffs, as vacuum needs to be able to perform pg_subtrans lookups when |
| 198 |
* determining visibility, but doesn't care about rows above its xmin to |
| 199 |
* be removed. |
| 200 |
* |
| 201 |
* This likely should only be needed to determine whether pg_subtrans can |
| 202 |
* be truncated. It currently includes the effects of replication slots, |
| 203 |
* for historical reasons. But that could likely be changed. |
| 204 |
*/ |
| 205 |
TransactionId oldest_considered_running; |
| 206 |
|
| 207 |
/* |
| 208 |
* Oldest xid for which deleted tuples need to be retained in shared |
| 209 |
* tables. |
| 210 |
* |
| 211 |
* This includes the effects of replication slots. If that's not desired, |
| 212 |
* look at shared_oldest_nonremovable_raw; |
| 213 |
*/ |
| 214 |
TransactionId shared_oldest_nonremovable; |
| 215 |
|
| 216 |
/* |
| 217 |
* Oldest xid that may be necessary to retain in shared tables. This is |
| 218 |
* the same as shared_oldest_nonremovable, except that is not affected by |
| 219 |
* replication slot's catalog_xmin. |
| 220 |
* |
| 221 |
* This is mainly useful to be able to send the catalog_xmin to upstream |
| 222 |
* streaming replication servers via hot_standby_feedback, so they can |
| 223 |
* apply the limit only when accessing catalog tables. |
| 224 |
*/ |
| 225 |
TransactionId shared_oldest_nonremovable_raw; |
| 226 |
|
| 227 |
/* |
| 228 |
* Oldest xid for which deleted tuples need to be retained in non-shared |
| 229 |
* catalog tables. |
| 230 |
*/ |
| 231 |
TransactionId catalog_oldest_nonremovable; |
| 232 |
|
| 233 |
/* |
| 234 |
* Oldest xid for which deleted tuples need to be retained in normal user |
| 235 |
* defined tables. |
| 236 |
*/ |
| 237 |
TransactionId data_oldest_nonremovable; |
| 238 |
|
| 239 |
/* |
| 240 |
* Oldest xid for which deleted tuples need to be retained in this |
| 241 |
* session's temporary tables. |
| 242 |
*/ |
| 243 |
TransactionId temp_oldest_nonremovable; |
| 244 |
} ComputeXidHorizonsResult; |
| 245 |
|
| 246 |
/* |
| 247 |
* Return value for GlobalVisHorizonKindForRel(). |
| 248 |
*/ |
| 249 |
typedef enum GlobalVisHorizonKind |
| 250 |
{ |
| 251 |
VISHORIZON_SHARED, |
| 252 |
VISHORIZON_CATALOG, |
| 253 |
VISHORIZON_DATA, |
| 254 |
VISHORIZON_TEMP, |
| 255 |
} GlobalVisHorizonKind; |
| 256 |
|
| 257 |
/* |
| 258 |
* Reason codes for KnownAssignedXidsCompress(). |
| 259 |
*/ |
| 260 |
typedef enum KAXCompressReason |
| 261 |
{ |
| 262 |
KAX_NO_SPACE, /* need to free up space at array end */ |
| 263 |
KAX_PRUNE, /* we just pruned old entries */ |
| 264 |
KAX_TRANSACTION_END, /* we just committed/removed some XIDs */ |
| 265 |
KAX_STARTUP_PROCESS_IDLE, /* startup process is about to sleep */ |
| 266 |
} KAXCompressReason; |
| 267 |
|
| 268 |
|
| 269 |
static ProcArrayStruct *procArray; |
| 270 |
|
| 271 |
static PGPROC *allProcs; |
| 272 |
|
| 273 |
/* |
| 274 |
* Cache to reduce overhead of repeated calls to TransactionIdIsInProgress() |
| 275 |
*/ |
| 276 |
static TransactionId cachedXidIsNotInProgress = InvalidTransactionId; |
| 277 |
|
| 278 |
/* |
| 279 |
* Bookkeeping for tracking emulated transactions in recovery |
| 280 |
*/ |
| 281 |
static TransactionId *KnownAssignedXids; |
| 282 |
static bool *KnownAssignedXidsValid; |
| 283 |
static TransactionId latestObservedXid = InvalidTransactionId; |
| 284 |
|
| 285 |
/* |
| 286 |
* If we're in STANDBY_SNAPSHOT_PENDING state, standbySnapshotPendingXmin is |
| 287 |
* the highest xid that might still be running that we don't have in |
| 288 |
* KnownAssignedXids. |
| 289 |
*/ |
| 290 |
static TransactionId standbySnapshotPendingXmin; |
| 291 |
|
| 292 |
/* |
| 293 |
* State for visibility checks on different types of relations. See struct |
| 294 |
* GlobalVisState for details. As shared, catalog, normal and temporary |
| 295 |
* relations can have different horizons, one such state exists for each. |
| 296 |
*/ |
| 297 |
static GlobalVisState GlobalVisSharedRels; |
| 298 |
static GlobalVisState GlobalVisCatalogRels; |
| 299 |
static GlobalVisState GlobalVisDataRels; |
| 300 |
static GlobalVisState GlobalVisTempRels; |
| 301 |
|
| 302 |
/* |
| 303 |
* This backend's RecentXmin at the last time the accurate xmin horizon was |
| 304 |
* recomputed, or InvalidTransactionId if it has not. Used to limit how many |
| 305 |
* times accurate horizons are recomputed. See GlobalVisTestShouldUpdate(). |
| 306 |
*/ |
| 307 |
static TransactionId ComputeXidHorizonsResultLastXmin; |
| 308 |
|
| 309 |
#ifdef XIDCACHE_DEBUG |
| 310 |
|
| 311 |
/* counters for XidCache measurement */ |
| 312 |
static long xc_by_recent_xmin = 0; |
| 313 |
static long xc_by_known_xact = 0; |
| 314 |
static long xc_by_my_xact = 0; |
| 315 |
static long xc_by_latest_xid = 0; |
| 316 |
static long xc_by_main_xid = 0; |
| 317 |
static long xc_by_child_xid = 0; |
| 318 |
static long xc_by_known_assigned = 0; |
| 319 |
static long xc_no_overflow = 0; |
| 320 |
static long xc_slow_answer = 0; |
| 321 |
|
| 322 |
#define xc_by_recent_xmin_inc() (xc_by_recent_xmin++) |
| 323 |
#define xc_by_known_xact_inc() (xc_by_known_xact++) |
| 324 |
#define xc_by_my_xact_inc() (xc_by_my_xact++) |
| 325 |
#define xc_by_latest_xid_inc() (xc_by_latest_xid++) |
| 326 |
#define xc_by_main_xid_inc() (xc_by_main_xid++) |
| 327 |
#define xc_by_child_xid_inc() (xc_by_child_xid++) |
| 328 |
#define xc_by_known_assigned_inc() (xc_by_known_assigned++) |
| 329 |
#define xc_no_overflow_inc() (xc_no_overflow++) |
| 330 |
#define xc_slow_answer_inc() (xc_slow_answer++) |
| 331 |
|
| 332 |
static void DisplayXidCache(void); |
| 333 |
#else /* !XIDCACHE_DEBUG */ |
| 334 |
|
| 335 |
#define xc_by_recent_xmin_inc() ((void) 0) |
| 336 |
#define xc_by_known_xact_inc() ((void) 0) |
| 337 |
#define xc_by_my_xact_inc() ((void) 0) |
| 338 |
#define xc_by_latest_xid_inc() ((void) 0) |
| 339 |
#define xc_by_main_xid_inc() ((void) 0) |
| 340 |
#define xc_by_child_xid_inc() ((void) 0) |
| 341 |
#define xc_by_known_assigned_inc() ((void) 0) |
| 342 |
#define xc_no_overflow_inc() ((void) 0) |
| 343 |
#define xc_slow_answer_inc() ((void) 0) |
| 344 |
#endif /* XIDCACHE_DEBUG */ |
| 345 |
|
| 346 |
/* Primitives for KnownAssignedXids array handling for standby */ |
| 347 |
static void KnownAssignedXidsCompress(KAXCompressReason reason, bool haveLock); |
| 348 |
static void KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid, |
| 349 |
bool exclusive_lock); |
| 350 |
static bool KnownAssignedXidsSearch(TransactionId xid, bool remove); |
| 351 |
static bool KnownAssignedXidExists(TransactionId xid); |
| 352 |
static void KnownAssignedXidsRemove(TransactionId xid); |
| 353 |
static void KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids, |
| 354 |
TransactionId *subxids); |
| 355 |
static void KnownAssignedXidsRemovePreceding(TransactionId removeXid); |
| 356 |
static int KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax); |
| 357 |
static int KnownAssignedXidsGetAndSetXmin(TransactionId *xarray, |
| 358 |
TransactionId *xmin, |
| 359 |
TransactionId xmax); |
| 360 |
static TransactionId KnownAssignedXidsGetOldestXmin(void); |
| 361 |
static void KnownAssignedXidsDisplay(int trace_level); |
| 362 |
static void KnownAssignedXidsReset(void); |
| 363 |
static inline void ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid); |
| 364 |
static void ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid); |
| 365 |
static void MaintainLatestCompletedXid(TransactionId latestXid); |
| 366 |
static void MaintainLatestCompletedXidRecovery(TransactionId latestXid); |
| 367 |
|
| 368 |
static inline FullTransactionId FullXidRelativeTo(FullTransactionId rel, |
| 369 |
TransactionId xid); |
| 370 |
static void GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons); |
| 371 |
|
| 372 |
/* |
| 373 |
* Report shared-memory space needed by ProcArrayShmemInit |
| 374 |
*/ |
| 375 |
Size |
| 376 |
ProcArrayShmemSize(void) |
| 377 |
{ |
| 378 |
Size size; |
| 379 |
|
| 380 |
/* Size of the ProcArray structure itself */ |
| 381 |
#define PROCARRAY_MAXPROCS (MaxBackends + max_prepared_xacts) |
| 382 |
|
| 383 |
size = offsetof(ProcArrayStruct, pgprocnos); |
| 384 |
size = add_size(size, mul_size(sizeof(int), PROCARRAY_MAXPROCS)); |
| 385 |
|
| 386 |
/* |
| 387 |
* During Hot Standby processing we have a data structure called |
| 388 |
* KnownAssignedXids, created in shared memory. Local data structures are |
| 389 |
* also created in various backends during GetSnapshotData(), |
| 390 |
* TransactionIdIsInProgress() and GetRunningTransactionData(). All of the |
| 391 |
* main structures created in those functions must be identically sized, |
| 392 |
* since we may at times copy the whole of the data structures around. We |
| 393 |
* refer to this size as TOTAL_MAX_CACHED_SUBXIDS. |
| 394 |
* |
| 395 |
* Ideally we'd only create this structure if we were actually doing hot |
| 396 |
* standby in the current run, but we don't know that yet at the time |
| 397 |
* shared memory is being set up. |
| 398 |
*/ |
| 399 |
#define TOTAL_MAX_CACHED_SUBXIDS \ |
| 400 |
((PGPROC_MAX_CACHED_SUBXIDS + 1) * PROCARRAY_MAXPROCS) |
| 401 |
|
| 402 |
if (EnableHotStandby) |
| 403 |
{ |
| 404 |
size = add_size(size, |
| 405 |
mul_size(sizeof(TransactionId), |
| 406 |
TOTAL_MAX_CACHED_SUBXIDS)); |
| 407 |
size = add_size(size, |
| 408 |
mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS)); |
| 409 |
} |
| 410 |
|
| 411 |
return size; |
| 412 |
} |
| 413 |
|
| 414 |
/* |
| 415 |
* Initialize the shared PGPROC array during postmaster startup. |
| 416 |
*/ |
| 417 |
void |
| 418 |
ProcArrayShmemInit(void) |
| 419 |
{ |
| 420 |
bool found; |
| 421 |
|
| 422 |
/* Create or attach to the ProcArray shared structure */ |
| 423 |
procArray = (ProcArrayStruct *) |
| 424 |
ShmemInitStruct("Proc Array", |
| 425 |
add_size(offsetof(ProcArrayStruct, pgprocnos), |
| 426 |
mul_size(sizeof(int), |
| 427 |
PROCARRAY_MAXPROCS)), |
| 428 |
&found); |
| 429 |
|
| 430 |
if (!found) |
| 431 |
{ |
| 432 |
/* |
| 433 |
* We're the first - initialize. |
| 434 |
*/ |
| 435 |
procArray->numProcs = 0; |
| 436 |
procArray->maxProcs = PROCARRAY_MAXPROCS; |
| 437 |
procArray->maxKnownAssignedXids = TOTAL_MAX_CACHED_SUBXIDS; |
| 438 |
procArray->numKnownAssignedXids = 0; |
| 439 |
procArray->tailKnownAssignedXids = 0; |
| 440 |
procArray->headKnownAssignedXids = 0; |
| 441 |
procArray->lastOverflowedXid = InvalidTransactionId; |
| 442 |
procArray->replication_slot_xmin = InvalidTransactionId; |
| 443 |
procArray->replication_slot_catalog_xmin = InvalidTransactionId; |
| 444 |
TransamVariables->xactCompletionCount = 1; |
| 445 |
} |
| 446 |
|
| 447 |
allProcs = ProcGlobal->allProcs; |
| 448 |
|
| 449 |
/* Create or attach to the KnownAssignedXids arrays too, if needed */ |
| 450 |
if (EnableHotStandby) |
| 451 |
{ |
| 452 |
KnownAssignedXids = (TransactionId *) |
| 453 |
ShmemInitStruct("KnownAssignedXids", |
| 454 |
mul_size(sizeof(TransactionId), |
| 455 |
TOTAL_MAX_CACHED_SUBXIDS), |
| 456 |
&found); |
| 457 |
KnownAssignedXidsValid = (bool *) |
| 458 |
ShmemInitStruct("KnownAssignedXidsValid", |
| 459 |
mul_size(sizeof(bool), TOTAL_MAX_CACHED_SUBXIDS), |
| 460 |
&found); |
| 461 |
} |
| 462 |
} |
| 463 |
|
| 464 |
/* |
| 465 |
* Add the specified PGPROC to the shared array. |
| 466 |
*/ |
| 467 |
void |
| 468 |
ProcArrayAdd(PGPROC *proc) |
| 469 |
{ |
| 470 |
int pgprocno = GetNumberFromPGProc(proc); |
| 471 |
ProcArrayStruct *arrayP = procArray; |
| 472 |
int index; |
| 473 |
int movecount; |
| 474 |
|
| 475 |
/* See ProcGlobal comment explaining why both locks are held */ |
| 476 |
LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE); |
| 477 |
LWLockAcquire(XidGenLock, LW_EXCLUSIVE); |
| 478 |
|
| 479 |
if (arrayP->numProcs >= arrayP->maxProcs) |
| 480 |
{ |
| 481 |
/* |
| 482 |
* Oops, no room. (This really shouldn't happen, since there is a |
| 483 |
* fixed supply of PGPROC structs too, and so we should have failed |
| 484 |
* earlier.) |
| 485 |
*/ |
| 486 |
ereport(FATAL, |
| 487 |
(errcode(ERRCODE_TOO_MANY_CONNECTIONS), |
| 488 |
errmsg("sorry, too many clients already"))); |
| 489 |
} |
| 490 |
|
| 491 |
/* |
| 492 |
* Keep the procs array sorted by (PGPROC *) so that we can utilize |
| 493 |
* locality of references much better. This is useful while traversing the |
| 494 |
* ProcArray because there is an increased likelihood of finding the next |
| 495 |
* PGPROC structure in the cache. |
| 496 |
* |
| 497 |
* Since the occurrence of adding/removing a proc is much lower than the |
| 498 |
* access to the ProcArray itself, the overhead should be marginal |
| 499 |
*/ |
| 500 |
for (index = 0; index < arrayP->numProcs; index++) |
| 501 |
{ |
| 502 |
int this_procno = arrayP->pgprocnos[index]; |
| 503 |
|
| 504 |
Assert(this_procno >= 0 && this_procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS)); |
| 505 |
Assert(allProcs[this_procno].pgxactoff == index); |
| 506 |
|
| 507 |
/* If we have found our right position in the array, break */ |
| 508 |
if (this_procno > pgprocno) |
| 509 |
break; |
| 510 |
} |
| 511 |
|
| 512 |
movecount = arrayP->numProcs - index; |
| 513 |
memmove(&arrayP->pgprocnos[index + 1], |
| 514 |
&arrayP->pgprocnos[index], |
| 515 |
movecount * sizeof(*arrayP->pgprocnos)); |
| 516 |
memmove(&ProcGlobal->xids[index + 1], |
| 517 |
&ProcGlobal->xids[index], |
| 518 |
movecount * sizeof(*ProcGlobal->xids)); |
| 519 |
memmove(&ProcGlobal->subxidStates[index + 1], |
| 520 |
&ProcGlobal->subxidStates[index], |
| 521 |
movecount * sizeof(*ProcGlobal->subxidStates)); |
| 522 |
memmove(&ProcGlobal->statusFlags[index + 1], |
| 523 |
&ProcGlobal->statusFlags[index], |
| 524 |
movecount * sizeof(*ProcGlobal->statusFlags)); |
| 525 |
|
| 526 |
arrayP->pgprocnos[index] = GetNumberFromPGProc(proc); |
| 527 |
proc->pgxactoff = index; |
| 528 |
ProcGlobal->xids[index] = proc->xid; |
| 529 |
ProcGlobal->subxidStates[index] = proc->subxidStatus; |
| 530 |
ProcGlobal->statusFlags[index] = proc->statusFlags; |
| 531 |
|
| 532 |
arrayP->numProcs++; |
| 533 |
|
| 534 |
/* adjust pgxactoff for all following PGPROCs */ |
| 535 |
index++; |
| 536 |
for (; index < arrayP->numProcs; index++) |
| 537 |
{ |
| 538 |
int procno = arrayP->pgprocnos[index]; |
| 539 |
|
| 540 |
Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS)); |
| 541 |
Assert(allProcs[procno].pgxactoff == index - 1); |
| 542 |
|
| 543 |
allProcs[procno].pgxactoff = index; |
| 544 |
} |
| 545 |
|
| 546 |
/* |
| 547 |
* Release in reversed acquisition order, to reduce frequency of having to |
| 548 |
* wait for XidGenLock while holding ProcArrayLock. |
| 549 |
*/ |
| 550 |
LWLockRelease(XidGenLock); |
| 551 |
LWLockRelease(ProcArrayLock); |
| 552 |
} |
| 553 |
|
| 554 |
/* |
| 555 |
* Remove the specified PGPROC from the shared array. |
| 556 |
* |
| 557 |
* When latestXid is a valid XID, we are removing a live 2PC gxact from the |
| 558 |
* array, and thus causing it to appear as "not running" anymore. In this |
| 559 |
* case we must advance latestCompletedXid. (This is essentially the same |
| 560 |
* as ProcArrayEndTransaction followed by removal of the PGPROC, but we take |
| 561 |
* the ProcArrayLock only once, and don't damage the content of the PGPROC; |
| 562 |
* twophase.c depends on the latter.) |
| 563 |
*/ |
| 564 |
void |
| 565 |
ProcArrayRemove(PGPROC *proc, TransactionId latestXid) |
| 566 |
{ |
| 567 |
ProcArrayStruct *arrayP = procArray; |
| 568 |
int myoff; |
| 569 |
int movecount; |
| 570 |
|
| 571 |
#ifdef XIDCACHE_DEBUG |
| 572 |
/* dump stats at backend shutdown, but not prepared-xact end */ |
| 573 |
if (proc->pid != 0) |
| 574 |
DisplayXidCache(); |
| 575 |
#endif |
| 576 |
|
| 577 |
/* See ProcGlobal comment explaining why both locks are held */ |
| 578 |
LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE); |
| 579 |
LWLockAcquire(XidGenLock, LW_EXCLUSIVE); |
| 580 |
|
| 581 |
myoff = proc->pgxactoff; |
| 582 |
|
| 583 |
Assert(myoff >= 0 && myoff < arrayP->numProcs); |
| 584 |
Assert(ProcGlobal->allProcs[arrayP->pgprocnos[myoff]].pgxactoff == myoff); |
| 585 |
|
| 586 |
if (TransactionIdIsValid(latestXid)) |
| 587 |
{ |
| 588 |
Assert(TransactionIdIsValid(ProcGlobal->xids[myoff])); |
| 589 |
|
| 590 |
/* Advance global latestCompletedXid while holding the lock */ |
| 591 |
MaintainLatestCompletedXid(latestXid); |
| 592 |
|
| 593 |
/* Same with xactCompletionCount */ |
| 594 |
TransamVariables->xactCompletionCount++; |
| 595 |
|
| 596 |
ProcGlobal->xids[myoff] = InvalidTransactionId; |
| 597 |
ProcGlobal->subxidStates[myoff].overflowed = false; |
| 598 |
ProcGlobal->subxidStates[myoff].count = 0; |
| 599 |
} |
| 600 |
else |
| 601 |
{ |
| 602 |
/* Shouldn't be trying to remove a live transaction here */ |
| 603 |
Assert(!TransactionIdIsValid(ProcGlobal->xids[myoff])); |
| 604 |
} |
| 605 |
|
| 606 |
Assert(!TransactionIdIsValid(ProcGlobal->xids[myoff])); |
| 607 |
Assert(ProcGlobal->subxidStates[myoff].count == 0); |
| 608 |
Assert(ProcGlobal->subxidStates[myoff].overflowed == false); |
| 609 |
|
| 610 |
ProcGlobal->statusFlags[myoff] = 0; |
| 611 |
|
| 612 |
/* Keep the PGPROC array sorted. See notes above */ |
| 613 |
movecount = arrayP->numProcs - myoff - 1; |
| 614 |
memmove(&arrayP->pgprocnos[myoff], |
| 615 |
&arrayP->pgprocnos[myoff + 1], |
| 616 |
movecount * sizeof(*arrayP->pgprocnos)); |
| 617 |
memmove(&ProcGlobal->xids[myoff], |
| 618 |
&ProcGlobal->xids[myoff + 1], |
| 619 |
movecount * sizeof(*ProcGlobal->xids)); |
| 620 |
memmove(&ProcGlobal->subxidStates[myoff], |
| 621 |
&ProcGlobal->subxidStates[myoff + 1], |
| 622 |
movecount * sizeof(*ProcGlobal->subxidStates)); |
| 623 |
memmove(&ProcGlobal->statusFlags[myoff], |
| 624 |
&ProcGlobal->statusFlags[myoff + 1], |
| 625 |
movecount * sizeof(*ProcGlobal->statusFlags)); |
| 626 |
|
| 627 |
arrayP->pgprocnos[arrayP->numProcs - 1] = -1; /* for debugging */ |
| 628 |
arrayP->numProcs--; |
| 629 |
|
| 630 |
/* |
| 631 |
* Adjust pgxactoff of following procs for removed PGPROC (note that |
| 632 |
* numProcs already has been decremented). |
| 633 |
*/ |
| 634 |
for (int index = myoff; index < arrayP->numProcs; index++) |
| 635 |
{ |
| 636 |
int procno = arrayP->pgprocnos[index]; |
| 637 |
|
| 638 |
Assert(procno >= 0 && procno < (arrayP->maxProcs + NUM_AUXILIARY_PROCS)); |
| 639 |
Assert(allProcs[procno].pgxactoff - 1 == index); |
| 640 |
|
| 641 |
allProcs[procno].pgxactoff = index; |
| 642 |
} |
| 643 |
|
| 644 |
/* |
| 645 |
* Release in reversed acquisition order, to reduce frequency of having to |
| 646 |
* wait for XidGenLock while holding ProcArrayLock. |
| 647 |
*/ |
| 648 |
LWLockRelease(XidGenLock); |
| 649 |
LWLockRelease(ProcArrayLock); |
| 650 |
} |
| 651 |
|
| 652 |
|
| 653 |
/* |
| 654 |
* ProcArrayEndTransaction -- mark a transaction as no longer running |
| 655 |
* |
| 656 |
* This is used interchangeably for commit and abort cases. The transaction |
| 657 |
* commit/abort must already be reported to WAL and pg_xact. |
| 658 |
* |
| 659 |
* proc is currently always MyProc, but we pass it explicitly for flexibility. |
| 660 |
* latestXid is the latest Xid among the transaction's main XID and |
| 661 |
* subtransactions, or InvalidTransactionId if it has no XID. (We must ask |
| 662 |
* the caller to pass latestXid, instead of computing it from the PGPROC's |
| 663 |
* contents, because the subxid information in the PGPROC might be |
| 664 |
* incomplete.) |
| 665 |
*/ |
| 666 |
void |
| 667 |
ProcArrayEndTransaction(PGPROC *proc, TransactionId latestXid) |
| 668 |
{ |
| 669 |
if (TransactionIdIsValid(latestXid)) |
| 670 |
{ |
| 671 |
/* |
| 672 |
* We must lock ProcArrayLock while clearing our advertised XID, so |
| 673 |
* that we do not exit the set of "running" transactions while someone |
| 674 |
* else is taking a snapshot. See discussion in |
| 675 |
* src/backend/access/transam/README. |
| 676 |
*/ |
| 677 |
Assert(TransactionIdIsValid(proc->xid)); |
| 678 |
|
| 679 |
/* |
| 680 |
* If we can immediately acquire ProcArrayLock, we clear our own XID |
| 681 |
* and release the lock. If not, use group XID clearing to improve |
| 682 |
* efficiency. |
| 683 |
*/ |
| 684 |
if (LWLockConditionalAcquire(ProcArrayLock, LW_EXCLUSIVE)) |
| 685 |
{ |
| 686 |
ProcArrayEndTransactionInternal(proc, latestXid); |
| 687 |
LWLockRelease(ProcArrayLock); |
| 688 |
} |
| 689 |
else |
| 690 |
ProcArrayGroupClearXid(proc, latestXid); |
| 691 |
} |
| 692 |
else |
| 693 |
{ |
| 694 |
/* |
| 695 |
* If we have no XID, we don't need to lock, since we won't affect |
| 696 |
* anyone else's calculation of a snapshot. We might change their |
| 697 |
* estimate of global xmin, but that's OK. |
| 698 |
*/ |
| 699 |
Assert(!TransactionIdIsValid(proc->xid)); |
| 700 |
Assert(proc->subxidStatus.count == 0); |
| 701 |
Assert(!proc->subxidStatus.overflowed); |
| 702 |
|
| 703 |
proc->vxid.lxid = InvalidLocalTransactionId; |
| 704 |
proc->xmin = InvalidTransactionId; |
| 705 |
|
| 706 |
/* be sure this is cleared in abort */ |
| 707 |
proc->delayChkptFlags = 0; |
| 708 |
|
| 709 |
proc->recoveryConflictPending = false; |
| 710 |
|
| 711 |
/* must be cleared with xid/xmin: */ |
| 712 |
/* avoid unnecessarily dirtying shared cachelines */ |
| 713 |
if (proc->statusFlags & PROC_VACUUM_STATE_MASK) |
| 714 |
{ |
| 715 |
Assert(!LWLockHeldByMe(ProcArrayLock)); |
| 716 |
LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE); |
| 717 |
Assert(proc->statusFlags == ProcGlobal->statusFlags[proc->pgxactoff]); |
| 718 |
proc->statusFlags &= ~PROC_VACUUM_STATE_MASK; |
| 719 |
ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags; |
| 720 |
LWLockRelease(ProcArrayLock); |
| 721 |
} |
| 722 |
} |
| 723 |
} |
| 724 |
|
| 725 |
/* |
| 726 |
* Mark a write transaction as no longer running. |
| 727 |
* |
| 728 |
* We don't do any locking here; caller must handle that. |
| 729 |
*/ |
| 730 |
static inline void |
| 731 |
ProcArrayEndTransactionInternal(PGPROC *proc, TransactionId latestXid) |
| 732 |
{ |
| 733 |
int pgxactoff = proc->pgxactoff; |
| 734 |
|
| 735 |
/* |
| 736 |
* Note: we need exclusive lock here because we're going to change other |
| 737 |
* processes' PGPROC entries. |
| 738 |
*/ |
| 739 |
Assert(LWLockHeldByMeInMode(ProcArrayLock, LW_EXCLUSIVE)); |
| 740 |
Assert(TransactionIdIsValid(ProcGlobal->xids[pgxactoff])); |
| 741 |
Assert(ProcGlobal->xids[pgxactoff] == proc->xid); |
| 742 |
|
| 743 |
ProcGlobal->xids[pgxactoff] = InvalidTransactionId; |
| 744 |
proc->xid = InvalidTransactionId; |
| 745 |
proc->vxid.lxid = InvalidLocalTransactionId; |
| 746 |
proc->xmin = InvalidTransactionId; |
| 747 |
|
| 748 |
/* be sure this is cleared in abort */ |
| 749 |
proc->delayChkptFlags = 0; |
| 750 |
|
| 751 |
proc->recoveryConflictPending = false; |
| 752 |
|
| 753 |
/* must be cleared with xid/xmin: */ |
| 754 |
/* avoid unnecessarily dirtying shared cachelines */ |
| 755 |
if (proc->statusFlags & PROC_VACUUM_STATE_MASK) |
| 756 |
{ |
| 757 |
proc->statusFlags &= ~PROC_VACUUM_STATE_MASK; |
| 758 |
ProcGlobal->statusFlags[proc->pgxactoff] = proc->statusFlags; |
| 759 |
} |
| 760 |
|
| 761 |
/* Clear the subtransaction-XID cache too while holding the lock */ |
| 762 |
Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count && |
| 763 |
ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed); |
| 764 |
if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed) |
| 765 |
{ |
| 766 |
ProcGlobal->subxidStates[pgxactoff].count = 0; |
| 767 |
ProcGlobal->subxidStates[pgxactoff].overflowed = false; |
| 768 |
proc->subxidStatus.count = 0; |
| 769 |
proc->subxidStatus.overflowed = false; |
| 770 |
} |
| 771 |
|
| 772 |
/* Also advance global latestCompletedXid while holding the lock */ |
| 773 |
MaintainLatestCompletedXid(latestXid); |
| 774 |
|
| 775 |
/* Same with xactCompletionCount */ |
| 776 |
TransamVariables->xactCompletionCount++; |
| 777 |
} |
| 778 |
|
| 779 |
/* |
| 780 |
* ProcArrayGroupClearXid -- group XID clearing |
| 781 |
* |
| 782 |
* When we cannot immediately acquire ProcArrayLock in exclusive mode at |
| 783 |
* commit time, add ourselves to a list of processes that need their XIDs |
| 784 |
* cleared. The first process to add itself to the list will acquire |
| 785 |
* ProcArrayLock in exclusive mode and perform ProcArrayEndTransactionInternal |
| 786 |
* on behalf of all group members. This avoids a great deal of contention |
| 787 |
* around ProcArrayLock when many processes are trying to commit at once, |
| 788 |
* since the lock need not be repeatedly handed off from one committing |
| 789 |
* process to the next. |
| 790 |
*/ |
| 791 |
static void |
| 792 |
ProcArrayGroupClearXid(PGPROC *proc, TransactionId latestXid) |
| 793 |
{ |
| 794 |
int pgprocno = GetNumberFromPGProc(proc); |
| 795 |
PROC_HDR *procglobal = ProcGlobal; |
| 796 |
uint32 nextidx; |
| 797 |
uint32 wakeidx; |
| 798 |
|
| 799 |
/* We should definitely have an XID to clear. */ |
| 800 |
Assert(TransactionIdIsValid(proc->xid)); |
| 801 |
|
| 802 |
/* Add ourselves to the list of processes needing a group XID clear. */ |
| 803 |
proc->procArrayGroupMember = true; |
| 804 |
proc->procArrayGroupMemberXid = latestXid; |
| 805 |
nextidx = pg_atomic_read_u32(&procglobal->procArrayGroupFirst); |
| 806 |
while (true) |
| 807 |
{ |
| 808 |
pg_atomic_write_u32(&proc->procArrayGroupNext, nextidx); |
| 809 |
|
| 810 |
if (pg_atomic_compare_exchange_u32(&procglobal->procArrayGroupFirst, |
| 811 |
&nextidx, |
| 812 |
(uint32) pgprocno)) |
| 813 |
break; |
| 814 |
} |
| 815 |
|
| 816 |
/* |
| 817 |
* If the list was not empty, the leader will clear our XID. It is |
| 818 |
* impossible to have followers without a leader because the first process |
| 819 |
* that has added itself to the list will always have nextidx as |
| 820 |
* INVALID_PROC_NUMBER. |
| 821 |
*/ |
| 822 |
if (nextidx != INVALID_PROC_NUMBER) |
| 823 |
{ |
| 824 |
int extraWaits = 0; |
| 825 |
|
| 826 |
/* Sleep until the leader clears our XID. */ |
| 827 |
pgstat_report_wait_start(WAIT_EVENT_PROCARRAY_GROUP_UPDATE); |
| 828 |
for (;;) |
| 829 |
{ |
| 830 |
/* acts as a read barrier */ |
| 831 |
PGSemaphoreLock(proc->sem); |
| 832 |
if (!proc->procArrayGroupMember) |
| 833 |
break; |
| 834 |
extraWaits++; |
| 835 |
} |
| 836 |
pgstat_report_wait_end(); |
| 837 |
|
| 838 |
Assert(pg_atomic_read_u32(&proc->procArrayGroupNext) == INVALID_PROC_NUMBER); |
| 839 |
|
| 840 |
/* Fix semaphore count for any absorbed wakeups */ |
| 841 |
while (extraWaits-- > 0) |
| 842 |
PGSemaphoreUnlock(proc->sem); |
| 843 |
return; |
| 844 |
} |
| 845 |
|
| 846 |
/* We are the leader. Acquire the lock on behalf of everyone. */ |
| 847 |
LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE); |
| 848 |
|
| 849 |
/* |
| 850 |
* Now that we've got the lock, clear the list of processes waiting for |
| 851 |
* group XID clearing, saving a pointer to the head of the list. Trying |
| 852 |
* to pop elements one at a time could lead to an ABA problem. |
| 853 |
*/ |
| 854 |
nextidx = pg_atomic_exchange_u32(&procglobal->procArrayGroupFirst, |
| 855 |
INVALID_PROC_NUMBER); |
| 856 |
|
| 857 |
/* Remember head of list so we can perform wakeups after dropping lock. */ |
| 858 |
wakeidx = nextidx; |
| 859 |
|
| 860 |
/* Walk the list and clear all XIDs. */ |
| 861 |
while (nextidx != INVALID_PROC_NUMBER) |
| 862 |
{ |
| 863 |
PGPROC *nextproc = &allProcs[nextidx]; |
| 864 |
|
| 865 |
ProcArrayEndTransactionInternal(nextproc, nextproc->procArrayGroupMemberXid); |
| 866 |
|
| 867 |
/* Move to next proc in list. */ |
| 868 |
nextidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext); |
| 869 |
} |
| 870 |
|
| 871 |
/* We're done with the lock now. */ |
| 872 |
LWLockRelease(ProcArrayLock); |
| 873 |
|
| 874 |
/* |
| 875 |
* Now that we've released the lock, go back and wake everybody up. We |
| 876 |
* don't do this under the lock so as to keep lock hold times to a |
| 877 |
* minimum. The system calls we need to perform to wake other processes |
| 878 |
* up are probably much slower than the simple memory writes we did while |
| 879 |
* holding the lock. |
| 880 |
*/ |
| 881 |
while (wakeidx != INVALID_PROC_NUMBER) |
| 882 |
{ |
| 883 |
PGPROC *nextproc = &allProcs[wakeidx]; |
| 884 |
|
| 885 |
wakeidx = pg_atomic_read_u32(&nextproc->procArrayGroupNext); |
| 886 |
pg_atomic_write_u32(&nextproc->procArrayGroupNext, INVALID_PROC_NUMBER); |
| 887 |
|
| 888 |
/* ensure all previous writes are visible before follower continues. */ |
| 889 |
pg_write_barrier(); |
| 890 |
|
| 891 |
nextproc->procArrayGroupMember = false; |
| 892 |
|
| 893 |
if (nextproc != MyProc) |
| 894 |
PGSemaphoreUnlock(nextproc->sem); |
| 895 |
} |
| 896 |
} |
| 897 |
|
| 898 |
/* |
| 899 |
* ProcArrayClearTransaction -- clear the transaction fields |
| 900 |
* |
| 901 |
* This is used after successfully preparing a 2-phase transaction. We are |
| 902 |
* not actually reporting the transaction's XID as no longer running --- it |
| 903 |
* will still appear as running because the 2PC's gxact is in the ProcArray |
| 904 |
* too. We just have to clear out our own PGPROC. |
| 905 |
*/ |
| 906 |
void |
| 907 |
ProcArrayClearTransaction(PGPROC *proc) |
| 908 |
{ |
| 909 |
int pgxactoff; |
| 910 |
|
| 911 |
/* |
| 912 |
* Currently we need to lock ProcArrayLock exclusively here, as we |
| 913 |
* increment xactCompletionCount below. We also need it at least in shared |
| 914 |
* mode for pgproc->pgxactoff to stay the same below. |
| 915 |
* |
| 916 |
* We could however, as this action does not actually change anyone's view |
| 917 |
* of the set of running XIDs (our entry is duplicate with the gxact that |
| 918 |
* has already been inserted into the ProcArray), lower the lock level to |
| 919 |
* shared if we were to make xactCompletionCount an atomic variable. But |
| 920 |
* that doesn't seem worth it currently, as a 2PC commit is heavyweight |
| 921 |
* enough for this not to be the bottleneck. If it ever becomes a |
| 922 |
* bottleneck it may also be worth considering to combine this with the |
| 923 |
* subsequent ProcArrayRemove() |
| 924 |
*/ |
| 925 |
LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE); |
| 926 |
|
| 927 |
pgxactoff = proc->pgxactoff; |
| 928 |
|
| 929 |
ProcGlobal->xids[pgxactoff] = InvalidTransactionId; |
| 930 |
proc->xid = InvalidTransactionId; |
| 931 |
|
| 932 |
proc->vxid.lxid = InvalidLocalTransactionId; |
| 933 |
proc->xmin = InvalidTransactionId; |
| 934 |
proc->recoveryConflictPending = false; |
| 935 |
|
| 936 |
Assert(!(proc->statusFlags & PROC_VACUUM_STATE_MASK)); |
| 937 |
Assert(!proc->delayChkptFlags); |
| 938 |
|
| 939 |
/* |
| 940 |
* Need to increment completion count even though transaction hasn't |
| 941 |
* really committed yet. The reason for that is that GetSnapshotData() |
| 942 |
* omits the xid of the current transaction, thus without the increment we |
| 943 |
* otherwise could end up reusing the snapshot later. Which would be bad, |
| 944 |
* because it might not count the prepared transaction as running. |
| 945 |
*/ |
| 946 |
TransamVariables->xactCompletionCount++; |
| 947 |
|
| 948 |
/* Clear the subtransaction-XID cache too */ |
| 949 |
Assert(ProcGlobal->subxidStates[pgxactoff].count == proc->subxidStatus.count && |
| 950 |
ProcGlobal->subxidStates[pgxactoff].overflowed == proc->subxidStatus.overflowed); |
| 951 |
if (proc->subxidStatus.count > 0 || proc->subxidStatus.overflowed) |
| 952 |
{ |
| 953 |
ProcGlobal->subxidStates[pgxactoff].count = 0; |
| 954 |
ProcGlobal->subxidStates[pgxactoff].overflowed = false; |
| 955 |
proc->subxidStatus.count = 0; |
| 956 |
proc->subxidStatus.overflowed = false; |
| 957 |
} |
| 958 |
|
| 959 |
LWLockRelease(ProcArrayLock); |
| 960 |
} |
| 961 |
|
| 962 |
/* |
| 963 |
* Update TransamVariables->latestCompletedXid to point to latestXid if |
| 964 |
* currently older. |
| 965 |
*/ |
| 966 |
static void |
| 967 |
MaintainLatestCompletedXid(TransactionId latestXid) |
| 968 |
{ |
| 969 |
FullTransactionId cur_latest = TransamVariables->latestCompletedXid; |
| 970 |
|
| 971 |
Assert(FullTransactionIdIsValid(cur_latest)); |
| 972 |
Assert(!RecoveryInProgress()); |
| 973 |
Assert(LWLockHeldByMe(ProcArrayLock)); |
| 974 |
|
| 975 |
if (TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid)) |
| 976 |
{ |
| 977 |
TransamVariables->latestCompletedXid = |
| 978 |
FullXidRelativeTo(cur_latest, latestXid); |
| 979 |
} |
| 980 |
|
| 981 |
Assert(IsBootstrapProcessingMode() || |
| 982 |
FullTransactionIdIsNormal(TransamVariables->latestCompletedXid)); |
| 983 |
} |
| 984 |
|
| 985 |
/* |
| 986 |
* Same as MaintainLatestCompletedXid, except for use during WAL replay. |
| 987 |
*/ |
| 988 |
static void |
| 989 |
MaintainLatestCompletedXidRecovery(TransactionId latestXid) |
| 990 |
{ |
| 991 |
FullTransactionId cur_latest = TransamVariables->latestCompletedXid; |
| 992 |
FullTransactionId rel; |
| 993 |
|
| 994 |
Assert(AmStartupProcess() || !IsUnderPostmaster); |
| 995 |
Assert(LWLockHeldByMe(ProcArrayLock)); |
| 996 |
|
| 997 |
/* |
| 998 |
* Need a FullTransactionId to compare latestXid with. Can't rely on |
| 999 |
* latestCompletedXid to be initialized in recovery. But in recovery it's |
| 1000 |
* safe to access nextXid without a lock for the startup process. |
| 1001 |
*/ |
| 1002 |
rel = TransamVariables->nextXid; |
| 1003 |
Assert(FullTransactionIdIsValid(TransamVariables->nextXid)); |
| 1004 |
|
| 1005 |
if (!FullTransactionIdIsValid(cur_latest) || |
| 1006 |
TransactionIdPrecedes(XidFromFullTransactionId(cur_latest), latestXid)) |
| 1007 |
{ |
| 1008 |
TransamVariables->latestCompletedXid = |
| 1009 |
FullXidRelativeTo(rel, latestXid); |
| 1010 |
} |
| 1011 |
|
| 1012 |
Assert(FullTransactionIdIsNormal(TransamVariables->latestCompletedXid)); |
| 1013 |
} |
| 1014 |
|
| 1015 |
/* |
| 1016 |
* ProcArrayInitRecovery -- initialize recovery xid mgmt environment |
| 1017 |
* |
| 1018 |
* Remember up to where the startup process initialized the CLOG and subtrans |
| 1019 |
* so we can ensure it's initialized gaplessly up to the point where necessary |
| 1020 |
* while in recovery. |
| 1021 |
*/ |
| 1022 |
void |
| 1023 |
ProcArrayInitRecovery(TransactionId initializedUptoXID) |
| 1024 |
{ |
| 1025 |
Assert(standbyState == STANDBY_INITIALIZED); |
| 1026 |
Assert(TransactionIdIsNormal(initializedUptoXID)); |
| 1027 |
|
| 1028 |
/* |
| 1029 |
* we set latestObservedXid to the xid SUBTRANS has been initialized up |
| 1030 |
* to, so we can extend it from that point onwards in |
| 1031 |
* RecordKnownAssignedTransactionIds, and when we get consistent in |
| 1032 |
* ProcArrayApplyRecoveryInfo(). |
| 1033 |
*/ |
| 1034 |
latestObservedXid = initializedUptoXID; |
| 1035 |
TransactionIdRetreat(latestObservedXid); |
| 1036 |
} |
| 1037 |
|
| 1038 |
/* |
| 1039 |
* ProcArrayApplyRecoveryInfo -- apply recovery info about xids |
| 1040 |
* |
| 1041 |
* Takes us through 3 states: Initialized, Pending and Ready. |
| 1042 |
* Normal case is to go all the way to Ready straight away, though there |
| 1043 |
* are atypical cases where we need to take it in steps. |
| 1044 |
* |
| 1045 |
* Use the data about running transactions on the primary to create the initial |
| 1046 |
* state of KnownAssignedXids. We also use these records to regularly prune |
| 1047 |
* KnownAssignedXids because we know it is possible that some transactions |
| 1048 |
* with FATAL errors fail to write abort records, which could cause eventual |
| 1049 |
* overflow. |
| 1050 |
* |
| 1051 |
* See comments for LogStandbySnapshot(). |
| 1052 |
*/ |
| 1053 |
void |
| 1054 |
ProcArrayApplyRecoveryInfo(RunningTransactions running) |
| 1055 |
{ |
| 1056 |
TransactionId *xids; |
| 1057 |
TransactionId advanceNextXid; |
| 1058 |
int nxids; |
| 1059 |
int i; |
| 1060 |
|
| 1061 |
Assert(standbyState >= STANDBY_INITIALIZED); |
| 1062 |
Assert(TransactionIdIsValid(running->nextXid)); |
| 1063 |
Assert(TransactionIdIsValid(running->oldestRunningXid)); |
| 1064 |
Assert(TransactionIdIsNormal(running->latestCompletedXid)); |
| 1065 |
|
| 1066 |
/* |
| 1067 |
* Remove stale transactions, if any. |
| 1068 |
*/ |
| 1069 |
ExpireOldKnownAssignedTransactionIds(running->oldestRunningXid); |
| 1070 |
|
| 1071 |
/* |
| 1072 |
* Adjust TransamVariables->nextXid before StandbyReleaseOldLocks(), |
| 1073 |
* because we will need it up to date for accessing two-phase transactions |
| 1074 |
* in StandbyReleaseOldLocks(). |
| 1075 |
*/ |
| 1076 |
advanceNextXid = running->nextXid; |
| 1077 |
TransactionIdRetreat(advanceNextXid); |
| 1078 |
AdvanceNextFullTransactionIdPastXid(advanceNextXid); |
| 1079 |
Assert(FullTransactionIdIsValid(TransamVariables->nextXid)); |
| 1080 |
|
| 1081 |
/* |
| 1082 |
* Remove stale locks, if any. |
| 1083 |
*/ |
| 1084 |
StandbyReleaseOldLocks(running->oldestRunningXid); |
| 1085 |
|
| 1086 |
/* |
| 1087 |
* If our snapshot is already valid, nothing else to do... |
| 1088 |
*/ |
| 1089 |
if (standbyState == STANDBY_SNAPSHOT_READY) |
| 1090 |
return; |
| 1091 |
|
| 1092 |
/* |
| 1093 |
* If our initial RunningTransactionsData had an overflowed snapshot then |
| 1094 |
* we knew we were missing some subxids from our snapshot. If we continue |
| 1095 |
* to see overflowed snapshots then we might never be able to start up, so |
| 1096 |
* we make another test to see if our snapshot is now valid. We know that |
| 1097 |
* the missing subxids are equal to or earlier than nextXid. After we |
| 1098 |
* initialise we continue to apply changes during recovery, so once the |
| 1099 |
* oldestRunningXid is later than the nextXid from the initial snapshot we |
| 1100 |
* know that we no longer have missing information and can mark the |
| 1101 |
* snapshot as valid. |
| 1102 |
*/ |
| 1103 |
if (standbyState == STANDBY_SNAPSHOT_PENDING) |
| 1104 |
{ |
| 1105 |
/* |
| 1106 |
* If the snapshot isn't overflowed or if its empty we can reset our |
| 1107 |
* pending state and use this snapshot instead. |
| 1108 |
*/ |
| 1109 |
if (running->subxid_status != SUBXIDS_MISSING || running->xcnt == 0) |
| 1110 |
{ |
| 1111 |
/* |
| 1112 |
* If we have already collected known assigned xids, we need to |
| 1113 |
* throw them away before we apply the recovery snapshot. |
| 1114 |
*/ |
| 1115 |
KnownAssignedXidsReset(); |
| 1116 |
standbyState = STANDBY_INITIALIZED; |
| 1117 |
} |
| 1118 |
else |
| 1119 |
{ |
| 1120 |
if (TransactionIdPrecedes(standbySnapshotPendingXmin, |
| 1121 |
running->oldestRunningXid)) |
| 1122 |
{ |
| 1123 |
standbyState = STANDBY_SNAPSHOT_READY; |
| 1124 |
elog(DEBUG1, |
| 1125 |
"recovery snapshots are now enabled"); |
| 1126 |
} |
| 1127 |
else |
| 1128 |
elog(DEBUG1, |
| 1129 |
"recovery snapshot waiting for non-overflowed snapshot or " |
| 1130 |
"until oldest active xid on standby is at least %u (now %u)", |
| 1131 |
standbySnapshotPendingXmin, |
| 1132 |
running->oldestRunningXid); |
| 1133 |
return; |
| 1134 |
} |
| 1135 |
} |
| 1136 |
|
| 1137 |
Assert(standbyState == STANDBY_INITIALIZED); |
| 1138 |
|
| 1139 |
/* |
| 1140 |
* NB: this can be reached at least twice, so make sure new code can deal |
| 1141 |
* with that. |
| 1142 |
*/ |
| 1143 |
|
| 1144 |
/* |
| 1145 |
* Nobody else is running yet, but take locks anyhow |
| 1146 |
*/ |
| 1147 |
LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE); |
| 1148 |
|
| 1149 |
/* |
| 1150 |
* KnownAssignedXids is sorted so we cannot just add the xids, we have to |
| 1151 |
* sort them first. |
| 1152 |
* |
| 1153 |
* Some of the new xids are top-level xids and some are subtransactions. |
| 1154 |
* We don't call SubTransSetParent because it doesn't matter yet. If we |
| 1155 |
* aren't overflowed then all xids will fit in snapshot and so we don't |
| 1156 |
* need subtrans. If we later overflow, an xid assignment record will add |
| 1157 |
* xids to subtrans. If RunningTransactionsData is overflowed then we |
| 1158 |
* don't have enough information to correctly update subtrans anyway. |
| 1159 |
*/ |
| 1160 |
|
| 1161 |
/* |
| 1162 |
* Allocate a temporary array to avoid modifying the array passed as |
| 1163 |
* argument. |
| 1164 |
*/ |
| 1165 |
xids = palloc(sizeof(TransactionId) * (running->xcnt + running->subxcnt)); |
| 1166 |
|
| 1167 |
/* |
| 1168 |
* Add to the temp array any xids which have not already completed. |
| 1169 |
*/ |
| 1170 |
nxids = 0; |
| 1171 |
for (i = 0; i < running->xcnt + running->subxcnt; i++) |
| 1172 |
{ |
| 1173 |
TransactionId xid = running->xids[i]; |
| 1174 |
|
| 1175 |
/* |
| 1176 |
* The running-xacts snapshot can contain xids that were still visible |
| 1177 |
* in the procarray when the snapshot was taken, but were already |
| 1178 |
* WAL-logged as completed. They're not running anymore, so ignore |
| 1179 |
* them. |
| 1180 |
*/ |
| 1181 |
if (TransactionIdDidCommit(xid) || TransactionIdDidAbort(xid)) |
| 1182 |
continue; |
| 1183 |
|
| 1184 |
xids[nxids++] = xid; |
| 1185 |
} |
| 1186 |
|
| 1187 |
if (nxids > 0) |
| 1188 |
{ |
| 1189 |
if (procArray->numKnownAssignedXids != 0) |
| 1190 |
{ |
| 1191 |
LWLockRelease(ProcArrayLock); |
| 1192 |
elog(ERROR, "KnownAssignedXids is not empty"); |
| 1193 |
} |
| 1194 |
|
| 1195 |
/* |
| 1196 |
* Sort the array so that we can add them safely into |
| 1197 |
* KnownAssignedXids. |
| 1198 |
* |
| 1199 |
* We have to sort them logically, because in KnownAssignedXidsAdd we |
| 1200 |
* call TransactionIdFollowsOrEquals and so on. But we know these XIDs |
| 1201 |
* come from RUNNING_XACTS, which means there are only normal XIDs |
| 1202 |
* from the same epoch, so this is safe. |
| 1203 |
*/ |
| 1204 |
qsort(xids, nxids, sizeof(TransactionId), xidLogicalComparator); |
| 1205 |
|
| 1206 |
/* |
| 1207 |
* Add the sorted snapshot into KnownAssignedXids. The running-xacts |
| 1208 |
* snapshot may include duplicated xids because of prepared |
| 1209 |
* transactions, so ignore them. |
| 1210 |
*/ |
| 1211 |
for (i = 0; i < nxids; i++) |
| 1212 |
{ |
| 1213 |
if (i > 0 && TransactionIdEquals(xids[i - 1], xids[i])) |
| 1214 |
{ |
| 1215 |
elog(DEBUG1, |
| 1216 |
"found duplicated transaction %u for KnownAssignedXids insertion", |
| 1217 |
xids[i]); |
| 1218 |
continue; |
| 1219 |
} |
| 1220 |
KnownAssignedXidsAdd(xids[i], xids[i], true); |
| 1221 |
} |
| 1222 |
|
| 1223 |
KnownAssignedXidsDisplay(DEBUG3); |
| 1224 |
} |
| 1225 |
|
| 1226 |
pfree(xids); |
| 1227 |
|
| 1228 |
/* |
| 1229 |
* latestObservedXid is at least set to the point where SUBTRANS was |
| 1230 |
* started up to (cf. ProcArrayInitRecovery()) or to the biggest xid |
| 1231 |
* RecordKnownAssignedTransactionIds() was called for. Initialize |
| 1232 |
* subtrans from thereon, up to nextXid - 1. |
| 1233 |
* |
| 1234 |
* We need to duplicate parts of RecordKnownAssignedTransactionId() here, |
| 1235 |
* because we've just added xids to the known assigned xids machinery that |
| 1236 |
* haven't gone through RecordKnownAssignedTransactionId(). |
| 1237 |
*/ |
| 1238 |
Assert(TransactionIdIsNormal(latestObservedXid)); |
| 1239 |
TransactionIdAdvance(latestObservedXid); |
| 1240 |
while (TransactionIdPrecedes(latestObservedXid, running->nextXid)) |
| 1241 |
{ |
| 1242 |
ExtendSUBTRANS(latestObservedXid); |
| 1243 |
TransactionIdAdvance(latestObservedXid); |
| 1244 |
} |
| 1245 |
TransactionIdRetreat(latestObservedXid); /* = running->nextXid - 1 */ |
| 1246 |
|
| 1247 |
/* ---------- |
| 1248 |
* Now we've got the running xids we need to set the global values that |
| 1249 |
* are used to track snapshots as they evolve further. |
| 1250 |
* |
| 1251 |
* - latestCompletedXid which will be the xmax for snapshots |
| 1252 |
* - lastOverflowedXid which shows whether snapshots overflow |
| 1253 |
* - nextXid |
| 1254 |
* |
| 1255 |
* If the snapshot overflowed, then we still initialise with what we know, |
| 1256 |
* but the recovery snapshot isn't fully valid yet because we know there |
| 1257 |
* are some subxids missing. We don't know the specific subxids that are |
| 1258 |
* missing, so conservatively assume the last one is latestObservedXid. |
| 1259 |
* ---------- |
| 1260 |
*/ |
| 1261 |
if (running->subxid_status == SUBXIDS_MISSING) |
| 1262 |
{ |
| 1263 |
standbyState = STANDBY_SNAPSHOT_PENDING; |
| 1264 |
|
| 1265 |
standbySnapshotPendingXmin = latestObservedXid; |
| 1266 |
procArray->lastOverflowedXid = latestObservedXid; |
| 1267 |
} |
| 1268 |
else |
| 1269 |
{ |
| 1270 |
standbyState = STANDBY_SNAPSHOT_READY; |
| 1271 |
|
| 1272 |
standbySnapshotPendingXmin = InvalidTransactionId; |
| 1273 |
|
| 1274 |
/* |
| 1275 |
* If the 'xids' array didn't include all subtransactions, we have to |
| 1276 |
* mark any snapshots taken as overflowed. |
| 1277 |
*/ |
| 1278 |
if (running->subxid_status == SUBXIDS_IN_SUBTRANS) |
| 1279 |
procArray->lastOverflowedXid = latestObservedXid; |
| 1280 |
else |
| 1281 |
{ |
| 1282 |
Assert(running->subxid_status == SUBXIDS_IN_ARRAY); |
| 1283 |
procArray->lastOverflowedXid = InvalidTransactionId; |
| 1284 |
} |
| 1285 |
} |
| 1286 |
|
| 1287 |
/* |
| 1288 |
* If a transaction wrote a commit record in the gap between taking and |
| 1289 |
* logging the snapshot then latestCompletedXid may already be higher than |
| 1290 |
* the value from the snapshot, so check before we use the incoming value. |
| 1291 |
* It also might not yet be set at all. |
| 1292 |
*/ |
| 1293 |
MaintainLatestCompletedXidRecovery(running->latestCompletedXid); |
| 1294 |
|
| 1295 |
/* |
| 1296 |
* NB: No need to increment TransamVariables->xactCompletionCount here, |
| 1297 |
* nobody can see it yet. |
| 1298 |
*/ |
| 1299 |
|
| 1300 |
LWLockRelease(ProcArrayLock); |
| 1301 |
|
| 1302 |
KnownAssignedXidsDisplay(DEBUG3); |
| 1303 |
if (standbyState == STANDBY_SNAPSHOT_READY) |
| 1304 |
elog(DEBUG1, "recovery snapshots are now enabled"); |
| 1305 |
else |
| 1306 |
elog(DEBUG1, |
| 1307 |
"recovery snapshot waiting for non-overflowed snapshot or " |
| 1308 |
"until oldest active xid on standby is at least %u (now %u)", |
| 1309 |
standbySnapshotPendingXmin, |
| 1310 |
running->oldestRunningXid); |
| 1311 |
} |
| 1312 |
|
| 1313 |
/* |
| 1314 |
* ProcArrayApplyXidAssignment |
| 1315 |
* Process an XLOG_XACT_ASSIGNMENT WAL record |
| 1316 |
*/ |
| 1317 |
void |
| 1318 |
ProcArrayApplyXidAssignment(TransactionId topxid, |
| 1319 |
int nsubxids, TransactionId *subxids) |
| 1320 |
{ |
| 1321 |
TransactionId max_xid; |
| 1322 |
int i; |
| 1323 |
|
| 1324 |
Assert(standbyState >= STANDBY_INITIALIZED); |
| 1325 |
|
| 1326 |
max_xid = TransactionIdLatest(topxid, nsubxids, subxids); |
| 1327 |
|
| 1328 |
/* |
| 1329 |
* Mark all the subtransactions as observed. |
| 1330 |
* |
| 1331 |
* NOTE: This will fail if the subxid contains too many previously |
| 1332 |
* unobserved xids to fit into known-assigned-xids. That shouldn't happen |
| 1333 |
* as the code stands, because xid-assignment records should never contain |
| 1334 |
* more than PGPROC_MAX_CACHED_SUBXIDS entries. |
| 1335 |
*/ |
| 1336 |
RecordKnownAssignedTransactionIds(max_xid); |
| 1337 |
|
| 1338 |
/* |
| 1339 |
* Notice that we update pg_subtrans with the top-level xid, rather than |
| 1340 |
* the parent xid. This is a difference between normal processing and |
| 1341 |
* recovery, yet is still correct in all cases. The reason is that |
| 1342 |
* subtransaction commit is not marked in clog until commit processing, so |
| 1343 |
* all aborted subtransactions have already been clearly marked in clog. |
| 1344 |
* As a result we are able to refer directly to the top-level |
| 1345 |
* transaction's state rather than skipping through all the intermediate |
| 1346 |
* states in the subtransaction tree. This should be the first time we |
| 1347 |
* have attempted to SubTransSetParent(). |
| 1348 |
*/ |
| 1349 |
for (i = 0; i < nsubxids; i++) |
| 1350 |
SubTransSetParent(subxids[i], topxid); |
| 1351 |
|
| 1352 |
/* KnownAssignedXids isn't maintained yet, so we're done for now */ |
| 1353 |
if (standbyState == STANDBY_INITIALIZED) |
| 1354 |
return; |
| 1355 |
|
| 1356 |
/* |
| 1357 |
* Uses same locking as transaction commit |
| 1358 |
*/ |
| 1359 |
LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE); |
| 1360 |
|
| 1361 |
/* |
| 1362 |
* Remove subxids from known-assigned-xacts. |
| 1363 |
*/ |
| 1364 |
KnownAssignedXidsRemoveTree(InvalidTransactionId, nsubxids, subxids); |
| 1365 |
|
| 1366 |
/* |
| 1367 |
* Advance lastOverflowedXid to be at least the last of these subxids. |
| 1368 |
*/ |
| 1369 |
if (TransactionIdPrecedes(procArray->lastOverflowedXid, max_xid)) |
| 1370 |
procArray->lastOverflowedXid = max_xid; |
| 1371 |
|
| 1372 |
LWLockRelease(ProcArrayLock); |
| 1373 |
} |
| 1374 |
|
| 1375 |
/* |
| 1376 |
* TransactionIdIsInProgress -- is given transaction running in some backend |
| 1377 |
* |
| 1378 |
* Aside from some shortcuts such as checking RecentXmin and our own Xid, |
| 1379 |
* there are four possibilities for finding a running transaction: |
| 1380 |
* |
| 1381 |
* 1. The given Xid is a main transaction Id. We will find this out cheaply |
| 1382 |
* by looking at ProcGlobal->xids. |
| 1383 |
* |
| 1384 |
* 2. The given Xid is one of the cached subxact Xids in the PGPROC array. |
| 1385 |
* We can find this out cheaply too. |
| 1386 |
* |
| 1387 |
* 3. In Hot Standby mode, we must search the KnownAssignedXids list to see |
| 1388 |
* if the Xid is running on the primary. |
| 1389 |
* |
| 1390 |
* 4. Search the SubTrans tree to find the Xid's topmost parent, and then see |
| 1391 |
* if that is running according to ProcGlobal->xids[] or KnownAssignedXids. |
| 1392 |
* This is the slowest way, but sadly it has to be done always if the others |
| 1393 |
* failed, unless we see that the cached subxact sets are complete (none have |
| 1394 |
* overflowed). |
| 1395 |
* |
| 1396 |
* ProcArrayLock has to be held while we do 1, 2, 3. If we save the top Xids |
| 1397 |
* while doing 1 and 3, we can release the ProcArrayLock while we do 4. |
| 1398 |
* This buys back some concurrency (and we can't retrieve the main Xids from |
| 1399 |
* ProcGlobal->xids[] again anyway; see GetNewTransactionId). |
| 1400 |
*/ |
| 1401 |
bool |
| 1402 |
TransactionIdIsInProgress(TransactionId xid) |
| 1403 |
{ |
| 1404 |
static TransactionId *xids = NULL; |
| 1405 |
static TransactionId *other_xids; |
| 1406 |
XidCacheStatus *other_subxidstates; |
| 1407 |
int nxids = 0; |
| 1408 |
ProcArrayStruct *arrayP = procArray; |
| 1409 |
TransactionId topxid; |
| 1410 |
TransactionId latestCompletedXid; |
| 1411 |
int mypgxactoff; |
| 1412 |
int numProcs; |
| 1413 |
int j; |
| 1414 |
|
| 1415 |
/* |
| 1416 |
* Don't bother checking a transaction older than RecentXmin; it could not |
| 1417 |
* possibly still be running. (Note: in particular, this guarantees that |
| 1418 |
* we reject InvalidTransactionId, FrozenTransactionId, etc as not |
| 1419 |
* running.) |
| 1420 |
*/ |
| 1421 |
if (TransactionIdPrecedes(xid, RecentXmin)) |
| 1422 |
{ |
| 1423 |
xc_by_recent_xmin_inc(); |
| 1424 |
return false; |
| 1425 |
} |
| 1426 |
|
| 1427 |
/* |
| 1428 |
* We may have just checked the status of this transaction, so if it is |
| 1429 |
* already known to be completed, we can fall out without any access to |
| 1430 |
* shared memory. |
| 1431 |
*/ |
| 1432 |
if (TransactionIdEquals(cachedXidIsNotInProgress, xid)) |
| 1433 |
{ |
| 1434 |
xc_by_known_xact_inc(); |
| 1435 |
return false; |
| 1436 |
} |
| 1437 |
|
| 1438 |
/* |
| 1439 |
* Also, we can handle our own transaction (and subtransactions) without |
| 1440 |
* any access to shared memory. |
| 1441 |
*/ |
| 1442 |
if (TransactionIdIsCurrentTransactionId(xid)) |
| 1443 |
{ |
| 1444 |
xc_by_my_xact_inc(); |
| 1445 |
return true; |
| 1446 |
} |
| 1447 |
|
| 1448 |
/* |
| 1449 |
* If first time through, get workspace to remember main XIDs in. We |
| 1450 |
* malloc it permanently to avoid repeated palloc/pfree overhead. |
| 1451 |
*/ |
| 1452 |
if (xids == NULL) |
| 1453 |
{ |
| 1454 |
/* |
| 1455 |
* In hot standby mode, reserve enough space to hold all xids in the |
| 1456 |
* known-assigned list. If we later finish recovery, we no longer need |
| 1457 |
* the bigger array, but we don't bother to shrink it. |
| 1458 |
*/ |
| 1459 |
int maxxids = RecoveryInProgress() ? TOTAL_MAX_CACHED_SUBXIDS : arrayP->maxProcs; |
| 1460 |
|
| 1461 |
xids = (TransactionId *) malloc(maxxids * sizeof(TransactionId)); |
| 1462 |
if (xids == NULL) |
| 1463 |
ereport(ERROR, |
| 1464 |
(errcode(ERRCODE_OUT_OF_MEMORY), |
| 1465 |
errmsg("out of memory"))); |
| 1466 |
} |
| 1467 |
|
| 1468 |
other_xids = ProcGlobal->xids; |
| 1469 |
other_subxidstates = ProcGlobal->subxidStates; |
| 1470 |
|
| 1471 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 1472 |
|
| 1473 |
/* |
| 1474 |
* Now that we have the lock, we can check latestCompletedXid; if the |
| 1475 |
* target Xid is after that, it's surely still running. |
| 1476 |
*/ |
| 1477 |
latestCompletedXid = |
| 1478 |
XidFromFullTransactionId(TransamVariables->latestCompletedXid); |
| 1479 |
if (TransactionIdPrecedes(latestCompletedXid, xid)) |
| 1480 |
{ |
| 1481 |
LWLockRelease(ProcArrayLock); |
| 1482 |
xc_by_latest_xid_inc(); |
| 1483 |
return true; |
| 1484 |
} |
| 1485 |
|
| 1486 |
/* No shortcuts, gotta grovel through the array */ |
| 1487 |
mypgxactoff = MyProc->pgxactoff; |
| 1488 |
numProcs = arrayP->numProcs; |
| 1489 |
for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++) |
| 1490 |
{ |
| 1491 |
int pgprocno; |
| 1492 |
PGPROC *proc; |
| 1493 |
TransactionId pxid; |
| 1494 |
int pxids; |
| 1495 |
|
| 1496 |
/* Ignore ourselves --- dealt with it above */ |
| 1497 |
if (pgxactoff == mypgxactoff) |
| 1498 |
continue; |
| 1499 |
|
| 1500 |
/* Fetch xid just once - see GetNewTransactionId */ |
| 1501 |
pxid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]); |
| 1502 |
|
| 1503 |
if (!TransactionIdIsValid(pxid)) |
| 1504 |
continue; |
| 1505 |
|
| 1506 |
/* |
| 1507 |
* Step 1: check the main Xid |
| 1508 |
*/ |
| 1509 |
if (TransactionIdEquals(pxid, xid)) |
| 1510 |
{ |
| 1511 |
LWLockRelease(ProcArrayLock); |
| 1512 |
xc_by_main_xid_inc(); |
| 1513 |
return true; |
| 1514 |
} |
| 1515 |
|
| 1516 |
/* |
| 1517 |
* We can ignore main Xids that are younger than the target Xid, since |
| 1518 |
* the target could not possibly be their child. |
| 1519 |
*/ |
| 1520 |
if (TransactionIdPrecedes(xid, pxid)) |
| 1521 |
continue; |
| 1522 |
|
| 1523 |
/* |
| 1524 |
* Step 2: check the cached child-Xids arrays |
| 1525 |
*/ |
| 1526 |
pxids = other_subxidstates[pgxactoff].count; |
| 1527 |
pg_read_barrier(); /* pairs with barrier in GetNewTransactionId() */ |
| 1528 |
pgprocno = arrayP->pgprocnos[pgxactoff]; |
| 1529 |
proc = &allProcs[pgprocno]; |
| 1530 |
for (j = pxids - 1; j >= 0; j--) |
| 1531 |
{ |
| 1532 |
/* Fetch xid just once - see GetNewTransactionId */ |
| 1533 |
TransactionId cxid = UINT32_ACCESS_ONCE(proc->subxids.xids[j]); |
| 1534 |
|
| 1535 |
if (TransactionIdEquals(cxid, xid)) |
| 1536 |
{ |
| 1537 |
LWLockRelease(ProcArrayLock); |
| 1538 |
xc_by_child_xid_inc(); |
| 1539 |
return true; |
| 1540 |
} |
| 1541 |
} |
| 1542 |
|
| 1543 |
/* |
| 1544 |
* Save the main Xid for step 4. We only need to remember main Xids |
| 1545 |
* that have uncached children. (Note: there is no race condition |
| 1546 |
* here because the overflowed flag cannot be cleared, only set, while |
| 1547 |
* we hold ProcArrayLock. So we can't miss an Xid that we need to |
| 1548 |
* worry about.) |
| 1549 |
*/ |
| 1550 |
if (other_subxidstates[pgxactoff].overflowed) |
| 1551 |
xids[nxids++] = pxid; |
| 1552 |
} |
| 1553 |
|
| 1554 |
/* |
| 1555 |
* Step 3: in hot standby mode, check the known-assigned-xids list. XIDs |
| 1556 |
* in the list must be treated as running. |
| 1557 |
*/ |
| 1558 |
if (RecoveryInProgress()) |
| 1559 |
{ |
| 1560 |
/* none of the PGPROC entries should have XIDs in hot standby mode */ |
| 1561 |
Assert(nxids == 0); |
| 1562 |
|
| 1563 |
if (KnownAssignedXidExists(xid)) |
| 1564 |
{ |
| 1565 |
LWLockRelease(ProcArrayLock); |
| 1566 |
xc_by_known_assigned_inc(); |
| 1567 |
return true; |
| 1568 |
} |
| 1569 |
|
| 1570 |
/* |
| 1571 |
* If the KnownAssignedXids overflowed, we have to check pg_subtrans |
| 1572 |
* too. Fetch all xids from KnownAssignedXids that are lower than |
| 1573 |
* xid, since if xid is a subtransaction its parent will always have a |
| 1574 |
* lower value. Note we will collect both main and subXIDs here, but |
| 1575 |
* there's no help for it. |
| 1576 |
*/ |
| 1577 |
if (TransactionIdPrecedesOrEquals(xid, procArray->lastOverflowedXid)) |
| 1578 |
nxids = KnownAssignedXidsGet(xids, xid); |
| 1579 |
} |
| 1580 |
|
| 1581 |
LWLockRelease(ProcArrayLock); |
| 1582 |
|
| 1583 |
/* |
| 1584 |
* If none of the relevant caches overflowed, we know the Xid is not |
| 1585 |
* running without even looking at pg_subtrans. |
| 1586 |
*/ |
| 1587 |
if (nxids == 0) |
| 1588 |
{ |
| 1589 |
xc_no_overflow_inc(); |
| 1590 |
cachedXidIsNotInProgress = xid; |
| 1591 |
return false; |
| 1592 |
} |
| 1593 |
|
| 1594 |
/* |
| 1595 |
* Step 4: have to check pg_subtrans. |
| 1596 |
* |
| 1597 |
* At this point, we know it's either a subtransaction of one of the Xids |
| 1598 |
* in xids[], or it's not running. If it's an already-failed |
| 1599 |
* subtransaction, we want to say "not running" even though its parent may |
| 1600 |
* still be running. So first, check pg_xact to see if it's been aborted. |
| 1601 |
*/ |
| 1602 |
xc_slow_answer_inc(); |
| 1603 |
|
| 1604 |
if (TransactionIdDidAbort(xid)) |
| 1605 |
{ |
| 1606 |
cachedXidIsNotInProgress = xid; |
| 1607 |
return false; |
| 1608 |
} |
| 1609 |
|
| 1610 |
/* |
| 1611 |
* It isn't aborted, so check whether the transaction tree it belongs to |
| 1612 |
* is still running (or, more precisely, whether it was running when we |
| 1613 |
* held ProcArrayLock). |
| 1614 |
*/ |
| 1615 |
topxid = SubTransGetTopmostTransaction(xid); |
| 1616 |
Assert(TransactionIdIsValid(topxid)); |
| 1617 |
if (!TransactionIdEquals(topxid, xid) && |
| 1618 |
pg_lfind32(topxid, xids, nxids)) |
| 1619 |
return true; |
| 1620 |
|
| 1621 |
cachedXidIsNotInProgress = xid; |
| 1622 |
return false; |
| 1623 |
} |
| 1624 |
|
| 1625 |
/* |
| 1626 |
* TransactionIdIsActive -- is xid the top-level XID of an active backend? |
| 1627 |
* |
| 1628 |
* This differs from TransactionIdIsInProgress in that it ignores prepared |
| 1629 |
* transactions, as well as transactions running on the primary if we're in |
| 1630 |
* hot standby. Also, we ignore subtransactions since that's not needed |
| 1631 |
* for current uses. |
| 1632 |
*/ |
| 1633 |
bool |
| 1634 |
TransactionIdIsActive(TransactionId xid) |
| 1635 |
{ |
| 1636 |
bool result = false; |
| 1637 |
ProcArrayStruct *arrayP = procArray; |
| 1638 |
TransactionId *other_xids = ProcGlobal->xids; |
| 1639 |
int i; |
| 1640 |
|
| 1641 |
/* |
| 1642 |
* Don't bother checking a transaction older than RecentXmin; it could not |
| 1643 |
* possibly still be running. |
| 1644 |
*/ |
| 1645 |
if (TransactionIdPrecedes(xid, RecentXmin)) |
| 1646 |
return false; |
| 1647 |
|
| 1648 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 1649 |
|
| 1650 |
for (i = 0; i < arrayP->numProcs; i++) |
| 1651 |
{ |
| 1652 |
int pgprocno = arrayP->pgprocnos[i]; |
| 1653 |
PGPROC *proc = &allProcs[pgprocno]; |
| 1654 |
TransactionId pxid; |
| 1655 |
|
| 1656 |
/* Fetch xid just once - see GetNewTransactionId */ |
| 1657 |
pxid = UINT32_ACCESS_ONCE(other_xids[i]); |
| 1658 |
|
| 1659 |
if (!TransactionIdIsValid(pxid)) |
| 1660 |
continue; |
| 1661 |
|
| 1662 |
if (proc->pid == 0) |
| 1663 |
continue; /* ignore prepared transactions */ |
| 1664 |
|
| 1665 |
if (TransactionIdEquals(pxid, xid)) |
| 1666 |
{ |
| 1667 |
result = true; |
| 1668 |
break; |
| 1669 |
} |
| 1670 |
} |
| 1671 |
|
| 1672 |
LWLockRelease(ProcArrayLock); |
| 1673 |
|
| 1674 |
return result; |
| 1675 |
} |
| 1676 |
|
| 1677 |
|
| 1678 |
/* |
| 1679 |
* Determine XID horizons. |
| 1680 |
* |
| 1681 |
* This is used by wrapper functions like GetOldestNonRemovableTransactionId() |
| 1682 |
* (for VACUUM), GetReplicationHorizons() (for hot_standby_feedback), etc as |
| 1683 |
* well as "internally" by GlobalVisUpdate() (see comment above struct |
| 1684 |
* GlobalVisState). |
| 1685 |
* |
| 1686 |
* See the definition of ComputeXidHorizonsResult for the various computed |
| 1687 |
* horizons. |
| 1688 |
* |
| 1689 |
* For VACUUM separate horizons (used to decide which deleted tuples must |
| 1690 |
* be preserved), for shared and non-shared tables are computed. For shared |
| 1691 |
* relations backends in all databases must be considered, but for non-shared |
| 1692 |
* relations that's not required, since only backends in my own database could |
| 1693 |
* ever see the tuples in them. Also, we can ignore concurrently running lazy |
| 1694 |
* VACUUMs because (a) they must be working on other tables, and (b) they |
| 1695 |
* don't need to do snapshot-based lookups. |
| 1696 |
* |
| 1697 |
* This also computes a horizon used to truncate pg_subtrans. For that |
| 1698 |
* backends in all databases have to be considered, and concurrently running |
| 1699 |
* lazy VACUUMs cannot be ignored, as they still may perform pg_subtrans |
| 1700 |
* accesses. |
| 1701 |
* |
| 1702 |
* Note: we include all currently running xids in the set of considered xids. |
| 1703 |
* This ensures that if a just-started xact has not yet set its snapshot, |
| 1704 |
* when it does set the snapshot it cannot set xmin less than what we compute. |
| 1705 |
* See notes in src/backend/access/transam/README. |
| 1706 |
* |
| 1707 |
* Note: despite the above, it's possible for the calculated values to move |
| 1708 |
* backwards on repeated calls. The calculated values are conservative, so |
| 1709 |
* that anything older is definitely not considered as running by anyone |
| 1710 |
* anymore, but the exact values calculated depend on a number of things. For |
| 1711 |
* example, if there are no transactions running in the current database, the |
| 1712 |
* horizon for normal tables will be latestCompletedXid. If a transaction |
| 1713 |
* begins after that, its xmin will include in-progress transactions in other |
| 1714 |
* databases that started earlier, so another call will return a lower value. |
| 1715 |
* Nonetheless it is safe to vacuum a table in the current database with the |
| 1716 |
* first result. There are also replication-related effects: a walsender |
| 1717 |
* process can set its xmin based on transactions that are no longer running |
| 1718 |
* on the primary but are still being replayed on the standby, thus possibly |
| 1719 |
* making the values go backwards. In this case there is a possibility that |
| 1720 |
* we lose data that the standby would like to have, but unless the standby |
| 1721 |
* uses a replication slot to make its xmin persistent there is little we can |
| 1722 |
* do about that --- data is only protected if the walsender runs continuously |
| 1723 |
* while queries are executed on the standby. (The Hot Standby code deals |
| 1724 |
* with such cases by failing standby queries that needed to access |
| 1725 |
* already-removed data, so there's no integrity bug.) |
| 1726 |
* |
| 1727 |
* Note: the approximate horizons (see definition of GlobalVisState) are |
| 1728 |
* updated by the computations done here. That's currently required for |
| 1729 |
* correctness and a small optimization. Without doing so it's possible that |
| 1730 |
* heap vacuum's call to heap_page_prune_and_freeze() uses a more conservative |
| 1731 |
* horizon than later when deciding which tuples can be removed - which the |
| 1732 |
* code doesn't expect (breaking HOT). |
| 1733 |
*/ |
| 1734 |
static void |
| 1735 |
ComputeXidHorizons(ComputeXidHorizonsResult *h) |
| 1736 |
{ |
| 1737 |
ProcArrayStruct *arrayP = procArray; |
| 1738 |
TransactionId kaxmin; |
| 1739 |
bool in_recovery = RecoveryInProgress(); |
| 1740 |
TransactionId *other_xids = ProcGlobal->xids; |
| 1741 |
|
| 1742 |
/* inferred after ProcArrayLock is released */ |
| 1743 |
h->catalog_oldest_nonremovable = InvalidTransactionId; |
| 1744 |
|
| 1745 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 1746 |
|
| 1747 |
h->latest_completed = TransamVariables->latestCompletedXid; |
| 1748 |
|
| 1749 |
/* |
| 1750 |
* We initialize the MIN() calculation with latestCompletedXid + 1. This |
| 1751 |
* is a lower bound for the XIDs that might appear in the ProcArray later, |
| 1752 |
* and so protects us against overestimating the result due to future |
| 1753 |
* additions. |
| 1754 |
*/ |
| 1755 |
{ |
| 1756 |
TransactionId initial; |
| 1757 |
|
| 1758 |
initial = XidFromFullTransactionId(h->latest_completed); |
| 1759 |
Assert(TransactionIdIsValid(initial)); |
| 1760 |
TransactionIdAdvance(initial); |
| 1761 |
|
| 1762 |
h->oldest_considered_running = initial; |
| 1763 |
h->shared_oldest_nonremovable = initial; |
| 1764 |
h->data_oldest_nonremovable = initial; |
| 1765 |
|
| 1766 |
/* |
| 1767 |
* Only modifications made by this backend affect the horizon for |
| 1768 |
* temporary relations. Instead of a check in each iteration of the |
| 1769 |
* loop over all PGPROCs it is cheaper to just initialize to the |
| 1770 |
* current top-level xid any. |
| 1771 |
* |
| 1772 |
* Without an assigned xid we could use a horizon as aggressive as |
| 1773 |
* GetNewTransactionId(), but we can get away with the much cheaper |
| 1774 |
* latestCompletedXid + 1: If this backend has no xid there, by |
| 1775 |
* definition, can't be any newer changes in the temp table than |
| 1776 |
* latestCompletedXid. |
| 1777 |
*/ |
| 1778 |
if (TransactionIdIsValid(MyProc->xid)) |
| 1779 |
h->temp_oldest_nonremovable = MyProc->xid; |
| 1780 |
else |
| 1781 |
h->temp_oldest_nonremovable = initial; |
| 1782 |
} |
| 1783 |
|
| 1784 |
/* |
| 1785 |
* Fetch slot horizons while ProcArrayLock is held - the |
| 1786 |
* LWLockAcquire/LWLockRelease are a barrier, ensuring this happens inside |
| 1787 |
* the lock. |
| 1788 |
*/ |
| 1789 |
h->slot_xmin = procArray->replication_slot_xmin; |
| 1790 |
h->slot_catalog_xmin = procArray->replication_slot_catalog_xmin; |
| 1791 |
|
| 1792 |
for (int index = 0; index < arrayP->numProcs; index++) |
| 1793 |
{ |
| 1794 |
int pgprocno = arrayP->pgprocnos[index]; |
| 1795 |
PGPROC *proc = &allProcs[pgprocno]; |
| 1796 |
int8 statusFlags = ProcGlobal->statusFlags[index]; |
| 1797 |
TransactionId xid; |
| 1798 |
TransactionId xmin; |
| 1799 |
|
| 1800 |
/* Fetch xid just once - see GetNewTransactionId */ |
| 1801 |
xid = UINT32_ACCESS_ONCE(other_xids[index]); |
| 1802 |
xmin = UINT32_ACCESS_ONCE(proc->xmin); |
| 1803 |
|
| 1804 |
/* |
| 1805 |
* Consider both the transaction's Xmin, and its Xid. |
| 1806 |
* |
| 1807 |
* We must check both because a transaction might have an Xmin but not |
| 1808 |
* (yet) an Xid; conversely, if it has an Xid, that could determine |
| 1809 |
* some not-yet-set Xmin. |
| 1810 |
*/ |
| 1811 |
xmin = TransactionIdOlder(xmin, xid); |
| 1812 |
|
| 1813 |
/* if neither is set, this proc doesn't influence the horizon */ |
| 1814 |
if (!TransactionIdIsValid(xmin)) |
| 1815 |
continue; |
| 1816 |
|
| 1817 |
/* |
| 1818 |
* Don't ignore any procs when determining which transactions might be |
| 1819 |
* considered running. While slots should ensure logical decoding |
| 1820 |
* backends are protected even without this check, it can't hurt to |
| 1821 |
* include them here as well.. |
| 1822 |
*/ |
| 1823 |
h->oldest_considered_running = |
| 1824 |
TransactionIdOlder(h->oldest_considered_running, xmin); |
| 1825 |
|
| 1826 |
/* |
| 1827 |
* Skip over backends either vacuuming (which is ok with rows being |
| 1828 |
* removed, as long as pg_subtrans is not truncated) or doing logical |
| 1829 |
* decoding (which manages xmin separately, check below). |
| 1830 |
*/ |
| 1831 |
if (statusFlags & (PROC_IN_VACUUM | PROC_IN_LOGICAL_DECODING)) |
| 1832 |
continue; |
| 1833 |
|
| 1834 |
/* shared tables need to take backends in all databases into account */ |
| 1835 |
h->shared_oldest_nonremovable = |
| 1836 |
TransactionIdOlder(h->shared_oldest_nonremovable, xmin); |
| 1837 |
|
| 1838 |
/* |
| 1839 |
* Normally sessions in other databases are ignored for anything but |
| 1840 |
* the shared horizon. |
| 1841 |
* |
| 1842 |
* However, include them when MyDatabaseId is not (yet) set. A |
| 1843 |
* backend in the process of starting up must not compute a "too |
| 1844 |
* aggressive" horizon, otherwise we could end up using it to prune |
| 1845 |
* still-needed data away. If the current backend never connects to a |
| 1846 |
* database this is harmless, because data_oldest_nonremovable will |
| 1847 |
* never be utilized. |
| 1848 |
* |
| 1849 |
* Also, sessions marked with PROC_AFFECTS_ALL_HORIZONS should always |
| 1850 |
* be included. (This flag is used for hot standby feedback, which |
| 1851 |
* can't be tied to a specific database.) |
| 1852 |
* |
| 1853 |
* Also, while in recovery we cannot compute an accurate per-database |
| 1854 |
* horizon, as all xids are managed via the KnownAssignedXids |
| 1855 |
* machinery. |
| 1856 |
*/ |
| 1857 |
if (proc->databaseId == MyDatabaseId || |
| 1858 |
MyDatabaseId == InvalidOid || |
| 1859 |
(statusFlags & PROC_AFFECTS_ALL_HORIZONS) || |
| 1860 |
in_recovery) |
| 1861 |
{ |
| 1862 |
h->data_oldest_nonremovable = |
| 1863 |
TransactionIdOlder(h->data_oldest_nonremovable, xmin); |
| 1864 |
} |
| 1865 |
} |
| 1866 |
|
| 1867 |
/* |
| 1868 |
* If in recovery fetch oldest xid in KnownAssignedXids, will be applied |
| 1869 |
* after lock is released. |
| 1870 |
*/ |
| 1871 |
if (in_recovery) |
| 1872 |
kaxmin = KnownAssignedXidsGetOldestXmin(); |
| 1873 |
|
| 1874 |
/* |
| 1875 |
* No other information from shared state is needed, release the lock |
| 1876 |
* immediately. The rest of the computations can be done without a lock. |
| 1877 |
*/ |
| 1878 |
LWLockRelease(ProcArrayLock); |
| 1879 |
|
| 1880 |
if (in_recovery) |
| 1881 |
{ |
| 1882 |
h->oldest_considered_running = |
| 1883 |
TransactionIdOlder(h->oldest_considered_running, kaxmin); |
| 1884 |
h->shared_oldest_nonremovable = |
| 1885 |
TransactionIdOlder(h->shared_oldest_nonremovable, kaxmin); |
| 1886 |
h->data_oldest_nonremovable = |
| 1887 |
TransactionIdOlder(h->data_oldest_nonremovable, kaxmin); |
| 1888 |
/* temp relations cannot be accessed in recovery */ |
| 1889 |
} |
| 1890 |
|
| 1891 |
Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running, |
| 1892 |
h->shared_oldest_nonremovable)); |
| 1893 |
Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable, |
| 1894 |
h->data_oldest_nonremovable)); |
| 1895 |
|
| 1896 |
/* |
| 1897 |
* Check whether there are replication slots requiring an older xmin. |
| 1898 |
*/ |
| 1899 |
h->shared_oldest_nonremovable = |
| 1900 |
TransactionIdOlder(h->shared_oldest_nonremovable, h->slot_xmin); |
| 1901 |
h->data_oldest_nonremovable = |
| 1902 |
TransactionIdOlder(h->data_oldest_nonremovable, h->slot_xmin); |
| 1903 |
|
| 1904 |
/* |
| 1905 |
* The only difference between catalog / data horizons is that the slot's |
| 1906 |
* catalog xmin is applied to the catalog one (so catalogs can be accessed |
| 1907 |
* for logical decoding). Initialize with data horizon, and then back up |
| 1908 |
* further if necessary. Have to back up the shared horizon as well, since |
| 1909 |
* that also can contain catalogs. |
| 1910 |
*/ |
| 1911 |
h->shared_oldest_nonremovable_raw = h->shared_oldest_nonremovable; |
| 1912 |
h->shared_oldest_nonremovable = |
| 1913 |
TransactionIdOlder(h->shared_oldest_nonremovable, |
| 1914 |
h->slot_catalog_xmin); |
| 1915 |
h->catalog_oldest_nonremovable = h->data_oldest_nonremovable; |
| 1916 |
h->catalog_oldest_nonremovable = |
| 1917 |
TransactionIdOlder(h->catalog_oldest_nonremovable, |
| 1918 |
h->slot_catalog_xmin); |
| 1919 |
|
| 1920 |
/* |
| 1921 |
* It's possible that slots backed up the horizons further than |
| 1922 |
* oldest_considered_running. Fix. |
| 1923 |
*/ |
| 1924 |
h->oldest_considered_running = |
| 1925 |
TransactionIdOlder(h->oldest_considered_running, |
| 1926 |
h->shared_oldest_nonremovable); |
| 1927 |
h->oldest_considered_running = |
| 1928 |
TransactionIdOlder(h->oldest_considered_running, |
| 1929 |
h->catalog_oldest_nonremovable); |
| 1930 |
h->oldest_considered_running = |
| 1931 |
TransactionIdOlder(h->oldest_considered_running, |
| 1932 |
h->data_oldest_nonremovable); |
| 1933 |
|
| 1934 |
/* |
| 1935 |
* shared horizons have to be at least as old as the oldest visible in |
| 1936 |
* current db |
| 1937 |
*/ |
| 1938 |
Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable, |
| 1939 |
h->data_oldest_nonremovable)); |
| 1940 |
Assert(TransactionIdPrecedesOrEquals(h->shared_oldest_nonremovable, |
| 1941 |
h->catalog_oldest_nonremovable)); |
| 1942 |
|
| 1943 |
/* |
| 1944 |
* Horizons need to ensure that pg_subtrans access is still possible for |
| 1945 |
* the relevant backends. |
| 1946 |
*/ |
| 1947 |
Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running, |
| 1948 |
h->shared_oldest_nonremovable)); |
| 1949 |
Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running, |
| 1950 |
h->catalog_oldest_nonremovable)); |
| 1951 |
Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running, |
| 1952 |
h->data_oldest_nonremovable)); |
| 1953 |
Assert(TransactionIdPrecedesOrEquals(h->oldest_considered_running, |
| 1954 |
h->temp_oldest_nonremovable)); |
| 1955 |
Assert(!TransactionIdIsValid(h->slot_xmin) || |
| 1956 |
TransactionIdPrecedesOrEquals(h->oldest_considered_running, |
| 1957 |
h->slot_xmin)); |
| 1958 |
Assert(!TransactionIdIsValid(h->slot_catalog_xmin) || |
| 1959 |
TransactionIdPrecedesOrEquals(h->oldest_considered_running, |
| 1960 |
h->slot_catalog_xmin)); |
| 1961 |
|
| 1962 |
/* update approximate horizons with the computed horizons */ |
| 1963 |
GlobalVisUpdateApply(h); |
| 1964 |
} |
| 1965 |
|
| 1966 |
/* |
| 1967 |
* Determine what kind of visibility horizon needs to be used for a |
| 1968 |
* relation. If rel is NULL, the most conservative horizon is used. |
| 1969 |
*/ |
| 1970 |
static inline GlobalVisHorizonKind |
| 1971 |
GlobalVisHorizonKindForRel(Relation rel) |
| 1972 |
{ |
| 1973 |
/* |
| 1974 |
* Other relkinds currently don't contain xids, nor always the necessary |
| 1975 |
* logical decoding markers. |
| 1976 |
*/ |
| 1977 |
Assert(!rel || |
| 1978 |
rel->rd_rel->relkind == RELKIND_RELATION || |
| 1979 |
rel->rd_rel->relkind == RELKIND_MATVIEW || |
| 1980 |
rel->rd_rel->relkind == RELKIND_TOASTVALUE); |
| 1981 |
|
| 1982 |
if (rel == NULL || rel->rd_rel->relisshared || RecoveryInProgress()) |
| 1983 |
return VISHORIZON_SHARED; |
| 1984 |
else if (IsCatalogRelation(rel) || |
| 1985 |
RelationIsAccessibleInLogicalDecoding(rel)) |
| 1986 |
return VISHORIZON_CATALOG; |
| 1987 |
else if (!RELATION_IS_LOCAL(rel)) |
| 1988 |
return VISHORIZON_DATA; |
| 1989 |
else |
| 1990 |
return VISHORIZON_TEMP; |
| 1991 |
} |
| 1992 |
|
| 1993 |
/* |
| 1994 |
* Return the oldest XID for which deleted tuples must be preserved in the |
| 1995 |
* passed table. |
| 1996 |
* |
| 1997 |
* If rel is not NULL the horizon may be considerably more recent than |
| 1998 |
* otherwise (i.e. fewer tuples will be removable). In the NULL case a horizon |
| 1999 |
* that is correct (but not optimal) for all relations will be returned. |
| 2000 |
* |
| 2001 |
* This is used by VACUUM to decide which deleted tuples must be preserved in |
| 2002 |
* the passed in table. |
| 2003 |
*/ |
| 2004 |
TransactionId |
| 2005 |
GetOldestNonRemovableTransactionId(Relation rel) |
| 2006 |
{ |
| 2007 |
ComputeXidHorizonsResult horizons; |
| 2008 |
|
| 2009 |
ComputeXidHorizons(&horizons); |
| 2010 |
|
| 2011 |
switch (GlobalVisHorizonKindForRel(rel)) |
| 2012 |
{ |
| 2013 |
case VISHORIZON_SHARED: |
| 2014 |
return horizons.shared_oldest_nonremovable; |
| 2015 |
case VISHORIZON_CATALOG: |
| 2016 |
return horizons.catalog_oldest_nonremovable; |
| 2017 |
case VISHORIZON_DATA: |
| 2018 |
return horizons.data_oldest_nonremovable; |
| 2019 |
case VISHORIZON_TEMP: |
| 2020 |
return horizons.temp_oldest_nonremovable; |
| 2021 |
} |
| 2022 |
|
| 2023 |
/* just to prevent compiler warnings */ |
| 2024 |
return InvalidTransactionId; |
| 2025 |
} |
| 2026 |
|
| 2027 |
/* |
| 2028 |
* Return the oldest transaction id any currently running backend might still |
| 2029 |
* consider running. This should not be used for visibility / pruning |
| 2030 |
* determinations (see GetOldestNonRemovableTransactionId()), but for |
| 2031 |
* decisions like up to where pg_subtrans can be truncated. |
| 2032 |
*/ |
| 2033 |
TransactionId |
| 2034 |
GetOldestTransactionIdConsideredRunning(void) |
| 2035 |
{ |
| 2036 |
ComputeXidHorizonsResult horizons; |
| 2037 |
|
| 2038 |
ComputeXidHorizons(&horizons); |
| 2039 |
|
| 2040 |
return horizons.oldest_considered_running; |
| 2041 |
} |
| 2042 |
|
| 2043 |
/* |
| 2044 |
* Return the visibility horizons for a hot standby feedback message. |
| 2045 |
*/ |
| 2046 |
void |
| 2047 |
GetReplicationHorizons(TransactionId *xmin, TransactionId *catalog_xmin) |
| 2048 |
{ |
| 2049 |
ComputeXidHorizonsResult horizons; |
| 2050 |
|
| 2051 |
ComputeXidHorizons(&horizons); |
| 2052 |
|
| 2053 |
/* |
| 2054 |
* Don't want to use shared_oldest_nonremovable here, as that contains the |
| 2055 |
* effect of replication slot's catalog_xmin. We want to send a separate |
| 2056 |
* feedback for the catalog horizon, so the primary can remove data table |
| 2057 |
* contents more aggressively. |
| 2058 |
*/ |
| 2059 |
*xmin = horizons.shared_oldest_nonremovable_raw; |
| 2060 |
*catalog_xmin = horizons.slot_catalog_xmin; |
| 2061 |
} |
| 2062 |
|
| 2063 |
/* |
| 2064 |
* GetMaxSnapshotXidCount -- get max size for snapshot XID array |
| 2065 |
* |
| 2066 |
* We have to export this for use by snapmgr.c. |
| 2067 |
*/ |
| 2068 |
int |
| 2069 |
GetMaxSnapshotXidCount(void) |
| 2070 |
{ |
| 2071 |
return procArray->maxProcs; |
| 2072 |
} |
| 2073 |
|
| 2074 |
/* |
| 2075 |
* GetMaxSnapshotSubxidCount -- get max size for snapshot sub-XID array |
| 2076 |
* |
| 2077 |
* We have to export this for use by snapmgr.c. |
| 2078 |
*/ |
| 2079 |
int |
| 2080 |
GetMaxSnapshotSubxidCount(void) |
| 2081 |
{ |
| 2082 |
return TOTAL_MAX_CACHED_SUBXIDS; |
| 2083 |
} |
| 2084 |
|
| 2085 |
/* |
| 2086 |
* Helper function for GetSnapshotData() that checks if the bulk of the |
| 2087 |
* visibility information in the snapshot is still valid. If so, it updates |
| 2088 |
* the fields that need to change and returns true. Otherwise it returns |
| 2089 |
* false. |
| 2090 |
* |
| 2091 |
* This very likely can be evolved to not need ProcArrayLock held (at very |
| 2092 |
* least in the case we already hold a snapshot), but that's for another day. |
| 2093 |
*/ |
| 2094 |
static bool |
| 2095 |
GetSnapshotDataReuse(Snapshot snapshot) |
| 2096 |
{ |
| 2097 |
uint64 curXactCompletionCount; |
| 2098 |
|
| 2099 |
Assert(LWLockHeldByMe(ProcArrayLock)); |
| 2100 |
|
| 2101 |
if (unlikely(snapshot->snapXactCompletionCount == 0)) |
| 2102 |
return false; |
| 2103 |
|
| 2104 |
curXactCompletionCount = TransamVariables->xactCompletionCount; |
| 2105 |
if (curXactCompletionCount != snapshot->snapXactCompletionCount) |
| 2106 |
return false; |
| 2107 |
|
| 2108 |
/* |
| 2109 |
* If the current xactCompletionCount is still the same as it was at the |
| 2110 |
* time the snapshot was built, we can be sure that rebuilding the |
| 2111 |
* contents of the snapshot the hard way would result in the same snapshot |
| 2112 |
* contents: |
| 2113 |
* |
| 2114 |
* As explained in transam/README, the set of xids considered running by |
| 2115 |
* GetSnapshotData() cannot change while ProcArrayLock is held. Snapshot |
| 2116 |
* contents only depend on transactions with xids and xactCompletionCount |
| 2117 |
* is incremented whenever a transaction with an xid finishes (while |
| 2118 |
* holding ProcArrayLock exclusively). Thus the xactCompletionCount check |
| 2119 |
* ensures we would detect if the snapshot would have changed. |
| 2120 |
* |
| 2121 |
* As the snapshot contents are the same as it was before, it is safe to |
| 2122 |
* re-enter the snapshot's xmin into the PGPROC array. None of the rows |
| 2123 |
* visible under the snapshot could already have been removed (that'd |
| 2124 |
* require the set of running transactions to change) and it fulfills the |
| 2125 |
* requirement that concurrent GetSnapshotData() calls yield the same |
| 2126 |
* xmin. |
| 2127 |
*/ |
| 2128 |
if (!TransactionIdIsValid(MyProc->xmin)) |
| 2129 |
MyProc->xmin = TransactionXmin = snapshot->xmin; |
| 2130 |
|
| 2131 |
RecentXmin = snapshot->xmin; |
| 2132 |
Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin)); |
| 2133 |
|
| 2134 |
snapshot->curcid = GetCurrentCommandId(false); |
| 2135 |
snapshot->active_count = 0; |
| 2136 |
snapshot->regd_count = 0; |
| 2137 |
snapshot->copied = false; |
| 2138 |
|
| 2139 |
return true; |
| 2140 |
} |
| 2141 |
|
| 2142 |
/* |
| 2143 |
* GetSnapshotData -- returns information about running transactions. |
| 2144 |
* |
| 2145 |
* The returned snapshot includes xmin (lowest still-running xact ID), |
| 2146 |
* xmax (highest completed xact ID + 1), and a list of running xact IDs |
| 2147 |
* in the range xmin <= xid < xmax. It is used as follows: |
| 2148 |
* All xact IDs < xmin are considered finished. |
| 2149 |
* All xact IDs >= xmax are considered still running. |
| 2150 |
* For an xact ID xmin <= xid < xmax, consult list to see whether |
| 2151 |
* it is considered running or not. |
| 2152 |
* This ensures that the set of transactions seen as "running" by the |
| 2153 |
* current xact will not change after it takes the snapshot. |
| 2154 |
* |
| 2155 |
* All running top-level XIDs are included in the snapshot, except for lazy |
| 2156 |
* VACUUM processes. We also try to include running subtransaction XIDs, |
| 2157 |
* but since PGPROC has only a limited cache area for subxact XIDs, full |
| 2158 |
* information may not be available. If we find any overflowed subxid arrays, |
| 2159 |
* we have to mark the snapshot's subxid data as overflowed, and extra work |
| 2160 |
* *may* need to be done to determine what's running (see XidInMVCCSnapshot()). |
| 2161 |
* |
| 2162 |
* We also update the following backend-global variables: |
| 2163 |
* TransactionXmin: the oldest xmin of any snapshot in use in the |
| 2164 |
* current transaction (this is the same as MyProc->xmin). |
| 2165 |
* RecentXmin: the xmin computed for the most recent snapshot. XIDs |
| 2166 |
* older than this are known not running any more. |
| 2167 |
* |
| 2168 |
* And try to advance the bounds of GlobalVis{Shared,Catalog,Data,Temp}Rels |
| 2169 |
* for the benefit of the GlobalVisTest* family of functions. |
| 2170 |
* |
| 2171 |
* Note: this function should probably not be called with an argument that's |
| 2172 |
* not statically allocated (see xip allocation below). |
| 2173 |
*/ |
| 2174 |
Snapshot |
| 2175 |
GetSnapshotData(Snapshot snapshot) |
| 2176 |
{ |
| 2177 |
ProcArrayStruct *arrayP = procArray; |
| 2178 |
TransactionId *other_xids = ProcGlobal->xids; |
| 2179 |
TransactionId xmin; |
| 2180 |
TransactionId xmax; |
| 2181 |
int count = 0; |
| 2182 |
int subcount = 0; |
| 2183 |
bool suboverflowed = false; |
| 2184 |
FullTransactionId latest_completed; |
| 2185 |
TransactionId oldestxid; |
| 2186 |
int mypgxactoff; |
| 2187 |
TransactionId myxid; |
| 2188 |
uint64 curXactCompletionCount; |
| 2189 |
|
| 2190 |
TransactionId replication_slot_xmin = InvalidTransactionId; |
| 2191 |
TransactionId replication_slot_catalog_xmin = InvalidTransactionId; |
| 2192 |
|
| 2193 |
Assert(snapshot != NULL); |
| 2194 |
|
| 2195 |
/* |
| 2196 |
* Allocating space for maxProcs xids is usually overkill; numProcs would |
| 2197 |
* be sufficient. But it seems better to do the malloc while not holding |
| 2198 |
* the lock, so we can't look at numProcs. Likewise, we allocate much |
| 2199 |
* more subxip storage than is probably needed. |
| 2200 |
* |
| 2201 |
* This does open a possibility for avoiding repeated malloc/free: since |
| 2202 |
* maxProcs does not change at runtime, we can simply reuse the previous |
| 2203 |
* xip arrays if any. (This relies on the fact that all callers pass |
| 2204 |
* static SnapshotData structs.) |
| 2205 |
*/ |
| 2206 |
if (snapshot->xip == NULL) |
| 2207 |
{ |
| 2208 |
/* |
| 2209 |
* First call for this snapshot. Snapshot is same size whether or not |
| 2210 |
* we are in recovery, see later comments. |
| 2211 |
*/ |
| 2212 |
snapshot->xip = (TransactionId *) |
| 2213 |
malloc(GetMaxSnapshotXidCount() * sizeof(TransactionId)); |
| 2214 |
if (snapshot->xip == NULL) |
| 2215 |
ereport(ERROR, |
| 2216 |
(errcode(ERRCODE_OUT_OF_MEMORY), |
| 2217 |
errmsg("out of memory"))); |
| 2218 |
Assert(snapshot->subxip == NULL); |
| 2219 |
snapshot->subxip = (TransactionId *) |
| 2220 |
malloc(GetMaxSnapshotSubxidCount() * sizeof(TransactionId)); |
| 2221 |
if (snapshot->subxip == NULL) |
| 2222 |
ereport(ERROR, |
| 2223 |
(errcode(ERRCODE_OUT_OF_MEMORY), |
| 2224 |
errmsg("out of memory"))); |
| 2225 |
} |
| 2226 |
|
| 2227 |
/* |
| 2228 |
* It is sufficient to get shared lock on ProcArrayLock, even if we are |
| 2229 |
* going to set MyProc->xmin. |
| 2230 |
*/ |
| 2231 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 2232 |
|
| 2233 |
if (GetSnapshotDataReuse(snapshot)) |
| 2234 |
{ |
| 2235 |
LWLockRelease(ProcArrayLock); |
| 2236 |
return snapshot; |
| 2237 |
} |
| 2238 |
|
| 2239 |
latest_completed = TransamVariables->latestCompletedXid; |
| 2240 |
mypgxactoff = MyProc->pgxactoff; |
| 2241 |
myxid = other_xids[mypgxactoff]; |
| 2242 |
Assert(myxid == MyProc->xid); |
| 2243 |
|
| 2244 |
oldestxid = TransamVariables->oldestXid; |
| 2245 |
curXactCompletionCount = TransamVariables->xactCompletionCount; |
| 2246 |
|
| 2247 |
/* xmax is always latestCompletedXid + 1 */ |
| 2248 |
xmax = XidFromFullTransactionId(latest_completed); |
| 2249 |
TransactionIdAdvance(xmax); |
| 2250 |
Assert(TransactionIdIsNormal(xmax)); |
| 2251 |
|
| 2252 |
/* initialize xmin calculation with xmax */ |
| 2253 |
xmin = xmax; |
| 2254 |
|
| 2255 |
/* take own xid into account, saves a check inside the loop */ |
| 2256 |
if (TransactionIdIsNormal(myxid) && NormalTransactionIdPrecedes(myxid, xmin)) |
| 2257 |
xmin = myxid; |
| 2258 |
|
| 2259 |
snapshot->takenDuringRecovery = RecoveryInProgress(); |
| 2260 |
|
| 2261 |
if (!snapshot->takenDuringRecovery) |
| 2262 |
{ |
| 2263 |
int numProcs = arrayP->numProcs; |
| 2264 |
TransactionId *xip = snapshot->xip; |
| 2265 |
int *pgprocnos = arrayP->pgprocnos; |
| 2266 |
XidCacheStatus *subxidStates = ProcGlobal->subxidStates; |
| 2267 |
uint8 *allStatusFlags = ProcGlobal->statusFlags; |
| 2268 |
|
| 2269 |
/* |
| 2270 |
* First collect set of pgxactoff/xids that need to be included in the |
| 2271 |
* snapshot. |
| 2272 |
*/ |
| 2273 |
for (int pgxactoff = 0; pgxactoff < numProcs; pgxactoff++) |
| 2274 |
{ |
| 2275 |
/* Fetch xid just once - see GetNewTransactionId */ |
| 2276 |
TransactionId xid = UINT32_ACCESS_ONCE(other_xids[pgxactoff]); |
| 2277 |
uint8 statusFlags; |
| 2278 |
|
| 2279 |
Assert(allProcs[arrayP->pgprocnos[pgxactoff]].pgxactoff == pgxactoff); |
| 2280 |
|
| 2281 |
/* |
| 2282 |
* If the transaction has no XID assigned, we can skip it; it |
| 2283 |
* won't have sub-XIDs either. |
| 2284 |
*/ |
| 2285 |
if (likely(xid == InvalidTransactionId)) |
| 2286 |
continue; |
| 2287 |
|
| 2288 |
/* |
| 2289 |
* We don't include our own XIDs (if any) in the snapshot. It |
| 2290 |
* needs to be included in the xmin computation, but we did so |
| 2291 |
* outside the loop. |
| 2292 |
*/ |
| 2293 |
if (pgxactoff == mypgxactoff) |
| 2294 |
continue; |
| 2295 |
|
| 2296 |
/* |
| 2297 |
* The only way we are able to get here with a non-normal xid is |
| 2298 |
* during bootstrap - with this backend using |
| 2299 |
* BootstrapTransactionId. But the above test should filter that |
| 2300 |
* out. |
| 2301 |
*/ |
| 2302 |
Assert(TransactionIdIsNormal(xid)); |
| 2303 |
|
| 2304 |
/* |
| 2305 |
* If the XID is >= xmax, we can skip it; such transactions will |
| 2306 |
* be treated as running anyway (and any sub-XIDs will also be >= |
| 2307 |
* xmax). |
| 2308 |
*/ |
| 2309 |
if (!NormalTransactionIdPrecedes(xid, xmax)) |
| 2310 |
continue; |
| 2311 |
|
| 2312 |
/* |
| 2313 |
* Skip over backends doing logical decoding which manages xmin |
| 2314 |
* separately (check below) and ones running LAZY VACUUM. |
| 2315 |
*/ |
| 2316 |
statusFlags = allStatusFlags[pgxactoff]; |
| 2317 |
if (statusFlags & (PROC_IN_LOGICAL_DECODING | PROC_IN_VACUUM)) |
| 2318 |
continue; |
| 2319 |
|
| 2320 |
if (NormalTransactionIdPrecedes(xid, xmin)) |
| 2321 |
xmin = xid; |
| 2322 |
|
| 2323 |
/* Add XID to snapshot. */ |
| 2324 |
xip[count++] = xid; |
| 2325 |
|
| 2326 |
/* |
| 2327 |
* Save subtransaction XIDs if possible (if we've already |
| 2328 |
* overflowed, there's no point). Note that the subxact XIDs must |
| 2329 |
* be later than their parent, so no need to check them against |
| 2330 |
* xmin. We could filter against xmax, but it seems better not to |
| 2331 |
* do that much work while holding the ProcArrayLock. |
| 2332 |
* |
| 2333 |
* The other backend can add more subxids concurrently, but cannot |
| 2334 |
* remove any. Hence it's important to fetch nxids just once. |
| 2335 |
* Should be safe to use memcpy, though. (We needn't worry about |
| 2336 |
* missing any xids added concurrently, because they must postdate |
| 2337 |
* xmax.) |
| 2338 |
* |
| 2339 |
* Again, our own XIDs are not included in the snapshot. |
| 2340 |
*/ |
| 2341 |
if (!suboverflowed) |
| 2342 |
{ |
| 2343 |
|
| 2344 |
if (subxidStates[pgxactoff].overflowed) |
| 2345 |
suboverflowed = true; |
| 2346 |
else |
| 2347 |
{ |
| 2348 |
int nsubxids = subxidStates[pgxactoff].count; |
| 2349 |
|
| 2350 |
if (nsubxids > 0) |
| 2351 |
{ |
| 2352 |
int pgprocno = pgprocnos[pgxactoff]; |
| 2353 |
PGPROC *proc = &allProcs[pgprocno]; |
| 2354 |
|
| 2355 |
pg_read_barrier(); /* pairs with GetNewTransactionId */ |
| 2356 |
|
| 2357 |
memcpy(snapshot->subxip + subcount, |
| 2358 |
proc->subxids.xids, |
| 2359 |
nsubxids * sizeof(TransactionId)); |
| 2360 |
subcount += nsubxids; |
| 2361 |
} |
| 2362 |
} |
| 2363 |
} |
| 2364 |
} |
| 2365 |
} |
| 2366 |
else |
| 2367 |
{ |
| 2368 |
/* |
| 2369 |
* We're in hot standby, so get XIDs from KnownAssignedXids. |
| 2370 |
* |
| 2371 |
* We store all xids directly into subxip[]. Here's why: |
| 2372 |
* |
| 2373 |
* In recovery we don't know which xids are top-level and which are |
| 2374 |
* subxacts, a design choice that greatly simplifies xid processing. |
| 2375 |
* |
| 2376 |
* It seems like we would want to try to put xids into xip[] only, but |
| 2377 |
* that is fairly small. We would either need to make that bigger or |
| 2378 |
* to increase the rate at which we WAL-log xid assignment; neither is |
| 2379 |
* an appealing choice. |
| 2380 |
* |
| 2381 |
* We could try to store xids into xip[] first and then into subxip[] |
| 2382 |
* if there are too many xids. That only works if the snapshot doesn't |
| 2383 |
* overflow because we do not search subxip[] in that case. A simpler |
| 2384 |
* way is to just store all xids in the subxip array because this is |
| 2385 |
* by far the bigger array. We just leave the xip array empty. |
| 2386 |
* |
| 2387 |
* Either way we need to change the way XidInMVCCSnapshot() works |
| 2388 |
* depending upon when the snapshot was taken, or change normal |
| 2389 |
* snapshot processing so it matches. |
| 2390 |
* |
| 2391 |
* Note: It is possible for recovery to end before we finish taking |
| 2392 |
* the snapshot, and for newly assigned transaction ids to be added to |
| 2393 |
* the ProcArray. xmax cannot change while we hold ProcArrayLock, so |
| 2394 |
* those newly added transaction ids would be filtered away, so we |
| 2395 |
* need not be concerned about them. |
| 2396 |
*/ |
| 2397 |
subcount = KnownAssignedXidsGetAndSetXmin(snapshot->subxip, &xmin, |
| 2398 |
xmax); |
| 2399 |
|
| 2400 |
if (TransactionIdPrecedesOrEquals(xmin, procArray->lastOverflowedXid)) |
| 2401 |
suboverflowed = true; |
| 2402 |
} |
| 2403 |
|
| 2404 |
|
| 2405 |
/* |
| 2406 |
* Fetch into local variable while ProcArrayLock is held - the |
| 2407 |
* LWLockRelease below is a barrier, ensuring this happens inside the |
| 2408 |
* lock. |
| 2409 |
*/ |
| 2410 |
replication_slot_xmin = procArray->replication_slot_xmin; |
| 2411 |
replication_slot_catalog_xmin = procArray->replication_slot_catalog_xmin; |
| 2412 |
|
| 2413 |
if (!TransactionIdIsValid(MyProc->xmin)) |
| 2414 |
MyProc->xmin = TransactionXmin = xmin; |
| 2415 |
|
| 2416 |
LWLockRelease(ProcArrayLock); |
| 2417 |
|
| 2418 |
/* maintain state for GlobalVis* */ |
| 2419 |
{ |
| 2420 |
TransactionId def_vis_xid; |
| 2421 |
TransactionId def_vis_xid_data; |
| 2422 |
FullTransactionId def_vis_fxid; |
| 2423 |
FullTransactionId def_vis_fxid_data; |
| 2424 |
FullTransactionId oldestfxid; |
| 2425 |
|
| 2426 |
/* |
| 2427 |
* Converting oldestXid is only safe when xid horizon cannot advance, |
| 2428 |
* i.e. holding locks. While we don't hold the lock anymore, all the |
| 2429 |
* necessary data has been gathered with lock held. |
| 2430 |
*/ |
| 2431 |
oldestfxid = FullXidRelativeTo(latest_completed, oldestxid); |
| 2432 |
|
| 2433 |
/* Check whether there's a replication slot requiring an older xmin. */ |
| 2434 |
def_vis_xid_data = |
| 2435 |
TransactionIdOlder(xmin, replication_slot_xmin); |
| 2436 |
|
| 2437 |
/* |
| 2438 |
* Rows in non-shared, non-catalog tables possibly could be vacuumed |
| 2439 |
* if older than this xid. |
| 2440 |
*/ |
| 2441 |
def_vis_xid = def_vis_xid_data; |
| 2442 |
|
| 2443 |
/* |
| 2444 |
* Check whether there's a replication slot requiring an older catalog |
| 2445 |
* xmin. |
| 2446 |
*/ |
| 2447 |
def_vis_xid = |
| 2448 |
TransactionIdOlder(replication_slot_catalog_xmin, def_vis_xid); |
| 2449 |
|
| 2450 |
def_vis_fxid = FullXidRelativeTo(latest_completed, def_vis_xid); |
| 2451 |
def_vis_fxid_data = FullXidRelativeTo(latest_completed, def_vis_xid_data); |
| 2452 |
|
| 2453 |
/* |
| 2454 |
* Check if we can increase upper bound. As a previous |
| 2455 |
* GlobalVisUpdate() might have computed more aggressive values, don't |
| 2456 |
* overwrite them if so. |
| 2457 |
*/ |
| 2458 |
GlobalVisSharedRels.definitely_needed = |
| 2459 |
FullTransactionIdNewer(def_vis_fxid, |
| 2460 |
GlobalVisSharedRels.definitely_needed); |
| 2461 |
GlobalVisCatalogRels.definitely_needed = |
| 2462 |
FullTransactionIdNewer(def_vis_fxid, |
| 2463 |
GlobalVisCatalogRels.definitely_needed); |
| 2464 |
GlobalVisDataRels.definitely_needed = |
| 2465 |
FullTransactionIdNewer(def_vis_fxid_data, |
| 2466 |
GlobalVisDataRels.definitely_needed); |
| 2467 |
/* See temp_oldest_nonremovable computation in ComputeXidHorizons() */ |
| 2468 |
if (TransactionIdIsNormal(myxid)) |
| 2469 |
GlobalVisTempRels.definitely_needed = |
| 2470 |
FullXidRelativeTo(latest_completed, myxid); |
| 2471 |
else |
| 2472 |
{ |
| 2473 |
GlobalVisTempRels.definitely_needed = latest_completed; |
| 2474 |
FullTransactionIdAdvance(&GlobalVisTempRels.definitely_needed); |
| 2475 |
} |
| 2476 |
|
| 2477 |
/* |
| 2478 |
* Check if we know that we can initialize or increase the lower |
| 2479 |
* bound. Currently the only cheap way to do so is to use |
| 2480 |
* TransamVariables->oldestXid as input. |
| 2481 |
* |
| 2482 |
* We should definitely be able to do better. We could e.g. put a |
| 2483 |
* global lower bound value into TransamVariables. |
| 2484 |
*/ |
| 2485 |
GlobalVisSharedRels.maybe_needed = |
| 2486 |
FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed, |
| 2487 |
oldestfxid); |
| 2488 |
GlobalVisCatalogRels.maybe_needed = |
| 2489 |
FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed, |
| 2490 |
oldestfxid); |
| 2491 |
GlobalVisDataRels.maybe_needed = |
| 2492 |
FullTransactionIdNewer(GlobalVisDataRels.maybe_needed, |
| 2493 |
oldestfxid); |
| 2494 |
/* accurate value known */ |
| 2495 |
GlobalVisTempRels.maybe_needed = GlobalVisTempRels.definitely_needed; |
| 2496 |
} |
| 2497 |
|
| 2498 |
RecentXmin = xmin; |
| 2499 |
Assert(TransactionIdPrecedesOrEquals(TransactionXmin, RecentXmin)); |
| 2500 |
|
| 2501 |
snapshot->xmin = xmin; |
| 2502 |
snapshot->xmax = xmax; |
| 2503 |
snapshot->xcnt = count; |
| 2504 |
snapshot->subxcnt = subcount; |
| 2505 |
snapshot->suboverflowed = suboverflowed; |
| 2506 |
snapshot->snapXactCompletionCount = curXactCompletionCount; |
| 2507 |
|
| 2508 |
snapshot->curcid = GetCurrentCommandId(false); |
| 2509 |
|
| 2510 |
/* |
| 2511 |
* This is a new snapshot, so set both refcounts are zero, and mark it as |
| 2512 |
* not copied in persistent memory. |
| 2513 |
*/ |
| 2514 |
snapshot->active_count = 0; |
| 2515 |
snapshot->regd_count = 0; |
| 2516 |
snapshot->copied = false; |
| 2517 |
|
| 2518 |
return snapshot; |
| 2519 |
} |
| 2520 |
|
| 2521 |
/* |
| 2522 |
* ProcArrayInstallImportedXmin -- install imported xmin into MyProc->xmin |
| 2523 |
* |
| 2524 |
* This is called when installing a snapshot imported from another |
| 2525 |
* transaction. To ensure that OldestXmin doesn't go backwards, we must |
| 2526 |
* check that the source transaction is still running, and we'd better do |
| 2527 |
* that atomically with installing the new xmin. |
| 2528 |
* |
| 2529 |
* Returns true if successful, false if source xact is no longer running. |
| 2530 |
*/ |
| 2531 |
bool |
| 2532 |
ProcArrayInstallImportedXmin(TransactionId xmin, |
| 2533 |
VirtualTransactionId *sourcevxid) |
| 2534 |
{ |
| 2535 |
bool result = false; |
| 2536 |
ProcArrayStruct *arrayP = procArray; |
| 2537 |
int index; |
| 2538 |
|
| 2539 |
Assert(TransactionIdIsNormal(xmin)); |
| 2540 |
if (!sourcevxid) |
| 2541 |
return false; |
| 2542 |
|
| 2543 |
/* Get lock so source xact can't end while we're doing this */ |
| 2544 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 2545 |
|
| 2546 |
/* |
| 2547 |
* Find the PGPROC entry of the source transaction. (This could use |
| 2548 |
* GetPGProcByNumber(), unless it's a prepared xact. But this isn't |
| 2549 |
* performance critical.) |
| 2550 |
*/ |
| 2551 |
for (index = 0; index < arrayP->numProcs; index++) |
| 2552 |
{ |
| 2553 |
int pgprocno = arrayP->pgprocnos[index]; |
| 2554 |
PGPROC *proc = &allProcs[pgprocno]; |
| 2555 |
int statusFlags = ProcGlobal->statusFlags[index]; |
| 2556 |
TransactionId xid; |
| 2557 |
|
| 2558 |
/* Ignore procs running LAZY VACUUM */ |
| 2559 |
if (statusFlags & PROC_IN_VACUUM) |
| 2560 |
continue; |
| 2561 |
|
| 2562 |
/* We are only interested in the specific virtual transaction. */ |
| 2563 |
if (proc->vxid.procNumber != sourcevxid->procNumber) |
| 2564 |
continue; |
| 2565 |
if (proc->vxid.lxid != sourcevxid->localTransactionId) |
| 2566 |
continue; |
| 2567 |
|
| 2568 |
/* |
| 2569 |
* We check the transaction's database ID for paranoia's sake: if it's |
| 2570 |
* in another DB then its xmin does not cover us. Caller should have |
| 2571 |
* detected this already, so we just treat any funny cases as |
| 2572 |
* "transaction not found". |
| 2573 |
*/ |
| 2574 |
if (proc->databaseId != MyDatabaseId) |
| 2575 |
continue; |
| 2576 |
|
| 2577 |
/* |
| 2578 |
* Likewise, let's just make real sure its xmin does cover us. |
| 2579 |
*/ |
| 2580 |
xid = UINT32_ACCESS_ONCE(proc->xmin); |
| 2581 |
if (!TransactionIdIsNormal(xid) || |
| 2582 |
!TransactionIdPrecedesOrEquals(xid, xmin)) |
| 2583 |
continue; |
| 2584 |
|
| 2585 |
/* |
| 2586 |
* We're good. Install the new xmin. As in GetSnapshotData, set |
| 2587 |
* TransactionXmin too. (Note that because snapmgr.c called |
| 2588 |
* GetSnapshotData first, we'll be overwriting a valid xmin here, so |
| 2589 |
* we don't check that.) |
| 2590 |
*/ |
| 2591 |
MyProc->xmin = TransactionXmin = xmin; |
| 2592 |
|
| 2593 |
result = true; |
| 2594 |
break; |
| 2595 |
} |
| 2596 |
|
| 2597 |
LWLockRelease(ProcArrayLock); |
| 2598 |
|
| 2599 |
return result; |
| 2600 |
} |
| 2601 |
|
| 2602 |
/* |
| 2603 |
* ProcArrayInstallRestoredXmin -- install restored xmin into MyProc->xmin |
| 2604 |
* |
| 2605 |
* This is like ProcArrayInstallImportedXmin, but we have a pointer to the |
| 2606 |
* PGPROC of the transaction from which we imported the snapshot, rather than |
| 2607 |
* an XID. |
| 2608 |
* |
| 2609 |
* Note that this function also copies statusFlags from the source `proc` in |
| 2610 |
* order to avoid the case where MyProc's xmin needs to be skipped for |
| 2611 |
* computing xid horizon. |
| 2612 |
* |
| 2613 |
* Returns true if successful, false if source xact is no longer running. |
| 2614 |
*/ |
| 2615 |
bool |
| 2616 |
ProcArrayInstallRestoredXmin(TransactionId xmin, PGPROC *proc) |
| 2617 |
{ |
| 2618 |
bool result = false; |
| 2619 |
TransactionId xid; |
| 2620 |
|
| 2621 |
Assert(TransactionIdIsNormal(xmin)); |
| 2622 |
Assert(proc != NULL); |
| 2623 |
|
| 2624 |
/* |
| 2625 |
* Get an exclusive lock so that we can copy statusFlags from source proc. |
| 2626 |
*/ |
| 2627 |
LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE); |
| 2628 |
|
| 2629 |
/* |
| 2630 |
* Be certain that the referenced PGPROC has an advertised xmin which is |
| 2631 |
* no later than the one we're installing, so that the system-wide xmin |
| 2632 |
* can't go backwards. Also, make sure it's running in the same database, |
| 2633 |
* so that the per-database xmin cannot go backwards. |
| 2634 |
*/ |
| 2635 |
xid = UINT32_ACCESS_ONCE(proc->xmin); |
| 2636 |
if (proc->databaseId == MyDatabaseId && |
| 2637 |
TransactionIdIsNormal(xid) && |
| 2638 |
TransactionIdPrecedesOrEquals(xid, xmin)) |
| 2639 |
{ |
| 2640 |
/* |
| 2641 |
* Install xmin and propagate the statusFlags that affect how the |
| 2642 |
* value is interpreted by vacuum. |
| 2643 |
*/ |
| 2644 |
MyProc->xmin = TransactionXmin = xmin; |
| 2645 |
MyProc->statusFlags = (MyProc->statusFlags & ~PROC_XMIN_FLAGS) | |
| 2646 |
(proc->statusFlags & PROC_XMIN_FLAGS); |
| 2647 |
ProcGlobal->statusFlags[MyProc->pgxactoff] = MyProc->statusFlags; |
| 2648 |
|
| 2649 |
result = true; |
| 2650 |
} |
| 2651 |
|
| 2652 |
LWLockRelease(ProcArrayLock); |
| 2653 |
|
| 2654 |
return result; |
| 2655 |
} |
| 2656 |
|
| 2657 |
/* |
| 2658 |
* GetRunningTransactionData -- returns information about running transactions. |
| 2659 |
* |
| 2660 |
* Similar to GetSnapshotData but returns more information. We include |
| 2661 |
* all PGPROCs with an assigned TransactionId, even VACUUM processes and |
| 2662 |
* prepared transactions. |
| 2663 |
* |
| 2664 |
* We acquire XidGenLock and ProcArrayLock, but the caller is responsible for |
| 2665 |
* releasing them. Acquiring XidGenLock ensures that no new XIDs enter the proc |
| 2666 |
* array until the caller has WAL-logged this snapshot, and releases the |
| 2667 |
* lock. Acquiring ProcArrayLock ensures that no transactions commit until the |
| 2668 |
* lock is released. |
| 2669 |
* |
| 2670 |
* The returned data structure is statically allocated; caller should not |
| 2671 |
* modify it, and must not assume it is valid past the next call. |
| 2672 |
* |
| 2673 |
* This is never executed during recovery so there is no need to look at |
| 2674 |
* KnownAssignedXids. |
| 2675 |
* |
| 2676 |
* Dummy PGPROCs from prepared transaction are included, meaning that this |
| 2677 |
* may return entries with duplicated TransactionId values coming from |
| 2678 |
* transaction finishing to prepare. Nothing is done about duplicated |
| 2679 |
* entries here to not hold on ProcArrayLock more than necessary. |
| 2680 |
* |
| 2681 |
* We don't worry about updating other counters, we want to keep this as |
| 2682 |
* simple as possible and leave GetSnapshotData() as the primary code for |
| 2683 |
* that bookkeeping. |
| 2684 |
* |
| 2685 |
* Note that if any transaction has overflowed its cached subtransactions |
| 2686 |
* then there is no real need include any subtransactions. |
| 2687 |
*/ |
| 2688 |
RunningTransactions |
| 2689 |
GetRunningTransactionData(void) |
| 2690 |
{ |
| 2691 |
/* result workspace */ |
| 2692 |
static RunningTransactionsData CurrentRunningXactsData; |
| 2693 |
|
| 2694 |
ProcArrayStruct *arrayP = procArray; |
| 2695 |
TransactionId *other_xids = ProcGlobal->xids; |
| 2696 |
RunningTransactions CurrentRunningXacts = &CurrentRunningXactsData; |
| 2697 |
TransactionId latestCompletedXid; |
| 2698 |
TransactionId oldestRunningXid; |
| 2699 |
TransactionId oldestDatabaseRunningXid; |
| 2700 |
TransactionId *xids; |
| 2701 |
int index; |
| 2702 |
int count; |
| 2703 |
int subcount; |
| 2704 |
bool suboverflowed; |
| 2705 |
|
| 2706 |
Assert(!RecoveryInProgress()); |
| 2707 |
|
| 2708 |
/* |
| 2709 |
* Allocating space for maxProcs xids is usually overkill; numProcs would |
| 2710 |
* be sufficient. But it seems better to do the malloc while not holding |
| 2711 |
* the lock, so we can't look at numProcs. Likewise, we allocate much |
| 2712 |
* more subxip storage than is probably needed. |
| 2713 |
* |
| 2714 |
* Should only be allocated in bgwriter, since only ever executed during |
| 2715 |
* checkpoints. |
| 2716 |
*/ |
| 2717 |
if (CurrentRunningXacts->xids == NULL) |
| 2718 |
{ |
| 2719 |
/* |
| 2720 |
* First call |
| 2721 |
*/ |
| 2722 |
CurrentRunningXacts->xids = (TransactionId *) |
| 2723 |
malloc(TOTAL_MAX_CACHED_SUBXIDS * sizeof(TransactionId)); |
| 2724 |
if (CurrentRunningXacts->xids == NULL) |
| 2725 |
ereport(ERROR, |
| 2726 |
(errcode(ERRCODE_OUT_OF_MEMORY), |
| 2727 |
errmsg("out of memory"))); |
| 2728 |
} |
| 2729 |
|
| 2730 |
xids = CurrentRunningXacts->xids; |
| 2731 |
|
| 2732 |
count = subcount = 0; |
| 2733 |
suboverflowed = false; |
| 2734 |
|
| 2735 |
/* |
| 2736 |
* Ensure that no xids enter or leave the procarray while we obtain |
| 2737 |
* snapshot. |
| 2738 |
*/ |
| 2739 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 2740 |
LWLockAcquire(XidGenLock, LW_SHARED); |
| 2741 |
|
| 2742 |
latestCompletedXid = |
| 2743 |
XidFromFullTransactionId(TransamVariables->latestCompletedXid); |
| 2744 |
oldestDatabaseRunningXid = oldestRunningXid = |
| 2745 |
XidFromFullTransactionId(TransamVariables->nextXid); |
| 2746 |
|
| 2747 |
/* |
| 2748 |
* Spin over procArray collecting all xids |
| 2749 |
*/ |
| 2750 |
for (index = 0; index < arrayP->numProcs; index++) |
| 2751 |
{ |
| 2752 |
TransactionId xid; |
| 2753 |
|
| 2754 |
/* Fetch xid just once - see GetNewTransactionId */ |
| 2755 |
xid = UINT32_ACCESS_ONCE(other_xids[index]); |
| 2756 |
|
| 2757 |
/* |
| 2758 |
* We don't need to store transactions that don't have a TransactionId |
| 2759 |
* yet because they will not show as running on a standby server. |
| 2760 |
*/ |
| 2761 |
if (!TransactionIdIsValid(xid)) |
| 2762 |
continue; |
| 2763 |
|
| 2764 |
/* |
| 2765 |
* Be careful not to exclude any xids before calculating the values of |
| 2766 |
* oldestRunningXid and suboverflowed, since these are used to clean |
| 2767 |
* up transaction information held on standbys. |
| 2768 |
*/ |
| 2769 |
if (TransactionIdPrecedes(xid, oldestRunningXid)) |
| 2770 |
oldestRunningXid = xid; |
| 2771 |
|
| 2772 |
/* |
| 2773 |
* Also, update the oldest running xid within the current database. As |
| 2774 |
* fetching pgprocno and PGPROC could cause cache misses, we do cheap |
| 2775 |
* TransactionId comparison first. |
| 2776 |
*/ |
| 2777 |
if (TransactionIdPrecedes(xid, oldestDatabaseRunningXid)) |
| 2778 |
{ |
| 2779 |
int pgprocno = arrayP->pgprocnos[index]; |
| 2780 |
PGPROC *proc = &allProcs[pgprocno]; |
| 2781 |
|
| 2782 |
if (proc->databaseId == MyDatabaseId) |
| 2783 |
oldestDatabaseRunningXid = xid; |
| 2784 |
} |
| 2785 |
|
| 2786 |
if (ProcGlobal->subxidStates[index].overflowed) |
| 2787 |
suboverflowed = true; |
| 2788 |
|
| 2789 |
/* |
| 2790 |
* If we wished to exclude xids this would be the right place for it. |
| 2791 |
* Procs with the PROC_IN_VACUUM flag set don't usually assign xids, |
| 2792 |
* but they do during truncation at the end when they get the lock and |
| 2793 |
* truncate, so it is not much of a problem to include them if they |
| 2794 |
* are seen and it is cleaner to include them. |
| 2795 |
*/ |
| 2796 |
|
| 2797 |
xids[count++] = xid; |
| 2798 |
} |
| 2799 |
|
| 2800 |
/* |
| 2801 |
* Spin over procArray collecting all subxids, but only if there hasn't |
| 2802 |
* been a suboverflow. |
| 2803 |
*/ |
| 2804 |
if (!suboverflowed) |
| 2805 |
{ |
| 2806 |
XidCacheStatus *other_subxidstates = ProcGlobal->subxidStates; |
| 2807 |
|
| 2808 |
for (index = 0; index < arrayP->numProcs; index++) |
| 2809 |
{ |
| 2810 |
int pgprocno = arrayP->pgprocnos[index]; |
| 2811 |
PGPROC *proc = &allProcs[pgprocno]; |
| 2812 |
int nsubxids; |
| 2813 |
|
| 2814 |
/* |
| 2815 |
* Save subtransaction XIDs. Other backends can't add or remove |
| 2816 |
* entries while we're holding XidGenLock. |
| 2817 |
*/ |
| 2818 |
nsubxids = other_subxidstates[index].count; |
| 2819 |
if (nsubxids > 0) |
| 2820 |
{ |
| 2821 |
/* barrier not really required, as XidGenLock is held, but ... */ |
| 2822 |
pg_read_barrier(); /* pairs with GetNewTransactionId */ |
| 2823 |
|
| 2824 |
memcpy(&xids[count], proc->subxids.xids, |
| 2825 |
nsubxids * sizeof(TransactionId)); |
| 2826 |
count += nsubxids; |
| 2827 |
subcount += nsubxids; |
| 2828 |
|
| 2829 |
/* |
| 2830 |
* Top-level XID of a transaction is always less than any of |
| 2831 |
* its subxids, so we don't need to check if any of the |
| 2832 |
* subxids are smaller than oldestRunningXid |
| 2833 |
*/ |
| 2834 |
} |
| 2835 |
} |
| 2836 |
} |
| 2837 |
|
| 2838 |
/* |
| 2839 |
* It's important *not* to include the limits set by slots here because |
| 2840 |
* snapbuild.c uses oldestRunningXid to manage its xmin horizon. If those |
| 2841 |
* were to be included here the initial value could never increase because |
| 2842 |
* of a circular dependency where slots only increase their limits when |
| 2843 |
* running xacts increases oldestRunningXid and running xacts only |
| 2844 |
* increases if slots do. |
| 2845 |
*/ |
| 2846 |
|
| 2847 |
CurrentRunningXacts->xcnt = count - subcount; |
| 2848 |
CurrentRunningXacts->subxcnt = subcount; |
| 2849 |
CurrentRunningXacts->subxid_status = suboverflowed ? SUBXIDS_IN_SUBTRANS : SUBXIDS_IN_ARRAY; |
| 2850 |
CurrentRunningXacts->nextXid = XidFromFullTransactionId(TransamVariables->nextXid); |
| 2851 |
CurrentRunningXacts->oldestRunningXid = oldestRunningXid; |
| 2852 |
CurrentRunningXacts->oldestDatabaseRunningXid = oldestDatabaseRunningXid; |
| 2853 |
CurrentRunningXacts->latestCompletedXid = latestCompletedXid; |
| 2854 |
|
| 2855 |
Assert(TransactionIdIsValid(CurrentRunningXacts->nextXid)); |
| 2856 |
Assert(TransactionIdIsValid(CurrentRunningXacts->oldestRunningXid)); |
| 2857 |
Assert(TransactionIdIsNormal(CurrentRunningXacts->latestCompletedXid)); |
| 2858 |
|
| 2859 |
/* We don't release the locks here, the caller is responsible for that */ |
| 2860 |
|
| 2861 |
return CurrentRunningXacts; |
| 2862 |
} |
| 2863 |
|
| 2864 |
/* |
| 2865 |
* GetOldestActiveTransactionId() |
| 2866 |
* |
| 2867 |
* Similar to GetSnapshotData but returns just oldestActiveXid. We include |
| 2868 |
* all PGPROCs with an assigned TransactionId, even VACUUM processes. |
| 2869 |
* We look at all databases, though there is no need to include WALSender |
| 2870 |
* since this has no effect on hot standby conflicts. |
| 2871 |
* |
| 2872 |
* This is never executed during recovery so there is no need to look at |
| 2873 |
* KnownAssignedXids. |
| 2874 |
* |
| 2875 |
* We don't worry about updating other counters, we want to keep this as |
| 2876 |
* simple as possible and leave GetSnapshotData() as the primary code for |
| 2877 |
* that bookkeeping. |
| 2878 |
*/ |
| 2879 |
TransactionId |
| 2880 |
GetOldestActiveTransactionId(void) |
| 2881 |
{ |
| 2882 |
ProcArrayStruct *arrayP = procArray; |
| 2883 |
TransactionId *other_xids = ProcGlobal->xids; |
| 2884 |
TransactionId oldestRunningXid; |
| 2885 |
int index; |
| 2886 |
|
| 2887 |
Assert(!RecoveryInProgress()); |
| 2888 |
|
| 2889 |
/* |
| 2890 |
* Read nextXid, as the upper bound of what's still active. |
| 2891 |
* |
| 2892 |
* Reading a TransactionId is atomic, but we must grab the lock to make |
| 2893 |
* sure that all XIDs < nextXid are already present in the proc array (or |
| 2894 |
* have already completed), when we spin over it. |
| 2895 |
*/ |
| 2896 |
LWLockAcquire(XidGenLock, LW_SHARED); |
| 2897 |
oldestRunningXid = XidFromFullTransactionId(TransamVariables->nextXid); |
| 2898 |
LWLockRelease(XidGenLock); |
| 2899 |
|
| 2900 |
/* |
| 2901 |
* Spin over procArray collecting all xids and subxids. |
| 2902 |
*/ |
| 2903 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 2904 |
for (index = 0; index < arrayP->numProcs; index++) |
| 2905 |
{ |
| 2906 |
TransactionId xid; |
| 2907 |
|
| 2908 |
/* Fetch xid just once - see GetNewTransactionId */ |
| 2909 |
xid = UINT32_ACCESS_ONCE(other_xids[index]); |
| 2910 |
|
| 2911 |
if (!TransactionIdIsNormal(xid)) |
| 2912 |
continue; |
| 2913 |
|
| 2914 |
if (TransactionIdPrecedes(xid, oldestRunningXid)) |
| 2915 |
oldestRunningXid = xid; |
| 2916 |
|
| 2917 |
/* |
| 2918 |
* Top-level XID of a transaction is always less than any of its |
| 2919 |
* subxids, so we don't need to check if any of the subxids are |
| 2920 |
* smaller than oldestRunningXid |
| 2921 |
*/ |
| 2922 |
} |
| 2923 |
LWLockRelease(ProcArrayLock); |
| 2924 |
|
| 2925 |
return oldestRunningXid; |
| 2926 |
} |
| 2927 |
|
| 2928 |
/* |
| 2929 |
* GetOldestSafeDecodingTransactionId -- lowest xid not affected by vacuum |
| 2930 |
* |
| 2931 |
* Returns the oldest xid that we can guarantee not to have been affected by |
| 2932 |
* vacuum, i.e. no rows >= that xid have been vacuumed away unless the |
| 2933 |
* transaction aborted. Note that the value can (and most of the time will) be |
| 2934 |
* much more conservative than what really has been affected by vacuum, but we |
| 2935 |
* currently don't have better data available. |
| 2936 |
* |
| 2937 |
* This is useful to initialize the cutoff xid after which a new changeset |
| 2938 |
* extraction replication slot can start decoding changes. |
| 2939 |
* |
| 2940 |
* Must be called with ProcArrayLock held either shared or exclusively, |
| 2941 |
* although most callers will want to use exclusive mode since it is expected |
| 2942 |
* that the caller will immediately use the xid to peg the xmin horizon. |
| 2943 |
*/ |
| 2944 |
TransactionId |
| 2945 |
GetOldestSafeDecodingTransactionId(bool catalogOnly) |
| 2946 |
{ |
| 2947 |
ProcArrayStruct *arrayP = procArray; |
| 2948 |
TransactionId oldestSafeXid; |
| 2949 |
int index; |
| 2950 |
bool recovery_in_progress = RecoveryInProgress(); |
| 2951 |
|
| 2952 |
Assert(LWLockHeldByMe(ProcArrayLock)); |
| 2953 |
|
| 2954 |
/* |
| 2955 |
* Acquire XidGenLock, so no transactions can acquire an xid while we're |
| 2956 |
* running. If no transaction with xid were running concurrently a new xid |
| 2957 |
* could influence the RecentXmin et al. |
| 2958 |
* |
| 2959 |
* We initialize the computation to nextXid since that's guaranteed to be |
| 2960 |
* a safe, albeit pessimal, value. |
| 2961 |
*/ |
| 2962 |
LWLockAcquire(XidGenLock, LW_SHARED); |
| 2963 |
oldestSafeXid = XidFromFullTransactionId(TransamVariables->nextXid); |
| 2964 |
|
| 2965 |
/* |
| 2966 |
* If there's already a slot pegging the xmin horizon, we can start with |
| 2967 |
* that value, it's guaranteed to be safe since it's computed by this |
| 2968 |
* routine initially and has been enforced since. We can always use the |
| 2969 |
* slot's general xmin horizon, but the catalog horizon is only usable |
| 2970 |
* when only catalog data is going to be looked at. |
| 2971 |
*/ |
| 2972 |
if (TransactionIdIsValid(procArray->replication_slot_xmin) && |
| 2973 |
TransactionIdPrecedes(procArray->replication_slot_xmin, |
| 2974 |
oldestSafeXid)) |
| 2975 |
oldestSafeXid = procArray->replication_slot_xmin; |
| 2976 |
|
| 2977 |
if (catalogOnly && |
| 2978 |
TransactionIdIsValid(procArray->replication_slot_catalog_xmin) && |
| 2979 |
TransactionIdPrecedes(procArray->replication_slot_catalog_xmin, |
| 2980 |
oldestSafeXid)) |
| 2981 |
oldestSafeXid = procArray->replication_slot_catalog_xmin; |
| 2982 |
|
| 2983 |
/* |
| 2984 |
* If we're not in recovery, we walk over the procarray and collect the |
| 2985 |
* lowest xid. Since we're called with ProcArrayLock held and have |
| 2986 |
* acquired XidGenLock, no entries can vanish concurrently, since |
| 2987 |
* ProcGlobal->xids[i] is only set with XidGenLock held and only cleared |
| 2988 |
* with ProcArrayLock held. |
| 2989 |
* |
| 2990 |
* In recovery we can't lower the safe value besides what we've computed |
| 2991 |
* above, so we'll have to wait a bit longer there. We unfortunately can |
| 2992 |
* *not* use KnownAssignedXidsGetOldestXmin() since the KnownAssignedXids |
| 2993 |
* machinery can miss values and return an older value than is safe. |
| 2994 |
*/ |
| 2995 |
if (!recovery_in_progress) |
| 2996 |
{ |
| 2997 |
TransactionId *other_xids = ProcGlobal->xids; |
| 2998 |
|
| 2999 |
/* |
| 3000 |
* Spin over procArray collecting min(ProcGlobal->xids[i]) |
| 3001 |
*/ |
| 3002 |
for (index = 0; index < arrayP->numProcs; index++) |
| 3003 |
{ |
| 3004 |
TransactionId xid; |
| 3005 |
|
| 3006 |
/* Fetch xid just once - see GetNewTransactionId */ |
| 3007 |
xid = UINT32_ACCESS_ONCE(other_xids[index]); |
| 3008 |
|
| 3009 |
if (!TransactionIdIsNormal(xid)) |
| 3010 |
continue; |
| 3011 |
|
| 3012 |
if (TransactionIdPrecedes(xid, oldestSafeXid)) |
| 3013 |
oldestSafeXid = xid; |
| 3014 |
} |
| 3015 |
} |
| 3016 |
|
| 3017 |
LWLockRelease(XidGenLock); |
| 3018 |
|
| 3019 |
return oldestSafeXid; |
| 3020 |
} |
| 3021 |
|
| 3022 |
/* |
| 3023 |
* GetVirtualXIDsDelayingChkpt -- Get the VXIDs of transactions that are |
| 3024 |
* delaying checkpoint because they have critical actions in progress. |
| 3025 |
* |
| 3026 |
* Constructs an array of VXIDs of transactions that are currently in commit |
| 3027 |
* critical sections, as shown by having specified delayChkptFlags bits set |
| 3028 |
* in their PGPROC. |
| 3029 |
* |
| 3030 |
* Returns a palloc'd array that should be freed by the caller. |
| 3031 |
* *nvxids is the number of valid entries. |
| 3032 |
* |
| 3033 |
* Note that because backends set or clear delayChkptFlags without holding any |
| 3034 |
* lock, the result is somewhat indeterminate, but we don't really care. Even |
| 3035 |
* in a multiprocessor with delayed writes to shared memory, it should be |
| 3036 |
* certain that setting of delayChkptFlags will propagate to shared memory |
| 3037 |
* when the backend takes a lock, so we cannot fail to see a virtual xact as |
| 3038 |
* delayChkptFlags if it's already inserted its commit record. Whether it |
| 3039 |
* takes a little while for clearing of delayChkptFlags to propagate is |
| 3040 |
* unimportant for correctness. |
| 3041 |
*/ |
| 3042 |
VirtualTransactionId * |
| 3043 |
GetVirtualXIDsDelayingChkpt(int *nvxids, int type) |
| 3044 |
{ |
| 3045 |
VirtualTransactionId *vxids; |
| 3046 |
ProcArrayStruct *arrayP = procArray; |
| 3047 |
int count = 0; |
| 3048 |
int index; |
| 3049 |
|
| 3050 |
Assert(type != 0); |
| 3051 |
|
| 3052 |
/* allocate what's certainly enough result space */ |
| 3053 |
vxids = (VirtualTransactionId *) |
| 3054 |
palloc(sizeof(VirtualTransactionId) * arrayP->maxProcs); |
| 3055 |
|
| 3056 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 3057 |
|
| 3058 |
for (index = 0; index < arrayP->numProcs; index++) |
| 3059 |
{ |
| 3060 |
int pgprocno = arrayP->pgprocnos[index]; |
| 3061 |
PGPROC *proc = &allProcs[pgprocno]; |
| 3062 |
|
| 3063 |
if ((proc->delayChkptFlags & type) != 0) |
| 3064 |
{ |
| 3065 |
VirtualTransactionId vxid; |
| 3066 |
|
| 3067 |
GET_VXID_FROM_PGPROC(vxid, *proc); |
| 3068 |
if (VirtualTransactionIdIsValid(vxid)) |
| 3069 |
vxids[count++] = vxid; |
| 3070 |
} |
| 3071 |
} |
| 3072 |
|
| 3073 |
LWLockRelease(ProcArrayLock); |
| 3074 |
|
| 3075 |
*nvxids = count; |
| 3076 |
return vxids; |
| 3077 |
} |
| 3078 |
|
| 3079 |
/* |
| 3080 |
* HaveVirtualXIDsDelayingChkpt -- Are any of the specified VXIDs delaying? |
| 3081 |
* |
| 3082 |
* This is used with the results of GetVirtualXIDsDelayingChkpt to see if any |
| 3083 |
* of the specified VXIDs are still in critical sections of code. |
| 3084 |
* |
| 3085 |
* Note: this is O(N^2) in the number of vxacts that are/were delaying, but |
| 3086 |
* those numbers should be small enough for it not to be a problem. |
| 3087 |
*/ |
| 3088 |
bool |
| 3089 |
HaveVirtualXIDsDelayingChkpt(VirtualTransactionId *vxids, int nvxids, int type) |
| 3090 |
{ |
| 3091 |
bool result = false; |
| 3092 |
ProcArrayStruct *arrayP = procArray; |
| 3093 |
int index; |
| 3094 |
|
| 3095 |
Assert(type != 0); |
| 3096 |
|
| 3097 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 3098 |
|
| 3099 |
for (index = 0; index < arrayP->numProcs; index++) |
| 3100 |
{ |
| 3101 |
int pgprocno = arrayP->pgprocnos[index]; |
| 3102 |
PGPROC *proc = &allProcs[pgprocno]; |
| 3103 |
VirtualTransactionId vxid; |
| 3104 |
|
| 3105 |
GET_VXID_FROM_PGPROC(vxid, *proc); |
| 3106 |
|
| 3107 |
if ((proc->delayChkptFlags & type) != 0 && |
| 3108 |
VirtualTransactionIdIsValid(vxid)) |
| 3109 |
{ |
| 3110 |
int i; |
| 3111 |
|
| 3112 |
for (i = 0; i < nvxids; i++) |
| 3113 |
{ |
| 3114 |
if (VirtualTransactionIdEquals(vxid, vxids[i])) |
| 3115 |
{ |
| 3116 |
result = true; |
| 3117 |
break; |
| 3118 |
} |
| 3119 |
} |
| 3120 |
if (result) |
| 3121 |
break; |
| 3122 |
} |
| 3123 |
} |
| 3124 |
|
| 3125 |
LWLockRelease(ProcArrayLock); |
| 3126 |
|
| 3127 |
return result; |
| 3128 |
} |
| 3129 |
|
| 3130 |
/* |
| 3131 |
* ProcNumberGetProc -- get a backend's PGPROC given its proc number |
| 3132 |
* |
| 3133 |
* The result may be out of date arbitrarily quickly, so the caller |
| 3134 |
* must be careful about how this information is used. NULL is |
| 3135 |
* returned if the backend is not active. |
| 3136 |
*/ |
| 3137 |
PGPROC * |
| 3138 |
ProcNumberGetProc(ProcNumber procNumber) |
| 3139 |
{ |
| 3140 |
PGPROC *result; |
| 3141 |
|
| 3142 |
if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount) |
| 3143 |
return NULL; |
| 3144 |
result = GetPGProcByNumber(procNumber); |
| 3145 |
|
| 3146 |
if (result->pid == 0) |
| 3147 |
return NULL; |
| 3148 |
|
| 3149 |
return result; |
| 3150 |
} |
| 3151 |
|
| 3152 |
/* |
| 3153 |
* ProcNumberGetTransactionIds -- get a backend's transaction status |
| 3154 |
* |
| 3155 |
* Get the xid, xmin, nsubxid and overflow status of the backend. The |
| 3156 |
* result may be out of date arbitrarily quickly, so the caller must be |
| 3157 |
* careful about how this information is used. |
| 3158 |
*/ |
| 3159 |
void |
| 3160 |
ProcNumberGetTransactionIds(ProcNumber procNumber, TransactionId *xid, |
| 3161 |
TransactionId *xmin, int *nsubxid, bool *overflowed) |
| 3162 |
{ |
| 3163 |
PGPROC *proc; |
| 3164 |
|
| 3165 |
*xid = InvalidTransactionId; |
| 3166 |
*xmin = InvalidTransactionId; |
| 3167 |
*nsubxid = 0; |
| 3168 |
*overflowed = false; |
| 3169 |
|
| 3170 |
if (procNumber < 0 || procNumber >= ProcGlobal->allProcCount) |
| 3171 |
return; |
| 3172 |
proc = GetPGProcByNumber(procNumber); |
| 3173 |
|
| 3174 |
/* Need to lock out additions/removals of backends */ |
| 3175 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 3176 |
|
| 3177 |
if (proc->pid != 0) |
| 3178 |
{ |
| 3179 |
*xid = proc->xid; |
| 3180 |
*xmin = proc->xmin; |
| 3181 |
*nsubxid = proc->subxidStatus.count; |
| 3182 |
*overflowed = proc->subxidStatus.overflowed; |
| 3183 |
} |
| 3184 |
|
| 3185 |
LWLockRelease(ProcArrayLock); |
| 3186 |
} |
| 3187 |
|
| 3188 |
/* |
| 3189 |
* BackendPidGetProc -- get a backend's PGPROC given its PID |
| 3190 |
* |
| 3191 |
* Returns NULL if not found. Note that it is up to the caller to be |
| 3192 |
* sure that the question remains meaningful for long enough for the |
| 3193 |
* answer to be used ... |
| 3194 |
*/ |
| 3195 |
PGPROC * |
| 3196 |
BackendPidGetProc(int pid) |
| 3197 |
{ |
| 3198 |
PGPROC *result; |
| 3199 |
|
| 3200 |
if (pid == 0) /* never match dummy PGPROCs */ |
| 3201 |
return NULL; |
| 3202 |
|
| 3203 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 3204 |
|
| 3205 |
result = BackendPidGetProcWithLock(pid); |
| 3206 |
|
| 3207 |
LWLockRelease(ProcArrayLock); |
| 3208 |
|
| 3209 |
return result; |
| 3210 |
} |
| 3211 |
|
| 3212 |
/* |
| 3213 |
* BackendPidGetProcWithLock -- get a backend's PGPROC given its PID |
| 3214 |
* |
| 3215 |
* Same as above, except caller must be holding ProcArrayLock. The found |
| 3216 |
* entry, if any, can be assumed to be valid as long as the lock remains held. |
| 3217 |
*/ |
| 3218 |
PGPROC * |
| 3219 |
BackendPidGetProcWithLock(int pid) |
| 3220 |
{ |
| 3221 |
PGPROC *result = NULL; |
| 3222 |
ProcArrayStruct *arrayP = procArray; |
| 3223 |
int index; |
| 3224 |
|
| 3225 |
if (pid == 0) /* never match dummy PGPROCs */ |
| 3226 |
return NULL; |
| 3227 |
|
| 3228 |
for (index = 0; index < arrayP->numProcs; index++) |
| 3229 |
{ |
| 3230 |
PGPROC *proc = &allProcs[arrayP->pgprocnos[index]]; |
| 3231 |
|
| 3232 |
if (proc->pid == pid) |
| 3233 |
{ |
| 3234 |
result = proc; |
| 3235 |
break; |
| 3236 |
} |
| 3237 |
} |
| 3238 |
|
| 3239 |
return result; |
| 3240 |
} |
| 3241 |
|
| 3242 |
/* |
| 3243 |
* BackendXidGetPid -- get a backend's pid given its XID |
| 3244 |
* |
| 3245 |
* Returns 0 if not found or it's a prepared transaction. Note that |
| 3246 |
* it is up to the caller to be sure that the question remains |
| 3247 |
* meaningful for long enough for the answer to be used ... |
| 3248 |
* |
| 3249 |
* Only main transaction Ids are considered. This function is mainly |
| 3250 |
* useful for determining what backend owns a lock. |
| 3251 |
* |
| 3252 |
* Beware that not every xact has an XID assigned. However, as long as you |
| 3253 |
* only call this using an XID found on disk, you're safe. |
| 3254 |
*/ |
| 3255 |
int |
| 3256 |
BackendXidGetPid(TransactionId xid) |
| 3257 |
{ |
| 3258 |
int result = 0; |
| 3259 |
ProcArrayStruct *arrayP = procArray; |
| 3260 |
TransactionId *other_xids = ProcGlobal->xids; |
| 3261 |
int index; |
| 3262 |
|
| 3263 |
if (xid == InvalidTransactionId) /* never match invalid xid */ |
| 3264 |
return 0; |
| 3265 |
|
| 3266 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 3267 |
|
| 3268 |
for (index = 0; index < arrayP->numProcs; index++) |
| 3269 |
{ |
| 3270 |
if (other_xids[index] == xid) |
| 3271 |
{ |
| 3272 |
int pgprocno = arrayP->pgprocnos[index]; |
| 3273 |
PGPROC *proc = &allProcs[pgprocno]; |
| 3274 |
|
| 3275 |
result = proc->pid; |
| 3276 |
break; |
| 3277 |
} |
| 3278 |
} |
| 3279 |
|
| 3280 |
LWLockRelease(ProcArrayLock); |
| 3281 |
|
| 3282 |
return result; |
| 3283 |
} |
| 3284 |
|
| 3285 |
/* |
| 3286 |
* IsBackendPid -- is a given pid a running backend |
| 3287 |
* |
| 3288 |
* This is not called by the backend, but is called by external modules. |
| 3289 |
*/ |
| 3290 |
bool |
| 3291 |
IsBackendPid(int pid) |
| 3292 |
{ |
| 3293 |
return (BackendPidGetProc(pid) != NULL); |
| 3294 |
} |
| 3295 |
|
| 3296 |
|
| 3297 |
/* |
| 3298 |
* GetCurrentVirtualXIDs -- returns an array of currently active VXIDs. |
| 3299 |
* |
| 3300 |
* The array is palloc'd. The number of valid entries is returned into *nvxids. |
| 3301 |
* |
| 3302 |
* The arguments allow filtering the set of VXIDs returned. Our own process |
| 3303 |
* is always skipped. In addition: |
| 3304 |
* If limitXmin is not InvalidTransactionId, skip processes with |
| 3305 |
* xmin > limitXmin. |
| 3306 |
* If excludeXmin0 is true, skip processes with xmin = 0. |
| 3307 |
* If allDbs is false, skip processes attached to other databases. |
| 3308 |
* If excludeVacuum isn't zero, skip processes for which |
| 3309 |
* (statusFlags & excludeVacuum) is not zero. |
| 3310 |
* |
| 3311 |
* Note: the purpose of the limitXmin and excludeXmin0 parameters is to |
| 3312 |
* allow skipping backends whose oldest live snapshot is no older than |
| 3313 |
* some snapshot we have. Since we examine the procarray with only shared |
| 3314 |
* lock, there are race conditions: a backend could set its xmin just after |
| 3315 |
* we look. Indeed, on multiprocessors with weak memory ordering, the |
| 3316 |
* other backend could have set its xmin *before* we look. We know however |
| 3317 |
* that such a backend must have held shared ProcArrayLock overlapping our |
| 3318 |
* own hold of ProcArrayLock, else we would see its xmin update. Therefore, |
| 3319 |
* any snapshot the other backend is taking concurrently with our scan cannot |
| 3320 |
* consider any transactions as still running that we think are committed |
| 3321 |
* (since backends must hold ProcArrayLock exclusive to commit). |
| 3322 |
*/ |
| 3323 |
VirtualTransactionId * |
| 3324 |
GetCurrentVirtualXIDs(TransactionId limitXmin, bool excludeXmin0, |
| 3325 |
bool allDbs, int excludeVacuum, |
| 3326 |
int *nvxids) |
| 3327 |
{ |
| 3328 |
VirtualTransactionId *vxids; |
| 3329 |
ProcArrayStruct *arrayP = procArray; |
| 3330 |
int count = 0; |
| 3331 |
int index; |
| 3332 |
|
| 3333 |
/* allocate what's certainly enough result space */ |
| 3334 |
vxids = (VirtualTransactionId *) |
| 3335 |
palloc(sizeof(VirtualTransactionId) * arrayP->maxProcs); |
| 3336 |
|
| 3337 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 3338 |
|
| 3339 |
for (index = 0; index < arrayP->numProcs; index++) |
| 3340 |
{ |
| 3341 |
int pgprocno = arrayP->pgprocnos[index]; |
| 3342 |
PGPROC *proc = &allProcs[pgprocno]; |
| 3343 |
uint8 statusFlags = ProcGlobal->statusFlags[index]; |
| 3344 |
|
| 3345 |
if (proc == MyProc) |
| 3346 |
continue; |
| 3347 |
|
| 3348 |
if (excludeVacuum & statusFlags) |
| 3349 |
continue; |
| 3350 |
|
| 3351 |
if (allDbs || proc->databaseId == MyDatabaseId) |
| 3352 |
{ |
| 3353 |
/* Fetch xmin just once - might change on us */ |
| 3354 |
TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin); |
| 3355 |
|
| 3356 |
if (excludeXmin0 && !TransactionIdIsValid(pxmin)) |
| 3357 |
continue; |
| 3358 |
|
| 3359 |
/* |
| 3360 |
* InvalidTransactionId precedes all other XIDs, so a proc that |
| 3361 |
* hasn't set xmin yet will not be rejected by this test. |
| 3362 |
*/ |
| 3363 |
if (!TransactionIdIsValid(limitXmin) || |
| 3364 |
TransactionIdPrecedesOrEquals(pxmin, limitXmin)) |
| 3365 |
{ |
| 3366 |
VirtualTransactionId vxid; |
| 3367 |
|
| 3368 |
GET_VXID_FROM_PGPROC(vxid, *proc); |
| 3369 |
if (VirtualTransactionIdIsValid(vxid)) |
| 3370 |
vxids[count++] = vxid; |
| 3371 |
} |
| 3372 |
} |
| 3373 |
} |
| 3374 |
|
| 3375 |
LWLockRelease(ProcArrayLock); |
| 3376 |
|
| 3377 |
*nvxids = count; |
| 3378 |
return vxids; |
| 3379 |
} |
| 3380 |
|
| 3381 |
/* |
| 3382 |
* GetConflictingVirtualXIDs -- returns an array of currently active VXIDs. |
| 3383 |
* |
| 3384 |
* Usage is limited to conflict resolution during recovery on standby servers. |
| 3385 |
* limitXmin is supplied as either a cutoff with snapshotConflictHorizon |
| 3386 |
* semantics, or InvalidTransactionId in cases where caller cannot accurately |
| 3387 |
* determine a safe snapshotConflictHorizon value. |
| 3388 |
* |
| 3389 |
* If limitXmin is InvalidTransactionId then we want to kill everybody, |
| 3390 |
* so we're not worried if they have a snapshot or not, nor does it really |
| 3391 |
* matter what type of lock we hold. Caller must avoid calling here with |
| 3392 |
* snapshotConflictHorizon style cutoffs that were set to InvalidTransactionId |
| 3393 |
* during original execution, since that actually indicates that there is |
| 3394 |
* definitely no need for a recovery conflict (the snapshotConflictHorizon |
| 3395 |
* convention for InvalidTransactionId values is the opposite of our own!). |
| 3396 |
* |
| 3397 |
* All callers that are checking xmins always now supply a valid and useful |
| 3398 |
* value for limitXmin. The limitXmin is always lower than the lowest |
| 3399 |
* numbered KnownAssignedXid that is not already a FATAL error. This is |
| 3400 |
* because we only care about cleanup records that are cleaning up tuple |
| 3401 |
* versions from committed transactions. In that case they will only occur |
| 3402 |
* at the point where the record is less than the lowest running xid. That |
| 3403 |
* allows us to say that if any backend takes a snapshot concurrently with |
| 3404 |
* us then the conflict assessment made here would never include the snapshot |
| 3405 |
* that is being derived. So we take LW_SHARED on the ProcArray and allow |
| 3406 |
* concurrent snapshots when limitXmin is valid. We might think about adding |
| 3407 |
* Assert(limitXmin < lowest(KnownAssignedXids)) |
| 3408 |
* but that would not be true in the case of FATAL errors lagging in array, |
| 3409 |
* but we already know those are bogus anyway, so we skip that test. |
| 3410 |
* |
| 3411 |
* If dbOid is valid we skip backends attached to other databases. |
| 3412 |
* |
| 3413 |
* Be careful to *not* pfree the result from this function. We reuse |
| 3414 |
* this array sufficiently often that we use malloc for the result. |
| 3415 |
*/ |
| 3416 |
VirtualTransactionId * |
| 3417 |
GetConflictingVirtualXIDs(TransactionId limitXmin, Oid dbOid) |
| 3418 |
{ |
| 3419 |
static VirtualTransactionId *vxids; |
| 3420 |
ProcArrayStruct *arrayP = procArray; |
| 3421 |
int count = 0; |
| 3422 |
int index; |
| 3423 |
|
| 3424 |
/* |
| 3425 |
* If first time through, get workspace to remember main XIDs in. We |
| 3426 |
* malloc it permanently to avoid repeated palloc/pfree overhead. Allow |
| 3427 |
* result space, remembering room for a terminator. |
| 3428 |
*/ |
| 3429 |
if (vxids == NULL) |
| 3430 |
{ |
| 3431 |
vxids = (VirtualTransactionId *) |
| 3432 |
malloc(sizeof(VirtualTransactionId) * (arrayP->maxProcs + 1)); |
| 3433 |
if (vxids == NULL) |
| 3434 |
ereport(ERROR, |
| 3435 |
(errcode(ERRCODE_OUT_OF_MEMORY), |
| 3436 |
errmsg("out of memory"))); |
| 3437 |
} |
| 3438 |
|
| 3439 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 3440 |
|
| 3441 |
for (index = 0; index < arrayP->numProcs; index++) |
| 3442 |
{ |
| 3443 |
int pgprocno = arrayP->pgprocnos[index]; |
| 3444 |
PGPROC *proc = &allProcs[pgprocno]; |
| 3445 |
|
| 3446 |
/* Exclude prepared transactions */ |
| 3447 |
if (proc->pid == 0) |
| 3448 |
continue; |
| 3449 |
|
| 3450 |
if (!OidIsValid(dbOid) || |
| 3451 |
proc->databaseId == dbOid) |
| 3452 |
{ |
| 3453 |
/* Fetch xmin just once - can't change on us, but good coding */ |
| 3454 |
TransactionId pxmin = UINT32_ACCESS_ONCE(proc->xmin); |
| 3455 |
|
| 3456 |
/* |
| 3457 |
* We ignore an invalid pxmin because this means that backend has |
| 3458 |
* no snapshot currently. We hold a Share lock to avoid contention |
| 3459 |
* with users taking snapshots. That is not a problem because the |
| 3460 |
* current xmin is always at least one higher than the latest |
| 3461 |
* removed xid, so any new snapshot would never conflict with the |
| 3462 |
* test here. |
| 3463 |
*/ |
| 3464 |
if (!TransactionIdIsValid(limitXmin) || |
| 3465 |
(TransactionIdIsValid(pxmin) && !TransactionIdFollows(pxmin, limitXmin))) |
| 3466 |
{ |
| 3467 |
VirtualTransactionId vxid; |
| 3468 |
|
| 3469 |
GET_VXID_FROM_PGPROC(vxid, *proc); |
| 3470 |
if (VirtualTransactionIdIsValid(vxid)) |
| 3471 |
vxids[count++] = vxid; |
| 3472 |
} |
| 3473 |
} |
| 3474 |
} |
| 3475 |
|
| 3476 |
LWLockRelease(ProcArrayLock); |
| 3477 |
|
| 3478 |
/* add the terminator */ |
| 3479 |
vxids[count].procNumber = INVALID_PROC_NUMBER; |
| 3480 |
vxids[count].localTransactionId = InvalidLocalTransactionId; |
| 3481 |
|
| 3482 |
return vxids; |
| 3483 |
} |
| 3484 |
|
| 3485 |
/* |
| 3486 |
* CancelVirtualTransaction - used in recovery conflict processing |
| 3487 |
* |
| 3488 |
* Returns pid of the process signaled, or 0 if not found. |
| 3489 |
*/ |
| 3490 |
pid_t |
| 3491 |
CancelVirtualTransaction(VirtualTransactionId vxid, ProcSignalReason sigmode) |
| 3492 |
{ |
| 3493 |
return SignalVirtualTransaction(vxid, sigmode, true); |
| 3494 |
} |
| 3495 |
|
| 3496 |
pid_t |
| 3497 |
SignalVirtualTransaction(VirtualTransactionId vxid, ProcSignalReason sigmode, |
| 3498 |
bool conflictPending) |
| 3499 |
{ |
| 3500 |
ProcArrayStruct *arrayP = procArray; |
| 3501 |
int index; |
| 3502 |
pid_t pid = 0; |
| 3503 |
|
| 3504 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 3505 |
|
| 3506 |
for (index = 0; index < arrayP->numProcs; index++) |
| 3507 |
{ |
| 3508 |
int pgprocno = arrayP->pgprocnos[index]; |
| 3509 |
PGPROC *proc = &allProcs[pgprocno]; |
| 3510 |
VirtualTransactionId procvxid; |
| 3511 |
|
| 3512 |
GET_VXID_FROM_PGPROC(procvxid, *proc); |
| 3513 |
|
| 3514 |
if (procvxid.procNumber == vxid.procNumber && |
| 3515 |
procvxid.localTransactionId == vxid.localTransactionId) |
| 3516 |
{ |
| 3517 |
proc->recoveryConflictPending = conflictPending; |
| 3518 |
pid = proc->pid; |
| 3519 |
if (pid != 0) |
| 3520 |
{ |
| 3521 |
/* |
| 3522 |
* Kill the pid if it's still here. If not, that's what we |
| 3523 |
* wanted so ignore any errors. |
| 3524 |
*/ |
| 3525 |
(void) SendProcSignal(pid, sigmode, vxid.procNumber); |
| 3526 |
} |
| 3527 |
break; |
| 3528 |
} |
| 3529 |
} |
| 3530 |
|
| 3531 |
LWLockRelease(ProcArrayLock); |
| 3532 |
|
| 3533 |
return pid; |
| 3534 |
} |
| 3535 |
|
| 3536 |
/* |
| 3537 |
* MinimumActiveBackends --- count backends (other than myself) that are |
| 3538 |
* in active transactions. Return true if the count exceeds the |
| 3539 |
* minimum threshold passed. This is used as a heuristic to decide if |
| 3540 |
* a pre-XLOG-flush delay is worthwhile during commit. |
| 3541 |
* |
| 3542 |
* Do not count backends that are blocked waiting for locks, since they are |
| 3543 |
* not going to get to run until someone else commits. |
| 3544 |
*/ |
| 3545 |
bool |
| 3546 |
MinimumActiveBackends(int min) |
| 3547 |
{ |
| 3548 |
ProcArrayStruct *arrayP = procArray; |
| 3549 |
int count = 0; |
| 3550 |
int index; |
| 3551 |
|
| 3552 |
/* Quick short-circuit if no minimum is specified */ |
| 3553 |
if (min == 0) |
| 3554 |
return true; |
| 3555 |
|
| 3556 |
/* |
| 3557 |
* Note: for speed, we don't acquire ProcArrayLock. This is a little bit |
| 3558 |
* bogus, but since we are only testing fields for zero or nonzero, it |
| 3559 |
* should be OK. The result is only used for heuristic purposes anyway... |
| 3560 |
*/ |
| 3561 |
for (index = 0; index < arrayP->numProcs; index++) |
| 3562 |
{ |
| 3563 |
int pgprocno = arrayP->pgprocnos[index]; |
| 3564 |
PGPROC *proc = &allProcs[pgprocno]; |
| 3565 |
|
| 3566 |
/* |
| 3567 |
* Since we're not holding a lock, need to be prepared to deal with |
| 3568 |
* garbage, as someone could have incremented numProcs but not yet |
| 3569 |
* filled the structure. |
| 3570 |
* |
| 3571 |
* If someone just decremented numProcs, 'proc' could also point to a |
| 3572 |
* PGPROC entry that's no longer in the array. It still points to a |
| 3573 |
* PGPROC struct, though, because freed PGPROC entries just go to the |
| 3574 |
* free list and are recycled. Its contents are nonsense in that case, |
| 3575 |
* but that's acceptable for this function. |
| 3576 |
*/ |
| 3577 |
if (pgprocno == -1) |
| 3578 |
continue; /* do not count deleted entries */ |
| 3579 |
if (proc == MyProc) |
| 3580 |
continue; /* do not count myself */ |
| 3581 |
if (proc->xid == InvalidTransactionId) |
| 3582 |
continue; /* do not count if no XID assigned */ |
| 3583 |
if (proc->pid == 0) |
| 3584 |
continue; /* do not count prepared xacts */ |
| 3585 |
if (proc->waitLock != NULL) |
| 3586 |
continue; /* do not count if blocked on a lock */ |
| 3587 |
count++; |
| 3588 |
if (count >= min) |
| 3589 |
break; |
| 3590 |
} |
| 3591 |
|
| 3592 |
return count >= min; |
| 3593 |
} |
| 3594 |
|
| 3595 |
/* |
| 3596 |
* CountDBBackends --- count backends that are using specified database |
| 3597 |
*/ |
| 3598 |
int |
| 3599 |
CountDBBackends(Oid databaseid) |
| 3600 |
{ |
| 3601 |
ProcArrayStruct *arrayP = procArray; |
| 3602 |
int count = 0; |
| 3603 |
int index; |
| 3604 |
|
| 3605 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 3606 |
|
| 3607 |
for (index = 0; index < arrayP->numProcs; index++) |
| 3608 |
{ |
| 3609 |
int pgprocno = arrayP->pgprocnos[index]; |
| 3610 |
PGPROC *proc = &allProcs[pgprocno]; |
| 3611 |
|
| 3612 |
if (proc->pid == 0) |
| 3613 |
continue; /* do not count prepared xacts */ |
| 3614 |
if (!OidIsValid(databaseid) || |
| 3615 |
proc->databaseId == databaseid) |
| 3616 |
count++; |
| 3617 |
} |
| 3618 |
|
| 3619 |
LWLockRelease(ProcArrayLock); |
| 3620 |
|
| 3621 |
return count; |
| 3622 |
} |
| 3623 |
|
| 3624 |
/* |
| 3625 |
* CountDBConnections --- counts database backends (only regular backends) |
| 3626 |
*/ |
| 3627 |
int |
| 3628 |
CountDBConnections(Oid databaseid) |
| 3629 |
{ |
| 3630 |
ProcArrayStruct *arrayP = procArray; |
| 3631 |
int count = 0; |
| 3632 |
int index; |
| 3633 |
|
| 3634 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 3635 |
|
| 3636 |
for (index = 0; index < arrayP->numProcs; index++) |
| 3637 |
{ |
| 3638 |
int pgprocno = arrayP->pgprocnos[index]; |
| 3639 |
PGPROC *proc = &allProcs[pgprocno]; |
| 3640 |
|
| 3641 |
if (proc->pid == 0) |
| 3642 |
continue; /* do not count prepared xacts */ |
| 3643 |
if (!proc->isRegularBackend) |
| 3644 |
continue; /* count only regular backend processes */ |
| 3645 |
if (!OidIsValid(databaseid) || |
| 3646 |
proc->databaseId == databaseid) |
| 3647 |
count++; |
| 3648 |
} |
| 3649 |
|
| 3650 |
LWLockRelease(ProcArrayLock); |
| 3651 |
|
| 3652 |
return count; |
| 3653 |
} |
| 3654 |
|
| 3655 |
/* |
| 3656 |
* CancelDBBackends --- cancel backends that are using specified database |
| 3657 |
*/ |
| 3658 |
void |
| 3659 |
CancelDBBackends(Oid databaseid, ProcSignalReason sigmode, bool conflictPending) |
| 3660 |
{ |
| 3661 |
ProcArrayStruct *arrayP = procArray; |
| 3662 |
int index; |
| 3663 |
|
| 3664 |
/* tell all backends to die */ |
| 3665 |
LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE); |
| 3666 |
|
| 3667 |
for (index = 0; index < arrayP->numProcs; index++) |
| 3668 |
{ |
| 3669 |
int pgprocno = arrayP->pgprocnos[index]; |
| 3670 |
PGPROC *proc = &allProcs[pgprocno]; |
| 3671 |
|
| 3672 |
if (databaseid == InvalidOid || proc->databaseId == databaseid) |
| 3673 |
{ |
| 3674 |
VirtualTransactionId procvxid; |
| 3675 |
pid_t pid; |
| 3676 |
|
| 3677 |
GET_VXID_FROM_PGPROC(procvxid, *proc); |
| 3678 |
|
| 3679 |
proc->recoveryConflictPending = conflictPending; |
| 3680 |
pid = proc->pid; |
| 3681 |
if (pid != 0) |
| 3682 |
{ |
| 3683 |
/* |
| 3684 |
* Kill the pid if it's still here. If not, that's what we |
| 3685 |
* wanted so ignore any errors. |
| 3686 |
*/ |
| 3687 |
(void) SendProcSignal(pid, sigmode, procvxid.procNumber); |
| 3688 |
} |
| 3689 |
} |
| 3690 |
} |
| 3691 |
|
| 3692 |
LWLockRelease(ProcArrayLock); |
| 3693 |
} |
| 3694 |
|
| 3695 |
/* |
| 3696 |
* CountUserBackends --- count backends that are used by specified user |
| 3697 |
* (only regular backends, not any type of background worker) |
| 3698 |
*/ |
| 3699 |
int |
| 3700 |
CountUserBackends(Oid roleid) |
| 3701 |
{ |
| 3702 |
ProcArrayStruct *arrayP = procArray; |
| 3703 |
int count = 0; |
| 3704 |
int index; |
| 3705 |
|
| 3706 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 3707 |
|
| 3708 |
for (index = 0; index < arrayP->numProcs; index++) |
| 3709 |
{ |
| 3710 |
int pgprocno = arrayP->pgprocnos[index]; |
| 3711 |
PGPROC *proc = &allProcs[pgprocno]; |
| 3712 |
|
| 3713 |
if (proc->pid == 0) |
| 3714 |
continue; /* do not count prepared xacts */ |
| 3715 |
if (!proc->isRegularBackend) |
| 3716 |
continue; /* count only regular backend processes */ |
| 3717 |
if (proc->roleId == roleid) |
| 3718 |
count++; |
| 3719 |
} |
| 3720 |
|
| 3721 |
LWLockRelease(ProcArrayLock); |
| 3722 |
|
| 3723 |
return count; |
| 3724 |
} |
| 3725 |
|
| 3726 |
/* |
| 3727 |
* CountOtherDBBackends -- check for other backends running in the given DB |
| 3728 |
* |
| 3729 |
* If there are other backends in the DB, we will wait a maximum of 5 seconds |
| 3730 |
* for them to exit. Autovacuum backends are encouraged to exit early by |
| 3731 |
* sending them SIGTERM, but normal user backends are just waited for. |
| 3732 |
* |
| 3733 |
* The current backend is always ignored; it is caller's responsibility to |
| 3734 |
* check whether the current backend uses the given DB, if it's important. |
| 3735 |
* |
| 3736 |
* Returns true if there are (still) other backends in the DB, false if not. |
| 3737 |
* Also, *nbackends and *nprepared are set to the number of other backends |
| 3738 |
* and prepared transactions in the DB, respectively. |
| 3739 |
* |
| 3740 |
* This function is used to interlock DROP DATABASE and related commands |
| 3741 |
* against there being any active backends in the target DB --- dropping the |
| 3742 |
* DB while active backends remain would be a Bad Thing. Note that we cannot |
| 3743 |
* detect here the possibility of a newly-started backend that is trying to |
| 3744 |
* connect to the doomed database, so additional interlocking is needed during |
| 3745 |
* backend startup. The caller should normally hold an exclusive lock on the |
| 3746 |
* target DB before calling this, which is one reason we mustn't wait |
| 3747 |
* indefinitely. |
| 3748 |
*/ |
| 3749 |
bool |
| 3750 |
CountOtherDBBackends(Oid databaseId, int *nbackends, int *nprepared) |
| 3751 |
{ |
| 3752 |
ProcArrayStruct *arrayP = procArray; |
| 3753 |
|
| 3754 |
#define MAXAUTOVACPIDS 10 /* max autovacs to SIGTERM per iteration */ |
| 3755 |
int autovac_pids[MAXAUTOVACPIDS]; |
| 3756 |
int tries; |
| 3757 |
|
| 3758 |
/* 50 tries with 100ms sleep between tries makes 5 sec total wait */ |
| 3759 |
for (tries = 0; tries < 50; tries++) |
| 3760 |
{ |
| 3761 |
int nautovacs = 0; |
| 3762 |
bool found = false; |
| 3763 |
int index; |
| 3764 |
|
| 3765 |
CHECK_FOR_INTERRUPTS(); |
| 3766 |
|
| 3767 |
*nbackends = *nprepared = 0; |
| 3768 |
|
| 3769 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 3770 |
|
| 3771 |
for (index = 0; index < arrayP->numProcs; index++) |
| 3772 |
{ |
| 3773 |
int pgprocno = arrayP->pgprocnos[index]; |
| 3774 |
PGPROC *proc = &allProcs[pgprocno]; |
| 3775 |
uint8 statusFlags = ProcGlobal->statusFlags[index]; |
| 3776 |
|
| 3777 |
if (proc->databaseId != databaseId) |
| 3778 |
continue; |
| 3779 |
if (proc == MyProc) |
| 3780 |
continue; |
| 3781 |
|
| 3782 |
found = true; |
| 3783 |
|
| 3784 |
if (proc->pid == 0) |
| 3785 |
(*nprepared)++; |
| 3786 |
else |
| 3787 |
{ |
| 3788 |
(*nbackends)++; |
| 3789 |
if ((statusFlags & PROC_IS_AUTOVACUUM) && |
| 3790 |
nautovacs < MAXAUTOVACPIDS) |
| 3791 |
autovac_pids[nautovacs++] = proc->pid; |
| 3792 |
} |
| 3793 |
} |
| 3794 |
|
| 3795 |
LWLockRelease(ProcArrayLock); |
| 3796 |
|
| 3797 |
if (!found) |
| 3798 |
return false; /* no conflicting backends, so done */ |
| 3799 |
|
| 3800 |
/* |
| 3801 |
* Send SIGTERM to any conflicting autovacuums before sleeping. We |
| 3802 |
* postpone this step until after the loop because we don't want to |
| 3803 |
* hold ProcArrayLock while issuing kill(). We have no idea what might |
| 3804 |
* block kill() inside the kernel... |
| 3805 |
*/ |
| 3806 |
for (index = 0; index < nautovacs; index++) |
| 3807 |
(void) kill(autovac_pids[index], SIGTERM); /* ignore any error */ |
| 3808 |
|
| 3809 |
/* sleep, then try again */ |
| 3810 |
pg_usleep(100 * 1000L); /* 100ms */ |
| 3811 |
} |
| 3812 |
|
| 3813 |
return true; /* timed out, still conflicts */ |
| 3814 |
} |
| 3815 |
|
| 3816 |
/* |
| 3817 |
* Terminate existing connections to the specified database. This routine |
| 3818 |
* is used by the DROP DATABASE command when user has asked to forcefully |
| 3819 |
* drop the database. |
| 3820 |
* |
| 3821 |
* The current backend is always ignored; it is caller's responsibility to |
| 3822 |
* check whether the current backend uses the given DB, if it's important. |
| 3823 |
* |
| 3824 |
* If the target database has a prepared transaction or permissions checks |
| 3825 |
* fail for a connection, this fails without terminating anything. |
| 3826 |
*/ |
| 3827 |
void |
| 3828 |
TerminateOtherDBBackends(Oid databaseId) |
| 3829 |
{ |
| 3830 |
ProcArrayStruct *arrayP = procArray; |
| 3831 |
List *pids = NIL; |
| 3832 |
int nprepared = 0; |
| 3833 |
int i; |
| 3834 |
|
| 3835 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 3836 |
|
| 3837 |
for (i = 0; i < procArray->numProcs; i++) |
| 3838 |
{ |
| 3839 |
int pgprocno = arrayP->pgprocnos[i]; |
| 3840 |
PGPROC *proc = &allProcs[pgprocno]; |
| 3841 |
|
| 3842 |
if (proc->databaseId != databaseId) |
| 3843 |
continue; |
| 3844 |
if (proc == MyProc) |
| 3845 |
continue; |
| 3846 |
|
| 3847 |
if (proc->pid != 0) |
| 3848 |
pids = lappend_int(pids, proc->pid); |
| 3849 |
else |
| 3850 |
nprepared++; |
| 3851 |
} |
| 3852 |
|
| 3853 |
LWLockRelease(ProcArrayLock); |
| 3854 |
|
| 3855 |
if (nprepared > 0) |
| 3856 |
ereport(ERROR, |
| 3857 |
(errcode(ERRCODE_OBJECT_IN_USE), |
| 3858 |
errmsg("database \"%s\" is being used by prepared transactions", |
| 3859 |
get_database_name(databaseId)), |
| 3860 |
errdetail_plural("There is %d prepared transaction using the database.", |
| 3861 |
"There are %d prepared transactions using the database.", |
| 3862 |
nprepared, |
| 3863 |
nprepared))); |
| 3864 |
|
| 3865 |
if (pids) |
| 3866 |
{ |
| 3867 |
ListCell *lc; |
| 3868 |
|
| 3869 |
/* |
| 3870 |
* Permissions checks relax the pg_terminate_backend checks in two |
| 3871 |
* ways, both by omitting the !OidIsValid(proc->roleId) check: |
| 3872 |
* |
| 3873 |
* - Accept terminating autovacuum workers, since DROP DATABASE |
| 3874 |
* without FORCE terminates them. |
| 3875 |
* |
| 3876 |
* - Accept terminating bgworkers. For bgworker authors, it's |
| 3877 |
* convenient to be able to recommend FORCE if a worker is blocking |
| 3878 |
* DROP DATABASE unexpectedly. |
| 3879 |
* |
| 3880 |
* Unlike pg_terminate_backend, we don't raise some warnings - like |
| 3881 |
* "PID %d is not a PostgreSQL server process", because for us already |
| 3882 |
* finished session is not a problem. |
| 3883 |
*/ |
| 3884 |
foreach(lc, pids) |
| 3885 |
{ |
| 3886 |
int pid = lfirst_int(lc); |
| 3887 |
PGPROC *proc = BackendPidGetProc(pid); |
| 3888 |
|
| 3889 |
if (proc != NULL) |
| 3890 |
{ |
| 3891 |
if (superuser_arg(proc->roleId) && !superuser()) |
| 3892 |
ereport(ERROR, |
| 3893 |
(errcode(ERRCODE_INSUFFICIENT_PRIVILEGE), |
| 3894 |
errmsg("permission denied to terminate process"), |
| 3895 |
errdetail("Only roles with the %s attribute may terminate processes of roles with the %s attribute.", |
| 3896 |
"SUPERUSER", "SUPERUSER"))); |
| 3897 |
|
| 3898 |
if (!has_privs_of_role(GetUserId(), proc->roleId) && |
| 3899 |
!has_privs_of_role(GetUserId(), ROLE_PG_SIGNAL_BACKEND)) |
| 3900 |
ereport(ERROR, |
| 3901 |
(errcode(ERRCODE_INSUFFICIENT_PRIVILEGE), |
| 3902 |
errmsg("permission denied to terminate process"), |
| 3903 |
errdetail("Only roles with privileges of the role whose process is being terminated or with privileges of the \"%s\" role may terminate this process.", |
| 3904 |
"pg_signal_backend"))); |
| 3905 |
} |
| 3906 |
} |
| 3907 |
|
| 3908 |
/* |
| 3909 |
* There's a race condition here: once we release the ProcArrayLock, |
| 3910 |
* it's possible for the session to exit before we issue kill. That |
| 3911 |
* race condition possibility seems too unlikely to worry about. See |
| 3912 |
* pg_signal_backend. |
| 3913 |
*/ |
| 3914 |
foreach(lc, pids) |
| 3915 |
{ |
| 3916 |
int pid = lfirst_int(lc); |
| 3917 |
PGPROC *proc = BackendPidGetProc(pid); |
| 3918 |
|
| 3919 |
if (proc != NULL) |
| 3920 |
{ |
| 3921 |
/* |
| 3922 |
* If we have setsid(), signal the backend's whole process |
| 3923 |
* group |
| 3924 |
*/ |
| 3925 |
#ifdef HAVE_SETSID |
| 3926 |
(void) kill(-pid, SIGTERM); |
| 3927 |
#else |
| 3928 |
(void) kill(pid, SIGTERM); |
| 3929 |
#endif |
| 3930 |
} |
| 3931 |
} |
| 3932 |
} |
| 3933 |
} |
| 3934 |
|
| 3935 |
/* |
| 3936 |
* ProcArraySetReplicationSlotXmin |
| 3937 |
* |
| 3938 |
* Install limits to future computations of the xmin horizon to prevent vacuum |
| 3939 |
* and HOT pruning from removing affected rows still needed by clients with |
| 3940 |
* replication slots. |
| 3941 |
*/ |
| 3942 |
void |
| 3943 |
ProcArraySetReplicationSlotXmin(TransactionId xmin, TransactionId catalog_xmin, |
| 3944 |
bool already_locked) |
| 3945 |
{ |
| 3946 |
Assert(!already_locked || LWLockHeldByMe(ProcArrayLock)); |
| 3947 |
|
| 3948 |
if (!already_locked) |
| 3949 |
LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE); |
| 3950 |
|
| 3951 |
procArray->replication_slot_xmin = xmin; |
| 3952 |
procArray->replication_slot_catalog_xmin = catalog_xmin; |
| 3953 |
|
| 3954 |
if (!already_locked) |
| 3955 |
LWLockRelease(ProcArrayLock); |
| 3956 |
|
| 3957 |
elog(DEBUG1, "xmin required by slots: data %u, catalog %u", |
| 3958 |
xmin, catalog_xmin); |
| 3959 |
} |
| 3960 |
|
| 3961 |
/* |
| 3962 |
* ProcArrayGetReplicationSlotXmin |
| 3963 |
* |
| 3964 |
* Return the current slot xmin limits. That's useful to be able to remove |
| 3965 |
* data that's older than those limits. |
| 3966 |
*/ |
| 3967 |
void |
| 3968 |
ProcArrayGetReplicationSlotXmin(TransactionId *xmin, |
| 3969 |
TransactionId *catalog_xmin) |
| 3970 |
{ |
| 3971 |
LWLockAcquire(ProcArrayLock, LW_SHARED); |
| 3972 |
|
| 3973 |
if (xmin != NULL) |
| 3974 |
*xmin = procArray->replication_slot_xmin; |
| 3975 |
|
| 3976 |
if (catalog_xmin != NULL) |
| 3977 |
*catalog_xmin = procArray->replication_slot_catalog_xmin; |
| 3978 |
|
| 3979 |
LWLockRelease(ProcArrayLock); |
| 3980 |
} |
| 3981 |
|
| 3982 |
/* |
| 3983 |
* XidCacheRemoveRunningXids |
| 3984 |
* |
| 3985 |
* Remove a bunch of TransactionIds from the list of known-running |
| 3986 |
* subtransactions for my backend. Both the specified xid and those in |
| 3987 |
* the xids[] array (of length nxids) are removed from the subxids cache. |
| 3988 |
* latestXid must be the latest XID among the group. |
| 3989 |
*/ |
| 3990 |
void |
| 3991 |
XidCacheRemoveRunningXids(TransactionId xid, |
| 3992 |
int nxids, const TransactionId *xids, |
| 3993 |
TransactionId latestXid) |
| 3994 |
{ |
| 3995 |
int i, |
| 3996 |
j; |
| 3997 |
XidCacheStatus *mysubxidstat; |
| 3998 |
|
| 3999 |
Assert(TransactionIdIsValid(xid)); |
| 4000 |
|
| 4001 |
/* |
| 4002 |
* We must hold ProcArrayLock exclusively in order to remove transactions |
| 4003 |
* from the PGPROC array. (See src/backend/access/transam/README.) It's |
| 4004 |
* possible this could be relaxed since we know this routine is only used |
| 4005 |
* to abort subtransactions, but pending closer analysis we'd best be |
| 4006 |
* conservative. |
| 4007 |
* |
| 4008 |
* Note that we do not have to be careful about memory ordering of our own |
| 4009 |
* reads wrt. GetNewTransactionId() here - only this process can modify |
| 4010 |
* relevant fields of MyProc/ProcGlobal->xids[]. But we do have to be |
| 4011 |
* careful about our own writes being well ordered. |
| 4012 |
*/ |
| 4013 |
LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE); |
| 4014 |
|
| 4015 |
mysubxidstat = &ProcGlobal->subxidStates[MyProc->pgxactoff]; |
| 4016 |
|
| 4017 |
/* |
| 4018 |
* Under normal circumstances xid and xids[] will be in increasing order, |
| 4019 |
* as will be the entries in subxids. Scan backwards to avoid O(N^2) |
| 4020 |
* behavior when removing a lot of xids. |
| 4021 |
*/ |
| 4022 |
for (i = nxids - 1; i >= 0; i--) |
| 4023 |
{ |
| 4024 |
TransactionId anxid = xids[i]; |
| 4025 |
|
| 4026 |
for (j = MyProc->subxidStatus.count - 1; j >= 0; j--) |
| 4027 |
{ |
| 4028 |
if (TransactionIdEquals(MyProc->subxids.xids[j], anxid)) |
| 4029 |
{ |
| 4030 |
MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1]; |
| 4031 |
pg_write_barrier(); |
| 4032 |
mysubxidstat->count--; |
| 4033 |
MyProc->subxidStatus.count--; |
| 4034 |
break; |
| 4035 |
} |
| 4036 |
} |
| 4037 |
|
| 4038 |
/* |
| 4039 |
* Ordinarily we should have found it, unless the cache has |
| 4040 |
* overflowed. However it's also possible for this routine to be |
| 4041 |
* invoked multiple times for the same subtransaction, in case of an |
| 4042 |
* error during AbortSubTransaction. So instead of Assert, emit a |
| 4043 |
* debug warning. |
| 4044 |
*/ |
| 4045 |
if (j < 0 && !MyProc->subxidStatus.overflowed) |
| 4046 |
elog(WARNING, "did not find subXID %u in MyProc", anxid); |
| 4047 |
} |
| 4048 |
|
| 4049 |
for (j = MyProc->subxidStatus.count - 1; j >= 0; j--) |
| 4050 |
{ |
| 4051 |
if (TransactionIdEquals(MyProc->subxids.xids[j], xid)) |
| 4052 |
{ |
| 4053 |
MyProc->subxids.xids[j] = MyProc->subxids.xids[MyProc->subxidStatus.count - 1]; |
| 4054 |
pg_write_barrier(); |
| 4055 |
mysubxidstat->count--; |
| 4056 |
MyProc->subxidStatus.count--; |
| 4057 |
break; |
| 4058 |
} |
| 4059 |
} |
| 4060 |
/* Ordinarily we should have found it, unless the cache has overflowed */ |
| 4061 |
if (j < 0 && !MyProc->subxidStatus.overflowed) |
| 4062 |
elog(WARNING, "did not find subXID %u in MyProc", xid); |
| 4063 |
|
| 4064 |
/* Also advance global latestCompletedXid while holding the lock */ |
| 4065 |
MaintainLatestCompletedXid(latestXid); |
| 4066 |
|
| 4067 |
/* ... and xactCompletionCount */ |
| 4068 |
TransamVariables->xactCompletionCount++; |
| 4069 |
|
| 4070 |
LWLockRelease(ProcArrayLock); |
| 4071 |
} |
| 4072 |
|
| 4073 |
#ifdef XIDCACHE_DEBUG |
| 4074 |
|
| 4075 |
/* |
| 4076 |
* Print stats about effectiveness of XID cache |
| 4077 |
*/ |
| 4078 |
static void |
| 4079 |
DisplayXidCache(void) |
| 4080 |
{ |
| 4081 |
fprintf(stderr, |
| 4082 |
"XidCache: xmin: %ld, known: %ld, myxact: %ld, latest: %ld, mainxid: %ld, childxid: %ld, knownassigned: %ld, nooflo: %ld, slow: %ld\n", |
| 4083 |
xc_by_recent_xmin, |
| 4084 |
xc_by_known_xact, |
| 4085 |
xc_by_my_xact, |
| 4086 |
xc_by_latest_xid, |
| 4087 |
xc_by_main_xid, |
| 4088 |
xc_by_child_xid, |
| 4089 |
xc_by_known_assigned, |
| 4090 |
xc_no_overflow, |
| 4091 |
xc_slow_answer); |
| 4092 |
} |
| 4093 |
#endif /* XIDCACHE_DEBUG */ |
| 4094 |
|
| 4095 |
/* |
| 4096 |
* If rel != NULL, return test state appropriate for relation, otherwise |
| 4097 |
* return state usable for all relations. The latter may consider XIDs as |
| 4098 |
* not-yet-visible-to-everyone that a state for a specific relation would |
| 4099 |
* already consider visible-to-everyone. |
| 4100 |
* |
| 4101 |
* This needs to be called while a snapshot is active or registered, otherwise |
| 4102 |
* there are wraparound and other dangers. |
| 4103 |
* |
| 4104 |
* See comment for GlobalVisState for details. |
| 4105 |
*/ |
| 4106 |
GlobalVisState * |
| 4107 |
GlobalVisTestFor(Relation rel) |
| 4108 |
{ |
| 4109 |
GlobalVisState *state = NULL; |
| 4110 |
|
| 4111 |
/* XXX: we should assert that a snapshot is pushed or registered */ |
| 4112 |
Assert(RecentXmin); |
| 4113 |
|
| 4114 |
switch (GlobalVisHorizonKindForRel(rel)) |
| 4115 |
{ |
| 4116 |
case VISHORIZON_SHARED: |
| 4117 |
state = &GlobalVisSharedRels; |
| 4118 |
break; |
| 4119 |
case VISHORIZON_CATALOG: |
| 4120 |
state = &GlobalVisCatalogRels; |
| 4121 |
break; |
| 4122 |
case VISHORIZON_DATA: |
| 4123 |
state = &GlobalVisDataRels; |
| 4124 |
break; |
| 4125 |
case VISHORIZON_TEMP: |
| 4126 |
state = &GlobalVisTempRels; |
| 4127 |
break; |
| 4128 |
} |
| 4129 |
|
| 4130 |
Assert(FullTransactionIdIsValid(state->definitely_needed) && |
| 4131 |
FullTransactionIdIsValid(state->maybe_needed)); |
| 4132 |
|
| 4133 |
return state; |
| 4134 |
} |
| 4135 |
|
| 4136 |
/* |
| 4137 |
* Return true if it's worth updating the accurate maybe_needed boundary. |
| 4138 |
* |
| 4139 |
* As it is somewhat expensive to determine xmin horizons, we don't want to |
| 4140 |
* repeatedly do so when there is a low likelihood of it being beneficial. |
| 4141 |
* |
| 4142 |
* The current heuristic is that we update only if RecentXmin has changed |
| 4143 |
* since the last update. If the oldest currently running transaction has not |
| 4144 |
* finished, it is unlikely that recomputing the horizon would be useful. |
| 4145 |
*/ |
| 4146 |
static bool |
| 4147 |
GlobalVisTestShouldUpdate(GlobalVisState *state) |
| 4148 |
{ |
| 4149 |
/* hasn't been updated yet */ |
| 4150 |
if (!TransactionIdIsValid(ComputeXidHorizonsResultLastXmin)) |
| 4151 |
return true; |
| 4152 |
|
| 4153 |
/* |
| 4154 |
* If the maybe_needed/definitely_needed boundaries are the same, it's |
| 4155 |
* unlikely to be beneficial to refresh boundaries. |
| 4156 |
*/ |
| 4157 |
if (FullTransactionIdFollowsOrEquals(state->maybe_needed, |
| 4158 |
state->definitely_needed)) |
| 4159 |
return false; |
| 4160 |
|
| 4161 |
/* does the last snapshot built have a different xmin? */ |
| 4162 |
return RecentXmin != ComputeXidHorizonsResultLastXmin; |
| 4163 |
} |
| 4164 |
|
| 4165 |
static void |
| 4166 |
GlobalVisUpdateApply(ComputeXidHorizonsResult *horizons) |
| 4167 |
{ |
| 4168 |
GlobalVisSharedRels.maybe_needed = |
| 4169 |
FullXidRelativeTo(horizons->latest_completed, |
| 4170 |
horizons->shared_oldest_nonremovable); |
| 4171 |
GlobalVisCatalogRels.maybe_needed = |
| 4172 |
FullXidRelativeTo(horizons->latest_completed, |
| 4173 |
horizons->catalog_oldest_nonremovable); |
| 4174 |
GlobalVisDataRels.maybe_needed = |
| 4175 |
FullXidRelativeTo(horizons->latest_completed, |
| 4176 |
horizons->data_oldest_nonremovable); |
| 4177 |
GlobalVisTempRels.maybe_needed = |
| 4178 |
FullXidRelativeTo(horizons->latest_completed, |
| 4179 |
horizons->temp_oldest_nonremovable); |
| 4180 |
|
| 4181 |
/* |
| 4182 |
* In longer running transactions it's possible that transactions we |
| 4183 |
* previously needed to treat as running aren't around anymore. So update |
| 4184 |
* definitely_needed to not be earlier than maybe_needed. |
| 4185 |
*/ |
| 4186 |
GlobalVisSharedRels.definitely_needed = |
| 4187 |
FullTransactionIdNewer(GlobalVisSharedRels.maybe_needed, |
| 4188 |
GlobalVisSharedRels.definitely_needed); |
| 4189 |
GlobalVisCatalogRels.definitely_needed = |
| 4190 |
FullTransactionIdNewer(GlobalVisCatalogRels.maybe_needed, |
| 4191 |
GlobalVisCatalogRels.definitely_needed); |
| 4192 |
GlobalVisDataRels.definitely_needed = |
| 4193 |
FullTransactionIdNewer(GlobalVisDataRels.maybe_needed, |
| 4194 |
GlobalVisDataRels.definitely_needed); |
| 4195 |
GlobalVisTempRels.definitely_needed = GlobalVisTempRels.maybe_needed; |
| 4196 |
|
| 4197 |
ComputeXidHorizonsResultLastXmin = RecentXmin; |
| 4198 |
} |
| 4199 |
|
| 4200 |
/* |
| 4201 |
* Update boundaries in GlobalVis{Shared,Catalog, Data}Rels |
| 4202 |
* using ComputeXidHorizons(). |
| 4203 |
*/ |
| 4204 |
static void |
| 4205 |
GlobalVisUpdate(void) |
| 4206 |
{ |
| 4207 |
ComputeXidHorizonsResult horizons; |
| 4208 |
|
| 4209 |
/* updates the horizons as a side-effect */ |
| 4210 |
ComputeXidHorizons(&horizons); |
| 4211 |
} |
| 4212 |
|
| 4213 |
/* |
| 4214 |
* Return true if no snapshot still considers fxid to be running. |
| 4215 |
* |
| 4216 |
* The state passed needs to have been initialized for the relation fxid is |
| 4217 |
* from (NULL is also OK), otherwise the result may not be correct. |
| 4218 |
* |
| 4219 |
* See comment for GlobalVisState for details. |
| 4220 |
*/ |
| 4221 |
bool |
| 4222 |
GlobalVisTestIsRemovableFullXid(GlobalVisState *state, |
| 4223 |
FullTransactionId fxid) |
| 4224 |
{ |
| 4225 |
/* |
| 4226 |
* If fxid is older than maybe_needed bound, it definitely is visible to |
| 4227 |
* everyone. |
| 4228 |
*/ |
| 4229 |
if (FullTransactionIdPrecedes(fxid, state->maybe_needed)) |
| 4230 |
return true; |
| 4231 |
|
| 4232 |
/* |
| 4233 |
* If fxid is >= definitely_needed bound, it is very likely to still be |
| 4234 |
* considered running. |
| 4235 |
*/ |
| 4236 |
if (FullTransactionIdFollowsOrEquals(fxid, state->definitely_needed)) |
| 4237 |
return false; |
| 4238 |
|
| 4239 |
/* |
| 4240 |
* fxid is between maybe_needed and definitely_needed, i.e. there might or |
| 4241 |
* might not exist a snapshot considering fxid running. If it makes sense, |
| 4242 |
* update boundaries and recheck. |
| 4243 |
*/ |
| 4244 |
if (GlobalVisTestShouldUpdate(state)) |
| 4245 |
{ |
| 4246 |
GlobalVisUpdate(); |
| 4247 |
|
| 4248 |
Assert(FullTransactionIdPrecedes(fxid, state->definitely_needed)); |
| 4249 |
|
| 4250 |
return FullTransactionIdPrecedes(fxid, state->maybe_needed); |
| 4251 |
} |
| 4252 |
else |
| 4253 |
return false; |
| 4254 |
} |
| 4255 |
|
| 4256 |
/* |
| 4257 |
* Wrapper around GlobalVisTestIsRemovableFullXid() for 32bit xids. |
| 4258 |
* |
| 4259 |
* It is crucial that this only gets called for xids from a source that |
| 4260 |
* protects against xid wraparounds (e.g. from a table and thus protected by |
| 4261 |
* relfrozenxid). |
| 4262 |
*/ |
| 4263 |
bool |
| 4264 |
GlobalVisTestIsRemovableXid(GlobalVisState *state, TransactionId xid) |
| 4265 |
{ |
| 4266 |
FullTransactionId fxid; |
| 4267 |
|
| 4268 |
/* |
| 4269 |
* Convert 32 bit argument to FullTransactionId. We can do so safely |
| 4270 |
* because we know the xid has to, at the very least, be between |
| 4271 |
* [oldestXid, nextXid), i.e. within 2 billion of xid. To avoid taking a |
| 4272 |
* lock to determine either, we can just compare with |
| 4273 |
* state->definitely_needed, which was based on those value at the time |
| 4274 |
* the current snapshot was built. |
| 4275 |
*/ |
| 4276 |
fxid = FullXidRelativeTo(state->definitely_needed, xid); |
| 4277 |
|
| 4278 |
return GlobalVisTestIsRemovableFullXid(state, fxid); |
| 4279 |
} |
| 4280 |
|
| 4281 |
/* |
| 4282 |
* Convenience wrapper around GlobalVisTestFor() and |
| 4283 |
* GlobalVisTestIsRemovableFullXid(), see their comments. |
| 4284 |
*/ |
| 4285 |
bool |
| 4286 |
GlobalVisCheckRemovableFullXid(Relation rel, FullTransactionId fxid) |
| 4287 |
{ |
| 4288 |
GlobalVisState *state; |
| 4289 |
|
| 4290 |
state = GlobalVisTestFor(rel); |
| 4291 |
|
| 4292 |
return GlobalVisTestIsRemovableFullXid(state, fxid); |
| 4293 |
} |
| 4294 |
|
| 4295 |
/* |
| 4296 |
* Convenience wrapper around GlobalVisTestFor() and |
| 4297 |
* GlobalVisTestIsRemovableXid(), see their comments. |
| 4298 |
*/ |
| 4299 |
bool |
| 4300 |
GlobalVisCheckRemovableXid(Relation rel, TransactionId xid) |
| 4301 |
{ |
| 4302 |
GlobalVisState *state; |
| 4303 |
|
| 4304 |
state = GlobalVisTestFor(rel); |
| 4305 |
|
| 4306 |
return GlobalVisTestIsRemovableXid(state, xid); |
| 4307 |
} |
| 4308 |
|
| 4309 |
/* |
| 4310 |
* Convert a 32 bit transaction id into 64 bit transaction id, by assuming it |
| 4311 |
* is within MaxTransactionId / 2 of XidFromFullTransactionId(rel). |
| 4312 |
* |
| 4313 |
* Be very careful about when to use this function. It can only safely be used |
| 4314 |
* when there is a guarantee that xid is within MaxTransactionId / 2 xids of |
| 4315 |
* rel. That e.g. can be guaranteed if the caller assures a snapshot is |
| 4316 |
* held by the backend and xid is from a table (where vacuum/freezing ensures |
| 4317 |
* the xid has to be within that range), or if xid is from the procarray and |
| 4318 |
* prevents xid wraparound that way. |
| 4319 |
*/ |
| 4320 |
static inline FullTransactionId |
| 4321 |
FullXidRelativeTo(FullTransactionId rel, TransactionId xid) |
| 4322 |
{ |
| 4323 |
TransactionId rel_xid = XidFromFullTransactionId(rel); |
| 4324 |
|
| 4325 |
Assert(TransactionIdIsValid(xid)); |
| 4326 |
Assert(TransactionIdIsValid(rel_xid)); |
| 4327 |
|
| 4328 |
/* not guaranteed to find issues, but likely to catch mistakes */ |
| 4329 |
AssertTransactionIdInAllowableRange(xid); |
| 4330 |
|
| 4331 |
return FullTransactionIdFromU64(U64FromFullTransactionId(rel) |
| 4332 |
+ (int32) (xid - rel_xid)); |
| 4333 |
} |
| 4334 |
|
| 4335 |
|
| 4336 |
/* ---------------------------------------------- |
| 4337 |
* KnownAssignedTransactionIds sub-module |
| 4338 |
* ---------------------------------------------- |
| 4339 |
*/ |
| 4340 |
|
| 4341 |
/* |
| 4342 |
* In Hot Standby mode, we maintain a list of transactions that are (or were) |
| 4343 |
* running on the primary at the current point in WAL. These XIDs must be |
| 4344 |
* treated as running by standby transactions, even though they are not in |
| 4345 |
* the standby server's PGPROC array. |
| 4346 |
* |
| 4347 |
* We record all XIDs that we know have been assigned. That includes all the |
| 4348 |
* XIDs seen in WAL records, plus all unobserved XIDs that we can deduce have |
| 4349 |
* been assigned. We can deduce the existence of unobserved XIDs because we |
| 4350 |
* know XIDs are assigned in sequence, with no gaps. The KnownAssignedXids |
| 4351 |
* list expands as new XIDs are observed or inferred, and contracts when |
| 4352 |
* transaction completion records arrive. |
| 4353 |
* |
| 4354 |
* During hot standby we do not fret too much about the distinction between |
| 4355 |
* top-level XIDs and subtransaction XIDs. We store both together in the |
| 4356 |
* KnownAssignedXids list. In backends, this is copied into snapshots in |
| 4357 |
* GetSnapshotData(), taking advantage of the fact that XidInMVCCSnapshot() |
| 4358 |
* doesn't care about the distinction either. Subtransaction XIDs are |
| 4359 |
* effectively treated as top-level XIDs and in the typical case pg_subtrans |
| 4360 |
* links are *not* maintained (which does not affect visibility). |
| 4361 |
* |
| 4362 |
* We have room in KnownAssignedXids and in snapshots to hold maxProcs * |
| 4363 |
* (1 + PGPROC_MAX_CACHED_SUBXIDS) XIDs, so every primary transaction must |
| 4364 |
* report its subtransaction XIDs in a WAL XLOG_XACT_ASSIGNMENT record at |
| 4365 |
* least every PGPROC_MAX_CACHED_SUBXIDS. When we receive one of these |
| 4366 |
* records, we mark the subXIDs as children of the top XID in pg_subtrans, |
| 4367 |
* and then remove them from KnownAssignedXids. This prevents overflow of |
| 4368 |
* KnownAssignedXids and snapshots, at the cost that status checks for these |
| 4369 |
* subXIDs will take a slower path through TransactionIdIsInProgress(). |
| 4370 |
* This means that KnownAssignedXids is not necessarily complete for subXIDs, |
| 4371 |
* though it should be complete for top-level XIDs; this is the same situation |
| 4372 |
* that holds with respect to the PGPROC entries in normal running. |
| 4373 |
* |
| 4374 |
* When we throw away subXIDs from KnownAssignedXids, we need to keep track of |
| 4375 |
* that, similarly to tracking overflow of a PGPROC's subxids array. We do |
| 4376 |
* that by remembering the lastOverflowedXid, ie the last thrown-away subXID. |
| 4377 |
* As long as that is within the range of interesting XIDs, we have to assume |
| 4378 |
* that subXIDs are missing from snapshots. (Note that subXID overflow occurs |
| 4379 |
* on primary when 65th subXID arrives, whereas on standby it occurs when 64th |
| 4380 |
* subXID arrives - that is not an error.) |
| 4381 |
* |
| 4382 |
* Should a backend on primary somehow disappear before it can write an abort |
| 4383 |
* record, then we just leave those XIDs in KnownAssignedXids. They actually |
| 4384 |
* aborted but we think they were running; the distinction is irrelevant |
| 4385 |
* because either way any changes done by the transaction are not visible to |
| 4386 |
* backends in the standby. We prune KnownAssignedXids when |
| 4387 |
* XLOG_RUNNING_XACTS arrives, to forestall possible overflow of the |
| 4388 |
* array due to such dead XIDs. |
| 4389 |
*/ |
| 4390 |
|
| 4391 |
/* |
| 4392 |
* RecordKnownAssignedTransactionIds |
| 4393 |
* Record the given XID in KnownAssignedXids, as well as any preceding |
| 4394 |
* unobserved XIDs. |
| 4395 |
* |
| 4396 |
* RecordKnownAssignedTransactionIds() should be run for *every* WAL record |
| 4397 |
* associated with a transaction. Must be called for each record after we |
| 4398 |
* have executed StartupCLOG() et al, since we must ExtendCLOG() etc.. |
| 4399 |
* |
| 4400 |
* Called during recovery in analogy with and in place of GetNewTransactionId() |
| 4401 |
*/ |
| 4402 |
void |
| 4403 |
RecordKnownAssignedTransactionIds(TransactionId xid) |
| 4404 |
{ |
| 4405 |
Assert(standbyState >= STANDBY_INITIALIZED); |
| 4406 |
Assert(TransactionIdIsValid(xid)); |
| 4407 |
Assert(TransactionIdIsValid(latestObservedXid)); |
| 4408 |
|
| 4409 |
elog(DEBUG4, "record known xact %u latestObservedXid %u", |
| 4410 |
xid, latestObservedXid); |
| 4411 |
|
| 4412 |
/* |
| 4413 |
* When a newly observed xid arrives, it is frequently the case that it is |
| 4414 |
* *not* the next xid in sequence. When this occurs, we must treat the |
| 4415 |
* intervening xids as running also. |
| 4416 |
*/ |
| 4417 |
if (TransactionIdFollows(xid, latestObservedXid)) |
| 4418 |
{ |
| 4419 |
TransactionId next_expected_xid; |
| 4420 |
|
| 4421 |
/* |
| 4422 |
* Extend subtrans like we do in GetNewTransactionId() during normal |
| 4423 |
* operation using individual extend steps. Note that we do not need |
| 4424 |
* to extend clog since its extensions are WAL logged. |
| 4425 |
* |
| 4426 |
* This part has to be done regardless of standbyState since we |
| 4427 |
* immediately start assigning subtransactions to their toplevel |
| 4428 |
* transactions. |
| 4429 |
*/ |
| 4430 |
next_expected_xid = latestObservedXid; |
| 4431 |
while (TransactionIdPrecedes(next_expected_xid, xid)) |
| 4432 |
{ |
| 4433 |
TransactionIdAdvance(next_expected_xid); |
| 4434 |
ExtendSUBTRANS(next_expected_xid); |
| 4435 |
} |
| 4436 |
Assert(next_expected_xid == xid); |
| 4437 |
|
| 4438 |
/* |
| 4439 |
* If the KnownAssignedXids machinery isn't up yet, there's nothing |
| 4440 |
* more to do since we don't track assigned xids yet. |
| 4441 |
*/ |
| 4442 |
if (standbyState <= STANDBY_INITIALIZED) |
| 4443 |
{ |
| 4444 |
latestObservedXid = xid; |
| 4445 |
return; |
| 4446 |
} |
| 4447 |
|
| 4448 |
/* |
| 4449 |
* Add (latestObservedXid, xid] onto the KnownAssignedXids array. |
| 4450 |
*/ |
| 4451 |
next_expected_xid = latestObservedXid; |
| 4452 |
TransactionIdAdvance(next_expected_xid); |
| 4453 |
KnownAssignedXidsAdd(next_expected_xid, xid, false); |
| 4454 |
|
| 4455 |
/* |
| 4456 |
* Now we can advance latestObservedXid |
| 4457 |
*/ |
| 4458 |
latestObservedXid = xid; |
| 4459 |
|
| 4460 |
/* TransamVariables->nextXid must be beyond any observed xid */ |
| 4461 |
AdvanceNextFullTransactionIdPastXid(latestObservedXid); |
| 4462 |
} |
| 4463 |
} |
| 4464 |
|
| 4465 |
/* |
| 4466 |
* ExpireTreeKnownAssignedTransactionIds |
| 4467 |
* Remove the given XIDs from KnownAssignedXids. |
| 4468 |
* |
| 4469 |
* Called during recovery in analogy with and in place of ProcArrayEndTransaction() |
| 4470 |
*/ |
| 4471 |
void |
| 4472 |
ExpireTreeKnownAssignedTransactionIds(TransactionId xid, int nsubxids, |
| 4473 |
TransactionId *subxids, TransactionId max_xid) |
| 4474 |
{ |
| 4475 |
Assert(standbyState >= STANDBY_INITIALIZED); |
| 4476 |
|
| 4477 |
/* |
| 4478 |
* Uses same locking as transaction commit |
| 4479 |
*/ |
| 4480 |
LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE); |
| 4481 |
|
| 4482 |
KnownAssignedXidsRemoveTree(xid, nsubxids, subxids); |
| 4483 |
|
| 4484 |
/* As in ProcArrayEndTransaction, advance latestCompletedXid */ |
| 4485 |
MaintainLatestCompletedXidRecovery(max_xid); |
| 4486 |
|
| 4487 |
/* ... and xactCompletionCount */ |
| 4488 |
TransamVariables->xactCompletionCount++; |
| 4489 |
|
| 4490 |
LWLockRelease(ProcArrayLock); |
| 4491 |
} |
| 4492 |
|
| 4493 |
/* |
| 4494 |
* ExpireAllKnownAssignedTransactionIds |
| 4495 |
* Remove all entries in KnownAssignedXids and reset lastOverflowedXid. |
| 4496 |
*/ |
| 4497 |
void |
| 4498 |
ExpireAllKnownAssignedTransactionIds(void) |
| 4499 |
{ |
| 4500 |
FullTransactionId latestXid; |
| 4501 |
|
| 4502 |
LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE); |
| 4503 |
KnownAssignedXidsRemovePreceding(InvalidTransactionId); |
| 4504 |
|
| 4505 |
/* Reset latestCompletedXid to nextXid - 1 */ |
| 4506 |
Assert(FullTransactionIdIsValid(TransamVariables->nextXid)); |
| 4507 |
latestXid = TransamVariables->nextXid; |
| 4508 |
FullTransactionIdRetreat(&latestXid); |
| 4509 |
TransamVariables->latestCompletedXid = latestXid; |
| 4510 |
|
| 4511 |
/* |
| 4512 |
* Any transactions that were in-progress were effectively aborted, so |
| 4513 |
* advance xactCompletionCount. |
| 4514 |
*/ |
| 4515 |
TransamVariables->xactCompletionCount++; |
| 4516 |
|
| 4517 |
/* |
| 4518 |
* Reset lastOverflowedXid. Currently, lastOverflowedXid has no use after |
| 4519 |
* the call of this function. But do this for unification with what |
| 4520 |
* ExpireOldKnownAssignedTransactionIds() do. |
| 4521 |
*/ |
| 4522 |
procArray->lastOverflowedXid = InvalidTransactionId; |
| 4523 |
LWLockRelease(ProcArrayLock); |
| 4524 |
} |
| 4525 |
|
| 4526 |
/* |
| 4527 |
* ExpireOldKnownAssignedTransactionIds |
| 4528 |
* Remove KnownAssignedXids entries preceding the given XID and |
| 4529 |
* potentially reset lastOverflowedXid. |
| 4530 |
*/ |
| 4531 |
void |
| 4532 |
ExpireOldKnownAssignedTransactionIds(TransactionId xid) |
| 4533 |
{ |
| 4534 |
TransactionId latestXid; |
| 4535 |
|
| 4536 |
LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE); |
| 4537 |
|
| 4538 |
/* As in ProcArrayEndTransaction, advance latestCompletedXid */ |
| 4539 |
latestXid = xid; |
| 4540 |
TransactionIdRetreat(latestXid); |
| 4541 |
MaintainLatestCompletedXidRecovery(latestXid); |
| 4542 |
|
| 4543 |
/* ... and xactCompletionCount */ |
| 4544 |
TransamVariables->xactCompletionCount++; |
| 4545 |
|
| 4546 |
/* |
| 4547 |
* Reset lastOverflowedXid if we know all transactions that have been |
| 4548 |
* possibly running are being gone. Not doing so could cause an incorrect |
| 4549 |
* lastOverflowedXid value, which makes extra snapshots be marked as |
| 4550 |
* suboverflowed. |
| 4551 |
*/ |
| 4552 |
if (TransactionIdPrecedes(procArray->lastOverflowedXid, xid)) |
| 4553 |
procArray->lastOverflowedXid = InvalidTransactionId; |
| 4554 |
KnownAssignedXidsRemovePreceding(xid); |
| 4555 |
LWLockRelease(ProcArrayLock); |
| 4556 |
} |
| 4557 |
|
| 4558 |
/* |
| 4559 |
* KnownAssignedTransactionIdsIdleMaintenance |
| 4560 |
* Opportunistically do maintenance work when the startup process |
| 4561 |
* is about to go idle. |
| 4562 |
*/ |
| 4563 |
void |
| 4564 |
KnownAssignedTransactionIdsIdleMaintenance(void) |
| 4565 |
{ |
| 4566 |
KnownAssignedXidsCompress(KAX_STARTUP_PROCESS_IDLE, false); |
| 4567 |
} |
| 4568 |
|
| 4569 |
|
| 4570 |
/* |
| 4571 |
* Private module functions to manipulate KnownAssignedXids |
| 4572 |
* |
| 4573 |
* There are 5 main uses of the KnownAssignedXids data structure: |
| 4574 |
* |
| 4575 |
* * backends taking snapshots - all valid XIDs need to be copied out |
| 4576 |
* * backends seeking to determine presence of a specific XID |
| 4577 |
* * startup process adding new known-assigned XIDs |
| 4578 |
* * startup process removing specific XIDs as transactions end |
| 4579 |
* * startup process pruning array when special WAL records arrive |
| 4580 |
* |
| 4581 |
* This data structure is known to be a hot spot during Hot Standby, so we |
| 4582 |
* go to some lengths to make these operations as efficient and as concurrent |
| 4583 |
* as possible. |
| 4584 |
* |
| 4585 |
* The XIDs are stored in an array in sorted order --- TransactionIdPrecedes |
| 4586 |
* order, to be exact --- to allow binary search for specific XIDs. Note: |
| 4587 |
* in general TransactionIdPrecedes would not provide a total order, but |
| 4588 |
* we know that the entries present at any instant should not extend across |
| 4589 |
* a large enough fraction of XID space to wrap around (the primary would |
| 4590 |
* shut down for fear of XID wrap long before that happens). So it's OK to |
| 4591 |
* use TransactionIdPrecedes as a binary-search comparator. |
| 4592 |
* |
| 4593 |
* It's cheap to maintain the sortedness during insertions, since new known |
| 4594 |
* XIDs are always reported in XID order; we just append them at the right. |
| 4595 |
* |
| 4596 |
* To keep individual deletions cheap, we need to allow gaps in the array. |
| 4597 |
* This is implemented by marking array elements as valid or invalid using |
| 4598 |
* the parallel boolean array KnownAssignedXidsValid[]. A deletion is done |
| 4599 |
* by setting KnownAssignedXidsValid[i] to false, *without* clearing the |
| 4600 |
* XID entry itself. This preserves the property that the XID entries are |
| 4601 |
* sorted, so we can do binary searches easily. Periodically we compress |
| 4602 |
* out the unused entries; that's much cheaper than having to compress the |
| 4603 |
* array immediately on every deletion. |
| 4604 |
* |
| 4605 |
* The actually valid items in KnownAssignedXids[] and KnownAssignedXidsValid[] |
| 4606 |
* are those with indexes tail <= i < head; items outside this subscript range |
| 4607 |
* have unspecified contents. When head reaches the end of the array, we |
| 4608 |
* force compression of unused entries rather than wrapping around, since |
| 4609 |
* allowing wraparound would greatly complicate the search logic. We maintain |
| 4610 |
* an explicit tail pointer so that pruning of old XIDs can be done without |
| 4611 |
* immediately moving the array contents. In most cases only a small fraction |
| 4612 |
* of the array contains valid entries at any instant. |
| 4613 |
* |
| 4614 |
* Although only the startup process can ever change the KnownAssignedXids |
| 4615 |
* data structure, we still need interlocking so that standby backends will |
| 4616 |
* not observe invalid intermediate states. The convention is that backends |
| 4617 |
* must hold shared ProcArrayLock to examine the array. To remove XIDs from |
| 4618 |
* the array, the startup process must hold ProcArrayLock exclusively, for |
| 4619 |
* the usual transactional reasons (compare commit/abort of a transaction |
| 4620 |
* during normal running). Compressing unused entries out of the array |
| 4621 |
* likewise requires exclusive lock. To add XIDs to the array, we just insert |
| 4622 |
* them into slots to the right of the head pointer and then advance the head |
| 4623 |
* pointer. This doesn't require any lock at all, but on machines with weak |
| 4624 |
* memory ordering, we need to be careful that other processors see the array |
| 4625 |
* element changes before they see the head pointer change. We handle this by |
| 4626 |
* using memory barriers when reading or writing the head/tail pointers (unless |
| 4627 |
* the caller holds ProcArrayLock exclusively). |
| 4628 |
* |
| 4629 |
* Algorithmic analysis: |
| 4630 |
* |
| 4631 |
* If we have a maximum of M slots, with N XIDs currently spread across |
| 4632 |
* S elements then we have N <= S <= M always. |
| 4633 |
* |
| 4634 |
* * Adding a new XID is O(1) and needs no lock (unless compression must |
| 4635 |
* happen) |
| 4636 |
* * Compressing the array is O(S) and requires exclusive lock |
| 4637 |
* * Removing an XID is O(logS) and requires exclusive lock |
| 4638 |
* * Taking a snapshot is O(S) and requires shared lock |
| 4639 |
* * Checking for an XID is O(logS) and requires shared lock |
| 4640 |
* |
| 4641 |
* In comparison, using a hash table for KnownAssignedXids would mean that |
| 4642 |
* taking snapshots would be O(M). If we can maintain S << M then the |
| 4643 |
* sorted array technique will deliver significantly faster snapshots. |
| 4644 |
* If we try to keep S too small then we will spend too much time compressing, |
| 4645 |
* so there is an optimal point for any workload mix. We use a heuristic to |
| 4646 |
* decide when to compress the array, though trimming also helps reduce |
| 4647 |
* frequency of compressing. The heuristic requires us to track the number of |
| 4648 |
* currently valid XIDs in the array (N). Except in special cases, we'll |
| 4649 |
* compress when S >= 2N. Bounding S at 2N in turn bounds the time for |
| 4650 |
* taking a snapshot to be O(N), which it would have to be anyway. |
| 4651 |
*/ |
| 4652 |
|
| 4653 |
|
| 4654 |
/* |
| 4655 |
* Compress KnownAssignedXids by shifting valid data down to the start of the |
| 4656 |
* array, removing any gaps. |
| 4657 |
* |
| 4658 |
* A compression step is forced if "reason" is KAX_NO_SPACE, otherwise |
| 4659 |
* we do it only if a heuristic indicates it's a good time to do it. |
| 4660 |
* |
| 4661 |
* Compression requires holding ProcArrayLock in exclusive mode. |
| 4662 |
* Caller must pass haveLock = true if it already holds the lock. |
| 4663 |
*/ |
| 4664 |
static void |
| 4665 |
KnownAssignedXidsCompress(KAXCompressReason reason, bool haveLock) |
| 4666 |
{ |
| 4667 |
ProcArrayStruct *pArray = procArray; |
| 4668 |
int head, |
| 4669 |
tail, |
| 4670 |
nelements; |
| 4671 |
int compress_index; |
| 4672 |
int i; |
| 4673 |
|
| 4674 |
/* Counters for compression heuristics */ |
| 4675 |
static unsigned int transactionEndsCounter; |
| 4676 |
static TimestampTz lastCompressTs; |
| 4677 |
|
| 4678 |
/* Tuning constants */ |
| 4679 |
#define KAX_COMPRESS_FREQUENCY 128 /* in transactions */ |
| 4680 |
#define KAX_COMPRESS_IDLE_INTERVAL 1000 /* in ms */ |
| 4681 |
|
| 4682 |
/* |
| 4683 |
* Since only the startup process modifies the head/tail pointers, we |
| 4684 |
* don't need a lock to read them here. |
| 4685 |
*/ |
| 4686 |
head = pArray->headKnownAssignedXids; |
| 4687 |
tail = pArray->tailKnownAssignedXids; |
| 4688 |
nelements = head - tail; |
| 4689 |
|
| 4690 |
/* |
| 4691 |
* If we can choose whether to compress, use a heuristic to avoid |
| 4692 |
* compressing too often or not often enough. "Compress" here simply |
| 4693 |
* means moving the values to the beginning of the array, so it is not as |
| 4694 |
* complex or costly as typical data compression algorithms. |
| 4695 |
*/ |
| 4696 |
if (nelements == pArray->numKnownAssignedXids) |
| 4697 |
{ |
| 4698 |
/* |
| 4699 |
* When there are no gaps between head and tail, don't bother to |
| 4700 |
* compress, except in the KAX_NO_SPACE case where we must compress to |
| 4701 |
* create some space after the head. |
| 4702 |
*/ |
| 4703 |
if (reason != KAX_NO_SPACE) |
| 4704 |
return; |
| 4705 |
} |
| 4706 |
else if (reason == KAX_TRANSACTION_END) |
| 4707 |
{ |
| 4708 |
/* |
| 4709 |
* Consider compressing only once every so many commits. Frequency |
| 4710 |
* determined by benchmarks. |
| 4711 |
*/ |
| 4712 |
if ((transactionEndsCounter++) % KAX_COMPRESS_FREQUENCY != 0) |
| 4713 |
return; |
| 4714 |
|
| 4715 |
/* |
| 4716 |
* Furthermore, compress only if the used part of the array is less |
| 4717 |
* than 50% full (see comments above). |
| 4718 |
*/ |
| 4719 |
if (nelements < 2 * pArray->numKnownAssignedXids) |
| 4720 |
return; |
| 4721 |
} |
| 4722 |
else if (reason == KAX_STARTUP_PROCESS_IDLE) |
| 4723 |
{ |
| 4724 |
/* |
| 4725 |
* We're about to go idle for lack of new WAL, so we might as well |
| 4726 |
* compress. But not too often, to avoid ProcArray lock contention |
| 4727 |
* with readers. |
| 4728 |
*/ |
| 4729 |
if (lastCompressTs != 0) |
| 4730 |
{ |
| 4731 |
TimestampTz compress_after; |
| 4732 |
|
| 4733 |
compress_after = TimestampTzPlusMilliseconds(lastCompressTs, |
| 4734 |
KAX_COMPRESS_IDLE_INTERVAL); |
| 4735 |
if (GetCurrentTimestamp() < compress_after) |
| 4736 |
return; |
| 4737 |
} |
| 4738 |
} |
| 4739 |
|
| 4740 |
/* Need to compress, so get the lock if we don't have it. */ |
| 4741 |
if (!haveLock) |
| 4742 |
LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE); |
| 4743 |
|
| 4744 |
/* |
| 4745 |
* We compress the array by reading the valid values from tail to head, |
| 4746 |
* re-aligning data to 0th element. |
| 4747 |
*/ |
| 4748 |
compress_index = 0; |
| 4749 |
for (i = tail; i < head; i++) |
| 4750 |
{ |
| 4751 |
if (KnownAssignedXidsValid[i]) |
| 4752 |
{ |
| 4753 |
KnownAssignedXids[compress_index] = KnownAssignedXids[i]; |
| 4754 |
KnownAssignedXidsValid[compress_index] = true; |
| 4755 |
compress_index++; |
| 4756 |
} |
| 4757 |
} |
| 4758 |
Assert(compress_index == pArray->numKnownAssignedXids); |
| 4759 |
|
| 4760 |
pArray->tailKnownAssignedXids = 0; |
| 4761 |
pArray->headKnownAssignedXids = compress_index; |
| 4762 |
|
| 4763 |
if (!haveLock) |
| 4764 |
LWLockRelease(ProcArrayLock); |
| 4765 |
|
| 4766 |
/* Update timestamp for maintenance. No need to hold lock for this. */ |
| 4767 |
lastCompressTs = GetCurrentTimestamp(); |
| 4768 |
} |
| 4769 |
|
| 4770 |
/* |
| 4771 |
* Add xids into KnownAssignedXids at the head of the array. |
| 4772 |
* |
| 4773 |
* xids from from_xid to to_xid, inclusive, are added to the array. |
| 4774 |
* |
| 4775 |
* If exclusive_lock is true then caller already holds ProcArrayLock in |
| 4776 |
* exclusive mode, so we need no extra locking here. Else caller holds no |
| 4777 |
* lock, so we need to be sure we maintain sufficient interlocks against |
| 4778 |
* concurrent readers. (Only the startup process ever calls this, so no need |
| 4779 |
* to worry about concurrent writers.) |
| 4780 |
*/ |
| 4781 |
static void |
| 4782 |
KnownAssignedXidsAdd(TransactionId from_xid, TransactionId to_xid, |
| 4783 |
bool exclusive_lock) |
| 4784 |
{ |
| 4785 |
ProcArrayStruct *pArray = procArray; |
| 4786 |
TransactionId next_xid; |
| 4787 |
int head, |
| 4788 |
tail; |
| 4789 |
int nxids; |
| 4790 |
int i; |
| 4791 |
|
| 4792 |
Assert(TransactionIdPrecedesOrEquals(from_xid, to_xid)); |
| 4793 |
|
| 4794 |
/* |
| 4795 |
* Calculate how many array slots we'll need. Normally this is cheap; in |
| 4796 |
* the unusual case where the XIDs cross the wrap point, we do it the hard |
| 4797 |
* way. |
| 4798 |
*/ |
| 4799 |
if (to_xid >= from_xid) |
| 4800 |
nxids = to_xid - from_xid + 1; |
| 4801 |
else |
| 4802 |
{ |
| 4803 |
nxids = 1; |
| 4804 |
next_xid = from_xid; |
| 4805 |
while (TransactionIdPrecedes(next_xid, to_xid)) |
| 4806 |
{ |
| 4807 |
nxids++; |
| 4808 |
TransactionIdAdvance(next_xid); |
| 4809 |
} |
| 4810 |
} |
| 4811 |
|
| 4812 |
/* |
| 4813 |
* Since only the startup process modifies the head/tail pointers, we |
| 4814 |
* don't need a lock to read them here. |
| 4815 |
*/ |
| 4816 |
head = pArray->headKnownAssignedXids; |
| 4817 |
tail = pArray->tailKnownAssignedXids; |
| 4818 |
|
| 4819 |
Assert(head >= 0 && head <= pArray->maxKnownAssignedXids); |
| 4820 |
Assert(tail >= 0 && tail < pArray->maxKnownAssignedXids); |
| 4821 |
|
| 4822 |
/* |
| 4823 |
* Verify that insertions occur in TransactionId sequence. Note that even |
| 4824 |
* if the last existing element is marked invalid, it must still have a |
| 4825 |
* correctly sequenced XID value. |
| 4826 |
*/ |
| 4827 |
if (head > tail && |
| 4828 |
TransactionIdFollowsOrEquals(KnownAssignedXids[head - 1], from_xid)) |
| 4829 |
{ |
| 4830 |
KnownAssignedXidsDisplay(LOG); |
| 4831 |
elog(ERROR, "out-of-order XID insertion in KnownAssignedXids"); |
| 4832 |
} |
| 4833 |
|
| 4834 |
/* |
| 4835 |
* If our xids won't fit in the remaining space, compress out free space |
| 4836 |
*/ |
| 4837 |
if (head + nxids > pArray->maxKnownAssignedXids) |
| 4838 |
{ |
| 4839 |
KnownAssignedXidsCompress(KAX_NO_SPACE, exclusive_lock); |
| 4840 |
|
| 4841 |
head = pArray->headKnownAssignedXids; |
| 4842 |
/* note: we no longer care about the tail pointer */ |
| 4843 |
|
| 4844 |
/* |
| 4845 |
* If it still won't fit then we're out of memory |
| 4846 |
*/ |
| 4847 |
if (head + nxids > pArray->maxKnownAssignedXids) |
| 4848 |
elog(ERROR, "too many KnownAssignedXids"); |
| 4849 |
} |
| 4850 |
|
| 4851 |
/* Now we can insert the xids into the space starting at head */ |
| 4852 |
next_xid = from_xid; |
| 4853 |
for (i = 0; i < nxids; i++) |
| 4854 |
{ |
| 4855 |
KnownAssignedXids[head] = next_xid; |
| 4856 |
KnownAssignedXidsValid[head] = true; |
| 4857 |
TransactionIdAdvance(next_xid); |
| 4858 |
head++; |
| 4859 |
} |
| 4860 |
|
| 4861 |
/* Adjust count of number of valid entries */ |
| 4862 |
pArray->numKnownAssignedXids += nxids; |
| 4863 |
|
| 4864 |
/* |
| 4865 |
* Now update the head pointer. We use a write barrier to ensure that |
| 4866 |
* other processors see the above array updates before they see the head |
| 4867 |
* pointer change. The barrier isn't required if we're holding |
| 4868 |
* ProcArrayLock exclusively. |
| 4869 |
*/ |
| 4870 |
if (!exclusive_lock) |
| 4871 |
pg_write_barrier(); |
| 4872 |
|
| 4873 |
pArray->headKnownAssignedXids = head; |
| 4874 |
} |
| 4875 |
|
| 4876 |
/* |
| 4877 |
* KnownAssignedXidsSearch |
| 4878 |
* |
| 4879 |
* Searches KnownAssignedXids for a specific xid and optionally removes it. |
| 4880 |
* Returns true if it was found, false if not. |
| 4881 |
* |
| 4882 |
* Caller must hold ProcArrayLock in shared or exclusive mode. |
| 4883 |
* Exclusive lock must be held for remove = true. |
| 4884 |
*/ |
| 4885 |
static bool |
| 4886 |
KnownAssignedXidsSearch(TransactionId xid, bool remove) |
| 4887 |
{ |
| 4888 |
ProcArrayStruct *pArray = procArray; |
| 4889 |
int first, |
| 4890 |
last; |
| 4891 |
int head; |
| 4892 |
int tail; |
| 4893 |
int result_index = -1; |
| 4894 |
|
| 4895 |
tail = pArray->tailKnownAssignedXids; |
| 4896 |
head = pArray->headKnownAssignedXids; |
| 4897 |
|
| 4898 |
/* |
| 4899 |
* Only the startup process removes entries, so we don't need the read |
| 4900 |
* barrier in that case. |
| 4901 |
*/ |
| 4902 |
if (!remove) |
| 4903 |
pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */ |
| 4904 |
|
| 4905 |
/* |
| 4906 |
* Standard binary search. Note we can ignore the KnownAssignedXidsValid |
| 4907 |
* array here, since even invalid entries will contain sorted XIDs. |
| 4908 |
*/ |
| 4909 |
first = tail; |
| 4910 |
last = head - 1; |
| 4911 |
while (first <= last) |
| 4912 |
{ |
| 4913 |
int mid_index; |
| 4914 |
TransactionId mid_xid; |
| 4915 |
|
| 4916 |
mid_index = (first + last) / 2; |
| 4917 |
mid_xid = KnownAssignedXids[mid_index]; |
| 4918 |
|
| 4919 |
if (xid == mid_xid) |
| 4920 |
{ |
| 4921 |
result_index = mid_index; |
| 4922 |
break; |
| 4923 |
} |
| 4924 |
else if (TransactionIdPrecedes(xid, mid_xid)) |
| 4925 |
last = mid_index - 1; |
| 4926 |
else |
| 4927 |
first = mid_index + 1; |
| 4928 |
} |
| 4929 |
|
| 4930 |
if (result_index < 0) |
| 4931 |
return false; /* not in array */ |
| 4932 |
|
| 4933 |
if (!KnownAssignedXidsValid[result_index]) |
| 4934 |
return false; /* in array, but invalid */ |
| 4935 |
|
| 4936 |
if (remove) |
| 4937 |
{ |
| 4938 |
KnownAssignedXidsValid[result_index] = false; |
| 4939 |
|
| 4940 |
pArray->numKnownAssignedXids--; |
| 4941 |
Assert(pArray->numKnownAssignedXids >= 0); |
| 4942 |
|
| 4943 |
/* |
| 4944 |
* If we're removing the tail element then advance tail pointer over |
| 4945 |
* any invalid elements. This will speed future searches. |
| 4946 |
*/ |
| 4947 |
if (result_index == tail) |
| 4948 |
{ |
| 4949 |
tail++; |
| 4950 |
while (tail < head && !KnownAssignedXidsValid[tail]) |
| 4951 |
tail++; |
| 4952 |
if (tail >= head) |
| 4953 |
{ |
| 4954 |
/* Array is empty, so we can reset both pointers */ |
| 4955 |
pArray->headKnownAssignedXids = 0; |
| 4956 |
pArray->tailKnownAssignedXids = 0; |
| 4957 |
} |
| 4958 |
else |
| 4959 |
{ |
| 4960 |
pArray->tailKnownAssignedXids = tail; |
| 4961 |
} |
| 4962 |
} |
| 4963 |
} |
| 4964 |
|
| 4965 |
return true; |
| 4966 |
} |
| 4967 |
|
| 4968 |
/* |
| 4969 |
* Is the specified XID present in KnownAssignedXids[]? |
| 4970 |
* |
| 4971 |
* Caller must hold ProcArrayLock in shared or exclusive mode. |
| 4972 |
*/ |
| 4973 |
static bool |
| 4974 |
KnownAssignedXidExists(TransactionId xid) |
| 4975 |
{ |
| 4976 |
Assert(TransactionIdIsValid(xid)); |
| 4977 |
|
| 4978 |
return KnownAssignedXidsSearch(xid, false); |
| 4979 |
} |
| 4980 |
|
| 4981 |
/* |
| 4982 |
* Remove the specified XID from KnownAssignedXids[]. |
| 4983 |
* |
| 4984 |
* Caller must hold ProcArrayLock in exclusive mode. |
| 4985 |
*/ |
| 4986 |
static void |
| 4987 |
KnownAssignedXidsRemove(TransactionId xid) |
| 4988 |
{ |
| 4989 |
Assert(TransactionIdIsValid(xid)); |
| 4990 |
|
| 4991 |
elog(DEBUG4, "remove KnownAssignedXid %u", xid); |
| 4992 |
|
| 4993 |
/* |
| 4994 |
* Note: we cannot consider it an error to remove an XID that's not |
| 4995 |
* present. We intentionally remove subxact IDs while processing |
| 4996 |
* XLOG_XACT_ASSIGNMENT, to avoid array overflow. Then those XIDs will be |
| 4997 |
* removed again when the top-level xact commits or aborts. |
| 4998 |
* |
| 4999 |
* It might be possible to track such XIDs to distinguish this case from |
| 5000 |
* actual errors, but it would be complicated and probably not worth it. |
| 5001 |
* So, just ignore the search result. |
| 5002 |
*/ |
| 5003 |
(void) KnownAssignedXidsSearch(xid, true); |
| 5004 |
} |
| 5005 |
|
| 5006 |
/* |
| 5007 |
* KnownAssignedXidsRemoveTree |
| 5008 |
* Remove xid (if it's not InvalidTransactionId) and all the subxids. |
| 5009 |
* |
| 5010 |
* Caller must hold ProcArrayLock in exclusive mode. |
| 5011 |
*/ |
| 5012 |
static void |
| 5013 |
KnownAssignedXidsRemoveTree(TransactionId xid, int nsubxids, |
| 5014 |
TransactionId *subxids) |
| 5015 |
{ |
| 5016 |
int i; |
| 5017 |
|
| 5018 |
if (TransactionIdIsValid(xid)) |
| 5019 |
KnownAssignedXidsRemove(xid); |
| 5020 |
|
| 5021 |
for (i = 0; i < nsubxids; i++) |
| 5022 |
KnownAssignedXidsRemove(subxids[i]); |
| 5023 |
|
| 5024 |
/* Opportunistically compress the array */ |
| 5025 |
KnownAssignedXidsCompress(KAX_TRANSACTION_END, true); |
| 5026 |
} |
| 5027 |
|
| 5028 |
/* |
| 5029 |
* Prune KnownAssignedXids up to, but *not* including xid. If xid is invalid |
| 5030 |
* then clear the whole table. |
| 5031 |
* |
| 5032 |
* Caller must hold ProcArrayLock in exclusive mode. |
| 5033 |
*/ |
| 5034 |
static void |
| 5035 |
KnownAssignedXidsRemovePreceding(TransactionId removeXid) |
| 5036 |
{ |
| 5037 |
ProcArrayStruct *pArray = procArray; |
| 5038 |
int count = 0; |
| 5039 |
int head, |
| 5040 |
tail, |
| 5041 |
i; |
| 5042 |
|
| 5043 |
if (!TransactionIdIsValid(removeXid)) |
| 5044 |
{ |
| 5045 |
elog(DEBUG4, "removing all KnownAssignedXids"); |
| 5046 |
pArray->numKnownAssignedXids = 0; |
| 5047 |
pArray->headKnownAssignedXids = pArray->tailKnownAssignedXids = 0; |
| 5048 |
return; |
| 5049 |
} |
| 5050 |
|
| 5051 |
elog(DEBUG4, "prune KnownAssignedXids to %u", removeXid); |
| 5052 |
|
| 5053 |
/* |
| 5054 |
* Mark entries invalid starting at the tail. Since array is sorted, we |
| 5055 |
* can stop as soon as we reach an entry >= removeXid. |
| 5056 |
*/ |
| 5057 |
tail = pArray->tailKnownAssignedXids; |
| 5058 |
head = pArray->headKnownAssignedXids; |
| 5059 |
|
| 5060 |
for (i = tail; i < head; i++) |
| 5061 |
{ |
| 5062 |
if (KnownAssignedXidsValid[i]) |
| 5063 |
{ |
| 5064 |
TransactionId knownXid = KnownAssignedXids[i]; |
| 5065 |
|
| 5066 |
if (TransactionIdFollowsOrEquals(knownXid, removeXid)) |
| 5067 |
break; |
| 5068 |
|
| 5069 |
if (!StandbyTransactionIdIsPrepared(knownXid)) |
| 5070 |
{ |
| 5071 |
KnownAssignedXidsValid[i] = false; |
| 5072 |
count++; |
| 5073 |
} |
| 5074 |
} |
| 5075 |
} |
| 5076 |
|
| 5077 |
pArray->numKnownAssignedXids -= count; |
| 5078 |
Assert(pArray->numKnownAssignedXids >= 0); |
| 5079 |
|
| 5080 |
/* |
| 5081 |
* Advance the tail pointer if we've marked the tail item invalid. |
| 5082 |
*/ |
| 5083 |
for (i = tail; i < head; i++) |
| 5084 |
{ |
| 5085 |
if (KnownAssignedXidsValid[i]) |
| 5086 |
break; |
| 5087 |
} |
| 5088 |
if (i >= head) |
| 5089 |
{ |
| 5090 |
/* Array is empty, so we can reset both pointers */ |
| 5091 |
pArray->headKnownAssignedXids = 0; |
| 5092 |
pArray->tailKnownAssignedXids = 0; |
| 5093 |
} |
| 5094 |
else |
| 5095 |
{ |
| 5096 |
pArray->tailKnownAssignedXids = i; |
| 5097 |
} |
| 5098 |
|
| 5099 |
/* Opportunistically compress the array */ |
| 5100 |
KnownAssignedXidsCompress(KAX_PRUNE, true); |
| 5101 |
} |
| 5102 |
|
| 5103 |
/* |
| 5104 |
* KnownAssignedXidsGet - Get an array of xids by scanning KnownAssignedXids. |
| 5105 |
* We filter out anything >= xmax. |
| 5106 |
* |
| 5107 |
* Returns the number of XIDs stored into xarray[]. Caller is responsible |
| 5108 |
* that array is large enough. |
| 5109 |
* |
| 5110 |
* Caller must hold ProcArrayLock in (at least) shared mode. |
| 5111 |
*/ |
| 5112 |
static int |
| 5113 |
KnownAssignedXidsGet(TransactionId *xarray, TransactionId xmax) |
| 5114 |
{ |
| 5115 |
TransactionId xtmp = InvalidTransactionId; |
| 5116 |
|
| 5117 |
return KnownAssignedXidsGetAndSetXmin(xarray, &xtmp, xmax); |
| 5118 |
} |
| 5119 |
|
| 5120 |
/* |
| 5121 |
* KnownAssignedXidsGetAndSetXmin - as KnownAssignedXidsGet, plus |
| 5122 |
* we reduce *xmin to the lowest xid value seen if not already lower. |
| 5123 |
* |
| 5124 |
* Caller must hold ProcArrayLock in (at least) shared mode. |
| 5125 |
*/ |
| 5126 |
static int |
| 5127 |
KnownAssignedXidsGetAndSetXmin(TransactionId *xarray, TransactionId *xmin, |
| 5128 |
TransactionId xmax) |
| 5129 |
{ |
| 5130 |
int count = 0; |
| 5131 |
int head, |
| 5132 |
tail; |
| 5133 |
int i; |
| 5134 |
|
| 5135 |
/* |
| 5136 |
* Fetch head just once, since it may change while we loop. We can stop |
| 5137 |
* once we reach the initially seen head, since we are certain that an xid |
| 5138 |
* cannot enter and then leave the array while we hold ProcArrayLock. We |
| 5139 |
* might miss newly-added xids, but they should be >= xmax so irrelevant |
| 5140 |
* anyway. |
| 5141 |
*/ |
| 5142 |
tail = procArray->tailKnownAssignedXids; |
| 5143 |
head = procArray->headKnownAssignedXids; |
| 5144 |
|
| 5145 |
pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */ |
| 5146 |
|
| 5147 |
for (i = tail; i < head; i++) |
| 5148 |
{ |
| 5149 |
/* Skip any gaps in the array */ |
| 5150 |
if (KnownAssignedXidsValid[i]) |
| 5151 |
{ |
| 5152 |
TransactionId knownXid = KnownAssignedXids[i]; |
| 5153 |
|
| 5154 |
/* |
| 5155 |
* Update xmin if required. Only the first XID need be checked, |
| 5156 |
* since the array is sorted. |
| 5157 |
*/ |
| 5158 |
if (count == 0 && |
| 5159 |
TransactionIdPrecedes(knownXid, *xmin)) |
| 5160 |
*xmin = knownXid; |
| 5161 |
|
| 5162 |
/* |
| 5163 |
* Filter out anything >= xmax, again relying on sorted property |
| 5164 |
* of array. |
| 5165 |
*/ |
| 5166 |
if (TransactionIdIsValid(xmax) && |
| 5167 |
TransactionIdFollowsOrEquals(knownXid, xmax)) |
| 5168 |
break; |
| 5169 |
|
| 5170 |
/* Add knownXid into output array */ |
| 5171 |
xarray[count++] = knownXid; |
| 5172 |
} |
| 5173 |
} |
| 5174 |
|
| 5175 |
return count; |
| 5176 |
} |
| 5177 |
|
| 5178 |
/* |
| 5179 |
* Get oldest XID in the KnownAssignedXids array, or InvalidTransactionId |
| 5180 |
* if nothing there. |
| 5181 |
*/ |
| 5182 |
static TransactionId |
| 5183 |
KnownAssignedXidsGetOldestXmin(void) |
| 5184 |
{ |
| 5185 |
int head, |
| 5186 |
tail; |
| 5187 |
int i; |
| 5188 |
|
| 5189 |
/* |
| 5190 |
* Fetch head just once, since it may change while we loop. |
| 5191 |
*/ |
| 5192 |
tail = procArray->tailKnownAssignedXids; |
| 5193 |
head = procArray->headKnownAssignedXids; |
| 5194 |
|
| 5195 |
pg_read_barrier(); /* pairs with KnownAssignedXidsAdd */ |
| 5196 |
|
| 5197 |
for (i = tail; i < head; i++) |
| 5198 |
{ |
| 5199 |
/* Skip any gaps in the array */ |
| 5200 |
if (KnownAssignedXidsValid[i]) |
| 5201 |
return KnownAssignedXids[i]; |
| 5202 |
} |
| 5203 |
|
| 5204 |
return InvalidTransactionId; |
| 5205 |
} |
| 5206 |
|
| 5207 |
/* |
| 5208 |
* Display KnownAssignedXids to provide debug trail |
| 5209 |
* |
| 5210 |
* Currently this is only called within startup process, so we need no |
| 5211 |
* special locking. |
| 5212 |
* |
| 5213 |
* Note this is pretty expensive, and much of the expense will be incurred |
| 5214 |
* even if the elog message will get discarded. It's not currently called |
| 5215 |
* in any performance-critical places, however, so no need to be tenser. |
| 5216 |
*/ |
| 5217 |
static void |
| 5218 |
KnownAssignedXidsDisplay(int trace_level) |
| 5219 |
{ |
| 5220 |
ProcArrayStruct *pArray = procArray; |
| 5221 |
StringInfoData buf; |
| 5222 |
int head, |
| 5223 |
tail, |
| 5224 |
i; |
| 5225 |
int nxids = 0; |
| 5226 |
|
| 5227 |
tail = pArray->tailKnownAssignedXids; |
| 5228 |
head = pArray->headKnownAssignedXids; |
| 5229 |
|
| 5230 |
initStringInfo(&buf); |
| 5231 |
|
| 5232 |
for (i = tail; i < head; i++) |
| 5233 |
{ |
| 5234 |
if (KnownAssignedXidsValid[i]) |
| 5235 |
{ |
| 5236 |
nxids++; |
| 5237 |
appendStringInfo(&buf, "[%d]=%u ", i, KnownAssignedXids[i]); |
| 5238 |
} |
| 5239 |
} |
| 5240 |
|
| 5241 |
elog(trace_level, "%d KnownAssignedXids (num=%d tail=%d head=%d) %s", |
| 5242 |
nxids, |
| 5243 |
pArray->numKnownAssignedXids, |
| 5244 |
pArray->tailKnownAssignedXids, |
| 5245 |
pArray->headKnownAssignedXids, |
| 5246 |
buf.data); |
| 5247 |
|
| 5248 |
pfree(buf.data); |
| 5249 |
} |
| 5250 |
|
| 5251 |
/* |
| 5252 |
* KnownAssignedXidsReset |
| 5253 |
* Resets KnownAssignedXids to be empty |
| 5254 |
*/ |
| 5255 |
static void |
| 5256 |
KnownAssignedXidsReset(void) |
| 5257 |
{ |
| 5258 |
ProcArrayStruct *pArray = procArray; |
| 5259 |
|
| 5260 |
LWLockAcquire(ProcArrayLock, LW_EXCLUSIVE); |
| 5261 |
|
| 5262 |
pArray->numKnownAssignedXids = 0; |
| 5263 |
pArray->tailKnownAssignedXids = 0; |
| 5264 |
pArray->headKnownAssignedXids = 0; |
| 5265 |
|
| 5266 |
LWLockRelease(ProcArrayLock); |
| 5267 |
} |