Age Owner Branch data TLA Line data Source code
1 : : /*-------------------------------------------------------------------------
2 : : *
3 : : * buf_internals.h
4 : : * Internal definitions for buffer manager and the buffer replacement
5 : : * strategy.
6 : : *
7 : : *
8 : : * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
9 : : * Portions Copyright (c) 1994, Regents of the University of California
10 : : *
11 : : * src/include/storage/buf_internals.h
12 : : *
13 : : *-------------------------------------------------------------------------
14 : : */
15 : : #ifndef BUFMGR_INTERNALS_H
16 : : #define BUFMGR_INTERNALS_H
17 : :
18 : : #include "pgstat.h"
19 : : #include "port/atomics.h"
20 : : #include "storage/buf.h"
21 : : #include "storage/bufmgr.h"
22 : : #include "storage/condition_variable.h"
23 : : #include "storage/latch.h"
24 : : #include "storage/lwlock.h"
25 : : #include "storage/shmem.h"
26 : : #include "storage/smgr.h"
27 : : #include "storage/spin.h"
28 : : #include "utils/relcache.h"
29 : : #include "utils/resowner.h"
30 : :
31 : : /*
32 : : * Buffer state is a single 32-bit variable where following data is combined.
33 : : *
34 : : * - 18 bits refcount
35 : : * - 4 bits usage count
36 : : * - 10 bits of flags
37 : : *
38 : : * Combining these values allows to perform some operations without locking
39 : : * the buffer header, by modifying them together with a CAS loop.
40 : : *
41 : : * The definition of buffer state components is below.
42 : : */
43 : : #define BUF_REFCOUNT_ONE 1
44 : : #define BUF_REFCOUNT_MASK ((1U << 18) - 1)
45 : : #define BUF_USAGECOUNT_MASK 0x003C0000U
46 : : #define BUF_USAGECOUNT_ONE (1U << 18)
47 : : #define BUF_USAGECOUNT_SHIFT 18
48 : : #define BUF_FLAG_MASK 0xFFC00000U
49 : :
50 : : /* Get refcount and usagecount from buffer state */
51 : : #define BUF_STATE_GET_REFCOUNT(state) ((state) & BUF_REFCOUNT_MASK)
52 : : #define BUF_STATE_GET_USAGECOUNT(state) (((state) & BUF_USAGECOUNT_MASK) >> BUF_USAGECOUNT_SHIFT)
53 : :
54 : : /*
55 : : * Flags for buffer descriptors
56 : : *
57 : : * Note: BM_TAG_VALID essentially means that there is a buffer hashtable
58 : : * entry associated with the buffer's tag.
59 : : */
60 : : #define BM_LOCKED (1U << 22) /* buffer header is locked */
61 : : #define BM_DIRTY (1U << 23) /* data needs writing */
62 : : #define BM_VALID (1U << 24) /* data is valid */
63 : : #define BM_TAG_VALID (1U << 25) /* tag is assigned */
64 : : #define BM_IO_IN_PROGRESS (1U << 26) /* read or write in progress */
65 : : #define BM_IO_ERROR (1U << 27) /* previous I/O failed */
66 : : #define BM_JUST_DIRTIED (1U << 28) /* dirtied since write started */
67 : : #define BM_PIN_COUNT_WAITER (1U << 29) /* have waiter for sole pin */
68 : : #define BM_CHECKPOINT_NEEDED (1U << 30) /* must write for checkpoint */
69 : : #define BM_PERMANENT (1U << 31) /* permanent buffer (not unlogged,
70 : : * or init fork) */
71 : : /*
72 : : * The maximum allowed value of usage_count represents a tradeoff between
73 : : * accuracy and speed of the clock-sweep buffer management algorithm. A
74 : : * large value (comparable to NBuffers) would approximate LRU semantics.
75 : : * But it can take as many as BM_MAX_USAGE_COUNT+1 complete cycles of
76 : : * clock sweeps to find a free buffer, so in practice we don't want the
77 : : * value to be very large.
78 : : */
79 : : #define BM_MAX_USAGE_COUNT 5
80 : :
81 : : /*
82 : : * Buffer tag identifies which disk block the buffer contains.
83 : : *
84 : : * Note: the BufferTag data must be sufficient to determine where to write the
85 : : * block, without reference to pg_class or pg_tablespace entries. It's
86 : : * possible that the backend flushing the buffer doesn't even believe the
87 : : * relation is visible yet (its xact may have started before the xact that
88 : : * created the rel). The storage manager must be able to cope anyway.
89 : : *
90 : : * Note: if there's any pad bytes in the struct, InitBufferTag will have
91 : : * to be fixed to zero them, since this struct is used as a hash key.
92 : : */
93 : : typedef struct buftag
94 : : {
95 : : Oid spcOid; /* tablespace oid */
96 : : Oid dbOid; /* database oid */
97 : : RelFileNumber relNumber; /* relation file number */
98 : : ForkNumber forkNum; /* fork number */
99 : : BlockNumber blockNum; /* blknum relative to begin of reln */
100 : : } BufferTag;
101 : :
102 : : static inline RelFileNumber
599 rhaas@postgresql.org 103 :CBC 142824303 : BufTagGetRelNumber(const BufferTag *tag)
104 : : {
564 105 : 142824303 : return tag->relNumber;
106 : : }
107 : :
108 : : static inline ForkNumber
599 109 : 22201668 : BufTagGetForkNum(const BufferTag *tag)
110 : : {
564 111 : 22201668 : return tag->forkNum;
112 : : }
113 : :
114 : : static inline void
599 115 : 62755850 : BufTagSetRelForkDetails(BufferTag *tag, RelFileNumber relnumber,
116 : : ForkNumber forknum)
117 : : {
564 118 : 62755850 : tag->relNumber = relnumber;
119 : 62755850 : tag->forkNum = forknum;
599 120 : 62755850 : }
121 : :
122 : : static inline RelFileLocator
123 : 18887906 : BufTagGetRelFileLocator(const BufferTag *tag)
124 : : {
125 : : RelFileLocator rlocator;
126 : :
127 : 18887906 : rlocator.spcOid = tag->spcOid;
128 : 18887906 : rlocator.dbOid = tag->dbOid;
129 : 18887906 : rlocator.relNumber = BufTagGetRelNumber(tag);
130 : :
131 : 18887906 : return rlocator;
132 : : }
133 : :
134 : : static inline void
627 135 : 9389886 : ClearBufferTag(BufferTag *tag)
136 : : {
599 137 : 9389886 : tag->spcOid = InvalidOid;
138 : 9389886 : tag->dbOid = InvalidOid;
139 : 9389886 : BufTagSetRelForkDetails(tag, InvalidRelFileNumber, InvalidForkNumber);
627 140 : 9389886 : tag->blockNum = InvalidBlockNumber;
141 : 9389886 : }
142 : :
143 : : static inline void
144 : 53365964 : InitBufferTag(BufferTag *tag, const RelFileLocator *rlocator,
145 : : ForkNumber forkNum, BlockNumber blockNum)
146 : : {
599 147 : 53365964 : tag->spcOid = rlocator->spcOid;
148 : 53365964 : tag->dbOid = rlocator->dbOid;
149 : 53365964 : BufTagSetRelForkDetails(tag, rlocator->relNumber, forkNum);
627 150 : 53365964 : tag->blockNum = blockNum;
151 : 53365964 : }
152 : :
153 : : static inline bool
154 : 2866334 : BufferTagsEqual(const BufferTag *tag1, const BufferTag *tag2)
155 : : {
599 156 : 5732602 : return (tag1->spcOid == tag2->spcOid) &&
157 [ + - ]: 2866268 : (tag1->dbOid == tag2->dbOid) &&
564 158 [ + + ]: 2866268 : (tag1->relNumber == tag2->relNumber) &&
159 [ + + + + ]: 8596591 : (tag1->blockNum == tag2->blockNum) &&
160 [ + - ]: 2863989 : (tag1->forkNum == tag2->forkNum);
161 : : }
162 : :
163 : : static inline bool
599 164 : 376645199 : BufTagMatchesRelFileLocator(const BufferTag *tag,
165 : : const RelFileLocator *rlocator)
166 : : {
167 : 539233156 : return (tag->spcOid == rlocator->spcOid) &&
168 [ + + + + ]: 500306323 : (tag->dbOid == rlocator->dbOid) &&
169 [ + + ]: 123661124 : (BufTagGetRelNumber(tag) == rlocator->relNumber);
170 : : }
171 : :
172 : :
173 : : /*
174 : : * The shared buffer mapping table is partitioned to reduce contention.
175 : : * To determine which partition lock a given tag requires, compute the tag's
176 : : * hash code with BufTableHashCode(), then apply BufMappingPartitionLock().
177 : : * NB: NUM_BUFFER_PARTITIONS must be a power of 2!
178 : : */
179 : : static inline uint32
627 180 : 52674195 : BufTableHashPartition(uint32 hashcode)
181 : : {
182 : 52674195 : return hashcode % NUM_BUFFER_PARTITIONS;
183 : : }
184 : :
185 : : static inline LWLock *
186 : 52674195 : BufMappingPartitionLock(uint32 hashcode)
187 : : {
188 : 52674195 : return &MainLWLockArray[BUFFER_MAPPING_LWLOCK_OFFSET +
189 : 52674195 : BufTableHashPartition(hashcode)].lock;
190 : : }
191 : :
192 : : static inline LWLock *
193 : : BufMappingPartitionLockByIndex(uint32 index)
194 : : {
195 : : return &MainLWLockArray[BUFFER_MAPPING_LWLOCK_OFFSET + index].lock;
196 : : }
197 : :
198 : : /*
199 : : * BufferDesc -- shared descriptor/state data for a single shared buffer.
200 : : *
201 : : * Note: Buffer header lock (BM_LOCKED flag) must be held to examine or change
202 : : * tag, state or wait_backend_pgprocno fields. In general, buffer header lock
203 : : * is a spinlock which is combined with flags, refcount and usagecount into
204 : : * single atomic variable. This layout allow us to do some operations in a
205 : : * single atomic operation, without actually acquiring and releasing spinlock;
206 : : * for instance, increase or decrease refcount. buf_id field never changes
207 : : * after initialization, so does not need locking. freeNext is protected by
208 : : * the buffer_strategy_lock not buffer header lock. The LWLock can take care
209 : : * of itself. The buffer header lock is *not* used to control access to the
210 : : * data in the buffer!
211 : : *
212 : : * It's assumed that nobody changes the state field while buffer header lock
213 : : * is held. Thus buffer header lock holder can do complex updates of the
214 : : * state variable in single write, simultaneously with lock release (cleaning
215 : : * BM_LOCKED flag). On the other hand, updating of state without holding
216 : : * buffer header lock is restricted to CAS, which ensures that BM_LOCKED flag
217 : : * is not set. Atomic increment/decrement, OR/AND etc. are not allowed.
218 : : *
219 : : * An exception is that if we have the buffer pinned, its tag can't change
220 : : * underneath us, so we can examine the tag without locking the buffer header.
221 : : * Also, in places we do one-time reads of the flags without bothering to
222 : : * lock the buffer header; this is generally for situations where we don't
223 : : * expect the flag bit being tested to be changing.
224 : : *
225 : : * We can't physically remove items from a disk page if another backend has
226 : : * the buffer pinned. Hence, a backend may need to wait for all other pins
227 : : * to go away. This is signaled by storing its own pgprocno into
228 : : * wait_backend_pgprocno and setting flag bit BM_PIN_COUNT_WAITER. At present,
229 : : * there can be only one such waiter per buffer.
230 : : *
231 : : * We use this same struct for local buffer headers, but the locks are not
232 : : * used and not all of the flag bits are useful either. To avoid unnecessary
233 : : * overhead, manipulations of the state field should be done without actual
234 : : * atomic operations (i.e. only pg_atomic_read_u32() and
235 : : * pg_atomic_unlocked_write_u32()).
236 : : *
237 : : * Be careful to avoid increasing the size of the struct when adding or
238 : : * reordering members. Keeping it below 64 bytes (the most common CPU
239 : : * cache line size) is fairly important for performance.
240 : : *
241 : : * Per-buffer I/O condition variables are currently kept outside this struct in
242 : : * a separate array. They could be moved in here and still fit within that
243 : : * limit on common systems, but for now that is not done.
244 : : */
245 : : typedef struct BufferDesc
246 : : {
247 : : BufferTag tag; /* ID of page contained in buffer */
248 : : int buf_id; /* buffer's index number (from 0) */
249 : :
250 : : /* state of the tag, containing flags, refcount and usagecount */
251 : : pg_atomic_uint32 state;
252 : :
253 : : int wait_backend_pgprocno; /* backend of pin-count waiter */
254 : : int freeNext; /* link in freelist chain */
255 : : LWLock content_lock; /* to lock access to buffer contents */
256 : : } BufferDesc;
257 : :
258 : : /*
259 : : * Concurrent access to buffer headers has proven to be more efficient if
260 : : * they're cache line aligned. So we force the start of the BufferDescriptors
261 : : * array to be on a cache line boundary and force the elements to be cache
262 : : * line sized.
263 : : *
264 : : * XXX: As this is primarily matters in highly concurrent workloads which
265 : : * probably all are 64bit these days, and the space wastage would be a bit
266 : : * more noticeable on 32bit systems, we don't force the stride to be cache
267 : : * line sized on those. If somebody does actual performance testing, we can
268 : : * reevaluate.
269 : : *
270 : : * Note that local buffer descriptors aren't forced to be aligned - as there's
271 : : * no concurrent access to those it's unlikely to be beneficial.
272 : : *
273 : : * We use a 64-byte cache line size here, because that's the most common
274 : : * size. Making it bigger would be a waste of memory. Even if running on a
275 : : * platform with either 32 or 128 byte line sizes, it's good to align to
276 : : * boundaries and avoid false sharing.
277 : : */
278 : : #define BUFFERDESC_PAD_TO_SIZE (SIZEOF_VOID_P == 8 ? 64 : 1)
279 : :
280 : : typedef union BufferDescPadded
281 : : {
282 : : BufferDesc bufferdesc;
283 : : char pad[BUFFERDESC_PAD_TO_SIZE];
284 : : } BufferDescPadded;
285 : :
286 : : /*
287 : : * The PendingWriteback & WritebackContext structure are used to keep
288 : : * information about pending flush requests to be issued to the OS.
289 : : */
290 : : typedef struct PendingWriteback
291 : : {
292 : : /* could store different types of pending flushes here */
293 : : BufferTag tag;
294 : : } PendingWriteback;
295 : :
296 : : /* struct forward declared in bufmgr.h */
297 : : typedef struct WritebackContext
298 : : {
299 : : /* pointer to the max number of writeback requests to coalesce */
300 : : int *max_pending;
301 : :
302 : : /* current number of pending writeback requests */
303 : : int nr_pending;
304 : :
305 : : /* pending requests */
306 : : PendingWriteback pending_writebacks[WRITEBACK_MAX_PENDING_FLUSHES];
307 : : } WritebackContext;
308 : :
309 : : /* in buf_init.c */
310 : : extern PGDLLIMPORT BufferDescPadded *BufferDescriptors;
311 : : extern PGDLLIMPORT ConditionVariableMinimallyPadded *BufferIOCVArray;
312 : : extern PGDLLIMPORT WritebackContext BackendWritebackContext;
313 : :
314 : : /* in localbuf.c */
315 : : extern PGDLLIMPORT BufferDesc *LocalBufferDescriptors;
316 : :
317 : :
318 : : static inline BufferDesc *
319 : 573470879 : GetBufferDescriptor(uint32 id)
320 : : {
321 : 573470879 : return &(BufferDescriptors[id]).bufferdesc;
322 : : }
323 : :
324 : : static inline BufferDesc *
325 : 8989917 : GetLocalBufferDescriptor(uint32 id)
326 : : {
327 : 8989917 : return &LocalBufferDescriptors[id];
328 : : }
329 : :
330 : : static inline Buffer
331 : 263702816 : BufferDescriptorGetBuffer(const BufferDesc *bdesc)
332 : : {
333 : 263702816 : return (Buffer) (bdesc->buf_id + 1);
334 : : }
335 : :
336 : : static inline ConditionVariable *
337 : 10156130 : BufferDescriptorGetIOCV(const BufferDesc *bdesc)
338 : : {
339 : 10156130 : return &(BufferIOCVArray[bdesc->buf_id]).cv;
340 : : }
341 : :
342 : : static inline LWLock *
343 : 256644096 : BufferDescriptorGetContentLock(const BufferDesc *bdesc)
344 : : {
345 : 256644096 : return (LWLock *) (&bdesc->content_lock);
346 : : }
347 : :
348 : : /*
349 : : * The freeNext field is either the index of the next freelist entry,
350 : : * or one of these special values:
351 : : */
352 : : #define FREENEXT_END_OF_LIST (-1)
353 : : #define FREENEXT_NOT_IN_LIST (-2)
354 : :
355 : : /*
356 : : * Functions for acquiring/releasing a shared buffer header's spinlock. Do
357 : : * not apply these to local buffers!
358 : : */
359 : : extern uint32 LockBufHdr(BufferDesc *desc);
360 : :
361 : : static inline void
362 : 26884984 : UnlockBufHdr(BufferDesc *desc, uint32 buf_state)
363 : : {
364 : 26884984 : pg_write_barrier();
365 : 26884984 : pg_atomic_write_u32(&desc->state, buf_state & (~BM_LOCKED));
366 : 26884984 : }
367 : :
368 : : /* in bufmgr.c */
369 : :
370 : : /*
371 : : * Structure to sort buffers per file on checkpoints.
372 : : *
373 : : * This structure is allocated per buffer in shared memory, so it should be
374 : : * kept as small as possible.
375 : : */
376 : : typedef struct CkptSortItem
377 : : {
378 : : Oid tsId;
379 : : RelFileNumber relNumber;
380 : : ForkNumber forkNum;
381 : : BlockNumber blockNum;
382 : : int buf_id;
383 : : } CkptSortItem;
384 : :
385 : : extern PGDLLIMPORT CkptSortItem *CkptBufferIds;
386 : :
387 : : /* ResourceOwner callbacks to hold buffer I/Os and pins */
388 : : extern PGDLLIMPORT const ResourceOwnerDesc buffer_io_resowner_desc;
389 : : extern PGDLLIMPORT const ResourceOwnerDesc buffer_pin_resowner_desc;
390 : :
391 : : /* Convenience wrappers over ResourceOwnerRemember/Forget */
392 : : static inline void
158 heikki.linnakangas@i 393 :GNC 62436614 : ResourceOwnerRememberBuffer(ResourceOwner owner, Buffer buffer)
394 : : {
395 : 62436614 : ResourceOwnerRemember(owner, Int32GetDatum(buffer), &buffer_pin_resowner_desc);
396 : 62436614 : }
397 : : static inline void
398 : 62432486 : ResourceOwnerForgetBuffer(ResourceOwner owner, Buffer buffer)
399 : : {
400 : 62432486 : ResourceOwnerForget(owner, Int32GetDatum(buffer), &buffer_pin_resowner_desc);
401 : 62432486 : }
402 : : static inline void
403 : 1875185 : ResourceOwnerRememberBufferIO(ResourceOwner owner, Buffer buffer)
404 : : {
405 : 1875185 : ResourceOwnerRemember(owner, Int32GetDatum(buffer), &buffer_io_resowner_desc);
406 : 1875185 : }
407 : : static inline void
408 : 1875170 : ResourceOwnerForgetBufferIO(ResourceOwner owner, Buffer buffer)
409 : : {
410 : 1875170 : ResourceOwnerForget(owner, Int32GetDatum(buffer), &buffer_io_resowner_desc);
411 : 1875170 : }
412 : :
413 : : /*
414 : : * Internal buffer management routines
415 : : */
416 : : /* bufmgr.c */
417 : : extern void WritebackContextInit(WritebackContext *context, int *max_pending);
418 : : extern void IssuePendingWritebacks(WritebackContext *wb_context, IOContext io_context);
419 : : extern void ScheduleBufferTagForWriteback(WritebackContext *wb_context,
420 : : IOContext io_context, BufferTag *tag);
421 : :
422 : : /* freelist.c */
423 : : extern IOContext IOContextForStrategy(BufferAccessStrategy strategy);
424 : : extern BufferDesc *StrategyGetBuffer(BufferAccessStrategy strategy,
425 : : uint32 *buf_state, bool *from_ring);
426 : : extern void StrategyFreeBuffer(BufferDesc *buf);
427 : : extern bool StrategyRejectBuffer(BufferAccessStrategy strategy,
428 : : BufferDesc *buf, bool from_ring);
429 : :
430 : : extern int StrategySyncStart(uint32 *complete_passes, uint32 *num_buf_alloc);
431 : : extern void StrategyNotifyBgWriter(int bgwprocno);
432 : :
433 : : extern Size StrategyShmemSize(void);
434 : : extern void StrategyInitialize(bool init);
435 : : extern bool have_free_buffer(void);
436 : :
437 : : /* buf_table.c */
438 : : extern Size BufTableShmemSize(int size);
439 : : extern void InitBufTable(int size);
440 : : extern uint32 BufTableHashCode(BufferTag *tagPtr);
441 : : extern int BufTableLookup(BufferTag *tagPtr, uint32 hashcode);
442 : : extern int BufTableInsert(BufferTag *tagPtr, uint32 hashcode, int buf_id);
443 : : extern void BufTableDelete(BufferTag *tagPtr, uint32 hashcode);
444 : :
445 : : /* localbuf.c */
446 : : extern bool PinLocalBuffer(BufferDesc *buf_hdr, bool adjust_usagecount);
447 : : extern void UnpinLocalBuffer(Buffer buffer);
448 : : extern void UnpinLocalBufferNoOwner(Buffer buffer);
449 : : extern PrefetchBufferResult PrefetchLocalBuffer(SMgrRelation smgr,
450 : : ForkNumber forkNum,
451 : : BlockNumber blockNum);
452 : : extern BufferDesc *LocalBufferAlloc(SMgrRelation smgr, ForkNumber forkNum,
453 : : BlockNumber blockNum, bool *foundPtr);
454 : : extern BlockNumber ExtendBufferedRelLocal(BufferManagerRelation bmr,
455 : : ForkNumber fork,
456 : : uint32 flags,
457 : : uint32 extend_by,
458 : : BlockNumber extend_upto,
459 : : Buffer *buffers,
460 : : uint32 *extended_by);
461 : : extern void MarkLocalBufferDirty(Buffer buffer);
462 : : extern void DropRelationLocalBuffers(RelFileLocator rlocator,
463 : : ForkNumber forkNum,
464 : : BlockNumber firstDelBlock);
465 : : extern void DropRelationAllLocalBuffers(RelFileLocator rlocator);
466 : : extern void AtEOXact_LocalBuffers(bool isCommit);
467 : :
468 : : #endif /* BUFMGR_INTERNALS_H */
|
