| 1 |
/*------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* freelist.c |
| 4 |
* routines for managing the buffer pool's replacement strategy. |
| 5 |
* |
| 6 |
* |
| 7 |
* Portions Copyright (c) 1996-2025, PostgreSQL Global Development Group |
| 8 |
* Portions Copyright (c) 1994, Regents of the University of California |
| 9 |
* |
| 10 |
* |
| 11 |
* IDENTIFICATION |
| 12 |
* src/backend/storage/buffer/freelist.c |
| 13 |
* |
| 14 |
*------------------------------------------------------------------------- |
| 15 |
*/ |
| 16 |
#include "postgres.h" |
| 17 |
|
| 18 |
#include "pgstat.h" |
| 19 |
#include "port/atomics.h" |
| 20 |
#include "storage/buf_internals.h" |
| 21 |
#include "storage/bufmgr.h" |
| 22 |
#include "storage/proc.h" |
| 23 |
|
| 24 |
#define INT_ACCESS_ONCE(var) ((int)(*((volatile int *)&(var)))) |
| 25 |
|
| 26 |
|
| 27 |
/* |
| 28 |
* The shared freelist control information. |
| 29 |
*/ |
| 30 |
typedef struct |
| 31 |
{ |
| 32 |
/* Spinlock: protects the values below */ |
| 33 |
slock_t buffer_strategy_lock; |
| 34 |
|
| 35 |
/* |
| 36 |
* Clock sweep hand: index of next buffer to consider grabbing. Note that |
| 37 |
* this isn't a concrete buffer - we only ever increase the value. So, to |
| 38 |
* get an actual buffer, it needs to be used modulo NBuffers. |
| 39 |
*/ |
| 40 |
pg_atomic_uint32 nextVictimBuffer; |
| 41 |
|
| 42 |
int firstFreeBuffer; /* Head of list of unused buffers */ |
| 43 |
int lastFreeBuffer; /* Tail of list of unused buffers */ |
| 44 |
|
| 45 |
/* |
| 46 |
* NOTE: lastFreeBuffer is undefined when firstFreeBuffer is -1 (that is, |
| 47 |
* when the list is empty) |
| 48 |
*/ |
| 49 |
|
| 50 |
/* |
| 51 |
* Statistics. These counters should be wide enough that they can't |
| 52 |
* overflow during a single bgwriter cycle. |
| 53 |
*/ |
| 54 |
uint32 completePasses; /* Complete cycles of the clock sweep */ |
| 55 |
pg_atomic_uint32 numBufferAllocs; /* Buffers allocated since last reset */ |
| 56 |
|
| 57 |
/* |
| 58 |
* Bgworker process to be notified upon activity or -1 if none. See |
| 59 |
* StrategyNotifyBgWriter. |
| 60 |
*/ |
| 61 |
int bgwprocno; |
| 62 |
} BufferStrategyControl; |
| 63 |
|
| 64 |
/* Pointers to shared state */ |
| 65 |
static BufferStrategyControl *StrategyControl = NULL; |
| 66 |
|
| 67 |
/* |
| 68 |
* Private (non-shared) state for managing a ring of shared buffers to re-use. |
| 69 |
* This is currently the only kind of BufferAccessStrategy object, but someday |
| 70 |
* we might have more kinds. |
| 71 |
*/ |
| 72 |
typedef struct BufferAccessStrategyData |
| 73 |
{ |
| 74 |
/* Overall strategy type */ |
| 75 |
BufferAccessStrategyType btype; |
| 76 |
/* Number of elements in buffers[] array */ |
| 77 |
int nbuffers; |
| 78 |
|
| 79 |
/* |
| 80 |
* Index of the "current" slot in the ring, ie, the one most recently |
| 81 |
* returned by GetBufferFromRing. |
| 82 |
*/ |
| 83 |
int current; |
| 84 |
|
| 85 |
/* |
| 86 |
* Array of buffer numbers. InvalidBuffer (that is, zero) indicates we |
| 87 |
* have not yet selected a buffer for this ring slot. For allocation |
| 88 |
* simplicity this is palloc'd together with the fixed fields of the |
| 89 |
* struct. |
| 90 |
*/ |
| 91 |
Buffer buffers[FLEXIBLE_ARRAY_MEMBER]; |
| 92 |
} BufferAccessStrategyData; |
| 93 |
|
| 94 |
|
| 95 |
/* Prototypes for internal functions */ |
| 96 |
static BufferDesc *GetBufferFromRing(BufferAccessStrategy strategy, |
| 97 |
uint32 *buf_state); |
| 98 |
static void AddBufferToRing(BufferAccessStrategy strategy, |
| 99 |
BufferDesc *buf); |
| 100 |
|
| 101 |
/* |
| 102 |
* ClockSweepTick - Helper routine for StrategyGetBuffer() |
| 103 |
* |
| 104 |
* Move the clock hand one buffer ahead of its current position and return the |
| 105 |
* id of the buffer now under the hand. |
| 106 |
*/ |
| 107 |
static inline uint32 |
| 108 |
ClockSweepTick(void) |
| 109 |
{ |
| 110 |
uint32 victim; |
| 111 |
|
| 112 |
/* |
| 113 |
* Atomically move hand ahead one buffer - if there's several processes |
| 114 |
* doing this, this can lead to buffers being returned slightly out of |
| 115 |
* apparent order. |
| 116 |
*/ |
| 117 |
victim = |
| 118 |
pg_atomic_fetch_add_u32(&StrategyControl->nextVictimBuffer, 1); |
| 119 |
|
| 120 |
if (victim >= NBuffers) |
| 121 |
{ |
| 122 |
uint32 originalVictim = victim; |
| 123 |
|
| 124 |
/* always wrap what we look up in BufferDescriptors */ |
| 125 |
victim = victim % NBuffers; |
| 126 |
|
| 127 |
/* |
| 128 |
* If we're the one that just caused a wraparound, force |
| 129 |
* completePasses to be incremented while holding the spinlock. We |
| 130 |
* need the spinlock so StrategySyncStart() can return a consistent |
| 131 |
* value consisting of nextVictimBuffer and completePasses. |
| 132 |
*/ |
| 133 |
if (victim == 0) |
| 134 |
{ |
| 135 |
uint32 expected; |
| 136 |
uint32 wrapped; |
| 137 |
bool success = false; |
| 138 |
|
| 139 |
expected = originalVictim + 1; |
| 140 |
|
| 141 |
while (!success) |
| 142 |
{ |
| 143 |
/* |
| 144 |
* Acquire the spinlock while increasing completePasses. That |
| 145 |
* allows other readers to read nextVictimBuffer and |
| 146 |
* completePasses in a consistent manner which is required for |
| 147 |
* StrategySyncStart(). In theory delaying the increment |
| 148 |
* could lead to an overflow of nextVictimBuffers, but that's |
| 149 |
* highly unlikely and wouldn't be particularly harmful. |
| 150 |
*/ |
| 151 |
SpinLockAcquire(&StrategyControl->buffer_strategy_lock); |
| 152 |
|
| 153 |
wrapped = expected % NBuffers; |
| 154 |
|
| 155 |
success = pg_atomic_compare_exchange_u32(&StrategyControl->nextVictimBuffer, |
| 156 |
&expected, wrapped); |
| 157 |
if (success) |
| 158 |
StrategyControl->completePasses++; |
| 159 |
SpinLockRelease(&StrategyControl->buffer_strategy_lock); |
| 160 |
} |
| 161 |
} |
| 162 |
} |
| 163 |
return victim; |
| 164 |
} |
| 165 |
|
| 166 |
/* |
| 167 |
* have_free_buffer -- a lockless check to see if there is a free buffer in |
| 168 |
* buffer pool. |
| 169 |
* |
| 170 |
* If the result is true that will become stale once free buffers are moved out |
| 171 |
* by other operations, so the caller who strictly want to use a free buffer |
| 172 |
* should not call this. |
| 173 |
*/ |
| 174 |
bool |
| 175 |
have_free_buffer(void) |
| 176 |
{ |
| 177 |
if (StrategyControl->firstFreeBuffer >= 0) |
| 178 |
return true; |
| 179 |
else |
| 180 |
return false; |
| 181 |
} |
| 182 |
|
| 183 |
/* |
| 184 |
* StrategyGetBuffer |
| 185 |
* |
| 186 |
* Called by the bufmgr to get the next candidate buffer to use in |
| 187 |
* BufferAlloc(). The only hard requirement BufferAlloc() has is that |
| 188 |
* the selected buffer must not currently be pinned by anyone. |
| 189 |
* |
| 190 |
* strategy is a BufferAccessStrategy object, or NULL for default strategy. |
| 191 |
* |
| 192 |
* To ensure that no one else can pin the buffer before we do, we must |
| 193 |
* return the buffer with the buffer header spinlock still held. |
| 194 |
*/ |
| 195 |
BufferDesc * |
| 196 |
StrategyGetBuffer(BufferAccessStrategy strategy, uint32 *buf_state, bool *from_ring) |
| 197 |
{ |
| 198 |
BufferDesc *buf; |
| 199 |
int bgwprocno; |
| 200 |
int trycounter; |
| 201 |
uint32 local_buf_state; /* to avoid repeated (de-)referencing */ |
| 202 |
|
| 203 |
*from_ring = false; |
| 204 |
|
| 205 |
/* |
| 206 |
* If given a strategy object, see whether it can select a buffer. We |
| 207 |
* assume strategy objects don't need buffer_strategy_lock. |
| 208 |
*/ |
| 209 |
if (strategy != NULL) |
| 210 |
{ |
| 211 |
buf = GetBufferFromRing(strategy, buf_state); |
| 212 |
if (buf != NULL) |
| 213 |
{ |
| 214 |
*from_ring = true; |
| 215 |
return buf; |
| 216 |
} |
| 217 |
} |
| 218 |
|
| 219 |
/* |
| 220 |
* If asked, we need to waken the bgwriter. Since we don't want to rely on |
| 221 |
* a spinlock for this we force a read from shared memory once, and then |
| 222 |
* set the latch based on that value. We need to go through that length |
| 223 |
* because otherwise bgwprocno might be reset while/after we check because |
| 224 |
* the compiler might just reread from memory. |
| 225 |
* |
| 226 |
* This can possibly set the latch of the wrong process if the bgwriter |
| 227 |
* dies in the wrong moment. But since PGPROC->procLatch is never |
| 228 |
* deallocated the worst consequence of that is that we set the latch of |
| 229 |
* some arbitrary process. |
| 230 |
*/ |
| 231 |
bgwprocno = INT_ACCESS_ONCE(StrategyControl->bgwprocno); |
| 232 |
if (bgwprocno != -1) |
| 233 |
{ |
| 234 |
/* reset bgwprocno first, before setting the latch */ |
| 235 |
StrategyControl->bgwprocno = -1; |
| 236 |
|
| 237 |
/* |
| 238 |
* Not acquiring ProcArrayLock here which is slightly icky. It's |
| 239 |
* actually fine because procLatch isn't ever freed, so we just can |
| 240 |
* potentially set the wrong process' (or no process') latch. |
| 241 |
*/ |
| 242 |
SetLatch(&ProcGlobal->allProcs[bgwprocno].procLatch); |
| 243 |
} |
| 244 |
|
| 245 |
/* |
| 246 |
* We count buffer allocation requests so that the bgwriter can estimate |
| 247 |
* the rate of buffer consumption. Note that buffers recycled by a |
| 248 |
* strategy object are intentionally not counted here. |
| 249 |
*/ |
| 250 |
pg_atomic_fetch_add_u32(&StrategyControl->numBufferAllocs, 1); |
| 251 |
|
| 252 |
/* |
| 253 |
* First check, without acquiring the lock, whether there's buffers in the |
| 254 |
* freelist. Since we otherwise don't require the spinlock in every |
| 255 |
* StrategyGetBuffer() invocation, it'd be sad to acquire it here - |
| 256 |
* uselessly in most cases. That obviously leaves a race where a buffer is |
| 257 |
* put on the freelist but we don't see the store yet - but that's pretty |
| 258 |
* harmless, it'll just get used during the next buffer acquisition. |
| 259 |
* |
| 260 |
* If there's buffers on the freelist, acquire the spinlock to pop one |
| 261 |
* buffer of the freelist. Then check whether that buffer is usable and |
| 262 |
* repeat if not. |
| 263 |
* |
| 264 |
* Note that the freeNext fields are considered to be protected by the |
| 265 |
* buffer_strategy_lock not the individual buffer spinlocks, so it's OK to |
| 266 |
* manipulate them without holding the spinlock. |
| 267 |
*/ |
| 268 |
if (StrategyControl->firstFreeBuffer >= 0) |
| 269 |
{ |
| 270 |
while (true) |
| 271 |
{ |
| 272 |
/* Acquire the spinlock to remove element from the freelist */ |
| 273 |
SpinLockAcquire(&StrategyControl->buffer_strategy_lock); |
| 274 |
|
| 275 |
if (StrategyControl->firstFreeBuffer < 0) |
| 276 |
{ |
| 277 |
SpinLockRelease(&StrategyControl->buffer_strategy_lock); |
| 278 |
break; |
| 279 |
} |
| 280 |
|
| 281 |
buf = GetBufferDescriptor(StrategyControl->firstFreeBuffer); |
| 282 |
Assert(buf->freeNext != FREENEXT_NOT_IN_LIST); |
| 283 |
|
| 284 |
/* Unconditionally remove buffer from freelist */ |
| 285 |
StrategyControl->firstFreeBuffer = buf->freeNext; |
| 286 |
buf->freeNext = FREENEXT_NOT_IN_LIST; |
| 287 |
|
| 288 |
/* |
| 289 |
* Release the lock so someone else can access the freelist while |
| 290 |
* we check out this buffer. |
| 291 |
*/ |
| 292 |
SpinLockRelease(&StrategyControl->buffer_strategy_lock); |
| 293 |
|
| 294 |
/* |
| 295 |
* If the buffer is pinned or has a nonzero usage_count, we cannot |
| 296 |
* use it; discard it and retry. (This can only happen if VACUUM |
| 297 |
* put a valid buffer in the freelist and then someone else used |
| 298 |
* it before we got to it. It's probably impossible altogether as |
| 299 |
* of 8.3, but we'd better check anyway.) |
| 300 |
*/ |
| 301 |
local_buf_state = LockBufHdr(buf); |
| 302 |
if (BUF_STATE_GET_REFCOUNT(local_buf_state) == 0 |
| 303 |
&& BUF_STATE_GET_USAGECOUNT(local_buf_state) == 0) |
| 304 |
{ |
| 305 |
if (strategy != NULL) |
| 306 |
AddBufferToRing(strategy, buf); |
| 307 |
*buf_state = local_buf_state; |
| 308 |
return buf; |
| 309 |
} |
| 310 |
UnlockBufHdr(buf, local_buf_state); |
| 311 |
} |
| 312 |
} |
| 313 |
|
| 314 |
/* Nothing on the freelist, so run the "clock sweep" algorithm */ |
| 315 |
trycounter = NBuffers; |
| 316 |
for (;;) |
| 317 |
{ |
| 318 |
buf = GetBufferDescriptor(ClockSweepTick()); |
| 319 |
|
| 320 |
/* |
| 321 |
* If the buffer is pinned or has a nonzero usage_count, we cannot use |
| 322 |
* it; decrement the usage_count (unless pinned) and keep scanning. |
| 323 |
*/ |
| 324 |
local_buf_state = LockBufHdr(buf); |
| 325 |
|
| 326 |
if (BUF_STATE_GET_REFCOUNT(local_buf_state) == 0) |
| 327 |
{ |
| 328 |
if (BUF_STATE_GET_USAGECOUNT(local_buf_state) != 0) |
| 329 |
{ |
| 330 |
local_buf_state -= BUF_USAGECOUNT_ONE; |
| 331 |
|
| 332 |
trycounter = NBuffers; |
| 333 |
} |
| 334 |
else |
| 335 |
{ |
| 336 |
/* Found a usable buffer */ |
| 337 |
if (strategy != NULL) |
| 338 |
AddBufferToRing(strategy, buf); |
| 339 |
*buf_state = local_buf_state; |
| 340 |
return buf; |
| 341 |
} |
| 342 |
} |
| 343 |
else if (--trycounter == 0) |
| 344 |
{ |
| 345 |
/* |
| 346 |
* We've scanned all the buffers without making any state changes, |
| 347 |
* so all the buffers are pinned (or were when we looked at them). |
| 348 |
* We could hope that someone will free one eventually, but it's |
| 349 |
* probably better to fail than to risk getting stuck in an |
| 350 |
* infinite loop. |
| 351 |
*/ |
| 352 |
UnlockBufHdr(buf, local_buf_state); |
| 353 |
elog(ERROR, "no unpinned buffers available"); |
| 354 |
} |
| 355 |
UnlockBufHdr(buf, local_buf_state); |
| 356 |
} |
| 357 |
} |
| 358 |
|
| 359 |
/* |
| 360 |
* StrategyFreeBuffer: put a buffer on the freelist |
| 361 |
*/ |
| 362 |
void |
| 363 |
StrategyFreeBuffer(BufferDesc *buf) |
| 364 |
{ |
| 365 |
SpinLockAcquire(&StrategyControl->buffer_strategy_lock); |
| 366 |
|
| 367 |
/* |
| 368 |
* It is possible that we are told to put something in the freelist that |
| 369 |
* is already in it; don't screw up the list if so. |
| 370 |
*/ |
| 371 |
if (buf->freeNext == FREENEXT_NOT_IN_LIST) |
| 372 |
{ |
| 373 |
buf->freeNext = StrategyControl->firstFreeBuffer; |
| 374 |
if (buf->freeNext < 0) |
| 375 |
StrategyControl->lastFreeBuffer = buf->buf_id; |
| 376 |
StrategyControl->firstFreeBuffer = buf->buf_id; |
| 377 |
} |
| 378 |
|
| 379 |
SpinLockRelease(&StrategyControl->buffer_strategy_lock); |
| 380 |
} |
| 381 |
|
| 382 |
/* |
| 383 |
* StrategySyncStart -- tell BgBufferSync where to start syncing |
| 384 |
* |
| 385 |
* The result is the buffer index of the best buffer to sync first. |
| 386 |
* BgBufferSync() will proceed circularly around the buffer array from there. |
| 387 |
* |
| 388 |
* In addition, we return the completed-pass count (which is effectively |
| 389 |
* the higher-order bits of nextVictimBuffer) and the count of recent buffer |
| 390 |
* allocs if non-NULL pointers are passed. The alloc count is reset after |
| 391 |
* being read. |
| 392 |
*/ |
| 393 |
int |
| 394 |
StrategySyncStart(uint32 *complete_passes, uint32 *num_buf_alloc) |
| 395 |
{ |
| 396 |
uint32 nextVictimBuffer; |
| 397 |
int result; |
| 398 |
|
| 399 |
SpinLockAcquire(&StrategyControl->buffer_strategy_lock); |
| 400 |
nextVictimBuffer = pg_atomic_read_u32(&StrategyControl->nextVictimBuffer); |
| 401 |
result = nextVictimBuffer % NBuffers; |
| 402 |
|
| 403 |
if (complete_passes) |
| 404 |
{ |
| 405 |
*complete_passes = StrategyControl->completePasses; |
| 406 |
|
| 407 |
/* |
| 408 |
* Additionally add the number of wraparounds that happened before |
| 409 |
* completePasses could be incremented. C.f. ClockSweepTick(). |
| 410 |
*/ |
| 411 |
*complete_passes += nextVictimBuffer / NBuffers; |
| 412 |
} |
| 413 |
|
| 414 |
if (num_buf_alloc) |
| 415 |
{ |
| 416 |
*num_buf_alloc = pg_atomic_exchange_u32(&StrategyControl->numBufferAllocs, 0); |
| 417 |
} |
| 418 |
SpinLockRelease(&StrategyControl->buffer_strategy_lock); |
| 419 |
return result; |
| 420 |
} |
| 421 |
|
| 422 |
/* |
| 423 |
* StrategyNotifyBgWriter -- set or clear allocation notification latch |
| 424 |
* |
| 425 |
* If bgwprocno isn't -1, the next invocation of StrategyGetBuffer will |
| 426 |
* set that latch. Pass -1 to clear the pending notification before it |
| 427 |
* happens. This feature is used by the bgwriter process to wake itself up |
| 428 |
* from hibernation, and is not meant for anybody else to use. |
| 429 |
*/ |
| 430 |
void |
| 431 |
StrategyNotifyBgWriter(int bgwprocno) |
| 432 |
{ |
| 433 |
/* |
| 434 |
* We acquire buffer_strategy_lock just to ensure that the store appears |
| 435 |
* atomic to StrategyGetBuffer. The bgwriter should call this rather |
| 436 |
* infrequently, so there's no performance penalty from being safe. |
| 437 |
*/ |
| 438 |
SpinLockAcquire(&StrategyControl->buffer_strategy_lock); |
| 439 |
StrategyControl->bgwprocno = bgwprocno; |
| 440 |
SpinLockRelease(&StrategyControl->buffer_strategy_lock); |
| 441 |
} |
| 442 |
|
| 443 |
|
| 444 |
/* |
| 445 |
* StrategyShmemSize |
| 446 |
* |
| 447 |
* estimate the size of shared memory used by the freelist-related structures. |
| 448 |
* |
| 449 |
* Note: for somewhat historical reasons, the buffer lookup hashtable size |
| 450 |
* is also determined here. |
| 451 |
*/ |
| 452 |
Size |
| 453 |
StrategyShmemSize(void) |
| 454 |
{ |
| 455 |
Size size = 0; |
| 456 |
|
| 457 |
/* size of lookup hash table ... see comment in StrategyInitialize */ |
| 458 |
size = add_size(size, BufTableShmemSize(NBuffers + NUM_BUFFER_PARTITIONS)); |
| 459 |
|
| 460 |
/* size of the shared replacement strategy control block */ |
| 461 |
size = add_size(size, MAXALIGN(sizeof(BufferStrategyControl))); |
| 462 |
|
| 463 |
return size; |
| 464 |
} |
| 465 |
|
| 466 |
/* |
| 467 |
* StrategyInitialize -- initialize the buffer cache replacement |
| 468 |
* strategy. |
| 469 |
* |
| 470 |
* Assumes: All of the buffers are already built into a linked list. |
| 471 |
* Only called by postmaster and only during initialization. |
| 472 |
*/ |
| 473 |
void |
| 474 |
StrategyInitialize(bool init) |
| 475 |
{ |
| 476 |
bool found; |
| 477 |
|
| 478 |
/* |
| 479 |
* Initialize the shared buffer lookup hashtable. |
| 480 |
* |
| 481 |
* Since we can't tolerate running out of lookup table entries, we must be |
| 482 |
* sure to specify an adequate table size here. The maximum steady-state |
| 483 |
* usage is of course NBuffers entries, but BufferAlloc() tries to insert |
| 484 |
* a new entry before deleting the old. In principle this could be |
| 485 |
* happening in each partition concurrently, so we could need as many as |
| 486 |
* NBuffers + NUM_BUFFER_PARTITIONS entries. |
| 487 |
*/ |
| 488 |
InitBufTable(NBuffers + NUM_BUFFER_PARTITIONS); |
| 489 |
|
| 490 |
/* |
| 491 |
* Get or create the shared strategy control block |
| 492 |
*/ |
| 493 |
StrategyControl = (BufferStrategyControl *) |
| 494 |
ShmemInitStruct("Buffer Strategy Status", |
| 495 |
sizeof(BufferStrategyControl), |
| 496 |
&found); |
| 497 |
|
| 498 |
if (!found) |
| 499 |
{ |
| 500 |
/* |
| 501 |
* Only done once, usually in postmaster |
| 502 |
*/ |
| 503 |
Assert(init); |
| 504 |
|
| 505 |
SpinLockInit(&StrategyControl->buffer_strategy_lock); |
| 506 |
|
| 507 |
/* |
| 508 |
* Grab the whole linked list of free buffers for our strategy. We |
| 509 |
* assume it was previously set up by BufferManagerShmemInit(). |
| 510 |
*/ |
| 511 |
StrategyControl->firstFreeBuffer = 0; |
| 512 |
StrategyControl->lastFreeBuffer = NBuffers - 1; |
| 513 |
|
| 514 |
/* Initialize the clock sweep pointer */ |
| 515 |
pg_atomic_init_u32(&StrategyControl->nextVictimBuffer, 0); |
| 516 |
|
| 517 |
/* Clear statistics */ |
| 518 |
StrategyControl->completePasses = 0; |
| 519 |
pg_atomic_init_u32(&StrategyControl->numBufferAllocs, 0); |
| 520 |
|
| 521 |
/* No pending notification */ |
| 522 |
StrategyControl->bgwprocno = -1; |
| 523 |
} |
| 524 |
else |
| 525 |
Assert(!init); |
| 526 |
} |
| 527 |
|
| 528 |
|
| 529 |
/* ---------------------------------------------------------------- |
| 530 |
* Backend-private buffer ring management |
| 531 |
* ---------------------------------------------------------------- |
| 532 |
*/ |
| 533 |
|
| 534 |
|
| 535 |
/* |
| 536 |
* GetAccessStrategy -- create a BufferAccessStrategy object |
| 537 |
* |
| 538 |
* The object is allocated in the current memory context. |
| 539 |
*/ |
| 540 |
BufferAccessStrategy |
| 541 |
GetAccessStrategy(BufferAccessStrategyType btype) |
| 542 |
{ |
| 543 |
int ring_size_kb; |
| 544 |
|
| 545 |
/* |
| 546 |
* Select ring size to use. See buffer/README for rationales. |
| 547 |
* |
| 548 |
* Note: if you change the ring size for BAS_BULKREAD, see also |
| 549 |
* SYNC_SCAN_REPORT_INTERVAL in access/heap/syncscan.c. |
| 550 |
*/ |
| 551 |
switch (btype) |
| 552 |
{ |
| 553 |
case BAS_NORMAL: |
| 554 |
/* if someone asks for NORMAL, just give 'em a "default" object */ |
| 555 |
return NULL; |
| 556 |
|
| 557 |
case BAS_BULKREAD: |
| 558 |
{ |
| 559 |
int ring_max_kb; |
| 560 |
|
| 561 |
/* |
| 562 |
* The ring always needs to be large enough to allow some |
| 563 |
* separation in time between providing a buffer to the user |
| 564 |
* of the strategy and that buffer being reused. Otherwise the |
| 565 |
* user's pin will prevent reuse of the buffer, even without |
| 566 |
* concurrent activity. |
| 567 |
* |
| 568 |
* We also need to ensure the ring always is large enough for |
| 569 |
* SYNC_SCAN_REPORT_INTERVAL, as noted above. |
| 570 |
* |
| 571 |
* Thus we start out a minimal size and increase the size |
| 572 |
* further if appropriate. |
| 573 |
*/ |
| 574 |
ring_size_kb = 256; |
| 575 |
|
| 576 |
/* |
| 577 |
* There's no point in a larger ring if we won't be allowed to |
| 578 |
* pin sufficiently many buffers. But we never limit to less |
| 579 |
* than the minimal size above. |
| 580 |
*/ |
| 581 |
ring_max_kb = GetPinLimit() * (BLCKSZ / 1024); |
| 582 |
ring_max_kb = Max(ring_size_kb, ring_max_kb); |
| 583 |
|
| 584 |
/* |
| 585 |
* We would like the ring to additionally have space for the |
| 586 |
* configured degree of IO concurrency. While being read in, |
| 587 |
* buffers can obviously not yet be reused. |
| 588 |
* |
| 589 |
* Each IO can be up to io_combine_limit blocks large, and we |
| 590 |
* want to start up to effective_io_concurrency IOs. |
| 591 |
* |
| 592 |
* Note that effective_io_concurrency may be 0, which disables |
| 593 |
* AIO. |
| 594 |
*/ |
| 595 |
ring_size_kb += (BLCKSZ / 1024) * |
| 596 |
io_combine_limit * effective_io_concurrency; |
| 597 |
|
| 598 |
if (ring_size_kb > ring_max_kb) |
| 599 |
ring_size_kb = ring_max_kb; |
| 600 |
break; |
| 601 |
} |
| 602 |
case BAS_BULKWRITE: |
| 603 |
ring_size_kb = 16 * 1024; |
| 604 |
break; |
| 605 |
case BAS_VACUUM: |
| 606 |
ring_size_kb = 2048; |
| 607 |
break; |
| 608 |
|
| 609 |
default: |
| 610 |
elog(ERROR, "unrecognized buffer access strategy: %d", |
| 611 |
(int) btype); |
| 612 |
return NULL; /* keep compiler quiet */ |
| 613 |
} |
| 614 |
|
| 615 |
return GetAccessStrategyWithSize(btype, ring_size_kb); |
| 616 |
} |
| 617 |
|
| 618 |
/* |
| 619 |
* GetAccessStrategyWithSize -- create a BufferAccessStrategy object with a |
| 620 |
* number of buffers equivalent to the passed in size. |
| 621 |
* |
| 622 |
* If the given ring size is 0, no BufferAccessStrategy will be created and |
| 623 |
* the function will return NULL. ring_size_kb must not be negative. |
| 624 |
*/ |
| 625 |
BufferAccessStrategy |
| 626 |
GetAccessStrategyWithSize(BufferAccessStrategyType btype, int ring_size_kb) |
| 627 |
{ |
| 628 |
int ring_buffers; |
| 629 |
BufferAccessStrategy strategy; |
| 630 |
|
| 631 |
Assert(ring_size_kb >= 0); |
| 632 |
|
| 633 |
/* Figure out how many buffers ring_size_kb is */ |
| 634 |
ring_buffers = ring_size_kb / (BLCKSZ / 1024); |
| 635 |
|
| 636 |
/* 0 means unlimited, so no BufferAccessStrategy required */ |
| 637 |
if (ring_buffers == 0) |
| 638 |
return NULL; |
| 639 |
|
| 640 |
/* Cap to 1/8th of shared_buffers */ |
| 641 |
ring_buffers = Min(NBuffers / 8, ring_buffers); |
| 642 |
|
| 643 |
/* NBuffers should never be less than 16, so this shouldn't happen */ |
| 644 |
Assert(ring_buffers > 0); |
| 645 |
|
| 646 |
/* Allocate the object and initialize all elements to zeroes */ |
| 647 |
strategy = (BufferAccessStrategy) |
| 648 |
palloc0(offsetof(BufferAccessStrategyData, buffers) + |
| 649 |
ring_buffers * sizeof(Buffer)); |
| 650 |
|
| 651 |
/* Set fields that don't start out zero */ |
| 652 |
strategy->btype = btype; |
| 653 |
strategy->nbuffers = ring_buffers; |
| 654 |
|
| 655 |
return strategy; |
| 656 |
} |
| 657 |
|
| 658 |
/* |
| 659 |
* GetAccessStrategyBufferCount -- an accessor for the number of buffers in |
| 660 |
* the ring |
| 661 |
* |
| 662 |
* Returns 0 on NULL input to match behavior of GetAccessStrategyWithSize() |
| 663 |
* returning NULL with 0 size. |
| 664 |
*/ |
| 665 |
int |
| 666 |
GetAccessStrategyBufferCount(BufferAccessStrategy strategy) |
| 667 |
{ |
| 668 |
if (strategy == NULL) |
| 669 |
return 0; |
| 670 |
|
| 671 |
return strategy->nbuffers; |
| 672 |
} |
| 673 |
|
| 674 |
/* |
| 675 |
* GetAccessStrategyPinLimit -- get cap of number of buffers that should be pinned |
| 676 |
* |
| 677 |
* When pinning extra buffers to look ahead, users of a ring-based strategy are |
| 678 |
* in danger of pinning too much of the ring at once while performing look-ahead. |
| 679 |
* For some strategies, that means "escaping" from the ring, and in others it |
| 680 |
* means forcing dirty data to disk very frequently with associated WAL |
| 681 |
* flushing. Since external code has no insight into any of that, allow |
| 682 |
* individual strategy types to expose a clamp that should be applied when |
| 683 |
* deciding on a maximum number of buffers to pin at once. |
| 684 |
* |
| 685 |
* Callers should combine this number with other relevant limits and take the |
| 686 |
* minimum. |
| 687 |
*/ |
| 688 |
int |
| 689 |
GetAccessStrategyPinLimit(BufferAccessStrategy strategy) |
| 690 |
{ |
| 691 |
if (strategy == NULL) |
| 692 |
return NBuffers; |
| 693 |
|
| 694 |
switch (strategy->btype) |
| 695 |
{ |
| 696 |
case BAS_BULKREAD: |
| 697 |
|
| 698 |
/* |
| 699 |
* Since BAS_BULKREAD uses StrategyRejectBuffer(), dirty buffers |
| 700 |
* shouldn't be a problem and the caller is free to pin up to the |
| 701 |
* entire ring at once. |
| 702 |
*/ |
| 703 |
return strategy->nbuffers; |
| 704 |
|
| 705 |
default: |
| 706 |
|
| 707 |
/* |
| 708 |
* Tell caller not to pin more than half the buffers in the ring. |
| 709 |
* This is a trade-off between look ahead distance and deferring |
| 710 |
* writeback and associated WAL traffic. |
| 711 |
*/ |
| 712 |
return strategy->nbuffers / 2; |
| 713 |
} |
| 714 |
} |
| 715 |
|
| 716 |
/* |
| 717 |
* FreeAccessStrategy -- release a BufferAccessStrategy object |
| 718 |
* |
| 719 |
* A simple pfree would do at the moment, but we would prefer that callers |
| 720 |
* don't assume that much about the representation of BufferAccessStrategy. |
| 721 |
*/ |
| 722 |
void |
| 723 |
FreeAccessStrategy(BufferAccessStrategy strategy) |
| 724 |
{ |
| 725 |
/* don't crash if called on a "default" strategy */ |
| 726 |
if (strategy != NULL) |
| 727 |
pfree(strategy); |
| 728 |
} |
| 729 |
|
| 730 |
/* |
| 731 |
* GetBufferFromRing -- returns a buffer from the ring, or NULL if the |
| 732 |
* ring is empty / not usable. |
| 733 |
* |
| 734 |
* The bufhdr spin lock is held on the returned buffer. |
| 735 |
*/ |
| 736 |
static BufferDesc * |
| 737 |
GetBufferFromRing(BufferAccessStrategy strategy, uint32 *buf_state) |
| 738 |
{ |
| 739 |
BufferDesc *buf; |
| 740 |
Buffer bufnum; |
| 741 |
uint32 local_buf_state; /* to avoid repeated (de-)referencing */ |
| 742 |
|
| 743 |
|
| 744 |
/* Advance to next ring slot */ |
| 745 |
if (++strategy->current >= strategy->nbuffers) |
| 746 |
strategy->current = 0; |
| 747 |
|
| 748 |
/* |
| 749 |
* If the slot hasn't been filled yet, tell the caller to allocate a new |
| 750 |
* buffer with the normal allocation strategy. He will then fill this |
| 751 |
* slot by calling AddBufferToRing with the new buffer. |
| 752 |
*/ |
| 753 |
bufnum = strategy->buffers[strategy->current]; |
| 754 |
if (bufnum == InvalidBuffer) |
| 755 |
return NULL; |
| 756 |
|
| 757 |
/* |
| 758 |
* If the buffer is pinned we cannot use it under any circumstances. |
| 759 |
* |
| 760 |
* If usage_count is 0 or 1 then the buffer is fair game (we expect 1, |
| 761 |
* since our own previous usage of the ring element would have left it |
| 762 |
* there, but it might've been decremented by clock sweep since then). A |
| 763 |
* higher usage_count indicates someone else has touched the buffer, so we |
| 764 |
* shouldn't re-use it. |
| 765 |
*/ |
| 766 |
buf = GetBufferDescriptor(bufnum - 1); |
| 767 |
local_buf_state = LockBufHdr(buf); |
| 768 |
if (BUF_STATE_GET_REFCOUNT(local_buf_state) == 0 |
| 769 |
&& BUF_STATE_GET_USAGECOUNT(local_buf_state) <= 1) |
| 770 |
{ |
| 771 |
*buf_state = local_buf_state; |
| 772 |
return buf; |
| 773 |
} |
| 774 |
UnlockBufHdr(buf, local_buf_state); |
| 775 |
|
| 776 |
/* |
| 777 |
* Tell caller to allocate a new buffer with the normal allocation |
| 778 |
* strategy. He'll then replace this ring element via AddBufferToRing. |
| 779 |
*/ |
| 780 |
return NULL; |
| 781 |
} |
| 782 |
|
| 783 |
/* |
| 784 |
* AddBufferToRing -- add a buffer to the buffer ring |
| 785 |
* |
| 786 |
* Caller must hold the buffer header spinlock on the buffer. Since this |
| 787 |
* is called with the spinlock held, it had better be quite cheap. |
| 788 |
*/ |
| 789 |
static void |
| 790 |
AddBufferToRing(BufferAccessStrategy strategy, BufferDesc *buf) |
| 791 |
{ |
| 792 |
strategy->buffers[strategy->current] = BufferDescriptorGetBuffer(buf); |
| 793 |
} |
| 794 |
|
| 795 |
/* |
| 796 |
* Utility function returning the IOContext of a given BufferAccessStrategy's |
| 797 |
* strategy ring. |
| 798 |
*/ |
| 799 |
IOContext |
| 800 |
IOContextForStrategy(BufferAccessStrategy strategy) |
| 801 |
{ |
| 802 |
if (!strategy) |
| 803 |
return IOCONTEXT_NORMAL; |
| 804 |
|
| 805 |
switch (strategy->btype) |
| 806 |
{ |
| 807 |
case BAS_NORMAL: |
| 808 |
|
| 809 |
/* |
| 810 |
* Currently, GetAccessStrategy() returns NULL for |
| 811 |
* BufferAccessStrategyType BAS_NORMAL, so this case is |
| 812 |
* unreachable. |
| 813 |
*/ |
| 814 |
pg_unreachable(); |
| 815 |
return IOCONTEXT_NORMAL; |
| 816 |
case BAS_BULKREAD: |
| 817 |
return IOCONTEXT_BULKREAD; |
| 818 |
case BAS_BULKWRITE: |
| 819 |
return IOCONTEXT_BULKWRITE; |
| 820 |
case BAS_VACUUM: |
| 821 |
return IOCONTEXT_VACUUM; |
| 822 |
} |
| 823 |
|
| 824 |
elog(ERROR, "unrecognized BufferAccessStrategyType: %d", strategy->btype); |
| 825 |
pg_unreachable(); |
| 826 |
} |
| 827 |
|
| 828 |
/* |
| 829 |
* StrategyRejectBuffer -- consider rejecting a dirty buffer |
| 830 |
* |
| 831 |
* When a nondefault strategy is used, the buffer manager calls this function |
| 832 |
* when it turns out that the buffer selected by StrategyGetBuffer needs to |
| 833 |
* be written out and doing so would require flushing WAL too. This gives us |
| 834 |
* a chance to choose a different victim. |
| 835 |
* |
| 836 |
* Returns true if buffer manager should ask for a new victim, and false |
| 837 |
* if this buffer should be written and re-used. |
| 838 |
*/ |
| 839 |
bool |
| 840 |
StrategyRejectBuffer(BufferAccessStrategy strategy, BufferDesc *buf, bool from_ring) |
| 841 |
{ |
| 842 |
/* We only do this in bulkread mode */ |
| 843 |
if (strategy->btype != BAS_BULKREAD) |
| 844 |
return false; |
| 845 |
|
| 846 |
/* Don't muck with behavior of normal buffer-replacement strategy */ |
| 847 |
if (!from_ring || |
| 848 |
strategy->buffers[strategy->current] != BufferDescriptorGetBuffer(buf)) |
| 849 |
return false; |
| 850 |
|
| 851 |
/* |
| 852 |
* Remove the dirty buffer from the ring; necessary to prevent infinite |
| 853 |
* loop if all ring members are dirty. |
| 854 |
*/ |
| 855 |
strategy->buffers[strategy->current] = InvalidBuffer; |
| 856 |
|
| 857 |
return true; |
| 858 |
} |