Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * slru.c
4 : * Simple LRU buffering for wrap-around-able permanent metadata
5 : *
6 : * This module is used to maintain various pieces of transaction status
7 : * indexed by TransactionId (such as commit status, parent transaction ID,
8 : * commit timestamp), as well as storage for multixacts, serializable
9 : * isolation locks and NOTIFY traffic. Extensions can define their own
10 : * SLRUs, too.
11 : *
12 : * Under ordinary circumstances we expect that write traffic will occur
13 : * mostly to the latest page (and to the just-prior page, soon after a
14 : * page transition). Read traffic will probably touch a larger span of
15 : * pages, but a relatively small number of buffers should be sufficient.
16 : *
17 : * We use a simple least-recently-used scheme to manage a pool of shared
18 : * page buffers, split in banks by the lowest bits of the page number, and
19 : * the management algorithm only processes the bank to which the desired
20 : * page belongs, so a linear search is sufficient; there's no need for a
21 : * hashtable or anything fancy. The algorithm is straight LRU except that
22 : * we will never swap out the latest page (since we know it's going to be
23 : * hit again eventually).
24 : *
25 : * We use per-bank control LWLocks to protect the shared data structures,
26 : * plus per-buffer LWLocks that synchronize I/O for each buffer. The
27 : * bank's control lock must be held to examine or modify any of the bank's
28 : * shared state. A process that is reading in or writing out a page
29 : * buffer does not hold the control lock, only the per-buffer lock for the
30 : * buffer it is working on. One exception is latest_page_number, which is
31 : * read and written using atomic ops.
32 : *
33 : * "Holding the bank control lock" means exclusive lock in all cases
34 : * except for SimpleLruReadPage_ReadOnly(); see comments for
35 : * SlruRecentlyUsed() for the implications of that.
36 : *
37 : * When initiating I/O on a buffer, we acquire the per-buffer lock exclusively
38 : * before releasing the control lock. The per-buffer lock is released after
39 : * completing the I/O, re-acquiring the control lock, and updating the shared
40 : * state. (Deadlock is not possible here, because we never try to initiate
41 : * I/O when someone else is already doing I/O on the same buffer.)
42 : * To wait for I/O to complete, release the control lock, acquire the
43 : * per-buffer lock in shared mode, immediately release the per-buffer lock,
44 : * reacquire the control lock, and then recheck state (since arbitrary things
45 : * could have happened while we didn't have the lock).
46 : *
47 : * As with the regular buffer manager, it is possible for another process
48 : * to re-dirty a page that is currently being written out. This is handled
49 : * by re-setting the page's page_dirty flag.
50 : *
51 : *
52 : * Portions Copyright (c) 1996-2024, PostgreSQL Global Development Group
53 : * Portions Copyright (c) 1994, Regents of the University of California
54 : *
55 : * src/backend/access/transam/slru.c
56 : *
57 : *-------------------------------------------------------------------------
58 : */
59 : #include "postgres.h"
60 :
61 : #include <fcntl.h>
62 : #include <sys/stat.h>
63 : #include <unistd.h>
64 :
65 : #include "access/slru.h"
66 : #include "access/transam.h"
67 : #include "access/xlog.h"
68 : #include "access/xlogutils.h"
69 : #include "miscadmin.h"
70 : #include "pgstat.h"
71 : #include "storage/fd.h"
72 : #include "storage/shmem.h"
73 : #include "utils/guc_hooks.h"
74 :
75 : static inline int
76 6858 : SlruFileName(SlruCtl ctl, char *path, int64 segno)
77 : {
78 6858 : if (ctl->long_segment_names)
79 : {
80 : /*
81 : * We could use 16 characters here but the disadvantage would be that
82 : * the SLRU segments will be hard to distinguish from WAL segments.
83 : *
84 : * For this reason we use 15 characters. It is enough but also means
85 : * that in the future we can't decrease SLRU_PAGES_PER_SEGMENT easily.
86 : */
87 : Assert(segno >= 0 && segno <= INT64CONST(0xFFFFFFFFFFFFFFF));
88 294 : return snprintf(path, MAXPGPATH, "%s/%015llX", ctl->Dir,
89 : (long long) segno);
90 : }
91 : else
92 : {
93 : /*
94 : * Despite the fact that %04X format string is used up to 24 bit
95 : * integers are allowed. See SlruCorrectSegmentFilenameLength()
96 : */
97 : Assert(segno >= 0 && segno <= INT64CONST(0xFFFFFF));
98 6564 : return snprintf(path, MAXPGPATH, "%s/%04X", (ctl)->Dir,
99 : (unsigned int) segno);
100 : }
101 : }
102 :
103 : /*
104 : * During SimpleLruWriteAll(), we will usually not need to write more than one
105 : * or two physical files, but we may need to write several pages per file. We
106 : * can consolidate the I/O requests by leaving files open until control returns
107 : * to SimpleLruWriteAll(). This data structure remembers which files are open.
108 : */
109 : #define MAX_WRITEALL_BUFFERS 16
110 :
111 : typedef struct SlruWriteAllData
112 : {
113 : int num_files; /* # files actually open */
114 : int fd[MAX_WRITEALL_BUFFERS]; /* their FD's */
115 : int64 segno[MAX_WRITEALL_BUFFERS]; /* their log seg#s */
116 : } SlruWriteAllData;
117 :
118 : typedef struct SlruWriteAllData *SlruWriteAll;
119 :
120 :
121 : /*
122 : * Bank size for the slot array. Pages are assigned a bank according to their
123 : * page number, with each bank being this size. We want a power of 2 so that
124 : * we can determine the bank number for a page with just bit shifting; we also
125 : * want to keep the bank size small so that LRU victim search is fast. 16
126 : * buffers per bank seems a good number.
127 : */
128 : #define SLRU_BANK_BITSHIFT 4
129 : #define SLRU_BANK_SIZE (1 << SLRU_BANK_BITSHIFT)
130 :
131 : /*
132 : * Macro to get the bank number to which the slot belongs.
133 : */
134 : #define SlotGetBankNumber(slotno) ((slotno) >> SLRU_BANK_BITSHIFT)
135 :
136 :
137 : /*
138 : * Populate a file tag describing a segment file. We only use the segment
139 : * number, since we can derive everything else we need by having separate
140 : * sync handler functions for clog, multixact etc.
141 : */
142 : #define INIT_SLRUFILETAG(a,xx_handler,xx_segno) \
143 : ( \
144 : memset(&(a), 0, sizeof(FileTag)), \
145 : (a).handler = (xx_handler), \
146 : (a).segno = (xx_segno) \
147 : )
148 :
149 : /* Saved info for SlruReportIOError */
150 : typedef enum
151 : {
152 : SLRU_OPEN_FAILED,
153 : SLRU_SEEK_FAILED,
154 : SLRU_READ_FAILED,
155 : SLRU_WRITE_FAILED,
156 : SLRU_FSYNC_FAILED,
157 : SLRU_CLOSE_FAILED,
158 : } SlruErrorCause;
159 :
160 : static SlruErrorCause slru_errcause;
161 : static int slru_errno;
162 :
163 :
164 : static void SimpleLruZeroLSNs(SlruCtl ctl, int slotno);
165 : static void SimpleLruWaitIO(SlruCtl ctl, int slotno);
166 : static void SlruInternalWritePage(SlruCtl ctl, int slotno, SlruWriteAll fdata);
167 : static bool SlruPhysicalReadPage(SlruCtl ctl, int64 pageno, int slotno);
168 : static bool SlruPhysicalWritePage(SlruCtl ctl, int64 pageno, int slotno,
169 : SlruWriteAll fdata);
170 : static void SlruReportIOError(SlruCtl ctl, int64 pageno, TransactionId xid);
171 : static int SlruSelectLRUPage(SlruCtl ctl, int64 pageno);
172 :
173 : static bool SlruScanDirCbDeleteCutoff(SlruCtl ctl, char *filename,
174 : int64 segpage, void *data);
175 : static void SlruInternalDeleteSegment(SlruCtl ctl, int64 segno);
176 : static inline void SlruRecentlyUsed(SlruShared shared, int slotno);
177 :
178 :
179 : /*
180 : * Initialization of shared memory
181 : */
182 :
183 : Size
184 35466 : SimpleLruShmemSize(int nslots, int nlsns)
185 : {
186 35466 : int nbanks = nslots / SLRU_BANK_SIZE;
187 : Size sz;
188 :
189 : Assert(nslots <= SLRU_MAX_ALLOWED_BUFFERS);
190 : Assert(nslots % SLRU_BANK_SIZE == 0);
191 :
192 : /* we assume nslots isn't so large as to risk overflow */
193 35466 : sz = MAXALIGN(sizeof(SlruSharedData));
194 35466 : sz += MAXALIGN(nslots * sizeof(char *)); /* page_buffer[] */
195 35466 : sz += MAXALIGN(nslots * sizeof(SlruPageStatus)); /* page_status[] */
196 35466 : sz += MAXALIGN(nslots * sizeof(bool)); /* page_dirty[] */
197 35466 : sz += MAXALIGN(nslots * sizeof(int64)); /* page_number[] */
198 35466 : sz += MAXALIGN(nslots * sizeof(int)); /* page_lru_count[] */
199 35466 : sz += MAXALIGN(nslots * sizeof(LWLockPadded)); /* buffer_locks[] */
200 35466 : sz += MAXALIGN(nbanks * sizeof(LWLockPadded)); /* bank_locks[] */
201 35466 : sz += MAXALIGN(nbanks * sizeof(int)); /* bank_cur_lru_count[] */
202 :
203 35466 : if (nlsns > 0)
204 5066 : sz += MAXALIGN(nslots * nlsns * sizeof(XLogRecPtr)); /* group_lsn[] */
205 :
206 35466 : return BUFFERALIGN(sz) + BLCKSZ * nslots;
207 : }
208 :
209 : /*
210 : * Determine a number of SLRU buffers to use.
211 : *
212 : * We simply divide shared_buffers by the divisor given and cap
213 : * that at the maximum given; but always at least SLRU_BANK_SIZE.
214 : * Round down to the nearest multiple of SLRU_BANK_SIZE.
215 : */
216 : int
217 15114 : SimpleLruAutotuneBuffers(int divisor, int max)
218 : {
219 15114 : return Min(max - (max % SLRU_BANK_SIZE),
220 : Max(SLRU_BANK_SIZE,
221 : NBuffers / divisor - (NBuffers / divisor) % SLRU_BANK_SIZE));
222 : }
223 :
224 : /*
225 : * Initialize, or attach to, a simple LRU cache in shared memory.
226 : *
227 : * ctl: address of local (unshared) control structure.
228 : * name: name of SLRU. (This is user-visible, pick with care!)
229 : * nslots: number of page slots to use.
230 : * nlsns: number of LSN groups per page (set to zero if not relevant).
231 : * ctllock: LWLock to use to control access to the shared control structure.
232 : * subdir: PGDATA-relative subdirectory that will contain the files.
233 : * buffer_tranche_id: tranche ID to use for the SLRU's per-buffer LWLocks.
234 : * bank_tranche_id: tranche ID to use for the bank LWLocks.
235 : * sync_handler: which set of functions to use to handle sync requests
236 : */
237 : void
238 12378 : SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
239 : const char *subdir, int buffer_tranche_id, int bank_tranche_id,
240 : SyncRequestHandler sync_handler, bool long_segment_names)
241 : {
242 : SlruShared shared;
243 : bool found;
244 12378 : int nbanks = nslots / SLRU_BANK_SIZE;
245 :
246 : Assert(nslots <= SLRU_MAX_ALLOWED_BUFFERS);
247 :
248 12378 : shared = (SlruShared) ShmemInitStruct(name,
249 : SimpleLruShmemSize(nslots, nlsns),
250 : &found);
251 :
252 12378 : if (!IsUnderPostmaster)
253 : {
254 : /* Initialize locks and shared memory area */
255 : char *ptr;
256 : Size offset;
257 :
258 : Assert(!found);
259 :
260 12378 : memset(shared, 0, sizeof(SlruSharedData));
261 :
262 12378 : shared->num_slots = nslots;
263 12378 : shared->lsn_groups_per_page = nlsns;
264 :
265 12378 : pg_atomic_init_u64(&shared->latest_page_number, 0);
266 :
267 12378 : shared->slru_stats_idx = pgstat_get_slru_index(name);
268 :
269 12378 : ptr = (char *) shared;
270 12378 : offset = MAXALIGN(sizeof(SlruSharedData));
271 12378 : shared->page_buffer = (char **) (ptr + offset);
272 12378 : offset += MAXALIGN(nslots * sizeof(char *));
273 12378 : shared->page_status = (SlruPageStatus *) (ptr + offset);
274 12378 : offset += MAXALIGN(nslots * sizeof(SlruPageStatus));
275 12378 : shared->page_dirty = (bool *) (ptr + offset);
276 12378 : offset += MAXALIGN(nslots * sizeof(bool));
277 12378 : shared->page_number = (int64 *) (ptr + offset);
278 12378 : offset += MAXALIGN(nslots * sizeof(int64));
279 12378 : shared->page_lru_count = (int *) (ptr + offset);
280 12378 : offset += MAXALIGN(nslots * sizeof(int));
281 :
282 : /* Initialize LWLocks */
283 12378 : shared->buffer_locks = (LWLockPadded *) (ptr + offset);
284 12378 : offset += MAXALIGN(nslots * sizeof(LWLockPadded));
285 12378 : shared->bank_locks = (LWLockPadded *) (ptr + offset);
286 12378 : offset += MAXALIGN(nbanks * sizeof(LWLockPadded));
287 12378 : shared->bank_cur_lru_count = (int *) (ptr + offset);
288 12378 : offset += MAXALIGN(nbanks * sizeof(int));
289 :
290 12378 : if (nlsns > 0)
291 : {
292 1768 : shared->group_lsn = (XLogRecPtr *) (ptr + offset);
293 1768 : offset += MAXALIGN(nslots * nlsns * sizeof(XLogRecPtr));
294 : }
295 :
296 12378 : ptr += BUFFERALIGN(offset);
297 314042 : for (int slotno = 0; slotno < nslots; slotno++)
298 : {
299 301664 : LWLockInitialize(&shared->buffer_locks[slotno].lock,
300 : buffer_tranche_id);
301 :
302 301664 : shared->page_buffer[slotno] = ptr;
303 301664 : shared->page_status[slotno] = SLRU_PAGE_EMPTY;
304 301664 : shared->page_dirty[slotno] = false;
305 301664 : shared->page_lru_count[slotno] = 0;
306 301664 : ptr += BLCKSZ;
307 : }
308 :
309 : /* Initialize the slot banks. */
310 31232 : for (int bankno = 0; bankno < nbanks; bankno++)
311 : {
312 18854 : LWLockInitialize(&shared->bank_locks[bankno].lock, bank_tranche_id);
313 18854 : shared->bank_cur_lru_count[bankno] = 0;
314 : }
315 :
316 : /* Should fit to estimated shmem size */
317 : Assert(ptr - (char *) shared <= SimpleLruShmemSize(nslots, nlsns));
318 : }
319 : else
320 : {
321 : Assert(found);
322 : Assert(shared->num_slots == nslots);
323 : }
324 :
325 : /*
326 : * Initialize the unshared control struct, including directory path. We
327 : * assume caller set PagePrecedes.
328 : */
329 12378 : ctl->shared = shared;
330 12378 : ctl->sync_handler = sync_handler;
331 12378 : ctl->long_segment_names = long_segment_names;
332 12378 : ctl->bank_mask = (nslots / SLRU_BANK_SIZE) - 1;
333 12378 : strlcpy(ctl->Dir, subdir, sizeof(ctl->Dir));
334 12378 : }
335 :
336 : /*
337 : * Helper function for GUC check_hook to check whether slru buffers are in
338 : * multiples of SLRU_BANK_SIZE.
339 : */
340 : bool
341 18112 : check_slru_buffers(const char *name, int *newval)
342 : {
343 : /* Valid values are multiples of SLRU_BANK_SIZE */
344 18112 : if (*newval % SLRU_BANK_SIZE == 0)
345 18112 : return true;
346 :
347 0 : GUC_check_errdetail("\"%s\" must be a multiple of %d", name,
348 : SLRU_BANK_SIZE);
349 0 : return false;
350 : }
351 :
352 : /*
353 : * Initialize (or reinitialize) a page to zeroes.
354 : *
355 : * The page is not actually written, just set up in shared memory.
356 : * The slot number of the new page is returned.
357 : *
358 : * Bank lock must be held at entry, and will be held at exit.
359 : */
360 : int
361 2420 : SimpleLruZeroPage(SlruCtl ctl, int64 pageno)
362 : {
363 2420 : SlruShared shared = ctl->shared;
364 : int slotno;
365 :
366 : Assert(LWLockHeldByMeInMode(SimpleLruGetBankLock(ctl, pageno), LW_EXCLUSIVE));
367 :
368 : /* Find a suitable buffer slot for the page */
369 2420 : slotno = SlruSelectLRUPage(ctl, pageno);
370 : Assert(shared->page_status[slotno] == SLRU_PAGE_EMPTY ||
371 : (shared->page_status[slotno] == SLRU_PAGE_VALID &&
372 : !shared->page_dirty[slotno]) ||
373 : shared->page_number[slotno] == pageno);
374 :
375 : /* Mark the slot as containing this page */
376 2420 : shared->page_number[slotno] = pageno;
377 2420 : shared->page_status[slotno] = SLRU_PAGE_VALID;
378 2420 : shared->page_dirty[slotno] = true;
379 2420 : SlruRecentlyUsed(shared, slotno);
380 :
381 : /* Set the buffer to zeroes */
382 2420 : MemSet(shared->page_buffer[slotno], 0, BLCKSZ);
383 :
384 : /* Set the LSNs for this new page to zero */
385 2420 : SimpleLruZeroLSNs(ctl, slotno);
386 :
387 : /*
388 : * Assume this page is now the latest active page.
389 : *
390 : * Note that because both this routine and SlruSelectLRUPage run with
391 : * ControlLock held, it is not possible for this to be zeroing a page that
392 : * SlruSelectLRUPage is going to evict simultaneously. Therefore, there's
393 : * no memory barrier here.
394 : */
395 2420 : pg_atomic_write_u64(&shared->latest_page_number, pageno);
396 :
397 : /* update the stats counter of zeroed pages */
398 2420 : pgstat_count_slru_page_zeroed(shared->slru_stats_idx);
399 :
400 2420 : return slotno;
401 : }
402 :
403 : /*
404 : * Zero all the LSNs we store for this slru page.
405 : *
406 : * This should be called each time we create a new page, and each time we read
407 : * in a page from disk into an existing buffer. (Such an old page cannot
408 : * have any interesting LSNs, since we'd have flushed them before writing
409 : * the page in the first place.)
410 : *
411 : * This assumes that InvalidXLogRecPtr is bitwise-all-0.
412 : */
413 : static void
414 5196 : SimpleLruZeroLSNs(SlruCtl ctl, int slotno)
415 : {
416 5196 : SlruShared shared = ctl->shared;
417 :
418 5196 : if (shared->lsn_groups_per_page > 0)
419 1556 : MemSet(&shared->group_lsn[slotno * shared->lsn_groups_per_page], 0,
420 : shared->lsn_groups_per_page * sizeof(XLogRecPtr));
421 5196 : }
422 :
423 : /*
424 : * Wait for any active I/O on a page slot to finish. (This does not
425 : * guarantee that new I/O hasn't been started before we return, though.
426 : * In fact the slot might not even contain the same page anymore.)
427 : *
428 : * Bank lock must be held at entry, and will be held at exit.
429 : */
430 : static void
431 0 : SimpleLruWaitIO(SlruCtl ctl, int slotno)
432 : {
433 0 : SlruShared shared = ctl->shared;
434 0 : int bankno = SlotGetBankNumber(slotno);
435 :
436 : Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
437 :
438 : /* See notes at top of file */
439 0 : LWLockRelease(&shared->bank_locks[bankno].lock);
440 0 : LWLockAcquire(&shared->buffer_locks[slotno].lock, LW_SHARED);
441 0 : LWLockRelease(&shared->buffer_locks[slotno].lock);
442 0 : LWLockAcquire(&shared->bank_locks[bankno].lock, LW_EXCLUSIVE);
443 :
444 : /*
445 : * If the slot is still in an io-in-progress state, then either someone
446 : * already started a new I/O on the slot, or a previous I/O failed and
447 : * neglected to reset the page state. That shouldn't happen, really, but
448 : * it seems worth a few extra cycles to check and recover from it. We can
449 : * cheaply test for failure by seeing if the buffer lock is still held (we
450 : * assume that transaction abort would release the lock).
451 : */
452 0 : if (shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS ||
453 0 : shared->page_status[slotno] == SLRU_PAGE_WRITE_IN_PROGRESS)
454 : {
455 0 : if (LWLockConditionalAcquire(&shared->buffer_locks[slotno].lock, LW_SHARED))
456 : {
457 : /* indeed, the I/O must have failed */
458 0 : if (shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS)
459 0 : shared->page_status[slotno] = SLRU_PAGE_EMPTY;
460 : else /* write_in_progress */
461 : {
462 0 : shared->page_status[slotno] = SLRU_PAGE_VALID;
463 0 : shared->page_dirty[slotno] = true;
464 : }
465 0 : LWLockRelease(&shared->buffer_locks[slotno].lock);
466 : }
467 : }
468 0 : }
469 :
470 : /*
471 : * Find a page in a shared buffer, reading it in if necessary.
472 : * The page number must correspond to an already-initialized page.
473 : *
474 : * If write_ok is true then it is OK to return a page that is in
475 : * WRITE_IN_PROGRESS state; it is the caller's responsibility to be sure
476 : * that modification of the page is safe. If write_ok is false then we
477 : * will not return the page until it is not undergoing active I/O.
478 : *
479 : * The passed-in xid is used only for error reporting, and may be
480 : * InvalidTransactionId if no specific xid is associated with the action.
481 : *
482 : * Return value is the shared-buffer slot number now holding the page.
483 : * The buffer's LRU access info is updated.
484 : *
485 : * The correct bank lock must be held at entry, and will be held at exit.
486 : */
487 : int
488 293244 : SimpleLruReadPage(SlruCtl ctl, int64 pageno, bool write_ok,
489 : TransactionId xid)
490 : {
491 293244 : SlruShared shared = ctl->shared;
492 293244 : LWLock *banklock = SimpleLruGetBankLock(ctl, pageno);
493 :
494 : Assert(LWLockHeldByMeInMode(banklock, LW_EXCLUSIVE));
495 :
496 : /* Outer loop handles restart if we must wait for someone else's I/O */
497 : for (;;)
498 0 : {
499 : int slotno;
500 : bool ok;
501 :
502 : /* See if page already is in memory; if not, pick victim slot */
503 293244 : slotno = SlruSelectLRUPage(ctl, pageno);
504 :
505 : /* Did we find the page in memory? */
506 293244 : if (shared->page_status[slotno] != SLRU_PAGE_EMPTY &&
507 290476 : shared->page_number[slotno] == pageno)
508 : {
509 : /*
510 : * If page is still being read in, we must wait for I/O. Likewise
511 : * if the page is being written and the caller said that's not OK.
512 : */
513 290468 : if (shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS ||
514 290468 : (shared->page_status[slotno] == SLRU_PAGE_WRITE_IN_PROGRESS &&
515 0 : !write_ok))
516 : {
517 0 : SimpleLruWaitIO(ctl, slotno);
518 : /* Now we must recheck state from the top */
519 0 : continue;
520 : }
521 : /* Otherwise, it's ready to use */
522 290468 : SlruRecentlyUsed(shared, slotno);
523 :
524 : /* update the stats counter of pages found in the SLRU */
525 290468 : pgstat_count_slru_page_hit(shared->slru_stats_idx);
526 :
527 290468 : return slotno;
528 : }
529 :
530 : /* We found no match; assert we selected a freeable slot */
531 : Assert(shared->page_status[slotno] == SLRU_PAGE_EMPTY ||
532 : (shared->page_status[slotno] == SLRU_PAGE_VALID &&
533 : !shared->page_dirty[slotno]));
534 :
535 : /* Mark the slot read-busy */
536 2776 : shared->page_number[slotno] = pageno;
537 2776 : shared->page_status[slotno] = SLRU_PAGE_READ_IN_PROGRESS;
538 2776 : shared->page_dirty[slotno] = false;
539 :
540 : /* Acquire per-buffer lock (cannot deadlock, see notes at top) */
541 2776 : LWLockAcquire(&shared->buffer_locks[slotno].lock, LW_EXCLUSIVE);
542 :
543 : /* Release bank lock while doing I/O */
544 2776 : LWLockRelease(banklock);
545 :
546 : /* Do the read */
547 2776 : ok = SlruPhysicalReadPage(ctl, pageno, slotno);
548 :
549 : /* Set the LSNs for this newly read-in page to zero */
550 2776 : SimpleLruZeroLSNs(ctl, slotno);
551 :
552 : /* Re-acquire bank control lock and update page state */
553 2776 : LWLockAcquire(banklock, LW_EXCLUSIVE);
554 :
555 : Assert(shared->page_number[slotno] == pageno &&
556 : shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS &&
557 : !shared->page_dirty[slotno]);
558 :
559 2776 : shared->page_status[slotno] = ok ? SLRU_PAGE_VALID : SLRU_PAGE_EMPTY;
560 :
561 2776 : LWLockRelease(&shared->buffer_locks[slotno].lock);
562 :
563 : /* Now it's okay to ereport if we failed */
564 2776 : if (!ok)
565 0 : SlruReportIOError(ctl, pageno, xid);
566 :
567 2776 : SlruRecentlyUsed(shared, slotno);
568 :
569 : /* update the stats counter of pages not found in SLRU */
570 2776 : pgstat_count_slru_page_read(shared->slru_stats_idx);
571 :
572 2776 : return slotno;
573 : }
574 : }
575 :
576 : /*
577 : * Find a page in a shared buffer, reading it in if necessary.
578 : * The page number must correspond to an already-initialized page.
579 : * The caller must intend only read-only access to the page.
580 : *
581 : * The passed-in xid is used only for error reporting, and may be
582 : * InvalidTransactionId if no specific xid is associated with the action.
583 : *
584 : * Return value is the shared-buffer slot number now holding the page.
585 : * The buffer's LRU access info is updated.
586 : *
587 : * Bank control lock must NOT be held at entry, but will be held at exit.
588 : * It is unspecified whether the lock will be shared or exclusive.
589 : */
590 : int
591 1325818 : SimpleLruReadPage_ReadOnly(SlruCtl ctl, int64 pageno, TransactionId xid)
592 : {
593 1325818 : SlruShared shared = ctl->shared;
594 1325818 : LWLock *banklock = SimpleLruGetBankLock(ctl, pageno);
595 1325818 : int bankno = pageno & ctl->bank_mask;
596 1325818 : int bankstart = bankno * SLRU_BANK_SIZE;
597 1325818 : int bankend = bankstart + SLRU_BANK_SIZE;
598 :
599 : /* Try to find the page while holding only shared lock */
600 1325818 : LWLockAcquire(banklock, LW_SHARED);
601 :
602 : /* See if page is already in a buffer */
603 1328366 : for (int slotno = bankstart; slotno < bankend; slotno++)
604 : {
605 1328264 : if (shared->page_status[slotno] != SLRU_PAGE_EMPTY &&
606 1326696 : shared->page_number[slotno] == pageno &&
607 1325716 : shared->page_status[slotno] != SLRU_PAGE_READ_IN_PROGRESS)
608 : {
609 : /* See comments for SlruRecentlyUsed macro */
610 1325716 : SlruRecentlyUsed(shared, slotno);
611 :
612 : /* update the stats counter of pages found in the SLRU */
613 1325716 : pgstat_count_slru_page_hit(shared->slru_stats_idx);
614 :
615 1325716 : return slotno;
616 : }
617 : }
618 :
619 : /* No luck, so switch to normal exclusive lock and do regular read */
620 102 : LWLockRelease(banklock);
621 102 : LWLockAcquire(banklock, LW_EXCLUSIVE);
622 :
623 102 : return SimpleLruReadPage(ctl, pageno, true, xid);
624 : }
625 :
626 : /*
627 : * Write a page from a shared buffer, if necessary.
628 : * Does nothing if the specified slot is not dirty.
629 : *
630 : * NOTE: only one write attempt is made here. Hence, it is possible that
631 : * the page is still dirty at exit (if someone else re-dirtied it during
632 : * the write). However, we *do* attempt a fresh write even if the page
633 : * is already being written; this is for checkpoints.
634 : *
635 : * Bank lock must be held at entry, and will be held at exit.
636 : */
637 : static void
638 5996 : SlruInternalWritePage(SlruCtl ctl, int slotno, SlruWriteAll fdata)
639 : {
640 5996 : SlruShared shared = ctl->shared;
641 5996 : int64 pageno = shared->page_number[slotno];
642 5996 : int bankno = SlotGetBankNumber(slotno);
643 : bool ok;
644 :
645 : Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
646 : Assert(LWLockHeldByMeInMode(SimpleLruGetBankLock(ctl, pageno), LW_EXCLUSIVE));
647 :
648 : /* If a write is in progress, wait for it to finish */
649 5996 : while (shared->page_status[slotno] == SLRU_PAGE_WRITE_IN_PROGRESS &&
650 0 : shared->page_number[slotno] == pageno)
651 : {
652 0 : SimpleLruWaitIO(ctl, slotno);
653 : }
654 :
655 : /*
656 : * Do nothing if page is not dirty, or if buffer no longer contains the
657 : * same page we were called for.
658 : */
659 5996 : if (!shared->page_dirty[slotno] ||
660 3976 : shared->page_status[slotno] != SLRU_PAGE_VALID ||
661 3976 : shared->page_number[slotno] != pageno)
662 2020 : return;
663 :
664 : /*
665 : * Mark the slot write-busy, and clear the dirtybit. After this point, a
666 : * transaction status update on this page will mark it dirty again.
667 : */
668 3976 : shared->page_status[slotno] = SLRU_PAGE_WRITE_IN_PROGRESS;
669 3976 : shared->page_dirty[slotno] = false;
670 :
671 : /* Acquire per-buffer lock (cannot deadlock, see notes at top) */
672 3976 : LWLockAcquire(&shared->buffer_locks[slotno].lock, LW_EXCLUSIVE);
673 :
674 : /* Release bank lock while doing I/O */
675 3976 : LWLockRelease(&shared->bank_locks[bankno].lock);
676 :
677 : /* Do the write */
678 3976 : ok = SlruPhysicalWritePage(ctl, pageno, slotno, fdata);
679 :
680 : /* If we failed, and we're in a flush, better close the files */
681 3976 : if (!ok && fdata)
682 : {
683 0 : for (int i = 0; i < fdata->num_files; i++)
684 0 : CloseTransientFile(fdata->fd[i]);
685 : }
686 :
687 : /* Re-acquire bank lock and update page state */
688 3976 : LWLockAcquire(&shared->bank_locks[bankno].lock, LW_EXCLUSIVE);
689 :
690 : Assert(shared->page_number[slotno] == pageno &&
691 : shared->page_status[slotno] == SLRU_PAGE_WRITE_IN_PROGRESS);
692 :
693 : /* If we failed to write, mark the page dirty again */
694 3976 : if (!ok)
695 0 : shared->page_dirty[slotno] = true;
696 :
697 3976 : shared->page_status[slotno] = SLRU_PAGE_VALID;
698 :
699 3976 : LWLockRelease(&shared->buffer_locks[slotno].lock);
700 :
701 : /* Now it's okay to ereport if we failed */
702 3976 : if (!ok)
703 0 : SlruReportIOError(ctl, pageno, InvalidTransactionId);
704 :
705 : /* If part of a checkpoint, count this as a buffer written. */
706 3976 : if (fdata)
707 3388 : CheckpointStats.ckpt_bufs_written++;
708 : }
709 :
710 : /*
711 : * Wrapper of SlruInternalWritePage, for external callers.
712 : * fdata is always passed a NULL here.
713 : */
714 : void
715 538 : SimpleLruWritePage(SlruCtl ctl, int slotno)
716 : {
717 : Assert(ctl->shared->page_status[slotno] != SLRU_PAGE_EMPTY);
718 :
719 538 : SlruInternalWritePage(ctl, slotno, NULL);
720 538 : }
721 :
722 : /*
723 : * Return whether the given page exists on disk.
724 : *
725 : * A false return means that either the file does not exist, or that it's not
726 : * large enough to contain the given page.
727 : */
728 : bool
729 138 : SimpleLruDoesPhysicalPageExist(SlruCtl ctl, int64 pageno)
730 : {
731 138 : int64 segno = pageno / SLRU_PAGES_PER_SEGMENT;
732 138 : int rpageno = pageno % SLRU_PAGES_PER_SEGMENT;
733 138 : int offset = rpageno * BLCKSZ;
734 : char path[MAXPGPATH];
735 : int fd;
736 : bool result;
737 : off_t endpos;
738 :
739 : /* update the stats counter of checked pages */
740 138 : pgstat_count_slru_page_exists(ctl->shared->slru_stats_idx);
741 :
742 138 : SlruFileName(ctl, path, segno);
743 :
744 138 : fd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
745 138 : if (fd < 0)
746 : {
747 : /* expected: file doesn't exist */
748 42 : if (errno == ENOENT)
749 42 : return false;
750 :
751 : /* report error normally */
752 0 : slru_errcause = SLRU_OPEN_FAILED;
753 0 : slru_errno = errno;
754 0 : SlruReportIOError(ctl, pageno, 0);
755 : }
756 :
757 96 : if ((endpos = lseek(fd, 0, SEEK_END)) < 0)
758 : {
759 0 : slru_errcause = SLRU_SEEK_FAILED;
760 0 : slru_errno = errno;
761 0 : SlruReportIOError(ctl, pageno, 0);
762 : }
763 :
764 96 : result = endpos >= (off_t) (offset + BLCKSZ);
765 :
766 96 : if (CloseTransientFile(fd) != 0)
767 : {
768 0 : slru_errcause = SLRU_CLOSE_FAILED;
769 0 : slru_errno = errno;
770 0 : return false;
771 : }
772 :
773 96 : return result;
774 : }
775 :
776 : /*
777 : * Physical read of a (previously existing) page into a buffer slot
778 : *
779 : * On failure, we cannot just ereport(ERROR) since caller has put state in
780 : * shared memory that must be undone. So, we return false and save enough
781 : * info in static variables to let SlruReportIOError make the report.
782 : *
783 : * For now, assume it's not worth keeping a file pointer open across
784 : * read/write operations. We could cache one virtual file pointer ...
785 : */
786 : static bool
787 2776 : SlruPhysicalReadPage(SlruCtl ctl, int64 pageno, int slotno)
788 : {
789 2776 : SlruShared shared = ctl->shared;
790 2776 : int64 segno = pageno / SLRU_PAGES_PER_SEGMENT;
791 2776 : int rpageno = pageno % SLRU_PAGES_PER_SEGMENT;
792 2776 : off_t offset = rpageno * BLCKSZ;
793 : char path[MAXPGPATH];
794 : int fd;
795 :
796 2776 : SlruFileName(ctl, path, segno);
797 :
798 : /*
799 : * In a crash-and-restart situation, it's possible for us to receive
800 : * commands to set the commit status of transactions whose bits are in
801 : * already-truncated segments of the commit log (see notes in
802 : * SlruPhysicalWritePage). Hence, if we are InRecovery, allow the case
803 : * where the file doesn't exist, and return zeroes instead.
804 : */
805 2776 : fd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
806 2776 : if (fd < 0)
807 : {
808 0 : if (errno != ENOENT || !InRecovery)
809 : {
810 0 : slru_errcause = SLRU_OPEN_FAILED;
811 0 : slru_errno = errno;
812 0 : return false;
813 : }
814 :
815 0 : ereport(LOG,
816 : (errmsg("file \"%s\" doesn't exist, reading as zeroes",
817 : path)));
818 0 : MemSet(shared->page_buffer[slotno], 0, BLCKSZ);
819 0 : return true;
820 : }
821 :
822 2776 : errno = 0;
823 2776 : pgstat_report_wait_start(WAIT_EVENT_SLRU_READ);
824 2776 : if (pg_pread(fd, shared->page_buffer[slotno], BLCKSZ, offset) != BLCKSZ)
825 : {
826 0 : pgstat_report_wait_end();
827 0 : slru_errcause = SLRU_READ_FAILED;
828 0 : slru_errno = errno;
829 0 : CloseTransientFile(fd);
830 0 : return false;
831 : }
832 2776 : pgstat_report_wait_end();
833 :
834 2776 : if (CloseTransientFile(fd) != 0)
835 : {
836 0 : slru_errcause = SLRU_CLOSE_FAILED;
837 0 : slru_errno = errno;
838 0 : return false;
839 : }
840 :
841 2776 : return true;
842 : }
843 :
844 : /*
845 : * Physical write of a page from a buffer slot
846 : *
847 : * On failure, we cannot just ereport(ERROR) since caller has put state in
848 : * shared memory that must be undone. So, we return false and save enough
849 : * info in static variables to let SlruReportIOError make the report.
850 : *
851 : * For now, assume it's not worth keeping a file pointer open across
852 : * independent read/write operations. We do batch operations during
853 : * SimpleLruWriteAll, though.
854 : *
855 : * fdata is NULL for a standalone write, pointer to open-file info during
856 : * SimpleLruWriteAll.
857 : */
858 : static bool
859 3976 : SlruPhysicalWritePage(SlruCtl ctl, int64 pageno, int slotno, SlruWriteAll fdata)
860 : {
861 3976 : SlruShared shared = ctl->shared;
862 3976 : int64 segno = pageno / SLRU_PAGES_PER_SEGMENT;
863 3976 : int rpageno = pageno % SLRU_PAGES_PER_SEGMENT;
864 3976 : off_t offset = rpageno * BLCKSZ;
865 : char path[MAXPGPATH];
866 3976 : int fd = -1;
867 :
868 : /* update the stats counter of written pages */
869 3976 : pgstat_count_slru_page_written(shared->slru_stats_idx);
870 :
871 : /*
872 : * Honor the write-WAL-before-data rule, if appropriate, so that we do not
873 : * write out data before associated WAL records. This is the same action
874 : * performed during FlushBuffer() in the main buffer manager.
875 : */
876 3976 : if (shared->group_lsn != NULL)
877 : {
878 : /*
879 : * We must determine the largest async-commit LSN for the page. This
880 : * is a bit tedious, but since this entire function is a slow path
881 : * anyway, it seems better to do this here than to maintain a per-page
882 : * LSN variable (which'd need an extra comparison in the
883 : * transaction-commit path).
884 : */
885 : XLogRecPtr max_lsn;
886 : int lsnindex;
887 :
888 1420 : lsnindex = slotno * shared->lsn_groups_per_page;
889 1420 : max_lsn = shared->group_lsn[lsnindex++];
890 1454080 : for (int lsnoff = 1; lsnoff < shared->lsn_groups_per_page; lsnoff++)
891 : {
892 1452660 : XLogRecPtr this_lsn = shared->group_lsn[lsnindex++];
893 :
894 1452660 : if (max_lsn < this_lsn)
895 10536 : max_lsn = this_lsn;
896 : }
897 :
898 1420 : if (!XLogRecPtrIsInvalid(max_lsn))
899 : {
900 : /*
901 : * As noted above, elog(ERROR) is not acceptable here, so if
902 : * XLogFlush were to fail, we must PANIC. This isn't much of a
903 : * restriction because XLogFlush is just about all critical
904 : * section anyway, but let's make sure.
905 : */
906 342 : START_CRIT_SECTION();
907 342 : XLogFlush(max_lsn);
908 342 : END_CRIT_SECTION();
909 : }
910 : }
911 :
912 : /*
913 : * During a SimpleLruWriteAll, we may already have the desired file open.
914 : */
915 3976 : if (fdata)
916 : {
917 3388 : for (int i = 0; i < fdata->num_files; i++)
918 : {
919 56 : if (fdata->segno[i] == segno)
920 : {
921 56 : fd = fdata->fd[i];
922 56 : break;
923 : }
924 : }
925 : }
926 :
927 3976 : if (fd < 0)
928 : {
929 : /*
930 : * If the file doesn't already exist, we should create it. It is
931 : * possible for this to need to happen when writing a page that's not
932 : * first in its segment; we assume the OS can cope with that. (Note:
933 : * it might seem that it'd be okay to create files only when
934 : * SimpleLruZeroPage is called for the first page of a segment.
935 : * However, if after a crash and restart the REDO logic elects to
936 : * replay the log from a checkpoint before the latest one, then it's
937 : * possible that we will get commands to set transaction status of
938 : * transactions that have already been truncated from the commit log.
939 : * Easiest way to deal with that is to accept references to
940 : * nonexistent files here and in SlruPhysicalReadPage.)
941 : *
942 : * Note: it is possible for more than one backend to be executing this
943 : * code simultaneously for different pages of the same file. Hence,
944 : * don't use O_EXCL or O_TRUNC or anything like that.
945 : */
946 3920 : SlruFileName(ctl, path, segno);
947 3920 : fd = OpenTransientFile(path, O_RDWR | O_CREAT | PG_BINARY);
948 3920 : if (fd < 0)
949 : {
950 0 : slru_errcause = SLRU_OPEN_FAILED;
951 0 : slru_errno = errno;
952 0 : return false;
953 : }
954 :
955 3920 : if (fdata)
956 : {
957 3332 : if (fdata->num_files < MAX_WRITEALL_BUFFERS)
958 : {
959 3332 : fdata->fd[fdata->num_files] = fd;
960 3332 : fdata->segno[fdata->num_files] = segno;
961 3332 : fdata->num_files++;
962 : }
963 : else
964 : {
965 : /*
966 : * In the unlikely event that we exceed MAX_WRITEALL_BUFFERS,
967 : * fall back to treating it as a standalone write.
968 : */
969 0 : fdata = NULL;
970 : }
971 : }
972 : }
973 :
974 3976 : errno = 0;
975 3976 : pgstat_report_wait_start(WAIT_EVENT_SLRU_WRITE);
976 3976 : if (pg_pwrite(fd, shared->page_buffer[slotno], BLCKSZ, offset) != BLCKSZ)
977 : {
978 0 : pgstat_report_wait_end();
979 : /* if write didn't set errno, assume problem is no disk space */
980 0 : if (errno == 0)
981 0 : errno = ENOSPC;
982 0 : slru_errcause = SLRU_WRITE_FAILED;
983 0 : slru_errno = errno;
984 0 : if (!fdata)
985 0 : CloseTransientFile(fd);
986 0 : return false;
987 : }
988 3976 : pgstat_report_wait_end();
989 :
990 : /* Queue up a sync request for the checkpointer. */
991 3976 : if (ctl->sync_handler != SYNC_HANDLER_NONE)
992 : {
993 : FileTag tag;
994 :
995 2576 : INIT_SLRUFILETAG(tag, ctl->sync_handler, segno);
996 2576 : if (!RegisterSyncRequest(&tag, SYNC_REQUEST, false))
997 : {
998 : /* No space to enqueue sync request. Do it synchronously. */
999 0 : pgstat_report_wait_start(WAIT_EVENT_SLRU_SYNC);
1000 0 : if (pg_fsync(fd) != 0)
1001 : {
1002 0 : pgstat_report_wait_end();
1003 0 : slru_errcause = SLRU_FSYNC_FAILED;
1004 0 : slru_errno = errno;
1005 0 : CloseTransientFile(fd);
1006 0 : return false;
1007 : }
1008 0 : pgstat_report_wait_end();
1009 : }
1010 : }
1011 :
1012 : /* Close file, unless part of flush request. */
1013 3976 : if (!fdata)
1014 : {
1015 588 : if (CloseTransientFile(fd) != 0)
1016 : {
1017 0 : slru_errcause = SLRU_CLOSE_FAILED;
1018 0 : slru_errno = errno;
1019 0 : return false;
1020 : }
1021 : }
1022 :
1023 3976 : return true;
1024 : }
1025 :
1026 : /*
1027 : * Issue the error message after failure of SlruPhysicalReadPage or
1028 : * SlruPhysicalWritePage. Call this after cleaning up shared-memory state.
1029 : */
1030 : static void
1031 0 : SlruReportIOError(SlruCtl ctl, int64 pageno, TransactionId xid)
1032 : {
1033 0 : int64 segno = pageno / SLRU_PAGES_PER_SEGMENT;
1034 0 : int rpageno = pageno % SLRU_PAGES_PER_SEGMENT;
1035 0 : int offset = rpageno * BLCKSZ;
1036 : char path[MAXPGPATH];
1037 :
1038 0 : SlruFileName(ctl, path, segno);
1039 0 : errno = slru_errno;
1040 0 : switch (slru_errcause)
1041 : {
1042 0 : case SLRU_OPEN_FAILED:
1043 0 : ereport(ERROR,
1044 : (errcode_for_file_access(),
1045 : errmsg("could not access status of transaction %u", xid),
1046 : errdetail("Could not open file \"%s\": %m.", path)));
1047 : break;
1048 0 : case SLRU_SEEK_FAILED:
1049 0 : ereport(ERROR,
1050 : (errcode_for_file_access(),
1051 : errmsg("could not access status of transaction %u", xid),
1052 : errdetail("Could not seek in file \"%s\" to offset %d: %m.",
1053 : path, offset)));
1054 : break;
1055 0 : case SLRU_READ_FAILED:
1056 0 : if (errno)
1057 0 : ereport(ERROR,
1058 : (errcode_for_file_access(),
1059 : errmsg("could not access status of transaction %u", xid),
1060 : errdetail("Could not read from file \"%s\" at offset %d: %m.",
1061 : path, offset)));
1062 : else
1063 0 : ereport(ERROR,
1064 : (errmsg("could not access status of transaction %u", xid),
1065 : errdetail("Could not read from file \"%s\" at offset %d: read too few bytes.", path, offset)));
1066 : break;
1067 0 : case SLRU_WRITE_FAILED:
1068 0 : if (errno)
1069 0 : ereport(ERROR,
1070 : (errcode_for_file_access(),
1071 : errmsg("could not access status of transaction %u", xid),
1072 : errdetail("Could not write to file \"%s\" at offset %d: %m.",
1073 : path, offset)));
1074 : else
1075 0 : ereport(ERROR,
1076 : (errmsg("could not access status of transaction %u", xid),
1077 : errdetail("Could not write to file \"%s\" at offset %d: wrote too few bytes.",
1078 : path, offset)));
1079 : break;
1080 0 : case SLRU_FSYNC_FAILED:
1081 0 : ereport(data_sync_elevel(ERROR),
1082 : (errcode_for_file_access(),
1083 : errmsg("could not access status of transaction %u", xid),
1084 : errdetail("Could not fsync file \"%s\": %m.",
1085 : path)));
1086 0 : break;
1087 0 : case SLRU_CLOSE_FAILED:
1088 0 : ereport(ERROR,
1089 : (errcode_for_file_access(),
1090 : errmsg("could not access status of transaction %u", xid),
1091 : errdetail("Could not close file \"%s\": %m.",
1092 : path)));
1093 : break;
1094 0 : default:
1095 : /* can't get here, we trust */
1096 0 : elog(ERROR, "unrecognized SimpleLru error cause: %d",
1097 : (int) slru_errcause);
1098 : break;
1099 : }
1100 0 : }
1101 :
1102 : /*
1103 : * Mark a buffer slot "most recently used".
1104 : */
1105 : static inline void
1106 1621380 : SlruRecentlyUsed(SlruShared shared, int slotno)
1107 : {
1108 1621380 : int bankno = SlotGetBankNumber(slotno);
1109 1621380 : int new_lru_count = shared->bank_cur_lru_count[bankno];
1110 :
1111 : Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
1112 :
1113 : /*
1114 : * The reason for the if-test is that there are often many consecutive
1115 : * accesses to the same page (particularly the latest page). By
1116 : * suppressing useless increments of bank_cur_lru_count, we reduce the
1117 : * probability that old pages' counts will "wrap around" and make them
1118 : * appear recently used.
1119 : *
1120 : * We allow this code to be executed concurrently by multiple processes
1121 : * within SimpleLruReadPage_ReadOnly(). As long as int reads and writes
1122 : * are atomic, this should not cause any completely-bogus values to enter
1123 : * the computation. However, it is possible for either bank_cur_lru_count
1124 : * or individual page_lru_count entries to be "reset" to lower values than
1125 : * they should have, in case a process is delayed while it executes this
1126 : * function. With care in SlruSelectLRUPage(), this does little harm, and
1127 : * in any case the absolute worst possible consequence is a nonoptimal
1128 : * choice of page to evict. The gain from allowing concurrent reads of
1129 : * SLRU pages seems worth it.
1130 : */
1131 1621380 : if (new_lru_count != shared->page_lru_count[slotno])
1132 : {
1133 5038 : shared->bank_cur_lru_count[bankno] = ++new_lru_count;
1134 5038 : shared->page_lru_count[slotno] = new_lru_count;
1135 : }
1136 1621380 : }
1137 :
1138 : /*
1139 : * Select the slot to re-use when we need a free slot for the given page.
1140 : *
1141 : * The target page number is passed not only because we need to know the
1142 : * correct bank to use, but also because we need to consider the possibility
1143 : * that some other process reads in the target page while we are doing I/O to
1144 : * free a slot. Hence, check or recheck to see if any slot already holds the
1145 : * target page, and return that slot if so. Thus, the returned slot is
1146 : * *either* a slot already holding the pageno (could be any state except
1147 : * EMPTY), *or* a freeable slot (state EMPTY or CLEAN).
1148 : *
1149 : * The correct bank lock must be held at entry, and will be held at exit.
1150 : */
1151 : static int
1152 295664 : SlruSelectLRUPage(SlruCtl ctl, int64 pageno)
1153 : {
1154 295664 : SlruShared shared = ctl->shared;
1155 :
1156 : /* Outer loop handles restart after I/O */
1157 : for (;;)
1158 40 : {
1159 : int cur_count;
1160 295704 : int bestvalidslot = 0; /* keep compiler quiet */
1161 295704 : int best_valid_delta = -1;
1162 295704 : int64 best_valid_page_number = 0; /* keep compiler quiet */
1163 295704 : int bestinvalidslot = 0; /* keep compiler quiet */
1164 295704 : int best_invalid_delta = -1;
1165 295704 : int64 best_invalid_page_number = 0; /* keep compiler quiet */
1166 295704 : int bankno = pageno & ctl->bank_mask;
1167 295704 : int bankstart = bankno * SLRU_BANK_SIZE;
1168 295704 : int bankend = bankstart + SLRU_BANK_SIZE;
1169 :
1170 : Assert(LWLockHeldByMe(SimpleLruGetBankLock(ctl, pageno)));
1171 :
1172 : /* See if page already has a buffer assigned */
1173 425744 : for (int slotno = 0; slotno < shared->num_slots; slotno++)
1174 : {
1175 420760 : if (shared->page_status[slotno] != SLRU_PAGE_EMPTY &&
1176 297478 : shared->page_number[slotno] == pageno)
1177 290720 : return slotno;
1178 : }
1179 :
1180 : /*
1181 : * If we find any EMPTY slot, just select that one. Else choose a
1182 : * victim page to replace. We normally take the least recently used
1183 : * valid page, but we will never take the slot containing
1184 : * latest_page_number, even if it appears least recently used. We
1185 : * will select a slot that is already I/O busy only if there is no
1186 : * other choice: a read-busy slot will not be least recently used once
1187 : * the read finishes, and waiting for an I/O on a write-busy slot is
1188 : * inferior to just picking some other slot. Testing shows the slot
1189 : * we pick instead will often be clean, allowing us to begin a read at
1190 : * once.
1191 : *
1192 : * Normally the page_lru_count values will all be different and so
1193 : * there will be a well-defined LRU page. But since we allow
1194 : * concurrent execution of SlruRecentlyUsed() within
1195 : * SimpleLruReadPage_ReadOnly(), it is possible that multiple pages
1196 : * acquire the same lru_count values. In that case we break ties by
1197 : * choosing the furthest-back page.
1198 : *
1199 : * Notice that this next line forcibly advances cur_lru_count to a
1200 : * value that is certainly beyond any value that will be in the
1201 : * page_lru_count array after the loop finishes. This ensures that
1202 : * the next execution of SlruRecentlyUsed will mark the page newly
1203 : * used, even if it's for a page that has the current counter value.
1204 : * That gets us back on the path to having good data when there are
1205 : * multiple pages with the same lru_count.
1206 : */
1207 4984 : cur_count = (shared->bank_cur_lru_count[bankno])++;
1208 9436 : for (int slotno = bankstart; slotno < bankend; slotno++)
1209 : {
1210 : int this_delta;
1211 : int64 this_page_number;
1212 :
1213 9206 : if (shared->page_status[slotno] == SLRU_PAGE_EMPTY)
1214 4754 : return slotno;
1215 :
1216 4452 : this_delta = cur_count - shared->page_lru_count[slotno];
1217 4452 : if (this_delta < 0)
1218 : {
1219 : /*
1220 : * Clean up in case shared updates have caused cur_count
1221 : * increments to get "lost". We back off the page counts,
1222 : * rather than trying to increase cur_count, to avoid any
1223 : * question of infinite loops or failure in the presence of
1224 : * wrapped-around counts.
1225 : */
1226 0 : shared->page_lru_count[slotno] = cur_count;
1227 0 : this_delta = 0;
1228 : }
1229 :
1230 : /*
1231 : * If this page is the one most recently zeroed, don't consider it
1232 : * an eviction candidate. See comments in SimpleLruZeroPage for an
1233 : * explanation about the lack of a memory barrier here.
1234 : */
1235 4452 : this_page_number = shared->page_number[slotno];
1236 4452 : if (this_page_number ==
1237 4452 : pg_atomic_read_u64(&shared->latest_page_number))
1238 340 : continue;
1239 :
1240 4112 : if (shared->page_status[slotno] == SLRU_PAGE_VALID)
1241 : {
1242 4112 : if (this_delta > best_valid_delta ||
1243 0 : (this_delta == best_valid_delta &&
1244 0 : ctl->PagePrecedes(this_page_number,
1245 : best_valid_page_number)))
1246 : {
1247 686 : bestvalidslot = slotno;
1248 686 : best_valid_delta = this_delta;
1249 686 : best_valid_page_number = this_page_number;
1250 : }
1251 : }
1252 : else
1253 : {
1254 0 : if (this_delta > best_invalid_delta ||
1255 0 : (this_delta == best_invalid_delta &&
1256 0 : ctl->PagePrecedes(this_page_number,
1257 : best_invalid_page_number)))
1258 : {
1259 0 : bestinvalidslot = slotno;
1260 0 : best_invalid_delta = this_delta;
1261 0 : best_invalid_page_number = this_page_number;
1262 : }
1263 : }
1264 : }
1265 :
1266 : /*
1267 : * If all pages (except possibly the latest one) are I/O busy, we'll
1268 : * have to wait for an I/O to complete and then retry. In that
1269 : * unhappy case, we choose to wait for the I/O on the least recently
1270 : * used slot, on the assumption that it was likely initiated first of
1271 : * all the I/Os in progress and may therefore finish first.
1272 : */
1273 230 : if (best_valid_delta < 0)
1274 : {
1275 0 : SimpleLruWaitIO(ctl, bestinvalidslot);
1276 0 : continue;
1277 : }
1278 :
1279 : /*
1280 : * If the selected page is clean, we're set.
1281 : */
1282 230 : if (!shared->page_dirty[bestvalidslot])
1283 190 : return bestvalidslot;
1284 :
1285 : /*
1286 : * Write the page.
1287 : */
1288 40 : SlruInternalWritePage(ctl, bestvalidslot, NULL);
1289 :
1290 : /*
1291 : * Now loop back and try again. This is the easiest way of dealing
1292 : * with corner cases such as the victim page being re-dirtied while we
1293 : * wrote it.
1294 : */
1295 : }
1296 : }
1297 :
1298 : /*
1299 : * Write dirty pages to disk during checkpoint or database shutdown. Flushing
1300 : * is deferred until the next call to ProcessSyncRequests(), though we do fsync
1301 : * the containing directory here to make sure that newly created directory
1302 : * entries are on disk.
1303 : */
1304 : void
1305 8574 : SimpleLruWriteAll(SlruCtl ctl, bool allow_redirtied)
1306 : {
1307 8574 : SlruShared shared = ctl->shared;
1308 : SlruWriteAllData fdata;
1309 8574 : int64 pageno = 0;
1310 8574 : int prevbank = SlotGetBankNumber(0);
1311 : bool ok;
1312 :
1313 : /* update the stats counter of flushes */
1314 8574 : pgstat_count_slru_flush(shared->slru_stats_idx);
1315 :
1316 : /*
1317 : * Find and write dirty pages
1318 : */
1319 8574 : fdata.num_files = 0;
1320 :
1321 8574 : LWLockAcquire(&shared->bank_locks[prevbank].lock, LW_EXCLUSIVE);
1322 :
1323 226782 : for (int slotno = 0; slotno < shared->num_slots; slotno++)
1324 : {
1325 218208 : int curbank = SlotGetBankNumber(slotno);
1326 :
1327 : /*
1328 : * If the current bank lock is not same as the previous bank lock then
1329 : * release the previous lock and acquire the new lock.
1330 : */
1331 218208 : if (curbank != prevbank)
1332 : {
1333 5064 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1334 5064 : LWLockAcquire(&shared->bank_locks[curbank].lock, LW_EXCLUSIVE);
1335 5064 : prevbank = curbank;
1336 : }
1337 :
1338 : /* Do nothing if slot is unused */
1339 218208 : if (shared->page_status[slotno] == SLRU_PAGE_EMPTY)
1340 212800 : continue;
1341 :
1342 5408 : SlruInternalWritePage(ctl, slotno, &fdata);
1343 :
1344 : /*
1345 : * In some places (e.g. checkpoints), we cannot assert that the slot
1346 : * is clean now, since another process might have re-dirtied it
1347 : * already. That's okay.
1348 : */
1349 : Assert(allow_redirtied ||
1350 : shared->page_status[slotno] == SLRU_PAGE_EMPTY ||
1351 : (shared->page_status[slotno] == SLRU_PAGE_VALID &&
1352 : !shared->page_dirty[slotno]));
1353 : }
1354 :
1355 8574 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1356 :
1357 : /*
1358 : * Now close any files that were open
1359 : */
1360 8574 : ok = true;
1361 11906 : for (int i = 0; i < fdata.num_files; i++)
1362 : {
1363 3332 : if (CloseTransientFile(fdata.fd[i]) != 0)
1364 : {
1365 0 : slru_errcause = SLRU_CLOSE_FAILED;
1366 0 : slru_errno = errno;
1367 0 : pageno = fdata.segno[i] * SLRU_PAGES_PER_SEGMENT;
1368 0 : ok = false;
1369 : }
1370 : }
1371 8574 : if (!ok)
1372 0 : SlruReportIOError(ctl, pageno, InvalidTransactionId);
1373 :
1374 : /* Ensure that directory entries for new files are on disk. */
1375 8574 : if (ctl->sync_handler != SYNC_HANDLER_NONE)
1376 6864 : fsync_fname(ctl->Dir, true);
1377 8574 : }
1378 :
1379 : /*
1380 : * Remove all segments before the one holding the passed page number
1381 : *
1382 : * All SLRUs prevent concurrent calls to this function, either with an LWLock
1383 : * or by calling it only as part of a checkpoint. Mutual exclusion must begin
1384 : * before computing cutoffPage. Mutual exclusion must end after any limit
1385 : * update that would permit other backends to write fresh data into the
1386 : * segment immediately preceding the one containing cutoffPage. Otherwise,
1387 : * when the SLRU is quite full, SimpleLruTruncate() might delete that segment
1388 : * after it has accrued freshly-written data.
1389 : */
1390 : void
1391 1654 : SimpleLruTruncate(SlruCtl ctl, int64 cutoffPage)
1392 : {
1393 1654 : SlruShared shared = ctl->shared;
1394 : int prevbank;
1395 :
1396 : /* update the stats counter of truncates */
1397 1654 : pgstat_count_slru_truncate(shared->slru_stats_idx);
1398 :
1399 : /*
1400 : * Scan shared memory and remove any pages preceding the cutoff page, to
1401 : * ensure we won't rewrite them later. (Since this is normally called in
1402 : * or just after a checkpoint, any dirty pages should have been flushed
1403 : * already ... we're just being extra careful here.)
1404 : */
1405 1664 : restart:
1406 :
1407 : /*
1408 : * An important safety check: the current endpoint page must not be
1409 : * eligible for removal. This check is just a backstop against wraparound
1410 : * bugs elsewhere in SLRU handling, so we don't care if we read a slightly
1411 : * outdated value; therefore we don't add a memory barrier.
1412 : */
1413 1664 : if (ctl->PagePrecedes(pg_atomic_read_u64(&shared->latest_page_number),
1414 : cutoffPage))
1415 : {
1416 0 : ereport(LOG,
1417 : (errmsg("could not truncate directory \"%s\": apparent wraparound",
1418 : ctl->Dir)));
1419 0 : return;
1420 : }
1421 :
1422 1664 : prevbank = SlotGetBankNumber(0);
1423 1664 : LWLockAcquire(&shared->bank_locks[prevbank].lock, LW_EXCLUSIVE);
1424 45888 : for (int slotno = 0; slotno < shared->num_slots; slotno++)
1425 : {
1426 44234 : int curbank = SlotGetBankNumber(slotno);
1427 :
1428 : /*
1429 : * If the current bank lock is not same as the previous bank lock then
1430 : * release the previous lock and acquire the new lock.
1431 : */
1432 44234 : if (curbank != prevbank)
1433 : {
1434 1110 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1435 1110 : LWLockAcquire(&shared->bank_locks[curbank].lock, LW_EXCLUSIVE);
1436 1110 : prevbank = curbank;
1437 : }
1438 :
1439 44234 : if (shared->page_status[slotno] == SLRU_PAGE_EMPTY)
1440 42422 : continue;
1441 1812 : if (!ctl->PagePrecedes(shared->page_number[slotno], cutoffPage))
1442 1718 : continue;
1443 :
1444 : /*
1445 : * If page is clean, just change state to EMPTY (expected case).
1446 : */
1447 94 : if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
1448 94 : !shared->page_dirty[slotno])
1449 : {
1450 84 : shared->page_status[slotno] = SLRU_PAGE_EMPTY;
1451 84 : continue;
1452 : }
1453 :
1454 : /*
1455 : * Hmm, we have (or may have) I/O operations acting on the page, so
1456 : * we've got to wait for them to finish and then start again. This is
1457 : * the same logic as in SlruSelectLRUPage. (XXX if page is dirty,
1458 : * wouldn't it be OK to just discard it without writing it?
1459 : * SlruMayDeleteSegment() uses a stricter qualification, so we might
1460 : * not delete this page in the end; even if we don't delete it, we
1461 : * won't have cause to read its data again. For now, keep the logic
1462 : * the same as it was.)
1463 : */
1464 10 : if (shared->page_status[slotno] == SLRU_PAGE_VALID)
1465 10 : SlruInternalWritePage(ctl, slotno, NULL);
1466 : else
1467 0 : SimpleLruWaitIO(ctl, slotno);
1468 :
1469 10 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1470 10 : goto restart;
1471 : }
1472 :
1473 1654 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1474 :
1475 : /* Now we can remove the old segment(s) */
1476 1654 : (void) SlruScanDirectory(ctl, SlruScanDirCbDeleteCutoff, &cutoffPage);
1477 : }
1478 :
1479 : /*
1480 : * Delete an individual SLRU segment.
1481 : *
1482 : * NB: This does not touch the SLRU buffers themselves, callers have to ensure
1483 : * they either can't yet contain anything, or have already been cleaned out.
1484 : */
1485 : static void
1486 20 : SlruInternalDeleteSegment(SlruCtl ctl, int64 segno)
1487 : {
1488 : char path[MAXPGPATH];
1489 :
1490 : /* Forget any fsync requests queued for this segment. */
1491 20 : if (ctl->sync_handler != SYNC_HANDLER_NONE)
1492 : {
1493 : FileTag tag;
1494 :
1495 10 : INIT_SLRUFILETAG(tag, ctl->sync_handler, segno);
1496 10 : RegisterSyncRequest(&tag, SYNC_FORGET_REQUEST, true);
1497 : }
1498 :
1499 : /* Unlink the file. */
1500 20 : SlruFileName(ctl, path, segno);
1501 20 : ereport(DEBUG2, (errmsg_internal("removing file \"%s\"", path)));
1502 20 : unlink(path);
1503 20 : }
1504 :
1505 : /*
1506 : * Delete an individual SLRU segment, identified by the segment number.
1507 : */
1508 : void
1509 4 : SlruDeleteSegment(SlruCtl ctl, int64 segno)
1510 : {
1511 4 : SlruShared shared = ctl->shared;
1512 4 : int prevbank = SlotGetBankNumber(0);
1513 : bool did_write;
1514 :
1515 : /* Clean out any possibly existing references to the segment. */
1516 4 : LWLockAcquire(&shared->bank_locks[prevbank].lock, LW_EXCLUSIVE);
1517 4 : restart:
1518 4 : did_write = false;
1519 68 : for (int slotno = 0; slotno < shared->num_slots; slotno++)
1520 : {
1521 : int pagesegno;
1522 64 : int curbank = SlotGetBankNumber(slotno);
1523 :
1524 : /*
1525 : * If the current bank lock is not same as the previous bank lock then
1526 : * release the previous lock and acquire the new lock.
1527 : */
1528 64 : if (curbank != prevbank)
1529 : {
1530 0 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1531 0 : LWLockAcquire(&shared->bank_locks[curbank].lock, LW_EXCLUSIVE);
1532 0 : prevbank = curbank;
1533 : }
1534 :
1535 64 : if (shared->page_status[slotno] == SLRU_PAGE_EMPTY)
1536 0 : continue;
1537 :
1538 64 : pagesegno = shared->page_number[slotno] / SLRU_PAGES_PER_SEGMENT;
1539 : /* not the segment we're looking for */
1540 64 : if (pagesegno != segno)
1541 44 : continue;
1542 :
1543 : /* If page is clean, just change state to EMPTY (expected case). */
1544 20 : if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
1545 20 : !shared->page_dirty[slotno])
1546 : {
1547 20 : shared->page_status[slotno] = SLRU_PAGE_EMPTY;
1548 20 : continue;
1549 : }
1550 :
1551 : /* Same logic as SimpleLruTruncate() */
1552 0 : if (shared->page_status[slotno] == SLRU_PAGE_VALID)
1553 0 : SlruInternalWritePage(ctl, slotno, NULL);
1554 : else
1555 0 : SimpleLruWaitIO(ctl, slotno);
1556 :
1557 0 : did_write = true;
1558 : }
1559 :
1560 : /*
1561 : * Be extra careful and re-check. The IO functions release the control
1562 : * lock, so new pages could have been read in.
1563 : */
1564 4 : if (did_write)
1565 0 : goto restart;
1566 :
1567 4 : SlruInternalDeleteSegment(ctl, segno);
1568 :
1569 4 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1570 4 : }
1571 :
1572 : /*
1573 : * Determine whether a segment is okay to delete.
1574 : *
1575 : * segpage is the first page of the segment, and cutoffPage is the oldest (in
1576 : * PagePrecedes order) page in the SLRU containing still-useful data. Since
1577 : * every core PagePrecedes callback implements "wrap around", check the
1578 : * segment's first and last pages:
1579 : *
1580 : * first<cutoff && last<cutoff: yes
1581 : * first<cutoff && last>=cutoff: no; cutoff falls inside this segment
1582 : * first>=cutoff && last<cutoff: no; wrap point falls inside this segment
1583 : * first>=cutoff && last>=cutoff: no; every page of this segment is too young
1584 : */
1585 : static bool
1586 1788 : SlruMayDeleteSegment(SlruCtl ctl, int64 segpage, int64 cutoffPage)
1587 : {
1588 1788 : int64 seg_last_page = segpage + SLRU_PAGES_PER_SEGMENT - 1;
1589 :
1590 : Assert(segpage % SLRU_PAGES_PER_SEGMENT == 0);
1591 :
1592 1910 : return (ctl->PagePrecedes(segpage, cutoffPage) &&
1593 122 : ctl->PagePrecedes(seg_last_page, cutoffPage));
1594 : }
1595 :
1596 : #ifdef USE_ASSERT_CHECKING
1597 : static void
1598 : SlruPagePrecedesTestOffset(SlruCtl ctl, int per_page, uint32 offset)
1599 : {
1600 : TransactionId lhs,
1601 : rhs;
1602 : int64 newestPage,
1603 : oldestPage;
1604 : TransactionId newestXact,
1605 : oldestXact;
1606 :
1607 : /*
1608 : * Compare an XID pair having undefined order (see RFC 1982), a pair at
1609 : * "opposite ends" of the XID space. TransactionIdPrecedes() treats each
1610 : * as preceding the other. If RHS is oldestXact, LHS is the first XID we
1611 : * must not assign.
1612 : */
1613 : lhs = per_page + offset; /* skip first page to avoid non-normal XIDs */
1614 : rhs = lhs + (1U << 31);
1615 : Assert(TransactionIdPrecedes(lhs, rhs));
1616 : Assert(TransactionIdPrecedes(rhs, lhs));
1617 : Assert(!TransactionIdPrecedes(lhs - 1, rhs));
1618 : Assert(TransactionIdPrecedes(rhs, lhs - 1));
1619 : Assert(TransactionIdPrecedes(lhs + 1, rhs));
1620 : Assert(!TransactionIdPrecedes(rhs, lhs + 1));
1621 : Assert(!TransactionIdFollowsOrEquals(lhs, rhs));
1622 : Assert(!TransactionIdFollowsOrEquals(rhs, lhs));
1623 : Assert(!ctl->PagePrecedes(lhs / per_page, lhs / per_page));
1624 : Assert(!ctl->PagePrecedes(lhs / per_page, rhs / per_page));
1625 : Assert(!ctl->PagePrecedes(rhs / per_page, lhs / per_page));
1626 : Assert(!ctl->PagePrecedes((lhs - per_page) / per_page, rhs / per_page));
1627 : Assert(ctl->PagePrecedes(rhs / per_page, (lhs - 3 * per_page) / per_page));
1628 : Assert(ctl->PagePrecedes(rhs / per_page, (lhs - 2 * per_page) / per_page));
1629 : Assert(ctl->PagePrecedes(rhs / per_page, (lhs - 1 * per_page) / per_page)
1630 : || (1U << 31) % per_page != 0); /* See CommitTsPagePrecedes() */
1631 : Assert(ctl->PagePrecedes((lhs + 1 * per_page) / per_page, rhs / per_page)
1632 : || (1U << 31) % per_page != 0);
1633 : Assert(ctl->PagePrecedes((lhs + 2 * per_page) / per_page, rhs / per_page));
1634 : Assert(ctl->PagePrecedes((lhs + 3 * per_page) / per_page, rhs / per_page));
1635 : Assert(!ctl->PagePrecedes(rhs / per_page, (lhs + per_page) / per_page));
1636 :
1637 : /*
1638 : * GetNewTransactionId() has assigned the last XID it can safely use, and
1639 : * that XID is in the *LAST* page of the second segment. We must not
1640 : * delete that segment.
1641 : */
1642 : newestPage = 2 * SLRU_PAGES_PER_SEGMENT - 1;
1643 : newestXact = newestPage * per_page + offset;
1644 : Assert(newestXact / per_page == newestPage);
1645 : oldestXact = newestXact + 1;
1646 : oldestXact -= 1U << 31;
1647 : oldestPage = oldestXact / per_page;
1648 : Assert(!SlruMayDeleteSegment(ctl,
1649 : (newestPage -
1650 : newestPage % SLRU_PAGES_PER_SEGMENT),
1651 : oldestPage));
1652 :
1653 : /*
1654 : * GetNewTransactionId() has assigned the last XID it can safely use, and
1655 : * that XID is in the *FIRST* page of the second segment. We must not
1656 : * delete that segment.
1657 : */
1658 : newestPage = SLRU_PAGES_PER_SEGMENT;
1659 : newestXact = newestPage * per_page + offset;
1660 : Assert(newestXact / per_page == newestPage);
1661 : oldestXact = newestXact + 1;
1662 : oldestXact -= 1U << 31;
1663 : oldestPage = oldestXact / per_page;
1664 : Assert(!SlruMayDeleteSegment(ctl,
1665 : (newestPage -
1666 : newestPage % SLRU_PAGES_PER_SEGMENT),
1667 : oldestPage));
1668 : }
1669 :
1670 : /*
1671 : * Unit-test a PagePrecedes function.
1672 : *
1673 : * This assumes every uint32 >= FirstNormalTransactionId is a valid key. It
1674 : * assumes each value occupies a contiguous, fixed-size region of SLRU bytes.
1675 : * (MultiXactMemberCtl separates flags from XIDs. NotifyCtl has
1676 : * variable-length entries, no keys, and no random access. These unit tests
1677 : * do not apply to them.)
1678 : */
1679 : void
1680 : SlruPagePrecedesUnitTests(SlruCtl ctl, int per_page)
1681 : {
1682 : /* Test first, middle and last entries of a page. */
1683 : SlruPagePrecedesTestOffset(ctl, per_page, 0);
1684 : SlruPagePrecedesTestOffset(ctl, per_page, per_page / 2);
1685 : SlruPagePrecedesTestOffset(ctl, per_page, per_page - 1);
1686 : }
1687 : #endif
1688 :
1689 : /*
1690 : * SlruScanDirectory callback
1691 : * This callback reports true if there's any segment wholly prior to the
1692 : * one containing the page passed as "data".
1693 : */
1694 : bool
1695 132 : SlruScanDirCbReportPresence(SlruCtl ctl, char *filename, int64 segpage,
1696 : void *data)
1697 : {
1698 132 : int64 cutoffPage = *(int64 *) data;
1699 :
1700 132 : if (SlruMayDeleteSegment(ctl, segpage, cutoffPage))
1701 0 : return true; /* found one; don't iterate any more */
1702 :
1703 132 : return false; /* keep going */
1704 : }
1705 :
1706 : /*
1707 : * SlruScanDirectory callback.
1708 : * This callback deletes segments prior to the one passed in as "data".
1709 : */
1710 : static bool
1711 1656 : SlruScanDirCbDeleteCutoff(SlruCtl ctl, char *filename, int64 segpage,
1712 : void *data)
1713 : {
1714 1656 : int64 cutoffPage = *(int64 *) data;
1715 :
1716 1656 : if (SlruMayDeleteSegment(ctl, segpage, cutoffPage))
1717 4 : SlruInternalDeleteSegment(ctl, segpage / SLRU_PAGES_PER_SEGMENT);
1718 :
1719 1656 : return false; /* keep going */
1720 : }
1721 :
1722 : /*
1723 : * SlruScanDirectory callback.
1724 : * This callback deletes all segments.
1725 : */
1726 : bool
1727 12 : SlruScanDirCbDeleteAll(SlruCtl ctl, char *filename, int64 segpage, void *data)
1728 : {
1729 12 : SlruInternalDeleteSegment(ctl, segpage / SLRU_PAGES_PER_SEGMENT);
1730 :
1731 12 : return false; /* keep going */
1732 : }
1733 :
1734 : /*
1735 : * An internal function used by SlruScanDirectory().
1736 : *
1737 : * Returns true if a file with a name of a given length may be a correct
1738 : * SLRU segment.
1739 : */
1740 : static inline bool
1741 11968 : SlruCorrectSegmentFilenameLength(SlruCtl ctl, size_t len)
1742 : {
1743 11968 : if (ctl->long_segment_names)
1744 3560 : return (len == 15); /* see SlruFileName() */
1745 : else
1746 :
1747 : /*
1748 : * Commit 638cf09e76d allowed 5-character lengths. Later commit
1749 : * 73c986adde5 allowed 6-character length.
1750 : *
1751 : * Note: There is an ongoing plan to migrate all SLRUs to 64-bit page
1752 : * numbers, and the corresponding 15-character file names, which may
1753 : * eventually deprecate the support for 4, 5, and 6-character names.
1754 : */
1755 8408 : return (len == 4 || len == 5 || len == 6);
1756 : }
1757 :
1758 : /*
1759 : * Scan the SimpleLru directory and apply a callback to each file found in it.
1760 : *
1761 : * If the callback returns true, the scan is stopped. The last return value
1762 : * from the callback is returned.
1763 : *
1764 : * The callback receives the following arguments: 1. the SlruCtl struct for the
1765 : * slru being truncated; 2. the filename being considered; 3. the page number
1766 : * for the first page of that file; 4. a pointer to the opaque data given to us
1767 : * by the caller.
1768 : *
1769 : * Note that the ordering in which the directory is scanned is not guaranteed.
1770 : *
1771 : * Note that no locking is applied.
1772 : */
1773 : bool
1774 5084 : SlruScanDirectory(SlruCtl ctl, SlruScanCallback callback, void *data)
1775 : {
1776 5084 : bool retval = false;
1777 : DIR *cldir;
1778 : struct dirent *clde;
1779 : int64 segno;
1780 : int64 segpage;
1781 :
1782 5084 : cldir = AllocateDir(ctl->Dir);
1783 17052 : while ((clde = ReadDir(cldir, ctl->Dir)) != NULL)
1784 : {
1785 : size_t len;
1786 :
1787 11968 : len = strlen(clde->d_name);
1788 :
1789 11968 : if (SlruCorrectSegmentFilenameLength(ctl, len) &&
1790 1800 : strspn(clde->d_name, "0123456789ABCDEF") == len)
1791 : {
1792 1800 : segno = strtoi64(clde->d_name, NULL, 16);
1793 1800 : segpage = segno * SLRU_PAGES_PER_SEGMENT;
1794 :
1795 1800 : elog(DEBUG2, "SlruScanDirectory invoking callback on %s/%s",
1796 : ctl->Dir, clde->d_name);
1797 1800 : retval = callback(ctl, clde->d_name, segpage, data);
1798 1800 : if (retval)
1799 0 : break;
1800 : }
1801 : }
1802 5084 : FreeDir(cldir);
1803 :
1804 5084 : return retval;
1805 : }
1806 :
1807 : /*
1808 : * Individual SLRUs (clog, ...) have to provide a sync.c handler function so
1809 : * that they can provide the correct "SlruCtl" (otherwise we don't know how to
1810 : * build the path), but they just forward to this common implementation that
1811 : * performs the fsync.
1812 : */
1813 : int
1814 4 : SlruSyncFileTag(SlruCtl ctl, const FileTag *ftag, char *path)
1815 : {
1816 : int fd;
1817 : int save_errno;
1818 : int result;
1819 :
1820 4 : SlruFileName(ctl, path, ftag->segno);
1821 :
1822 4 : fd = OpenTransientFile(path, O_RDWR | PG_BINARY);
1823 4 : if (fd < 0)
1824 0 : return -1;
1825 :
1826 4 : pgstat_report_wait_start(WAIT_EVENT_SLRU_FLUSH_SYNC);
1827 4 : result = pg_fsync(fd);
1828 4 : pgstat_report_wait_end();
1829 4 : save_errno = errno;
1830 :
1831 4 : CloseTransientFile(fd);
1832 :
1833 4 : errno = save_errno;
1834 4 : return result;
1835 : }
|