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) 1996-2025, 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.h"
74 :
75 : /*
76 : * Converts segment number to the filename of the segment.
77 : *
78 : * "path" should point to a buffer at least MAXPGPATH characters long.
79 : *
80 : * If ctl->long_segment_names is true, segno can be in the range [0, 2^60-1].
81 : * The resulting file name is made of 15 characters, e.g. dir/123456789ABCDEF.
82 : *
83 : * If ctl->long_segment_names is false, segno can be in the range [0, 2^24-1].
84 : * The resulting file name is made of 4 to 6 characters, as of:
85 : *
86 : * dir/1234 for [0, 2^16-1]
87 : * dir/12345 for [2^16, 2^20-1]
88 : * dir/123456 for [2^20, 2^24-1]
89 : */
90 : static inline int
91 14970026 : SlruFileName(SlruCtl ctl, char *path, int64 segno)
92 : {
93 14970026 : if (ctl->long_segment_names)
94 : {
95 : /*
96 : * We could use 16 characters here but the disadvantage would be that
97 : * the SLRU segments will be hard to distinguish from WAL segments.
98 : *
99 : * For this reason we use 15 characters. It is enough but also means
100 : * that in the future we can't decrease SLRU_PAGES_PER_SEGMENT easily.
101 : */
102 : Assert(segno >= 0 && segno <= INT64CONST(0xFFFFFFFFFFFFFFF));
103 294 : return snprintf(path, MAXPGPATH, "%s/%015" PRIX64, ctl->Dir, segno);
104 : }
105 : else
106 : {
107 : /*
108 : * Despite the fact that %04X format string is used up to 24 bit
109 : * integers are allowed. See SlruCorrectSegmentFilenameLength()
110 : */
111 : Assert(segno >= 0 && segno <= INT64CONST(0xFFFFFF));
112 14969732 : return snprintf(path, MAXPGPATH, "%s/%04X", (ctl)->Dir,
113 : (unsigned int) segno);
114 : }
115 : }
116 :
117 : /*
118 : * During SimpleLruWriteAll(), we will usually not need to write more than one
119 : * or two physical files, but we may need to write several pages per file. We
120 : * can consolidate the I/O requests by leaving files open until control returns
121 : * to SimpleLruWriteAll(). This data structure remembers which files are open.
122 : */
123 : #define MAX_WRITEALL_BUFFERS 16
124 :
125 : typedef struct SlruWriteAllData
126 : {
127 : int num_files; /* # files actually open */
128 : int fd[MAX_WRITEALL_BUFFERS]; /* their FD's */
129 : int64 segno[MAX_WRITEALL_BUFFERS]; /* their log seg#s */
130 : } SlruWriteAllData;
131 :
132 : typedef struct SlruWriteAllData *SlruWriteAll;
133 :
134 :
135 : /*
136 : * Bank size for the slot array. Pages are assigned a bank according to their
137 : * page number, with each bank being this size. We want a power of 2 so that
138 : * we can determine the bank number for a page with just bit shifting; we also
139 : * want to keep the bank size small so that LRU victim search is fast. 16
140 : * buffers per bank seems a good number.
141 : */
142 : #define SLRU_BANK_BITSHIFT 4
143 : #define SLRU_BANK_SIZE (1 << SLRU_BANK_BITSHIFT)
144 :
145 : /*
146 : * Macro to get the bank number to which the slot belongs.
147 : */
148 : #define SlotGetBankNumber(slotno) ((slotno) >> SLRU_BANK_BITSHIFT)
149 :
150 :
151 : /*
152 : * Populate a file tag describing a segment file. We only use the segment
153 : * number, since we can derive everything else we need by having separate
154 : * sync handler functions for clog, multixact etc.
155 : */
156 : #define INIT_SLRUFILETAG(a,xx_handler,xx_segno) \
157 : ( \
158 : memset(&(a), 0, sizeof(FileTag)), \
159 : (a).handler = (xx_handler), \
160 : (a).segno = (xx_segno) \
161 : )
162 :
163 : /* Saved info for SlruReportIOError */
164 : typedef enum
165 : {
166 : SLRU_OPEN_FAILED,
167 : SLRU_SEEK_FAILED,
168 : SLRU_READ_FAILED,
169 : SLRU_WRITE_FAILED,
170 : SLRU_FSYNC_FAILED,
171 : SLRU_CLOSE_FAILED,
172 : } SlruErrorCause;
173 :
174 : static SlruErrorCause slru_errcause;
175 : static int slru_errno;
176 :
177 :
178 : static void SimpleLruZeroLSNs(SlruCtl ctl, int slotno);
179 : static void SimpleLruWaitIO(SlruCtl ctl, int slotno);
180 : static void SlruInternalWritePage(SlruCtl ctl, int slotno, SlruWriteAll fdata);
181 : static bool SlruPhysicalReadPage(SlruCtl ctl, int64 pageno, int slotno);
182 : static bool SlruPhysicalWritePage(SlruCtl ctl, int64 pageno, int slotno,
183 : SlruWriteAll fdata);
184 : static void SlruReportIOError(SlruCtl ctl, int64 pageno, TransactionId xid);
185 : static int SlruSelectLRUPage(SlruCtl ctl, int64 pageno);
186 :
187 : static bool SlruScanDirCbDeleteCutoff(SlruCtl ctl, char *filename,
188 : int64 segpage, void *data);
189 : static void SlruInternalDeleteSegment(SlruCtl ctl, int64 segno);
190 : static inline void SlruRecentlyUsed(SlruShared shared, int slotno);
191 :
192 :
193 : /*
194 : * Initialization of shared memory
195 : */
196 :
197 : Size
198 43562 : SimpleLruShmemSize(int nslots, int nlsns)
199 : {
200 43562 : int nbanks = nslots / SLRU_BANK_SIZE;
201 : Size sz;
202 :
203 : Assert(nslots <= SLRU_MAX_ALLOWED_BUFFERS);
204 : Assert(nslots % SLRU_BANK_SIZE == 0);
205 :
206 : /* we assume nslots isn't so large as to risk overflow */
207 43562 : sz = MAXALIGN(sizeof(SlruSharedData));
208 43562 : sz += MAXALIGN(nslots * sizeof(char *)); /* page_buffer[] */
209 43562 : sz += MAXALIGN(nslots * sizeof(SlruPageStatus)); /* page_status[] */
210 43562 : sz += MAXALIGN(nslots * sizeof(bool)); /* page_dirty[] */
211 43562 : sz += MAXALIGN(nslots * sizeof(int64)); /* page_number[] */
212 43562 : sz += MAXALIGN(nslots * sizeof(int)); /* page_lru_count[] */
213 43562 : sz += MAXALIGN(nslots * sizeof(LWLockPadded)); /* buffer_locks[] */
214 43562 : sz += MAXALIGN(nbanks * sizeof(LWLockPadded)); /* bank_locks[] */
215 43562 : sz += MAXALIGN(nbanks * sizeof(int)); /* bank_cur_lru_count[] */
216 :
217 43562 : if (nlsns > 0)
218 6222 : sz += MAXALIGN(nslots * nlsns * sizeof(XLogRecPtr)); /* group_lsn[] */
219 :
220 43562 : return BUFFERALIGN(sz) + BLCKSZ * nslots;
221 : }
222 :
223 : /*
224 : * Determine a number of SLRU buffers to use.
225 : *
226 : * We simply divide shared_buffers by the divisor given and cap
227 : * that at the maximum given; but always at least SLRU_BANK_SIZE.
228 : * Round down to the nearest multiple of SLRU_BANK_SIZE.
229 : */
230 : int
231 18564 : SimpleLruAutotuneBuffers(int divisor, int max)
232 : {
233 18564 : return Min(max - (max % SLRU_BANK_SIZE),
234 : Max(SLRU_BANK_SIZE,
235 : NBuffers / divisor - (NBuffers / divisor) % SLRU_BANK_SIZE));
236 : }
237 :
238 : /*
239 : * Initialize, or attach to, a simple LRU cache in shared memory.
240 : *
241 : * ctl: address of local (unshared) control structure.
242 : * name: name of SLRU. (This is user-visible, pick with care!)
243 : * nslots: number of page slots to use.
244 : * nlsns: number of LSN groups per page (set to zero if not relevant).
245 : * subdir: PGDATA-relative subdirectory that will contain the files.
246 : * buffer_tranche_id: tranche ID to use for the SLRU's per-buffer LWLocks.
247 : * bank_tranche_id: tranche ID to use for the bank LWLocks.
248 : * sync_handler: which set of functions to use to handle sync requests
249 : * long_segment_names: use short or long segment names
250 : */
251 : void
252 15222 : SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
253 : const char *subdir, int buffer_tranche_id, int bank_tranche_id,
254 : SyncRequestHandler sync_handler, bool long_segment_names)
255 : {
256 : SlruShared shared;
257 : bool found;
258 15222 : int nbanks = nslots / SLRU_BANK_SIZE;
259 :
260 : Assert(nslots <= SLRU_MAX_ALLOWED_BUFFERS);
261 :
262 15222 : shared = (SlruShared) ShmemInitStruct(name,
263 : SimpleLruShmemSize(nslots, nlsns),
264 : &found);
265 :
266 15222 : if (!IsUnderPostmaster)
267 : {
268 : /* Initialize locks and shared memory area */
269 : char *ptr;
270 : Size offset;
271 :
272 : Assert(!found);
273 :
274 15222 : memset(shared, 0, sizeof(SlruSharedData));
275 :
276 15222 : shared->num_slots = nslots;
277 15222 : shared->lsn_groups_per_page = nlsns;
278 :
279 15222 : pg_atomic_init_u64(&shared->latest_page_number, 0);
280 :
281 15222 : shared->slru_stats_idx = pgstat_get_slru_index(name);
282 :
283 15222 : ptr = (char *) shared;
284 15222 : offset = MAXALIGN(sizeof(SlruSharedData));
285 15222 : shared->page_buffer = (char **) (ptr + offset);
286 15222 : offset += MAXALIGN(nslots * sizeof(char *));
287 15222 : shared->page_status = (SlruPageStatus *) (ptr + offset);
288 15222 : offset += MAXALIGN(nslots * sizeof(SlruPageStatus));
289 15222 : shared->page_dirty = (bool *) (ptr + offset);
290 15222 : offset += MAXALIGN(nslots * sizeof(bool));
291 15222 : shared->page_number = (int64 *) (ptr + offset);
292 15222 : offset += MAXALIGN(nslots * sizeof(int64));
293 15222 : shared->page_lru_count = (int *) (ptr + offset);
294 15222 : offset += MAXALIGN(nslots * sizeof(int));
295 :
296 : /* Initialize LWLocks */
297 15222 : shared->buffer_locks = (LWLockPadded *) (ptr + offset);
298 15222 : offset += MAXALIGN(nslots * sizeof(LWLockPadded));
299 15222 : shared->bank_locks = (LWLockPadded *) (ptr + offset);
300 15222 : offset += MAXALIGN(nbanks * sizeof(LWLockPadded));
301 15222 : shared->bank_cur_lru_count = (int *) (ptr + offset);
302 15222 : offset += MAXALIGN(nbanks * sizeof(int));
303 :
304 15222 : if (nlsns > 0)
305 : {
306 2174 : shared->group_lsn = (XLogRecPtr *) (ptr + offset);
307 2174 : offset += MAXALIGN(nslots * nlsns * sizeof(XLogRecPtr));
308 : }
309 :
310 15222 : ptr += BUFFERALIGN(offset);
311 387318 : for (int slotno = 0; slotno < nslots; slotno++)
312 : {
313 372096 : LWLockInitialize(&shared->buffer_locks[slotno].lock,
314 : buffer_tranche_id);
315 :
316 372096 : shared->page_buffer[slotno] = ptr;
317 372096 : shared->page_status[slotno] = SLRU_PAGE_EMPTY;
318 372096 : shared->page_dirty[slotno] = false;
319 372096 : shared->page_lru_count[slotno] = 0;
320 372096 : ptr += BLCKSZ;
321 : }
322 :
323 : /* Initialize the slot banks. */
324 38478 : for (int bankno = 0; bankno < nbanks; bankno++)
325 : {
326 23256 : LWLockInitialize(&shared->bank_locks[bankno].lock, bank_tranche_id);
327 23256 : shared->bank_cur_lru_count[bankno] = 0;
328 : }
329 :
330 : /* Should fit to estimated shmem size */
331 : Assert(ptr - (char *) shared <= SimpleLruShmemSize(nslots, nlsns));
332 : }
333 : else
334 : {
335 : Assert(found);
336 : Assert(shared->num_slots == nslots);
337 : }
338 :
339 : /*
340 : * Initialize the unshared control struct, including directory path. We
341 : * assume caller set PagePrecedes.
342 : */
343 15222 : ctl->shared = shared;
344 15222 : ctl->sync_handler = sync_handler;
345 15222 : ctl->long_segment_names = long_segment_names;
346 15222 : ctl->nbanks = nbanks;
347 15222 : strlcpy(ctl->Dir, subdir, sizeof(ctl->Dir));
348 15222 : }
349 :
350 : /*
351 : * Helper function for GUC check_hook to check whether slru buffers are in
352 : * multiples of SLRU_BANK_SIZE.
353 : */
354 : bool
355 22266 : check_slru_buffers(const char *name, int *newval)
356 : {
357 : /* Valid values are multiples of SLRU_BANK_SIZE */
358 22266 : if (*newval % SLRU_BANK_SIZE == 0)
359 22266 : return true;
360 :
361 0 : GUC_check_errdetail("\"%s\" must be a multiple of %d.", name,
362 : SLRU_BANK_SIZE);
363 0 : return false;
364 : }
365 :
366 : /*
367 : * Initialize (or reinitialize) a page to zeroes.
368 : *
369 : * The page is not actually written, just set up in shared memory.
370 : * The slot number of the new page is returned.
371 : *
372 : * Bank lock must be held at entry, and will be held at exit.
373 : */
374 : int
375 14679210 : SimpleLruZeroPage(SlruCtl ctl, int64 pageno)
376 : {
377 14679210 : SlruShared shared = ctl->shared;
378 : int slotno;
379 :
380 : Assert(LWLockHeldByMeInMode(SimpleLruGetBankLock(ctl, pageno), LW_EXCLUSIVE));
381 :
382 : /* Find a suitable buffer slot for the page */
383 14679210 : slotno = SlruSelectLRUPage(ctl, pageno);
384 : Assert(shared->page_status[slotno] == SLRU_PAGE_EMPTY ||
385 : (shared->page_status[slotno] == SLRU_PAGE_VALID &&
386 : !shared->page_dirty[slotno]) ||
387 : shared->page_number[slotno] == pageno);
388 :
389 : /* Mark the slot as containing this page */
390 14679210 : shared->page_number[slotno] = pageno;
391 14679210 : shared->page_status[slotno] = SLRU_PAGE_VALID;
392 14679210 : shared->page_dirty[slotno] = true;
393 14679210 : SlruRecentlyUsed(shared, slotno);
394 :
395 : /* Set the buffer to zeroes */
396 14679210 : MemSet(shared->page_buffer[slotno], 0, BLCKSZ);
397 :
398 : /* Set the LSNs for this new page to zero */
399 14679210 : SimpleLruZeroLSNs(ctl, slotno);
400 :
401 : /*
402 : * Assume this page is now the latest active page.
403 : *
404 : * Note that because both this routine and SlruSelectLRUPage run with a
405 : * SLRU bank lock held, it is not possible for this to be zeroing a page
406 : * that SlruSelectLRUPage is going to evict simultaneously. Therefore,
407 : * there's no memory barrier here.
408 : */
409 14679210 : pg_atomic_write_u64(&shared->latest_page_number, pageno);
410 :
411 : /* update the stats counter of zeroed pages */
412 14679210 : pgstat_count_slru_blocks_zeroed(shared->slru_stats_idx);
413 :
414 14679210 : return slotno;
415 : }
416 :
417 : /*
418 : * Zero all the LSNs we store for this slru page.
419 : *
420 : * This should be called each time we create a new page, and each time we read
421 : * in a page from disk into an existing buffer. (Such an old page cannot
422 : * have any interesting LSNs, since we'd have flushed them before writing
423 : * the page in the first place.)
424 : *
425 : * This assumes that InvalidXLogRecPtr is bitwise-all-0.
426 : */
427 : static void
428 14683296 : SimpleLruZeroLSNs(SlruCtl ctl, int slotno)
429 : {
430 14683296 : SlruShared shared = ctl->shared;
431 :
432 14683296 : if (shared->lsn_groups_per_page > 0)
433 865794 : MemSet(&shared->group_lsn[slotno * shared->lsn_groups_per_page], 0,
434 : shared->lsn_groups_per_page * sizeof(XLogRecPtr));
435 14683296 : }
436 :
437 : /*
438 : * This is a convenience wrapper for the common case of zeroing a page and
439 : * immediately flushing it to disk.
440 : *
441 : * SLRU bank lock is acquired and released here.
442 : */
443 : void
444 432 : SimpleLruZeroAndWritePage(SlruCtl ctl, int64 pageno)
445 : {
446 : int slotno;
447 : LWLock *lock;
448 :
449 432 : lock = SimpleLruGetBankLock(ctl, pageno);
450 432 : LWLockAcquire(lock, LW_EXCLUSIVE);
451 :
452 : /* Create and zero the page */
453 432 : slotno = SimpleLruZeroPage(ctl, pageno);
454 :
455 : /* Make sure it's written out */
456 432 : SimpleLruWritePage(ctl, slotno);
457 : Assert(!ctl->shared->page_dirty[slotno]);
458 :
459 432 : LWLockRelease(lock);
460 432 : }
461 :
462 : /*
463 : * Wait for any active I/O on a page slot to finish. (This does not
464 : * guarantee that new I/O hasn't been started before we return, though.
465 : * In fact the slot might not even contain the same page anymore.)
466 : *
467 : * Bank lock must be held at entry, and will be held at exit.
468 : */
469 : static void
470 4 : SimpleLruWaitIO(SlruCtl ctl, int slotno)
471 : {
472 4 : SlruShared shared = ctl->shared;
473 4 : int bankno = SlotGetBankNumber(slotno);
474 :
475 : Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
476 :
477 : /* See notes at top of file */
478 4 : LWLockRelease(&shared->bank_locks[bankno].lock);
479 4 : LWLockAcquire(&shared->buffer_locks[slotno].lock, LW_SHARED);
480 4 : LWLockRelease(&shared->buffer_locks[slotno].lock);
481 4 : LWLockAcquire(&shared->bank_locks[bankno].lock, LW_EXCLUSIVE);
482 :
483 : /*
484 : * If the slot is still in an io-in-progress state, then either someone
485 : * already started a new I/O on the slot, or a previous I/O failed and
486 : * neglected to reset the page state. That shouldn't happen, really, but
487 : * it seems worth a few extra cycles to check and recover from it. We can
488 : * cheaply test for failure by seeing if the buffer lock is still held (we
489 : * assume that transaction abort would release the lock).
490 : */
491 4 : if (shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS ||
492 4 : shared->page_status[slotno] == SLRU_PAGE_WRITE_IN_PROGRESS)
493 : {
494 0 : if (LWLockConditionalAcquire(&shared->buffer_locks[slotno].lock, LW_SHARED))
495 : {
496 : /* indeed, the I/O must have failed */
497 0 : if (shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS)
498 0 : shared->page_status[slotno] = SLRU_PAGE_EMPTY;
499 : else /* write_in_progress */
500 : {
501 0 : shared->page_status[slotno] = SLRU_PAGE_VALID;
502 0 : shared->page_dirty[slotno] = true;
503 : }
504 0 : LWLockRelease(&shared->buffer_locks[slotno].lock);
505 : }
506 : }
507 4 : }
508 :
509 : /*
510 : * Find a page in a shared buffer, reading it in if necessary.
511 : * The page number must correspond to an already-initialized page.
512 : *
513 : * If write_ok is true then it is OK to return a page that is in
514 : * WRITE_IN_PROGRESS state; it is the caller's responsibility to be sure
515 : * that modification of the page is safe. If write_ok is false then we
516 : * will not return the page until it is not undergoing active I/O.
517 : *
518 : * The passed-in xid is used only for error reporting, and may be
519 : * InvalidTransactionId if no specific xid is associated with the action.
520 : *
521 : * Return value is the shared-buffer slot number now holding the page.
522 : * The buffer's LRU access info is updated.
523 : *
524 : * The correct bank lock must be held at entry, and will be held at exit.
525 : */
526 : int
527 328388 : SimpleLruReadPage(SlruCtl ctl, int64 pageno, bool write_ok,
528 : TransactionId xid)
529 : {
530 328388 : SlruShared shared = ctl->shared;
531 328388 : LWLock *banklock = SimpleLruGetBankLock(ctl, pageno);
532 :
533 : Assert(LWLockHeldByMeInMode(banklock, LW_EXCLUSIVE));
534 :
535 : /* Outer loop handles restart if we must wait for someone else's I/O */
536 : for (;;)
537 0 : {
538 : int slotno;
539 : bool ok;
540 :
541 : /* See if page already is in memory; if not, pick victim slot */
542 328388 : slotno = SlruSelectLRUPage(ctl, pageno);
543 :
544 : /* Did we find the page in memory? */
545 328388 : if (shared->page_status[slotno] != SLRU_PAGE_EMPTY &&
546 324928 : shared->page_number[slotno] == pageno)
547 : {
548 : /*
549 : * If page is still being read in, we must wait for I/O. Likewise
550 : * if the page is being written and the caller said that's not OK.
551 : */
552 324302 : if (shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS ||
553 324302 : (shared->page_status[slotno] == SLRU_PAGE_WRITE_IN_PROGRESS &&
554 0 : !write_ok))
555 : {
556 0 : SimpleLruWaitIO(ctl, slotno);
557 : /* Now we must recheck state from the top */
558 0 : continue;
559 : }
560 : /* Otherwise, it's ready to use */
561 324302 : SlruRecentlyUsed(shared, slotno);
562 :
563 : /* update the stats counter of pages found in the SLRU */
564 324302 : pgstat_count_slru_blocks_hit(shared->slru_stats_idx);
565 :
566 324302 : return slotno;
567 : }
568 :
569 : /* We found no match; assert we selected a freeable slot */
570 : Assert(shared->page_status[slotno] == SLRU_PAGE_EMPTY ||
571 : (shared->page_status[slotno] == SLRU_PAGE_VALID &&
572 : !shared->page_dirty[slotno]));
573 :
574 : /* Mark the slot read-busy */
575 4086 : shared->page_number[slotno] = pageno;
576 4086 : shared->page_status[slotno] = SLRU_PAGE_READ_IN_PROGRESS;
577 4086 : shared->page_dirty[slotno] = false;
578 :
579 : /* Acquire per-buffer lock (cannot deadlock, see notes at top) */
580 4086 : LWLockAcquire(&shared->buffer_locks[slotno].lock, LW_EXCLUSIVE);
581 :
582 : /* Release bank lock while doing I/O */
583 4086 : LWLockRelease(banklock);
584 :
585 : /* Do the read */
586 4086 : ok = SlruPhysicalReadPage(ctl, pageno, slotno);
587 :
588 : /* Set the LSNs for this newly read-in page to zero */
589 4086 : SimpleLruZeroLSNs(ctl, slotno);
590 :
591 : /* Re-acquire bank control lock and update page state */
592 4086 : LWLockAcquire(banklock, LW_EXCLUSIVE);
593 :
594 : Assert(shared->page_number[slotno] == pageno &&
595 : shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS &&
596 : !shared->page_dirty[slotno]);
597 :
598 4086 : shared->page_status[slotno] = ok ? SLRU_PAGE_VALID : SLRU_PAGE_EMPTY;
599 :
600 4086 : LWLockRelease(&shared->buffer_locks[slotno].lock);
601 :
602 : /* Now it's okay to ereport if we failed */
603 4086 : if (!ok)
604 0 : SlruReportIOError(ctl, pageno, xid);
605 :
606 4086 : SlruRecentlyUsed(shared, slotno);
607 :
608 : /* update the stats counter of pages not found in SLRU */
609 4086 : pgstat_count_slru_blocks_read(shared->slru_stats_idx);
610 :
611 4086 : return slotno;
612 : }
613 : }
614 :
615 : /*
616 : * Find a page in a shared buffer, reading it in if necessary.
617 : * The page number must correspond to an already-initialized page.
618 : * The caller must intend only read-only access to the page.
619 : *
620 : * The passed-in xid is used only for error reporting, and may be
621 : * InvalidTransactionId if no specific xid is associated with the action.
622 : *
623 : * Return value is the shared-buffer slot number now holding the page.
624 : * The buffer's LRU access info is updated.
625 : *
626 : * Bank control lock must NOT be held at entry, but will be held at exit.
627 : * It is unspecified whether the lock will be shared or exclusive.
628 : */
629 : int
630 1379716 : SimpleLruReadPage_ReadOnly(SlruCtl ctl, int64 pageno, TransactionId xid)
631 : {
632 1379716 : SlruShared shared = ctl->shared;
633 1379716 : LWLock *banklock = SimpleLruGetBankLock(ctl, pageno);
634 1379716 : int bankno = pageno % ctl->nbanks;
635 1379716 : int bankstart = bankno * SLRU_BANK_SIZE;
636 1379716 : int bankend = bankstart + SLRU_BANK_SIZE;
637 :
638 : /* Try to find the page while holding only shared lock */
639 1379716 : LWLockAcquire(banklock, LW_SHARED);
640 :
641 : /* See if page is already in a buffer */
642 1393450 : for (int slotno = bankstart; slotno < bankend; slotno++)
643 : {
644 1392954 : if (shared->page_status[slotno] != SLRU_PAGE_EMPTY &&
645 1391280 : shared->page_number[slotno] == pageno &&
646 1379220 : shared->page_status[slotno] != SLRU_PAGE_READ_IN_PROGRESS)
647 : {
648 : /* See comments for SlruRecentlyUsed() */
649 1379220 : SlruRecentlyUsed(shared, slotno);
650 :
651 : /* update the stats counter of pages found in the SLRU */
652 1379220 : pgstat_count_slru_blocks_hit(shared->slru_stats_idx);
653 :
654 1379220 : return slotno;
655 : }
656 : }
657 :
658 : /* No luck, so switch to normal exclusive lock and do regular read */
659 496 : LWLockRelease(banklock);
660 496 : LWLockAcquire(banklock, LW_EXCLUSIVE);
661 :
662 496 : return SimpleLruReadPage(ctl, pageno, true, xid);
663 : }
664 :
665 : /*
666 : * Write a page from a shared buffer, if necessary.
667 : * Does nothing if the specified slot is not dirty.
668 : *
669 : * NOTE: only one write attempt is made here. Hence, it is possible that
670 : * the page is still dirty at exit (if someone else re-dirtied it during
671 : * the write). However, we *do* attempt a fresh write even if the page
672 : * is already being written; this is for checkpoints.
673 : *
674 : * Bank lock must be held at entry, and will be held at exit.
675 : */
676 : static void
677 14687684 : SlruInternalWritePage(SlruCtl ctl, int slotno, SlruWriteAll fdata)
678 : {
679 14687684 : SlruShared shared = ctl->shared;
680 14687684 : int64 pageno = shared->page_number[slotno];
681 14687684 : int bankno = SlotGetBankNumber(slotno);
682 : bool ok;
683 :
684 : Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
685 : Assert(LWLockHeldByMeInMode(SimpleLruGetBankLock(ctl, pageno), LW_EXCLUSIVE));
686 :
687 : /* If a write is in progress, wait for it to finish */
688 14687688 : while (shared->page_status[slotno] == SLRU_PAGE_WRITE_IN_PROGRESS &&
689 4 : shared->page_number[slotno] == pageno)
690 : {
691 4 : SimpleLruWaitIO(ctl, slotno);
692 : }
693 :
694 : /*
695 : * Do nothing if page is not dirty, or if buffer no longer contains the
696 : * same page we were called for.
697 : */
698 14687684 : if (!shared->page_dirty[slotno] ||
699 14682176 : shared->page_status[slotno] != SLRU_PAGE_VALID ||
700 14682176 : shared->page_number[slotno] != pageno)
701 5508 : return;
702 :
703 : /*
704 : * Mark the slot write-busy, and clear the dirtybit. After this point, a
705 : * transaction status update on this page will mark it dirty again.
706 : */
707 14682176 : shared->page_status[slotno] = SLRU_PAGE_WRITE_IN_PROGRESS;
708 14682176 : shared->page_dirty[slotno] = false;
709 :
710 : /* Acquire per-buffer lock (cannot deadlock, see notes at top) */
711 14682176 : LWLockAcquire(&shared->buffer_locks[slotno].lock, LW_EXCLUSIVE);
712 :
713 : /* Release bank lock while doing I/O */
714 14682176 : LWLockRelease(&shared->bank_locks[bankno].lock);
715 :
716 : /* Do the write */
717 14682176 : ok = SlruPhysicalWritePage(ctl, pageno, slotno, fdata);
718 :
719 : /* If we failed, and we're in a flush, better close the files */
720 14682176 : if (!ok && fdata)
721 : {
722 0 : for (int i = 0; i < fdata->num_files; i++)
723 0 : CloseTransientFile(fdata->fd[i]);
724 : }
725 :
726 : /* Re-acquire bank lock and update page state */
727 14682176 : LWLockAcquire(&shared->bank_locks[bankno].lock, LW_EXCLUSIVE);
728 :
729 : Assert(shared->page_number[slotno] == pageno &&
730 : shared->page_status[slotno] == SLRU_PAGE_WRITE_IN_PROGRESS);
731 :
732 : /* If we failed to write, mark the page dirty again */
733 14682176 : if (!ok)
734 0 : shared->page_dirty[slotno] = true;
735 :
736 14682176 : shared->page_status[slotno] = SLRU_PAGE_VALID;
737 :
738 14682176 : LWLockRelease(&shared->buffer_locks[slotno].lock);
739 :
740 : /* Now it's okay to ereport if we failed */
741 14682176 : if (!ok)
742 0 : SlruReportIOError(ctl, pageno, InvalidTransactionId);
743 :
744 : /* If part of a checkpoint, count this as a SLRU buffer written. */
745 14682176 : if (fdata)
746 : {
747 5550 : CheckpointStats.ckpt_slru_written++;
748 5550 : PendingCheckpointerStats.slru_written++;
749 : }
750 : }
751 :
752 : /*
753 : * Wrapper of SlruInternalWritePage, for external callers.
754 : * fdata is always passed a NULL here.
755 : */
756 : void
757 628 : SimpleLruWritePage(SlruCtl ctl, int slotno)
758 : {
759 : Assert(ctl->shared->page_status[slotno] != SLRU_PAGE_EMPTY);
760 :
761 628 : SlruInternalWritePage(ctl, slotno, NULL);
762 628 : }
763 :
764 : /*
765 : * Return whether the given page exists on disk.
766 : *
767 : * A false return means that either the file does not exist, or that it's not
768 : * large enough to contain the given page.
769 : */
770 : bool
771 214 : SimpleLruDoesPhysicalPageExist(SlruCtl ctl, int64 pageno)
772 : {
773 214 : int64 segno = pageno / SLRU_PAGES_PER_SEGMENT;
774 214 : int rpageno = pageno % SLRU_PAGES_PER_SEGMENT;
775 214 : int offset = rpageno * BLCKSZ;
776 : char path[MAXPGPATH];
777 : int fd;
778 : bool result;
779 : off_t endpos;
780 :
781 : /* update the stats counter of checked pages */
782 214 : pgstat_count_slru_blocks_exists(ctl->shared->slru_stats_idx);
783 :
784 214 : SlruFileName(ctl, path, segno);
785 :
786 214 : fd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
787 214 : if (fd < 0)
788 : {
789 : /* expected: file doesn't exist */
790 52 : if (errno == ENOENT)
791 52 : return false;
792 :
793 : /* report error normally */
794 0 : slru_errcause = SLRU_OPEN_FAILED;
795 0 : slru_errno = errno;
796 0 : SlruReportIOError(ctl, pageno, 0);
797 : }
798 :
799 162 : if ((endpos = lseek(fd, 0, SEEK_END)) < 0)
800 : {
801 0 : slru_errcause = SLRU_SEEK_FAILED;
802 0 : slru_errno = errno;
803 0 : SlruReportIOError(ctl, pageno, 0);
804 : }
805 :
806 162 : result = endpos >= (off_t) (offset + BLCKSZ);
807 :
808 162 : if (CloseTransientFile(fd) != 0)
809 : {
810 0 : slru_errcause = SLRU_CLOSE_FAILED;
811 0 : slru_errno = errno;
812 0 : return false;
813 : }
814 :
815 162 : return result;
816 : }
817 :
818 : /*
819 : * Physical read of a (previously existing) page into a buffer slot
820 : *
821 : * On failure, we cannot just ereport(ERROR) since caller has put state in
822 : * shared memory that must be undone. So, we return false and save enough
823 : * info in static variables to let SlruReportIOError make the report.
824 : *
825 : * For now, assume it's not worth keeping a file pointer open across
826 : * read/write operations. We could cache one virtual file pointer ...
827 : */
828 : static bool
829 4086 : SlruPhysicalReadPage(SlruCtl ctl, int64 pageno, int slotno)
830 : {
831 4086 : SlruShared shared = ctl->shared;
832 4086 : int64 segno = pageno / SLRU_PAGES_PER_SEGMENT;
833 4086 : int rpageno = pageno % SLRU_PAGES_PER_SEGMENT;
834 4086 : off_t offset = rpageno * BLCKSZ;
835 : char path[MAXPGPATH];
836 : int fd;
837 :
838 4086 : SlruFileName(ctl, path, segno);
839 :
840 : /*
841 : * In a crash-and-restart situation, it's possible for us to receive
842 : * commands to set the commit status of transactions whose bits are in
843 : * already-truncated segments of the commit log (see notes in
844 : * SlruPhysicalWritePage). Hence, if we are InRecovery, allow the case
845 : * where the file doesn't exist, and return zeroes instead.
846 : */
847 4086 : fd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
848 4086 : if (fd < 0)
849 : {
850 0 : if (errno != ENOENT || !InRecovery)
851 : {
852 0 : slru_errcause = SLRU_OPEN_FAILED;
853 0 : slru_errno = errno;
854 0 : return false;
855 : }
856 :
857 0 : ereport(LOG,
858 : (errmsg("file \"%s\" doesn't exist, reading as zeroes",
859 : path)));
860 0 : MemSet(shared->page_buffer[slotno], 0, BLCKSZ);
861 0 : return true;
862 : }
863 :
864 4086 : errno = 0;
865 4086 : pgstat_report_wait_start(WAIT_EVENT_SLRU_READ);
866 4086 : if (pg_pread(fd, shared->page_buffer[slotno], BLCKSZ, offset) != BLCKSZ)
867 : {
868 0 : pgstat_report_wait_end();
869 0 : slru_errcause = SLRU_READ_FAILED;
870 0 : slru_errno = errno;
871 0 : CloseTransientFile(fd);
872 0 : return false;
873 : }
874 4086 : pgstat_report_wait_end();
875 :
876 4086 : if (CloseTransientFile(fd) != 0)
877 : {
878 0 : slru_errcause = SLRU_CLOSE_FAILED;
879 0 : slru_errno = errno;
880 0 : return false;
881 : }
882 :
883 4086 : return true;
884 : }
885 :
886 : /*
887 : * Physical write of a page from a buffer slot
888 : *
889 : * On failure, we cannot just ereport(ERROR) since caller has put state in
890 : * shared memory that must be undone. So, we return false and save enough
891 : * info in static variables to let SlruReportIOError make the report.
892 : *
893 : * For now, assume it's not worth keeping a file pointer open across
894 : * independent read/write operations. We do batch operations during
895 : * SimpleLruWriteAll, though.
896 : *
897 : * fdata is NULL for a standalone write, pointer to open-file info during
898 : * SimpleLruWriteAll.
899 : */
900 : static bool
901 14682176 : SlruPhysicalWritePage(SlruCtl ctl, int64 pageno, int slotno, SlruWriteAll fdata)
902 : {
903 14682176 : SlruShared shared = ctl->shared;
904 14682176 : int64 segno = pageno / SLRU_PAGES_PER_SEGMENT;
905 14682176 : int rpageno = pageno % SLRU_PAGES_PER_SEGMENT;
906 14682176 : off_t offset = rpageno * BLCKSZ;
907 : char path[MAXPGPATH];
908 14682176 : int fd = -1;
909 :
910 : /* update the stats counter of written pages */
911 14682176 : pgstat_count_slru_blocks_written(shared->slru_stats_idx);
912 :
913 : /*
914 : * Honor the write-WAL-before-data rule, if appropriate, so that we do not
915 : * write out data before associated WAL records. This is the same action
916 : * performed during FlushBuffer() in the main buffer manager.
917 : */
918 14682176 : if (shared->group_lsn != NULL)
919 : {
920 : /*
921 : * We must determine the largest async-commit LSN for the page. This
922 : * is a bit tedious, but since this entire function is a slow path
923 : * anyway, it seems better to do this here than to maintain a per-page
924 : * LSN variable (which'd need an extra comparison in the
925 : * transaction-commit path).
926 : */
927 : XLogRecPtr max_lsn;
928 : int lsnindex;
929 :
930 865936 : lsnindex = slotno * shared->lsn_groups_per_page;
931 865936 : max_lsn = shared->group_lsn[lsnindex++];
932 886718464 : for (int lsnoff = 1; lsnoff < shared->lsn_groups_per_page; lsnoff++)
933 : {
934 885852528 : XLogRecPtr this_lsn = shared->group_lsn[lsnindex++];
935 :
936 885852528 : if (max_lsn < this_lsn)
937 82254 : max_lsn = this_lsn;
938 : }
939 :
940 865936 : if (!XLogRecPtrIsInvalid(max_lsn))
941 : {
942 : /*
943 : * As noted above, elog(ERROR) is not acceptable here, so if
944 : * XLogFlush were to fail, we must PANIC. This isn't much of a
945 : * restriction because XLogFlush is just about all critical
946 : * section anyway, but let's make sure.
947 : */
948 1006 : START_CRIT_SECTION();
949 1006 : XLogFlush(max_lsn);
950 1006 : END_CRIT_SECTION();
951 : }
952 : }
953 :
954 : /*
955 : * During a SimpleLruWriteAll, we may already have the desired file open.
956 : */
957 14682176 : if (fdata)
958 : {
959 5774 : for (int i = 0; i < fdata->num_files; i++)
960 : {
961 778 : if (fdata->segno[i] == segno)
962 : {
963 554 : fd = fdata->fd[i];
964 554 : break;
965 : }
966 : }
967 : }
968 :
969 14682176 : if (fd < 0)
970 : {
971 : /*
972 : * If the file doesn't already exist, we should create it. It is
973 : * possible for this to need to happen when writing a page that's not
974 : * first in its segment; we assume the OS can cope with that. (Note:
975 : * it might seem that it'd be okay to create files only when
976 : * SimpleLruZeroPage is called for the first page of a segment.
977 : * However, if after a crash and restart the REDO logic elects to
978 : * replay the log from a checkpoint before the latest one, then it's
979 : * possible that we will get commands to set transaction status of
980 : * transactions that have already been truncated from the commit log.
981 : * Easiest way to deal with that is to accept references to
982 : * nonexistent files here and in SlruPhysicalReadPage.)
983 : *
984 : * Note: it is possible for more than one backend to be executing this
985 : * code simultaneously for different pages of the same file. Hence,
986 : * don't use O_EXCL or O_TRUNC or anything like that.
987 : */
988 14681622 : SlruFileName(ctl, path, segno);
989 14681622 : fd = OpenTransientFile(path, O_RDWR | O_CREAT | PG_BINARY);
990 14681622 : if (fd < 0)
991 : {
992 0 : slru_errcause = SLRU_OPEN_FAILED;
993 0 : slru_errno = errno;
994 0 : return false;
995 : }
996 :
997 14681622 : if (fdata)
998 : {
999 4996 : if (fdata->num_files < MAX_WRITEALL_BUFFERS)
1000 : {
1001 4996 : fdata->fd[fdata->num_files] = fd;
1002 4996 : fdata->segno[fdata->num_files] = segno;
1003 4996 : fdata->num_files++;
1004 : }
1005 : else
1006 : {
1007 : /*
1008 : * In the unlikely event that we exceed MAX_WRITEALL_BUFFERS,
1009 : * fall back to treating it as a standalone write.
1010 : */
1011 0 : fdata = NULL;
1012 : }
1013 : }
1014 : }
1015 :
1016 14682176 : errno = 0;
1017 14682176 : pgstat_report_wait_start(WAIT_EVENT_SLRU_WRITE);
1018 14682176 : if (pg_pwrite(fd, shared->page_buffer[slotno], BLCKSZ, offset) != BLCKSZ)
1019 : {
1020 0 : pgstat_report_wait_end();
1021 : /* if write didn't set errno, assume problem is no disk space */
1022 0 : if (errno == 0)
1023 0 : errno = ENOSPC;
1024 0 : slru_errcause = SLRU_WRITE_FAILED;
1025 0 : slru_errno = errno;
1026 0 : if (!fdata)
1027 0 : CloseTransientFile(fd);
1028 0 : return false;
1029 : }
1030 14682176 : pgstat_report_wait_end();
1031 :
1032 : /* Queue up a sync request for the checkpointer. */
1033 14682176 : if (ctl->sync_handler != SYNC_HANDLER_NONE)
1034 : {
1035 : FileTag tag;
1036 :
1037 867448 : INIT_SLRUFILETAG(tag, ctl->sync_handler, segno);
1038 867448 : if (!RegisterSyncRequest(&tag, SYNC_REQUEST, false))
1039 : {
1040 : /* No space to enqueue sync request. Do it synchronously. */
1041 24 : pgstat_report_wait_start(WAIT_EVENT_SLRU_SYNC);
1042 24 : if (pg_fsync(fd) != 0)
1043 : {
1044 0 : pgstat_report_wait_end();
1045 0 : slru_errcause = SLRU_FSYNC_FAILED;
1046 0 : slru_errno = errno;
1047 0 : CloseTransientFile(fd);
1048 0 : return false;
1049 : }
1050 24 : pgstat_report_wait_end();
1051 : }
1052 : }
1053 :
1054 : /* Close file, unless part of flush request. */
1055 14682176 : if (!fdata)
1056 : {
1057 14676626 : if (CloseTransientFile(fd) != 0)
1058 : {
1059 0 : slru_errcause = SLRU_CLOSE_FAILED;
1060 0 : slru_errno = errno;
1061 0 : return false;
1062 : }
1063 : }
1064 :
1065 14682176 : return true;
1066 : }
1067 :
1068 : /*
1069 : * Issue the error message after failure of SlruPhysicalReadPage or
1070 : * SlruPhysicalWritePage. Call this after cleaning up shared-memory state.
1071 : */
1072 : static void
1073 0 : SlruReportIOError(SlruCtl ctl, int64 pageno, TransactionId xid)
1074 : {
1075 0 : int64 segno = pageno / SLRU_PAGES_PER_SEGMENT;
1076 0 : int rpageno = pageno % SLRU_PAGES_PER_SEGMENT;
1077 0 : int offset = rpageno * BLCKSZ;
1078 : char path[MAXPGPATH];
1079 :
1080 0 : SlruFileName(ctl, path, segno);
1081 0 : errno = slru_errno;
1082 0 : switch (slru_errcause)
1083 : {
1084 0 : case SLRU_OPEN_FAILED:
1085 0 : ereport(ERROR,
1086 : (errcode_for_file_access(),
1087 : errmsg("could not access status of transaction %u", xid),
1088 : errdetail("Could not open file \"%s\": %m.", path)));
1089 : break;
1090 0 : case SLRU_SEEK_FAILED:
1091 0 : ereport(ERROR,
1092 : (errcode_for_file_access(),
1093 : errmsg("could not access status of transaction %u", xid),
1094 : errdetail("Could not seek in file \"%s\" to offset %d: %m.",
1095 : path, offset)));
1096 : break;
1097 0 : case SLRU_READ_FAILED:
1098 0 : if (errno)
1099 0 : ereport(ERROR,
1100 : (errcode_for_file_access(),
1101 : errmsg("could not access status of transaction %u", xid),
1102 : errdetail("Could not read from file \"%s\" at offset %d: %m.",
1103 : path, offset)));
1104 : else
1105 0 : ereport(ERROR,
1106 : (errmsg("could not access status of transaction %u", xid),
1107 : errdetail("Could not read from file \"%s\" at offset %d: read too few bytes.", path, offset)));
1108 : break;
1109 0 : case SLRU_WRITE_FAILED:
1110 0 : if (errno)
1111 0 : ereport(ERROR,
1112 : (errcode_for_file_access(),
1113 : errmsg("could not access status of transaction %u", xid),
1114 : errdetail("Could not write to file \"%s\" at offset %d: %m.",
1115 : path, offset)));
1116 : else
1117 0 : ereport(ERROR,
1118 : (errmsg("could not access status of transaction %u", xid),
1119 : errdetail("Could not write to file \"%s\" at offset %d: wrote too few bytes.",
1120 : path, offset)));
1121 : break;
1122 0 : case SLRU_FSYNC_FAILED:
1123 0 : ereport(data_sync_elevel(ERROR),
1124 : (errcode_for_file_access(),
1125 : errmsg("could not access status of transaction %u", xid),
1126 : errdetail("Could not fsync file \"%s\": %m.",
1127 : path)));
1128 0 : break;
1129 0 : case SLRU_CLOSE_FAILED:
1130 0 : ereport(ERROR,
1131 : (errcode_for_file_access(),
1132 : errmsg("could not access status of transaction %u", xid),
1133 : errdetail("Could not close file \"%s\": %m.",
1134 : path)));
1135 : break;
1136 0 : default:
1137 : /* can't get here, we trust */
1138 0 : elog(ERROR, "unrecognized SimpleLru error cause: %d",
1139 : (int) slru_errcause);
1140 : break;
1141 : }
1142 0 : }
1143 :
1144 : /*
1145 : * Mark a buffer slot "most recently used".
1146 : */
1147 : static inline void
1148 16386818 : SlruRecentlyUsed(SlruShared shared, int slotno)
1149 : {
1150 16386818 : int bankno = SlotGetBankNumber(slotno);
1151 16386818 : int new_lru_count = shared->bank_cur_lru_count[bankno];
1152 :
1153 : Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
1154 :
1155 : /*
1156 : * The reason for the if-test is that there are often many consecutive
1157 : * accesses to the same page (particularly the latest page). By
1158 : * suppressing useless increments of bank_cur_lru_count, we reduce the
1159 : * probability that old pages' counts will "wrap around" and make them
1160 : * appear recently used.
1161 : *
1162 : * We allow this code to be executed concurrently by multiple processes
1163 : * within SimpleLruReadPage_ReadOnly(). As long as int reads and writes
1164 : * are atomic, this should not cause any completely-bogus values to enter
1165 : * the computation. However, it is possible for either bank_cur_lru_count
1166 : * or individual page_lru_count entries to be "reset" to lower values than
1167 : * they should have, in case a process is delayed while it executes this
1168 : * function. With care in SlruSelectLRUPage(), this does little harm, and
1169 : * in any case the absolute worst possible consequence is a nonoptimal
1170 : * choice of page to evict. The gain from allowing concurrent reads of
1171 : * SLRU pages seems worth it.
1172 : */
1173 16386818 : if (new_lru_count != shared->page_lru_count[slotno])
1174 : {
1175 14683566 : shared->bank_cur_lru_count[bankno] = ++new_lru_count;
1176 14683566 : shared->page_lru_count[slotno] = new_lru_count;
1177 : }
1178 16386818 : }
1179 :
1180 : /*
1181 : * Select the slot to re-use when we need a free slot for the given page.
1182 : *
1183 : * The target page number is passed not only because we need to know the
1184 : * correct bank to use, but also because we need to consider the possibility
1185 : * that some other process reads in the target page while we are doing I/O to
1186 : * free a slot. Hence, check or recheck to see if any slot already holds the
1187 : * target page, and return that slot if so. Thus, the returned slot is
1188 : * *either* a slot already holding the pageno (could be any state except
1189 : * EMPTY), *or* a freeable slot (state EMPTY or CLEAN).
1190 : *
1191 : * The correct bank lock must be held at entry, and will be held at exit.
1192 : */
1193 : static int
1194 15007598 : SlruSelectLRUPage(SlruCtl ctl, int64 pageno)
1195 : {
1196 15007598 : SlruShared shared = ctl->shared;
1197 :
1198 : /* Outer loop handles restart after I/O */
1199 : for (;;)
1200 14675808 : {
1201 : int cur_count;
1202 29683406 : int bestvalidslot = 0; /* keep compiler quiet */
1203 29683406 : int best_valid_delta = -1;
1204 29683406 : int64 best_valid_page_number = 0; /* keep compiler quiet */
1205 29683406 : int bestinvalidslot = 0; /* keep compiler quiet */
1206 29683406 : int best_invalid_delta = -1;
1207 29683406 : int64 best_invalid_page_number = 0; /* keep compiler quiet */
1208 29683406 : int bankno = pageno % ctl->nbanks;
1209 29683406 : int bankstart = bankno * SLRU_BANK_SIZE;
1210 29683406 : int bankend = bankstart + SLRU_BANK_SIZE;
1211 :
1212 : Assert(LWLockHeldByMe(SimpleLruGetBankLock(ctl, pageno)));
1213 :
1214 : /* See if page already has a buffer assigned */
1215 499427388 : for (int slotno = bankstart; slotno < bankend; slotno++)
1216 : {
1217 470068600 : if (shared->page_status[slotno] != SLRU_PAGE_EMPTY &&
1218 469971860 : shared->page_number[slotno] == pageno)
1219 324618 : return slotno;
1220 : }
1221 :
1222 : /*
1223 : * If we find any EMPTY slot, just select that one. Else choose a
1224 : * victim page to replace. We normally take the least recently used
1225 : * valid page, but we will never take the slot containing
1226 : * latest_page_number, even if it appears least recently used. We
1227 : * will select a slot that is already I/O busy only if there is no
1228 : * other choice: a read-busy slot will not be least recently used once
1229 : * the read finishes, and waiting for an I/O on a write-busy slot is
1230 : * inferior to just picking some other slot. Testing shows the slot
1231 : * we pick instead will often be clean, allowing us to begin a read at
1232 : * once.
1233 : *
1234 : * Normally the page_lru_count values will all be different and so
1235 : * there will be a well-defined LRU page. But since we allow
1236 : * concurrent execution of SlruRecentlyUsed() within
1237 : * SimpleLruReadPage_ReadOnly(), it is possible that multiple pages
1238 : * acquire the same lru_count values. In that case we break ties by
1239 : * choosing the furthest-back page.
1240 : *
1241 : * Notice that this next line forcibly advances cur_lru_count to a
1242 : * value that is certainly beyond any value that will be in the
1243 : * page_lru_count array after the loop finishes. This ensures that
1244 : * the next execution of SlruRecentlyUsed will mark the page newly
1245 : * used, even if it's for a page that has the current counter value.
1246 : * That gets us back on the path to having good data when there are
1247 : * multiple pages with the same lru_count.
1248 : */
1249 29358788 : cur_count = (shared->bank_cur_lru_count[bankno])++;
1250 499002764 : for (int slotno = bankstart; slotno < bankend; slotno++)
1251 : {
1252 : int this_delta;
1253 : int64 this_page_number;
1254 :
1255 469650242 : if (shared->page_status[slotno] == SLRU_PAGE_EMPTY)
1256 6266 : return slotno;
1257 :
1258 469643976 : this_delta = cur_count - shared->page_lru_count[slotno];
1259 469643976 : if (this_delta < 0)
1260 : {
1261 : /*
1262 : * Clean up in case shared updates have caused cur_count
1263 : * increments to get "lost". We back off the page counts,
1264 : * rather than trying to increase cur_count, to avoid any
1265 : * question of infinite loops or failure in the presence of
1266 : * wrapped-around counts.
1267 : */
1268 0 : shared->page_lru_count[slotno] = cur_count;
1269 0 : this_delta = 0;
1270 : }
1271 :
1272 : /*
1273 : * If this page is the one most recently zeroed, don't consider it
1274 : * an eviction candidate. See comments in SimpleLruZeroPage for an
1275 : * explanation about the lack of a memory barrier here.
1276 : */
1277 469643976 : this_page_number = shared->page_number[slotno];
1278 469643976 : if (this_page_number ==
1279 469643976 : pg_atomic_read_u64(&shared->latest_page_number))
1280 1024 : continue;
1281 :
1282 469642952 : if (shared->page_status[slotno] == SLRU_PAGE_VALID)
1283 : {
1284 469641372 : if (this_delta > best_valid_delta ||
1285 0 : (this_delta == best_valid_delta &&
1286 0 : ctl->PagePrecedes(this_page_number,
1287 : best_valid_page_number)))
1288 : {
1289 75786974 : bestvalidslot = slotno;
1290 75786974 : best_valid_delta = this_delta;
1291 75786974 : best_valid_page_number = this_page_number;
1292 : }
1293 : }
1294 : else
1295 : {
1296 1580 : if (this_delta > best_invalid_delta ||
1297 0 : (this_delta == best_invalid_delta &&
1298 0 : ctl->PagePrecedes(this_page_number,
1299 : best_invalid_page_number)))
1300 : {
1301 1580 : bestinvalidslot = slotno;
1302 1580 : best_invalid_delta = this_delta;
1303 1580 : best_invalid_page_number = this_page_number;
1304 : }
1305 : }
1306 : }
1307 :
1308 : /*
1309 : * If all pages (except possibly the latest one) are I/O busy, we'll
1310 : * have to wait for an I/O to complete and then retry. In that
1311 : * unhappy case, we choose to wait for the I/O on the least recently
1312 : * used slot, on the assumption that it was likely initiated first of
1313 : * all the I/Os in progress and may therefore finish first.
1314 : */
1315 29352522 : if (best_valid_delta < 0)
1316 : {
1317 0 : SimpleLruWaitIO(ctl, bestinvalidslot);
1318 0 : continue;
1319 : }
1320 :
1321 : /*
1322 : * If the selected page is clean, we're set.
1323 : */
1324 29352522 : if (!shared->page_dirty[bestvalidslot])
1325 14676714 : return bestvalidslot;
1326 :
1327 : /*
1328 : * Write the page.
1329 : */
1330 14675808 : SlruInternalWritePage(ctl, bestvalidslot, NULL);
1331 :
1332 : /*
1333 : * Now loop back and try again. This is the easiest way of dealing
1334 : * with corner cases such as the victim page being re-dirtied while we
1335 : * wrote it.
1336 : */
1337 : }
1338 : }
1339 :
1340 : /*
1341 : * Write dirty pages to disk during checkpoint or database shutdown. Flushing
1342 : * is deferred until the next call to ProcessSyncRequests(), though we do fsync
1343 : * the containing directory here to make sure that newly created directory
1344 : * entries are on disk.
1345 : */
1346 : void
1347 17130 : SimpleLruWriteAll(SlruCtl ctl, bool allow_redirtied)
1348 : {
1349 17130 : SlruShared shared = ctl->shared;
1350 : SlruWriteAllData fdata;
1351 17130 : int64 pageno = 0;
1352 17130 : int prevbank = SlotGetBankNumber(0);
1353 : bool ok;
1354 :
1355 : /* update the stats counter of flushes */
1356 17130 : pgstat_count_slru_flush(shared->slru_stats_idx);
1357 :
1358 : /*
1359 : * Find and write dirty pages
1360 : */
1361 17130 : fdata.num_files = 0;
1362 :
1363 17130 : LWLockAcquire(&shared->bank_locks[prevbank].lock, LW_EXCLUSIVE);
1364 :
1365 415498 : for (int slotno = 0; slotno < shared->num_slots; slotno++)
1366 : {
1367 398368 : int curbank = SlotGetBankNumber(slotno);
1368 :
1369 : /*
1370 : * If the current bank lock is not same as the previous bank lock then
1371 : * release the previous lock and acquire the new lock.
1372 : */
1373 398368 : if (curbank != prevbank)
1374 : {
1375 7768 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1376 7768 : LWLockAcquire(&shared->bank_locks[curbank].lock, LW_EXCLUSIVE);
1377 7768 : prevbank = curbank;
1378 : }
1379 :
1380 : /* Do nothing if slot is unused */
1381 398368 : if (shared->page_status[slotno] == SLRU_PAGE_EMPTY)
1382 387310 : continue;
1383 :
1384 11058 : SlruInternalWritePage(ctl, slotno, &fdata);
1385 :
1386 : /*
1387 : * In some places (e.g. checkpoints), we cannot assert that the slot
1388 : * is clean now, since another process might have re-dirtied it
1389 : * already. That's okay.
1390 : */
1391 : Assert(allow_redirtied ||
1392 : shared->page_status[slotno] == SLRU_PAGE_EMPTY ||
1393 : (shared->page_status[slotno] == SLRU_PAGE_VALID &&
1394 : !shared->page_dirty[slotno]));
1395 : }
1396 :
1397 17130 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1398 :
1399 : /*
1400 : * Now close any files that were open
1401 : */
1402 17130 : ok = true;
1403 22126 : for (int i = 0; i < fdata.num_files; i++)
1404 : {
1405 4996 : if (CloseTransientFile(fdata.fd[i]) != 0)
1406 : {
1407 0 : slru_errcause = SLRU_CLOSE_FAILED;
1408 0 : slru_errno = errno;
1409 0 : pageno = fdata.segno[i] * SLRU_PAGES_PER_SEGMENT;
1410 0 : ok = false;
1411 : }
1412 : }
1413 17130 : if (!ok)
1414 0 : SlruReportIOError(ctl, pageno, InvalidTransactionId);
1415 :
1416 : /* Ensure that directory entries for new files are on disk. */
1417 17130 : if (ctl->sync_handler != SYNC_HANDLER_NONE)
1418 13712 : fsync_fname(ctl->Dir, true);
1419 17130 : }
1420 :
1421 : /*
1422 : * Remove all segments before the one holding the passed page number
1423 : *
1424 : * All SLRUs prevent concurrent calls to this function, either with an LWLock
1425 : * or by calling it only as part of a checkpoint. Mutual exclusion must begin
1426 : * before computing cutoffPage. Mutual exclusion must end after any limit
1427 : * update that would permit other backends to write fresh data into the
1428 : * segment immediately preceding the one containing cutoffPage. Otherwise,
1429 : * when the SLRU is quite full, SimpleLruTruncate() might delete that segment
1430 : * after it has accrued freshly-written data.
1431 : */
1432 : void
1433 3562 : SimpleLruTruncate(SlruCtl ctl, int64 cutoffPage)
1434 : {
1435 3562 : SlruShared shared = ctl->shared;
1436 : int prevbank;
1437 :
1438 : /* update the stats counter of truncates */
1439 3562 : pgstat_count_slru_truncate(shared->slru_stats_idx);
1440 :
1441 : /*
1442 : * Scan shared memory and remove any pages preceding the cutoff page, to
1443 : * ensure we won't rewrite them later. (Since this is normally called in
1444 : * or just after a checkpoint, any dirty pages should have been flushed
1445 : * already ... we're just being extra careful here.)
1446 : */
1447 3752 : restart:
1448 :
1449 : /*
1450 : * An important safety check: the current endpoint page must not be
1451 : * eligible for removal. This check is just a backstop against wraparound
1452 : * bugs elsewhere in SLRU handling, so we don't care if we read a slightly
1453 : * outdated value; therefore we don't add a memory barrier.
1454 : */
1455 3752 : if (ctl->PagePrecedes(pg_atomic_read_u64(&shared->latest_page_number),
1456 : cutoffPage))
1457 : {
1458 0 : ereport(LOG,
1459 : (errmsg("could not truncate directory \"%s\": apparent wraparound",
1460 : ctl->Dir)));
1461 0 : return;
1462 : }
1463 :
1464 3752 : prevbank = SlotGetBankNumber(0);
1465 3752 : LWLockAcquire(&shared->bank_locks[prevbank].lock, LW_EXCLUSIVE);
1466 89768 : for (int slotno = 0; slotno < shared->num_slots; slotno++)
1467 : {
1468 86206 : int curbank = SlotGetBankNumber(slotno);
1469 :
1470 : /*
1471 : * If the current bank lock is not same as the previous bank lock then
1472 : * release the previous lock and acquire the new lock.
1473 : */
1474 86206 : if (curbank != prevbank)
1475 : {
1476 1730 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1477 1730 : LWLockAcquire(&shared->bank_locks[curbank].lock, LW_EXCLUSIVE);
1478 1730 : prevbank = curbank;
1479 : }
1480 :
1481 86206 : if (shared->page_status[slotno] == SLRU_PAGE_EMPTY)
1482 75692 : continue;
1483 10514 : if (!ctl->PagePrecedes(shared->page_number[slotno], cutoffPage))
1484 9978 : continue;
1485 :
1486 : /*
1487 : * If page is clean, just change state to EMPTY (expected case).
1488 : */
1489 536 : if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
1490 536 : !shared->page_dirty[slotno])
1491 : {
1492 346 : shared->page_status[slotno] = SLRU_PAGE_EMPTY;
1493 346 : continue;
1494 : }
1495 :
1496 : /*
1497 : * Hmm, we have (or may have) I/O operations acting on the page, so
1498 : * we've got to wait for them to finish and then start again. This is
1499 : * the same logic as in SlruSelectLRUPage. (XXX if page is dirty,
1500 : * wouldn't it be OK to just discard it without writing it?
1501 : * SlruMayDeleteSegment() uses a stricter qualification, so we might
1502 : * not delete this page in the end; even if we don't delete it, we
1503 : * won't have cause to read its data again. For now, keep the logic
1504 : * the same as it was.)
1505 : */
1506 190 : if (shared->page_status[slotno] == SLRU_PAGE_VALID)
1507 190 : SlruInternalWritePage(ctl, slotno, NULL);
1508 : else
1509 0 : SimpleLruWaitIO(ctl, slotno);
1510 :
1511 190 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1512 190 : goto restart;
1513 : }
1514 :
1515 3562 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1516 :
1517 : /* Now we can remove the old segment(s) */
1518 3562 : (void) SlruScanDirectory(ctl, SlruScanDirCbDeleteCutoff, &cutoffPage);
1519 : }
1520 :
1521 : /*
1522 : * Delete an individual SLRU segment.
1523 : *
1524 : * NB: This does not touch the SLRU buffers themselves, callers have to ensure
1525 : * they either can't yet contain anything, or have already been cleaned out.
1526 : */
1527 : static void
1528 284100 : SlruInternalDeleteSegment(SlruCtl ctl, int64 segno)
1529 : {
1530 : char path[MAXPGPATH];
1531 :
1532 : /* Forget any fsync requests queued for this segment. */
1533 284100 : if (ctl->sync_handler != SYNC_HANDLER_NONE)
1534 : {
1535 : FileTag tag;
1536 :
1537 26544 : INIT_SLRUFILETAG(tag, ctl->sync_handler, segno);
1538 26544 : RegisterSyncRequest(&tag, SYNC_FORGET_REQUEST, true);
1539 : }
1540 :
1541 : /* Unlink the file. */
1542 284100 : SlruFileName(ctl, path, segno);
1543 284100 : ereport(DEBUG2, (errmsg_internal("removing file \"%s\"", path)));
1544 284100 : unlink(path);
1545 284100 : }
1546 :
1547 : /*
1548 : * Delete an individual SLRU segment, identified by the segment number.
1549 : */
1550 : void
1551 4 : SlruDeleteSegment(SlruCtl ctl, int64 segno)
1552 : {
1553 4 : SlruShared shared = ctl->shared;
1554 4 : int prevbank = SlotGetBankNumber(0);
1555 : bool did_write;
1556 :
1557 : /* Clean out any possibly existing references to the segment. */
1558 4 : LWLockAcquire(&shared->bank_locks[prevbank].lock, LW_EXCLUSIVE);
1559 4 : restart:
1560 4 : did_write = false;
1561 68 : for (int slotno = 0; slotno < shared->num_slots; slotno++)
1562 : {
1563 : int64 pagesegno;
1564 64 : int curbank = SlotGetBankNumber(slotno);
1565 :
1566 : /*
1567 : * If the current bank lock is not same as the previous bank lock then
1568 : * release the previous lock and acquire the new lock.
1569 : */
1570 64 : if (curbank != prevbank)
1571 : {
1572 0 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1573 0 : LWLockAcquire(&shared->bank_locks[curbank].lock, LW_EXCLUSIVE);
1574 0 : prevbank = curbank;
1575 : }
1576 :
1577 64 : if (shared->page_status[slotno] == SLRU_PAGE_EMPTY)
1578 0 : continue;
1579 :
1580 64 : pagesegno = shared->page_number[slotno] / SLRU_PAGES_PER_SEGMENT;
1581 : /* not the segment we're looking for */
1582 64 : if (pagesegno != segno)
1583 14 : continue;
1584 :
1585 : /* If page is clean, just change state to EMPTY (expected case). */
1586 50 : if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
1587 50 : !shared->page_dirty[slotno])
1588 : {
1589 50 : shared->page_status[slotno] = SLRU_PAGE_EMPTY;
1590 50 : continue;
1591 : }
1592 :
1593 : /* Same logic as SimpleLruTruncate() */
1594 0 : if (shared->page_status[slotno] == SLRU_PAGE_VALID)
1595 0 : SlruInternalWritePage(ctl, slotno, NULL);
1596 : else
1597 0 : SimpleLruWaitIO(ctl, slotno);
1598 :
1599 0 : did_write = true;
1600 : }
1601 :
1602 : /*
1603 : * Be extra careful and re-check. The IO functions release the control
1604 : * lock, so new pages could have been read in.
1605 : */
1606 4 : if (did_write)
1607 0 : goto restart;
1608 :
1609 4 : SlruInternalDeleteSegment(ctl, segno);
1610 :
1611 4 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1612 4 : }
1613 :
1614 : /*
1615 : * Determine whether a segment is okay to delete.
1616 : *
1617 : * segpage is the first page of the segment, and cutoffPage is the oldest (in
1618 : * PagePrecedes order) page in the SLRU containing still-useful data. Since
1619 : * every core PagePrecedes callback implements "wrap around", check the
1620 : * segment's first and last pages:
1621 : *
1622 : * first<cutoff && last<cutoff: yes
1623 : * first<cutoff && last>=cutoff: no; cutoff falls inside this segment
1624 : * first>=cutoff && last<cutoff: no; wrap point falls inside this segment
1625 : * first>=cutoff && last>=cutoff: no; every page of this segment is too young
1626 : */
1627 : static bool
1628 2891066 : SlruMayDeleteSegment(SlruCtl ctl, int64 segpage, int64 cutoffPage)
1629 : {
1630 2891066 : int64 seg_last_page = segpage + SLRU_PAGES_PER_SEGMENT - 1;
1631 :
1632 : Assert(segpage % SLRU_PAGES_PER_SEGMENT == 0);
1633 :
1634 3176506 : return (ctl->PagePrecedes(segpage, cutoffPage) &&
1635 285440 : ctl->PagePrecedes(seg_last_page, cutoffPage));
1636 : }
1637 :
1638 : #ifdef USE_ASSERT_CHECKING
1639 : static void
1640 : SlruPagePrecedesTestOffset(SlruCtl ctl, int per_page, uint32 offset)
1641 : {
1642 : TransactionId lhs,
1643 : rhs;
1644 : int64 newestPage,
1645 : oldestPage;
1646 : TransactionId newestXact,
1647 : oldestXact;
1648 :
1649 : /*
1650 : * Compare an XID pair having undefined order (see RFC 1982), a pair at
1651 : * "opposite ends" of the XID space. TransactionIdPrecedes() treats each
1652 : * as preceding the other. If RHS is oldestXact, LHS is the first XID we
1653 : * must not assign.
1654 : */
1655 : lhs = per_page + offset; /* skip first page to avoid non-normal XIDs */
1656 : rhs = lhs + (1U << 31);
1657 : Assert(TransactionIdPrecedes(lhs, rhs));
1658 : Assert(TransactionIdPrecedes(rhs, lhs));
1659 : Assert(!TransactionIdPrecedes(lhs - 1, rhs));
1660 : Assert(TransactionIdPrecedes(rhs, lhs - 1));
1661 : Assert(TransactionIdPrecedes(lhs + 1, rhs));
1662 : Assert(!TransactionIdPrecedes(rhs, lhs + 1));
1663 : Assert(!TransactionIdFollowsOrEquals(lhs, rhs));
1664 : Assert(!TransactionIdFollowsOrEquals(rhs, lhs));
1665 : Assert(!ctl->PagePrecedes(lhs / per_page, lhs / per_page));
1666 : Assert(!ctl->PagePrecedes(lhs / per_page, rhs / per_page));
1667 : Assert(!ctl->PagePrecedes(rhs / per_page, lhs / per_page));
1668 : Assert(!ctl->PagePrecedes((lhs - per_page) / per_page, rhs / per_page));
1669 : Assert(ctl->PagePrecedes(rhs / per_page, (lhs - 3 * per_page) / per_page));
1670 : Assert(ctl->PagePrecedes(rhs / per_page, (lhs - 2 * per_page) / per_page));
1671 : Assert(ctl->PagePrecedes(rhs / per_page, (lhs - 1 * per_page) / per_page)
1672 : || (1U << 31) % per_page != 0); /* See CommitTsPagePrecedes() */
1673 : Assert(ctl->PagePrecedes((lhs + 1 * per_page) / per_page, rhs / per_page)
1674 : || (1U << 31) % per_page != 0);
1675 : Assert(ctl->PagePrecedes((lhs + 2 * per_page) / per_page, rhs / per_page));
1676 : Assert(ctl->PagePrecedes((lhs + 3 * per_page) / per_page, rhs / per_page));
1677 : Assert(!ctl->PagePrecedes(rhs / per_page, (lhs + per_page) / per_page));
1678 :
1679 : /*
1680 : * GetNewTransactionId() has assigned the last XID it can safely use, and
1681 : * that XID is in the *LAST* page of the second segment. We must not
1682 : * delete that segment.
1683 : */
1684 : newestPage = 2 * SLRU_PAGES_PER_SEGMENT - 1;
1685 : newestXact = newestPage * per_page + offset;
1686 : Assert(newestXact / per_page == newestPage);
1687 : oldestXact = newestXact + 1;
1688 : oldestXact -= 1U << 31;
1689 : oldestPage = oldestXact / per_page;
1690 : Assert(!SlruMayDeleteSegment(ctl,
1691 : (newestPage -
1692 : newestPage % SLRU_PAGES_PER_SEGMENT),
1693 : oldestPage));
1694 :
1695 : /*
1696 : * GetNewTransactionId() has assigned the last XID it can safely use, and
1697 : * that XID is in the *FIRST* page of the second segment. We must not
1698 : * delete that segment.
1699 : */
1700 : newestPage = SLRU_PAGES_PER_SEGMENT;
1701 : newestXact = newestPage * per_page + offset;
1702 : Assert(newestXact / per_page == newestPage);
1703 : oldestXact = newestXact + 1;
1704 : oldestXact -= 1U << 31;
1705 : oldestPage = oldestXact / per_page;
1706 : Assert(!SlruMayDeleteSegment(ctl,
1707 : (newestPage -
1708 : newestPage % SLRU_PAGES_PER_SEGMENT),
1709 : oldestPage));
1710 : }
1711 :
1712 : /*
1713 : * Unit-test a PagePrecedes function.
1714 : *
1715 : * This assumes every uint32 >= FirstNormalTransactionId is a valid key. It
1716 : * assumes each value occupies a contiguous, fixed-size region of SLRU bytes.
1717 : * (MultiXactMemberCtl separates flags from XIDs. NotifyCtl has
1718 : * variable-length entries, no keys, and no random access. These unit tests
1719 : * do not apply to them.)
1720 : */
1721 : void
1722 : SlruPagePrecedesUnitTests(SlruCtl ctl, int per_page)
1723 : {
1724 : /* Test first, middle and last entries of a page. */
1725 : SlruPagePrecedesTestOffset(ctl, per_page, 0);
1726 : SlruPagePrecedesTestOffset(ctl, per_page, per_page / 2);
1727 : SlruPagePrecedesTestOffset(ctl, per_page, per_page - 1);
1728 : }
1729 : #endif
1730 :
1731 : /*
1732 : * SlruScanDirectory callback
1733 : * This callback reports true if there's any segment wholly prior to the
1734 : * one containing the page passed as "data".
1735 : */
1736 : bool
1737 2492168 : SlruScanDirCbReportPresence(SlruCtl ctl, char *filename, int64 segpage,
1738 : void *data)
1739 : {
1740 2492168 : int64 cutoffPage = *(int64 *) data;
1741 :
1742 2492168 : if (SlruMayDeleteSegment(ctl, segpage, cutoffPage))
1743 202 : return true; /* found one; don't iterate any more */
1744 :
1745 2491966 : return false; /* keep going */
1746 : }
1747 :
1748 : /*
1749 : * SlruScanDirectory callback.
1750 : * This callback deletes segments prior to the one passed in as "data".
1751 : */
1752 : static bool
1753 398898 : SlruScanDirCbDeleteCutoff(SlruCtl ctl, char *filename, int64 segpage,
1754 : void *data)
1755 : {
1756 398898 : int64 cutoffPage = *(int64 *) data;
1757 :
1758 398898 : if (SlruMayDeleteSegment(ctl, segpage, cutoffPage))
1759 284080 : SlruInternalDeleteSegment(ctl, segpage / SLRU_PAGES_PER_SEGMENT);
1760 :
1761 398898 : return false; /* keep going */
1762 : }
1763 :
1764 : /*
1765 : * SlruScanDirectory callback.
1766 : * This callback deletes all segments.
1767 : */
1768 : bool
1769 16 : SlruScanDirCbDeleteAll(SlruCtl ctl, char *filename, int64 segpage, void *data)
1770 : {
1771 16 : SlruInternalDeleteSegment(ctl, segpage / SLRU_PAGES_PER_SEGMENT);
1772 :
1773 16 : return false; /* keep going */
1774 : }
1775 :
1776 : /*
1777 : * An internal function used by SlruScanDirectory().
1778 : *
1779 : * Returns true if a file with a name of a given length may be a correct
1780 : * SLRU segment.
1781 : */
1782 : static inline bool
1783 2916086 : SlruCorrectSegmentFilenameLength(SlruCtl ctl, size_t len)
1784 : {
1785 2916086 : if (ctl->long_segment_names)
1786 4372 : return (len == 15); /* see SlruFileName() */
1787 : else
1788 :
1789 : /*
1790 : * Commit 638cf09e76d allowed 5-character lengths. Later commit
1791 : * 73c986adde5 allowed 6-character length.
1792 : *
1793 : * Note: There is an ongoing plan to migrate all SLRUs to 64-bit page
1794 : * numbers, and the corresponding 15-character file names, which may
1795 : * eventually deprecate the support for 4, 5, and 6-character names.
1796 : */
1797 2911714 : return (len == 4 || len == 5 || len == 6);
1798 : }
1799 :
1800 : /*
1801 : * Scan the SimpleLru directory and apply a callback to each file found in it.
1802 : *
1803 : * If the callback returns true, the scan is stopped. The last return value
1804 : * from the callback is returned.
1805 : *
1806 : * The callback receives the following arguments: 1. the SlruCtl struct for the
1807 : * slru being truncated; 2. the filename being considered; 3. the page number
1808 : * for the first page of that file; 4. a pointer to the opaque data given to us
1809 : * by the caller.
1810 : *
1811 : * Note that the ordering in which the directory is scanned is not guaranteed.
1812 : *
1813 : * Note that no locking is applied.
1814 : */
1815 : bool
1816 12562 : SlruScanDirectory(SlruCtl ctl, SlruScanCallback callback, void *data)
1817 : {
1818 12562 : bool retval = false;
1819 : DIR *cldir;
1820 : struct dirent *clde;
1821 : int64 segno;
1822 : int64 segpage;
1823 :
1824 12562 : cldir = AllocateDir(ctl->Dir);
1825 2928446 : while ((clde = ReadDir(cldir, ctl->Dir)) != NULL)
1826 : {
1827 : size_t len;
1828 :
1829 2916086 : len = strlen(clde->d_name);
1830 :
1831 2916086 : if (SlruCorrectSegmentFilenameLength(ctl, len) &&
1832 2891082 : strspn(clde->d_name, "0123456789ABCDEF") == len)
1833 : {
1834 2891082 : segno = strtoi64(clde->d_name, NULL, 16);
1835 2891082 : segpage = segno * SLRU_PAGES_PER_SEGMENT;
1836 :
1837 2891082 : elog(DEBUG2, "SlruScanDirectory invoking callback on %s/%s",
1838 : ctl->Dir, clde->d_name);
1839 2891082 : retval = callback(ctl, clde->d_name, segpage, data);
1840 2891082 : if (retval)
1841 202 : break;
1842 : }
1843 : }
1844 12562 : FreeDir(cldir);
1845 :
1846 12562 : return retval;
1847 : }
1848 :
1849 : /*
1850 : * Individual SLRUs (clog, ...) have to provide a sync.c handler function so
1851 : * that they can provide the correct "SlruCtl" (otherwise we don't know how to
1852 : * build the path), but they just forward to this common implementation that
1853 : * performs the fsync.
1854 : */
1855 : int
1856 4 : SlruSyncFileTag(SlruCtl ctl, const FileTag *ftag, char *path)
1857 : {
1858 : int fd;
1859 : int save_errno;
1860 : int result;
1861 :
1862 4 : SlruFileName(ctl, path, ftag->segno);
1863 :
1864 4 : fd = OpenTransientFile(path, O_RDWR | PG_BINARY);
1865 4 : if (fd < 0)
1866 0 : return -1;
1867 :
1868 4 : pgstat_report_wait_start(WAIT_EVENT_SLRU_FLUSH_SYNC);
1869 4 : result = pg_fsync(fd);
1870 4 : pgstat_report_wait_end();
1871 4 : save_errno = errno;
1872 :
1873 4 : CloseTransientFile(fd);
1874 :
1875 4 : errno = save_errno;
1876 4 : return result;
1877 : }
|
