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) 1996-2026, 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 : #include "utils/wait_event.h"
75 :
76 : /*
77 : * Converts segment number to the filename of the segment.
78 : *
79 : * "path" should point to a buffer at least MAXPGPATH characters long.
80 : *
81 : * If ctl->long_segment_names is true, segno can be in the range [0, 2^60-1].
82 : * The resulting file name is made of 15 characters, e.g. dir/123456789ABCDEF.
83 : *
84 : * If ctl->long_segment_names is false, segno can be in the range [0, 2^24-1].
85 : * The resulting file name is made of 4 to 6 characters, as of:
86 : *
87 : * dir/1234 for [0, 2^16-1]
88 : * dir/12345 for [2^16, 2^20-1]
89 : * dir/123456 for [2^20, 2^24-1]
90 : */
91 : static inline int
92 7506023 : SlruFileName(SlruCtl ctl, char *path, int64 segno)
93 : {
94 7506023 : if (ctl->long_segment_names)
95 : {
96 : /*
97 : * We could use 16 characters here but the disadvantage would be that
98 : * the SLRU segments will be hard to distinguish from WAL segments.
99 : *
100 : * For this reason we use 15 characters. It is enough but also means
101 : * that in the future we can't decrease SLRU_PAGES_PER_SEGMENT easily.
102 : */
103 : Assert(segno >= 0 && segno <= INT64CONST(0xFFFFFFFFFFFFFFF));
104 16616 : return snprintf(path, MAXPGPATH, "%s/%015" PRIX64, ctl->Dir, segno);
105 : }
106 : else
107 : {
108 : /*
109 : * Despite the fact that %04X format string is used up to 24 bit
110 : * integers are allowed. See SlruCorrectSegmentFilenameLength()
111 : */
112 : Assert(segno >= 0 && segno <= INT64CONST(0xFFFFFF));
113 7489407 : return snprintf(path, MAXPGPATH, "%s/%04X", (ctl)->Dir,
114 : (unsigned int) segno);
115 : }
116 : }
117 :
118 : /*
119 : * During SimpleLruWriteAll(), we will usually not need to write more than one
120 : * or two physical files, but we may need to write several pages per file. We
121 : * can consolidate the I/O requests by leaving files open until control returns
122 : * to SimpleLruWriteAll(). This data structure remembers which files are open.
123 : */
124 : #define MAX_WRITEALL_BUFFERS 16
125 :
126 : typedef struct SlruWriteAllData
127 : {
128 : int num_files; /* # files actually open */
129 : int fd[MAX_WRITEALL_BUFFERS]; /* their FD's */
130 : int64 segno[MAX_WRITEALL_BUFFERS]; /* their log seg#s */
131 : } SlruWriteAllData;
132 :
133 : typedef struct SlruWriteAllData *SlruWriteAll;
134 :
135 :
136 : /*
137 : * Bank size for the slot array. Pages are assigned a bank according to their
138 : * page number, with each bank being this size. We want a power of 2 so that
139 : * we can determine the bank number for a page with just bit shifting; we also
140 : * want to keep the bank size small so that LRU victim search is fast. 16
141 : * buffers per bank seems a good number.
142 : */
143 : #define SLRU_BANK_BITSHIFT 4
144 : #define SLRU_BANK_SIZE (1 << SLRU_BANK_BITSHIFT)
145 :
146 : /*
147 : * Macro to get the bank number to which the slot belongs.
148 : */
149 : #define SlotGetBankNumber(slotno) ((slotno) >> SLRU_BANK_BITSHIFT)
150 :
151 :
152 : /*
153 : * Populate a file tag describing a segment file. We only use the segment
154 : * number, since we can derive everything else we need by having separate
155 : * sync handler functions for clog, multixact etc.
156 : */
157 : #define INIT_SLRUFILETAG(a,xx_handler,xx_segno) \
158 : ( \
159 : memset(&(a), 0, sizeof(FileTag)), \
160 : (a).handler = (xx_handler), \
161 : (a).segno = (xx_segno) \
162 : )
163 :
164 : /* Saved info for SlruReportIOError */
165 : typedef enum
166 : {
167 : SLRU_OPEN_FAILED,
168 : SLRU_SEEK_FAILED,
169 : SLRU_READ_FAILED,
170 : SLRU_WRITE_FAILED,
171 : SLRU_FSYNC_FAILED,
172 : SLRU_CLOSE_FAILED,
173 : } SlruErrorCause;
174 :
175 : static SlruErrorCause slru_errcause;
176 : static int slru_errno;
177 :
178 :
179 : static void SimpleLruZeroLSNs(SlruCtl ctl, int slotno);
180 : static void SimpleLruWaitIO(SlruCtl ctl, int slotno);
181 : static void SlruInternalWritePage(SlruCtl ctl, int slotno, SlruWriteAll fdata);
182 : static bool SlruPhysicalReadPage(SlruCtl ctl, int64 pageno, int slotno);
183 : static bool SlruPhysicalWritePage(SlruCtl ctl, int64 pageno, int slotno,
184 : SlruWriteAll fdata);
185 : static void SlruReportIOError(SlruCtl ctl, int64 pageno,
186 : const void *opaque_data);
187 : static int SlruSelectLRUPage(SlruCtl ctl, int64 pageno);
188 :
189 : static bool SlruScanDirCbDeleteCutoff(SlruCtl ctl, char *filename,
190 : int64 segpage, void *data);
191 : static void SlruInternalDeleteSegment(SlruCtl ctl, int64 segno);
192 : static inline void SlruRecentlyUsed(SlruShared shared, int slotno);
193 :
194 :
195 : /*
196 : * Initialization of shared memory
197 : */
198 :
199 : Size
200 23717 : SimpleLruShmemSize(int nslots, int nlsns)
201 : {
202 23717 : int nbanks = nslots / SLRU_BANK_SIZE;
203 : Size sz;
204 :
205 : Assert(nslots <= SLRU_MAX_ALLOWED_BUFFERS);
206 : Assert(nslots % SLRU_BANK_SIZE == 0);
207 :
208 : /* we assume nslots isn't so large as to risk overflow */
209 23717 : sz = MAXALIGN(sizeof(SlruSharedData));
210 23717 : sz += MAXALIGN(nslots * sizeof(char *)); /* page_buffer[] */
211 23717 : sz += MAXALIGN(nslots * sizeof(SlruPageStatus)); /* page_status[] */
212 23717 : sz += MAXALIGN(nslots * sizeof(bool)); /* page_dirty[] */
213 23717 : sz += MAXALIGN(nslots * sizeof(int64)); /* page_number[] */
214 23717 : sz += MAXALIGN(nslots * sizeof(int)); /* page_lru_count[] */
215 23717 : sz += MAXALIGN(nslots * sizeof(LWLockPadded)); /* buffer_locks[] */
216 23717 : sz += MAXALIGN(nbanks * sizeof(LWLockPadded)); /* bank_locks[] */
217 23717 : sz += MAXALIGN(nbanks * sizeof(int)); /* bank_cur_lru_count[] */
218 :
219 23717 : if (nlsns > 0)
220 3387 : sz += MAXALIGN(nslots * nlsns * sizeof(XLogRecPtr)); /* group_lsn[] */
221 :
222 23717 : return BUFFERALIGN(sz) + BLCKSZ * nslots;
223 : }
224 :
225 : /*
226 : * Determine a number of SLRU buffers to use.
227 : *
228 : * We simply divide shared_buffers by the divisor given and cap
229 : * that at the maximum given; but always at least SLRU_BANK_SIZE.
230 : * Round down to the nearest multiple of SLRU_BANK_SIZE.
231 : */
232 : int
233 10110 : SimpleLruAutotuneBuffers(int divisor, int max)
234 : {
235 10110 : return Min(max - (max % SLRU_BANK_SIZE),
236 : Max(SLRU_BANK_SIZE,
237 : NBuffers / divisor - (NBuffers / divisor) % SLRU_BANK_SIZE));
238 : }
239 :
240 : /*
241 : * Initialize, or attach to, a simple LRU cache in shared memory.
242 : *
243 : * ctl: address of local (unshared) control structure.
244 : * name: name of SLRU. (This is user-visible, pick with care!)
245 : * nslots: number of page slots to use.
246 : * nlsns: number of LSN groups per page (set to zero if not relevant).
247 : * subdir: PGDATA-relative subdirectory that will contain the files.
248 : * buffer_tranche_id: tranche ID to use for the SLRU's per-buffer LWLocks.
249 : * bank_tranche_id: tranche ID to use for the bank LWLocks.
250 : * sync_handler: which set of functions to use to handle sync requests
251 : * long_segment_names: use short or long segment names
252 : */
253 : void
254 8264 : SimpleLruInit(SlruCtl ctl, const char *name, int nslots, int nlsns,
255 : const char *subdir, int buffer_tranche_id, int bank_tranche_id,
256 : SyncRequestHandler sync_handler, bool long_segment_names)
257 : {
258 : SlruShared shared;
259 : bool found;
260 8264 : int nbanks = nslots / SLRU_BANK_SIZE;
261 :
262 : Assert(nslots <= SLRU_MAX_ALLOWED_BUFFERS);
263 :
264 : Assert(ctl->PagePrecedes != NULL);
265 : Assert(ctl->errdetail_for_io_error != NULL);
266 :
267 8264 : shared = (SlruShared) ShmemInitStruct(name,
268 : SimpleLruShmemSize(nslots, nlsns),
269 : &found);
270 :
271 8264 : if (!IsUnderPostmaster)
272 : {
273 : /* Initialize locks and shared memory area */
274 : char *ptr;
275 : Size offset;
276 :
277 : Assert(!found);
278 :
279 8264 : memset(shared, 0, sizeof(SlruSharedData));
280 :
281 8264 : shared->num_slots = nslots;
282 8264 : shared->lsn_groups_per_page = nlsns;
283 :
284 8264 : pg_atomic_init_u64(&shared->latest_page_number, 0);
285 :
286 8264 : shared->slru_stats_idx = pgstat_get_slru_index(name);
287 :
288 8264 : ptr = (char *) shared;
289 8264 : offset = MAXALIGN(sizeof(SlruSharedData));
290 8264 : shared->page_buffer = (char **) (ptr + offset);
291 8264 : offset += MAXALIGN(nslots * sizeof(char *));
292 8264 : shared->page_status = (SlruPageStatus *) (ptr + offset);
293 8264 : offset += MAXALIGN(nslots * sizeof(SlruPageStatus));
294 8264 : shared->page_dirty = (bool *) (ptr + offset);
295 8264 : offset += MAXALIGN(nslots * sizeof(bool));
296 8264 : shared->page_number = (int64 *) (ptr + offset);
297 8264 : offset += MAXALIGN(nslots * sizeof(int64));
298 8264 : shared->page_lru_count = (int *) (ptr + offset);
299 8264 : offset += MAXALIGN(nslots * sizeof(int));
300 :
301 : /* Initialize LWLocks */
302 8264 : shared->buffer_locks = (LWLockPadded *) (ptr + offset);
303 8264 : offset += MAXALIGN(nslots * sizeof(LWLockPadded));
304 8264 : shared->bank_locks = (LWLockPadded *) (ptr + offset);
305 8264 : offset += MAXALIGN(nbanks * sizeof(LWLockPadded));
306 8264 : shared->bank_cur_lru_count = (int *) (ptr + offset);
307 8264 : offset += MAXALIGN(nbanks * sizeof(int));
308 :
309 8264 : if (nlsns > 0)
310 : {
311 1180 : shared->group_lsn = (XLogRecPtr *) (ptr + offset);
312 1180 : offset += MAXALIGN(nslots * nlsns * sizeof(XLogRecPtr));
313 : }
314 :
315 8264 : ptr += BUFFERALIGN(offset);
316 210904 : for (int slotno = 0; slotno < nslots; slotno++)
317 : {
318 202640 : LWLockInitialize(&shared->buffer_locks[slotno].lock,
319 : buffer_tranche_id);
320 :
321 202640 : shared->page_buffer[slotno] = ptr;
322 202640 : shared->page_status[slotno] = SLRU_PAGE_EMPTY;
323 202640 : shared->page_dirty[slotno] = false;
324 202640 : shared->page_lru_count[slotno] = 0;
325 202640 : ptr += BLCKSZ;
326 : }
327 :
328 : /* Initialize the slot banks. */
329 20929 : for (int bankno = 0; bankno < nbanks; bankno++)
330 : {
331 12665 : LWLockInitialize(&shared->bank_locks[bankno].lock, bank_tranche_id);
332 12665 : shared->bank_cur_lru_count[bankno] = 0;
333 : }
334 :
335 : /* Should fit to estimated shmem size */
336 : Assert(ptr - (char *) shared <= SimpleLruShmemSize(nslots, nlsns));
337 : }
338 : else
339 : {
340 : Assert(found);
341 : Assert(shared->num_slots == nslots);
342 : }
343 :
344 : /*
345 : * Initialize the unshared control struct, including directory path. We
346 : * assume caller set PagePrecedes.
347 : */
348 8264 : ctl->shared = shared;
349 8264 : ctl->sync_handler = sync_handler;
350 8264 : ctl->long_segment_names = long_segment_names;
351 8264 : ctl->nbanks = nbanks;
352 8264 : strlcpy(ctl->Dir, subdir, sizeof(ctl->Dir));
353 8264 : }
354 :
355 : /*
356 : * Helper function for GUC check_hook to check whether slru buffers are in
357 : * multiples of SLRU_BANK_SIZE.
358 : */
359 : bool
360 12098 : check_slru_buffers(const char *name, int *newval)
361 : {
362 : /* Valid values are multiples of SLRU_BANK_SIZE */
363 12098 : if (*newval % SLRU_BANK_SIZE == 0)
364 12098 : return true;
365 :
366 0 : GUC_check_errdetail("\"%s\" must be a multiple of %d.", name,
367 : SLRU_BANK_SIZE);
368 0 : return false;
369 : }
370 :
371 : /*
372 : * Initialize (or reinitialize) a page to zeroes.
373 : *
374 : * The page is not actually written, just set up in shared memory.
375 : * The slot number of the new page is returned.
376 : *
377 : * Bank lock must be held at entry, and will be held at exit.
378 : */
379 : int
380 7345894 : SimpleLruZeroPage(SlruCtl ctl, int64 pageno)
381 : {
382 7345894 : SlruShared shared = ctl->shared;
383 : int slotno;
384 :
385 : Assert(LWLockHeldByMeInMode(SimpleLruGetBankLock(ctl, pageno), LW_EXCLUSIVE));
386 :
387 : /* Find a suitable buffer slot for the page */
388 7345894 : slotno = SlruSelectLRUPage(ctl, pageno);
389 : Assert(shared->page_status[slotno] == SLRU_PAGE_EMPTY ||
390 : (shared->page_status[slotno] == SLRU_PAGE_VALID &&
391 : !shared->page_dirty[slotno]) ||
392 : shared->page_number[slotno] == pageno);
393 :
394 : /* Mark the slot as containing this page */
395 7345894 : shared->page_number[slotno] = pageno;
396 7345894 : shared->page_status[slotno] = SLRU_PAGE_VALID;
397 7345894 : shared->page_dirty[slotno] = true;
398 7345894 : SlruRecentlyUsed(shared, slotno);
399 :
400 : /* Set the buffer to zeroes */
401 7345894 : MemSet(shared->page_buffer[slotno], 0, BLCKSZ);
402 :
403 : /* Set the LSNs for this new page to zero */
404 7345894 : SimpleLruZeroLSNs(ctl, slotno);
405 :
406 : /*
407 : * Assume this page is now the latest active page.
408 : *
409 : * Note that because both this routine and SlruSelectLRUPage run with a
410 : * SLRU bank lock held, it is not possible for this to be zeroing a page
411 : * that SlruSelectLRUPage is going to evict simultaneously. Therefore,
412 : * there's no memory barrier here.
413 : */
414 7345894 : pg_atomic_write_u64(&shared->latest_page_number, pageno);
415 :
416 : /* update the stats counter of zeroed pages */
417 7345894 : pgstat_count_slru_blocks_zeroed(shared->slru_stats_idx);
418 :
419 7345894 : return slotno;
420 : }
421 :
422 : /*
423 : * Zero all the LSNs we store for this slru page.
424 : *
425 : * This should be called each time we create a new page, and each time we read
426 : * in a page from disk into an existing buffer. (Such an old page cannot
427 : * have any interesting LSNs, since we'd have flushed them before writing
428 : * the page in the first place.)
429 : *
430 : * This assumes that InvalidXLogRecPtr is bitwise-all-0.
431 : */
432 : static void
433 7363490 : SimpleLruZeroLSNs(SlruCtl ctl, int slotno)
434 : {
435 7363490 : SlruShared shared = ctl->shared;
436 :
437 7363490 : if (shared->lsn_groups_per_page > 0)
438 433260 : MemSet(&shared->group_lsn[slotno * shared->lsn_groups_per_page], 0,
439 : shared->lsn_groups_per_page * sizeof(XLogRecPtr));
440 7363490 : }
441 :
442 : /*
443 : * This is a convenience wrapper for the common case of zeroing a page and
444 : * immediately flushing it to disk.
445 : *
446 : * SLRU bank lock is acquired and released here.
447 : */
448 : void
449 218 : SimpleLruZeroAndWritePage(SlruCtl ctl, int64 pageno)
450 : {
451 : int slotno;
452 : LWLock *lock;
453 :
454 218 : lock = SimpleLruGetBankLock(ctl, pageno);
455 218 : LWLockAcquire(lock, LW_EXCLUSIVE);
456 :
457 : /* Create and zero the page */
458 218 : slotno = SimpleLruZeroPage(ctl, pageno);
459 :
460 : /* Make sure it's written out */
461 218 : SimpleLruWritePage(ctl, slotno);
462 : Assert(!ctl->shared->page_dirty[slotno]);
463 :
464 218 : LWLockRelease(lock);
465 218 : }
466 :
467 : /*
468 : * Wait for any active I/O on a page slot to finish. (This does not
469 : * guarantee that new I/O hasn't been started before we return, though.
470 : * In fact the slot might not even contain the same page anymore.)
471 : *
472 : * Bank lock must be held at entry, and will be held at exit.
473 : */
474 : static void
475 2 : SimpleLruWaitIO(SlruCtl ctl, int slotno)
476 : {
477 2 : SlruShared shared = ctl->shared;
478 2 : int bankno = SlotGetBankNumber(slotno);
479 :
480 : Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
481 :
482 : /* See notes at top of file */
483 2 : LWLockRelease(&shared->bank_locks[bankno].lock);
484 2 : LWLockAcquire(&shared->buffer_locks[slotno].lock, LW_SHARED);
485 2 : LWLockRelease(&shared->buffer_locks[slotno].lock);
486 2 : LWLockAcquire(&shared->bank_locks[bankno].lock, LW_EXCLUSIVE);
487 :
488 : /*
489 : * If the slot is still in an io-in-progress state, then either someone
490 : * already started a new I/O on the slot, or a previous I/O failed and
491 : * neglected to reset the page state. That shouldn't happen, really, but
492 : * it seems worth a few extra cycles to check and recover from it. We can
493 : * cheaply test for failure by seeing if the buffer lock is still held (we
494 : * assume that transaction abort would release the lock).
495 : */
496 2 : if (shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS ||
497 2 : shared->page_status[slotno] == SLRU_PAGE_WRITE_IN_PROGRESS)
498 : {
499 0 : if (LWLockConditionalAcquire(&shared->buffer_locks[slotno].lock, LW_SHARED))
500 : {
501 : /* indeed, the I/O must have failed */
502 0 : if (shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS)
503 0 : shared->page_status[slotno] = SLRU_PAGE_EMPTY;
504 : else /* write_in_progress */
505 : {
506 0 : shared->page_status[slotno] = SLRU_PAGE_VALID;
507 0 : shared->page_dirty[slotno] = true;
508 : }
509 0 : LWLockRelease(&shared->buffer_locks[slotno].lock);
510 : }
511 : }
512 2 : }
513 :
514 : /*
515 : * Find a page in a shared buffer, reading it in if necessary.
516 : * The page number must correspond to an already-initialized page.
517 : *
518 : * If write_ok is true then it is OK to return a page that is in
519 : * WRITE_IN_PROGRESS state; it is the caller's responsibility to be sure
520 : * that modification of the page is safe. If write_ok is false then we
521 : * will not return the page until it is not undergoing active I/O.
522 : *
523 : * On error, the passed-in 'opaque_data' is passed to the
524 : * 'errdetail_for_io_error' callback, to provide details on the operation that
525 : * failed. It is only used for error reporting.
526 : *
527 : * Return value is the shared-buffer slot number now holding the page.
528 : * The buffer's LRU access info is updated.
529 : *
530 : * The correct bank lock must be held at entry, and will be held at exit.
531 : */
532 : int
533 388906 : SimpleLruReadPage(SlruCtl ctl, int64 pageno, bool write_ok,
534 : const void *opaque_data)
535 : {
536 388906 : SlruShared shared = ctl->shared;
537 388906 : LWLock *banklock = SimpleLruGetBankLock(ctl, pageno);
538 :
539 : Assert(LWLockHeldByMeInMode(banklock, LW_EXCLUSIVE));
540 :
541 : /* Outer loop handles restart if we must wait for someone else's I/O */
542 : for (;;)
543 2 : {
544 : int slotno;
545 : bool ok;
546 :
547 : /* See if page already is in memory; if not, pick victim slot */
548 388908 : slotno = SlruSelectLRUPage(ctl, pageno);
549 :
550 : /* Did we find the page in memory? */
551 388908 : if (shared->page_status[slotno] != SLRU_PAGE_EMPTY &&
552 387816 : shared->page_number[slotno] == pageno)
553 : {
554 : /*
555 : * If page is still being read in, we must wait for I/O. Likewise
556 : * if the page is being written and the caller said that's not OK.
557 : */
558 371312 : if (shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS ||
559 371312 : (shared->page_status[slotno] == SLRU_PAGE_WRITE_IN_PROGRESS &&
560 3 : !write_ok))
561 : {
562 2 : SimpleLruWaitIO(ctl, slotno);
563 : /* Now we must recheck state from the top */
564 2 : continue;
565 : }
566 : /* Otherwise, it's ready to use */
567 371310 : SlruRecentlyUsed(shared, slotno);
568 :
569 : /* update the stats counter of pages found in the SLRU */
570 371310 : pgstat_count_slru_blocks_hit(shared->slru_stats_idx);
571 :
572 371310 : return slotno;
573 : }
574 :
575 : /* We found no match; assert we selected a freeable slot */
576 : Assert(shared->page_status[slotno] == SLRU_PAGE_EMPTY ||
577 : (shared->page_status[slotno] == SLRU_PAGE_VALID &&
578 : !shared->page_dirty[slotno]));
579 :
580 : /* Mark the slot read-busy */
581 17596 : shared->page_number[slotno] = pageno;
582 17596 : shared->page_status[slotno] = SLRU_PAGE_READ_IN_PROGRESS;
583 17596 : shared->page_dirty[slotno] = false;
584 :
585 : /* Acquire per-buffer lock (cannot deadlock, see notes at top) */
586 17596 : LWLockAcquire(&shared->buffer_locks[slotno].lock, LW_EXCLUSIVE);
587 :
588 : /* Release bank lock while doing I/O */
589 17596 : LWLockRelease(banklock);
590 :
591 : /* Do the read */
592 17596 : ok = SlruPhysicalReadPage(ctl, pageno, slotno);
593 :
594 : /* Set the LSNs for this newly read-in page to zero */
595 17596 : SimpleLruZeroLSNs(ctl, slotno);
596 :
597 : /* Re-acquire bank control lock and update page state */
598 17596 : LWLockAcquire(banklock, LW_EXCLUSIVE);
599 :
600 : Assert(shared->page_number[slotno] == pageno &&
601 : shared->page_status[slotno] == SLRU_PAGE_READ_IN_PROGRESS &&
602 : !shared->page_dirty[slotno]);
603 :
604 17596 : shared->page_status[slotno] = ok ? SLRU_PAGE_VALID : SLRU_PAGE_EMPTY;
605 :
606 17596 : LWLockRelease(&shared->buffer_locks[slotno].lock);
607 :
608 : /* Now it's okay to ereport if we failed */
609 17596 : if (!ok)
610 1 : SlruReportIOError(ctl, pageno, opaque_data);
611 :
612 17595 : SlruRecentlyUsed(shared, slotno);
613 :
614 : /* update the stats counter of pages not found in SLRU */
615 17595 : pgstat_count_slru_blocks_read(shared->slru_stats_idx);
616 :
617 17595 : return slotno;
618 : }
619 : }
620 :
621 : /*
622 : * Find a page in a shared buffer, reading it in if necessary.
623 : * The page number must correspond to an already-initialized page.
624 : * The caller must intend only read-only access to the page.
625 : *
626 : * On error, the passed-in 'opaque_data' is passed to the
627 : * 'errdetail_for_io_error' callback, to provide details on the operation that
628 : * failed. It is only used for error reporting.
629 : *
630 : * Return value is the shared-buffer slot number now holding the page.
631 : * The buffer's LRU access info is updated.
632 : *
633 : * Bank control lock must NOT be held at entry, but will be held at exit.
634 : * It is unspecified whether the lock will be shared or exclusive.
635 : */
636 : int
637 869510 : SimpleLruReadPage_ReadOnly(SlruCtl ctl, int64 pageno, const void *opaque_data)
638 : {
639 869510 : SlruShared shared = ctl->shared;
640 869510 : LWLock *banklock = SimpleLruGetBankLock(ctl, pageno);
641 869510 : int bankno = pageno % ctl->nbanks;
642 869510 : int bankstart = bankno * SLRU_BANK_SIZE;
643 869510 : int bankend = bankstart + SLRU_BANK_SIZE;
644 :
645 : /* Try to find the page while holding only shared lock */
646 869510 : LWLockAcquire(banklock, LW_SHARED);
647 :
648 : /* See if page is already in a buffer */
649 876218 : for (int slotno = bankstart; slotno < bankend; slotno++)
650 : {
651 875996 : if (shared->page_status[slotno] != SLRU_PAGE_EMPTY &&
652 875022 : shared->page_number[slotno] == pageno &&
653 869288 : shared->page_status[slotno] != SLRU_PAGE_READ_IN_PROGRESS)
654 : {
655 : /* See comments for SlruRecentlyUsed() */
656 869288 : SlruRecentlyUsed(shared, slotno);
657 :
658 : /* update the stats counter of pages found in the SLRU */
659 869288 : pgstat_count_slru_blocks_hit(shared->slru_stats_idx);
660 :
661 869288 : return slotno;
662 : }
663 : }
664 :
665 : /* No luck, so switch to normal exclusive lock and do regular read */
666 222 : LWLockRelease(banklock);
667 222 : LWLockAcquire(banklock, LW_EXCLUSIVE);
668 :
669 222 : return SimpleLruReadPage(ctl, pageno, true, opaque_data);
670 : }
671 :
672 : /*
673 : * Write a page from a shared buffer, if necessary.
674 : * Does nothing if the specified slot is not dirty.
675 : *
676 : * NOTE: only one write attempt is made here. Hence, it is possible that
677 : * the page is still dirty at exit (if someone else re-dirtied it during
678 : * the write). However, we *do* attempt a fresh write even if the page
679 : * is already being written; this is for checkpoints.
680 : *
681 : * Bank lock must be held at entry, and will be held at exit.
682 : */
683 : static void
684 7349916 : SlruInternalWritePage(SlruCtl ctl, int slotno, SlruWriteAll fdata)
685 : {
686 7349916 : SlruShared shared = ctl->shared;
687 7349916 : int64 pageno = shared->page_number[slotno];
688 7349916 : int bankno = SlotGetBankNumber(slotno);
689 : bool ok;
690 :
691 : Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
692 : Assert(LWLockHeldByMeInMode(SimpleLruGetBankLock(ctl, pageno), LW_EXCLUSIVE));
693 :
694 : /* If a write is in progress, wait for it to finish */
695 7349916 : while (shared->page_status[slotno] == SLRU_PAGE_WRITE_IN_PROGRESS &&
696 0 : shared->page_number[slotno] == pageno)
697 : {
698 0 : SimpleLruWaitIO(ctl, slotno);
699 : }
700 :
701 : /*
702 : * Do nothing if page is not dirty, or if buffer no longer contains the
703 : * same page we were called for.
704 : */
705 7349916 : if (!shared->page_dirty[slotno] ||
706 7346594 : shared->page_status[slotno] != SLRU_PAGE_VALID ||
707 7346594 : shared->page_number[slotno] != pageno)
708 3322 : return;
709 :
710 : /*
711 : * Mark the slot write-busy, and clear the dirtybit. After this point, a
712 : * transaction status update on this page will mark it dirty again.
713 : */
714 7346594 : shared->page_status[slotno] = SLRU_PAGE_WRITE_IN_PROGRESS;
715 7346594 : shared->page_dirty[slotno] = false;
716 :
717 : /* Acquire per-buffer lock (cannot deadlock, see notes at top) */
718 7346594 : LWLockAcquire(&shared->buffer_locks[slotno].lock, LW_EXCLUSIVE);
719 :
720 : /* Release bank lock while doing I/O */
721 7346594 : LWLockRelease(&shared->bank_locks[bankno].lock);
722 :
723 : /* Do the write */
724 7346594 : ok = SlruPhysicalWritePage(ctl, pageno, slotno, fdata);
725 :
726 : /* If we failed, and we're in a flush, better close the files */
727 7346594 : if (!ok && fdata)
728 : {
729 0 : for (int i = 0; i < fdata->num_files; i++)
730 0 : CloseTransientFile(fdata->fd[i]);
731 : }
732 :
733 : /* Re-acquire bank lock and update page state */
734 7346594 : LWLockAcquire(&shared->bank_locks[bankno].lock, LW_EXCLUSIVE);
735 :
736 : Assert(shared->page_number[slotno] == pageno &&
737 : shared->page_status[slotno] == SLRU_PAGE_WRITE_IN_PROGRESS);
738 :
739 : /* If we failed to write, mark the page dirty again */
740 7346594 : if (!ok)
741 0 : shared->page_dirty[slotno] = true;
742 :
743 7346594 : shared->page_status[slotno] = SLRU_PAGE_VALID;
744 :
745 7346594 : LWLockRelease(&shared->buffer_locks[slotno].lock);
746 :
747 : /* Now it's okay to ereport if we failed */
748 7346594 : if (!ok)
749 0 : SlruReportIOError(ctl, pageno, NULL);
750 :
751 : /* If part of a checkpoint, count this as a SLRU buffer written. */
752 7346594 : if (fdata)
753 : {
754 3004 : CheckpointStats.ckpt_slru_written++;
755 3004 : PendingCheckpointerStats.slru_written++;
756 : }
757 : }
758 :
759 : /*
760 : * Wrapper of SlruInternalWritePage, for external callers.
761 : * fdata is always passed a NULL here.
762 : */
763 : void
764 316 : SimpleLruWritePage(SlruCtl ctl, int slotno)
765 : {
766 : Assert(ctl->shared->page_status[slotno] != SLRU_PAGE_EMPTY);
767 :
768 316 : SlruInternalWritePage(ctl, slotno, NULL);
769 316 : }
770 :
771 : /*
772 : * Return whether the given page exists on disk.
773 : *
774 : * A false return means that either the file does not exist, or that it's not
775 : * large enough to contain the given page.
776 : */
777 : bool
778 66 : SimpleLruDoesPhysicalPageExist(SlruCtl ctl, int64 pageno)
779 : {
780 66 : int64 segno = pageno / SLRU_PAGES_PER_SEGMENT;
781 66 : int rpageno = pageno % SLRU_PAGES_PER_SEGMENT;
782 66 : int offset = rpageno * BLCKSZ;
783 : char path[MAXPGPATH];
784 : int fd;
785 : bool result;
786 : off_t endpos;
787 :
788 : /* update the stats counter of checked pages */
789 66 : pgstat_count_slru_blocks_exists(ctl->shared->slru_stats_idx);
790 :
791 66 : SlruFileName(ctl, path, segno);
792 :
793 66 : fd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
794 66 : if (fd < 0)
795 : {
796 : /* expected: file doesn't exist */
797 26 : if (errno == ENOENT)
798 26 : return false;
799 :
800 : /* report error normally */
801 0 : slru_errcause = SLRU_OPEN_FAILED;
802 0 : slru_errno = errno;
803 0 : SlruReportIOError(ctl, pageno, NULL);
804 : }
805 :
806 40 : if ((endpos = lseek(fd, 0, SEEK_END)) < 0)
807 : {
808 0 : slru_errcause = SLRU_SEEK_FAILED;
809 0 : slru_errno = errno;
810 0 : SlruReportIOError(ctl, pageno, NULL);
811 : }
812 :
813 40 : result = endpos >= (off_t) (offset + BLCKSZ);
814 :
815 40 : if (CloseTransientFile(fd) != 0)
816 : {
817 0 : slru_errcause = SLRU_CLOSE_FAILED;
818 0 : slru_errno = errno;
819 0 : return false;
820 : }
821 :
822 40 : return result;
823 : }
824 :
825 : /*
826 : * Physical read of a (previously existing) page into a buffer slot
827 : *
828 : * On failure, we cannot just ereport(ERROR) since caller has put state in
829 : * shared memory that must be undone. So, we return false and save enough
830 : * info in static variables to let SlruReportIOError make the report.
831 : *
832 : * For now, assume it's not worth keeping a file pointer open across
833 : * read/write operations. We could cache one virtual file pointer ...
834 : */
835 : static bool
836 17596 : SlruPhysicalReadPage(SlruCtl ctl, int64 pageno, int slotno)
837 : {
838 17596 : SlruShared shared = ctl->shared;
839 17596 : int64 segno = pageno / SLRU_PAGES_PER_SEGMENT;
840 17596 : int rpageno = pageno % SLRU_PAGES_PER_SEGMENT;
841 17596 : off_t offset = rpageno * BLCKSZ;
842 : char path[MAXPGPATH];
843 : int fd;
844 :
845 17596 : SlruFileName(ctl, path, segno);
846 :
847 : /*
848 : * In a crash-and-restart situation, it's possible for us to receive
849 : * commands to set the commit status of transactions whose bits are in
850 : * already-truncated segments of the commit log (see notes in
851 : * SlruPhysicalWritePage). Hence, if we are InRecovery, allow the case
852 : * where the file doesn't exist, and return zeroes instead.
853 : */
854 17596 : fd = OpenTransientFile(path, O_RDONLY | PG_BINARY);
855 17596 : if (fd < 0)
856 : {
857 1 : if (errno != ENOENT || !InRecovery)
858 : {
859 1 : slru_errcause = SLRU_OPEN_FAILED;
860 1 : slru_errno = errno;
861 1 : return false;
862 : }
863 :
864 0 : ereport(LOG,
865 : (errmsg("file \"%s\" doesn't exist, reading as zeroes",
866 : path)));
867 0 : MemSet(shared->page_buffer[slotno], 0, BLCKSZ);
868 0 : return true;
869 : }
870 :
871 17595 : errno = 0;
872 17595 : pgstat_report_wait_start(WAIT_EVENT_SLRU_READ);
873 17595 : if (pg_pread(fd, shared->page_buffer[slotno], BLCKSZ, offset) != BLCKSZ)
874 : {
875 0 : pgstat_report_wait_end();
876 0 : slru_errcause = SLRU_READ_FAILED;
877 0 : slru_errno = errno;
878 0 : CloseTransientFile(fd);
879 0 : return false;
880 : }
881 17595 : pgstat_report_wait_end();
882 :
883 17595 : if (CloseTransientFile(fd) != 0)
884 : {
885 0 : slru_errcause = SLRU_CLOSE_FAILED;
886 0 : slru_errno = errno;
887 0 : return false;
888 : }
889 :
890 17595 : return true;
891 : }
892 :
893 : /*
894 : * Physical write of a page from a buffer slot
895 : *
896 : * On failure, we cannot just ereport(ERROR) since caller has put state in
897 : * shared memory that must be undone. So, we return false and save enough
898 : * info in static variables to let SlruReportIOError make the report.
899 : *
900 : * For now, assume it's not worth keeping a file pointer open across
901 : * independent read/write operations. We do batch operations during
902 : * SimpleLruWriteAll, though.
903 : *
904 : * fdata is NULL for a standalone write, pointer to open-file info during
905 : * SimpleLruWriteAll.
906 : */
907 : static bool
908 7346594 : SlruPhysicalWritePage(SlruCtl ctl, int64 pageno, int slotno, SlruWriteAll fdata)
909 : {
910 7346594 : SlruShared shared = ctl->shared;
911 7346594 : int64 segno = pageno / SLRU_PAGES_PER_SEGMENT;
912 7346594 : int rpageno = pageno % SLRU_PAGES_PER_SEGMENT;
913 7346594 : off_t offset = rpageno * BLCKSZ;
914 : char path[MAXPGPATH];
915 7346594 : int fd = -1;
916 :
917 : /* update the stats counter of written pages */
918 7346594 : pgstat_count_slru_blocks_written(shared->slru_stats_idx);
919 :
920 : /*
921 : * Honor the write-WAL-before-data rule, if appropriate, so that we do not
922 : * write out data before associated WAL records. This is the same action
923 : * performed during FlushBuffer() in the main buffer manager.
924 : */
925 7346594 : if (shared->group_lsn != NULL)
926 : {
927 : /*
928 : * We must determine the largest async-commit LSN for the page. This
929 : * is a bit tedious, but since this entire function is a slow path
930 : * anyway, it seems better to do this here than to maintain a per-page
931 : * LSN variable (which'd need an extra comparison in the
932 : * transaction-commit path).
933 : */
934 : XLogRecPtr max_lsn;
935 : int lsnindex;
936 :
937 433377 : lsnindex = slotno * shared->lsn_groups_per_page;
938 433377 : max_lsn = shared->group_lsn[lsnindex++];
939 443778048 : for (int lsnoff = 1; lsnoff < shared->lsn_groups_per_page; lsnoff++)
940 : {
941 443344671 : XLogRecPtr this_lsn = shared->group_lsn[lsnindex++];
942 :
943 443344671 : if (max_lsn < this_lsn)
944 50772 : max_lsn = this_lsn;
945 : }
946 :
947 433377 : if (XLogRecPtrIsValid(max_lsn))
948 : {
949 : /*
950 : * As noted above, elog(ERROR) is not acceptable here, so if
951 : * XLogFlush were to fail, we must PANIC. This isn't much of a
952 : * restriction because XLogFlush is just about all critical
953 : * section anyway, but let's make sure.
954 : */
955 546 : START_CRIT_SECTION();
956 546 : XLogFlush(max_lsn);
957 546 : END_CRIT_SECTION();
958 : }
959 : }
960 :
961 : /*
962 : * During a SimpleLruWriteAll, we may already have the desired file open.
963 : */
964 7346594 : if (fdata)
965 : {
966 3090 : for (int i = 0; i < fdata->num_files; i++)
967 : {
968 325 : if (fdata->segno[i] == segno)
969 : {
970 239 : fd = fdata->fd[i];
971 239 : break;
972 : }
973 : }
974 : }
975 :
976 7346594 : if (fd < 0)
977 : {
978 : /*
979 : * If the file doesn't already exist, we should create it. It is
980 : * possible for this to need to happen when writing a page that's not
981 : * first in its segment; we assume the OS can cope with that. (Note:
982 : * it might seem that it'd be okay to create files only when
983 : * SimpleLruZeroPage is called for the first page of a segment.
984 : * However, if after a crash and restart the REDO logic elects to
985 : * replay the log from a checkpoint before the latest one, then it's
986 : * possible that we will get commands to set transaction status of
987 : * transactions that have already been truncated from the commit log.
988 : * Easiest way to deal with that is to accept references to
989 : * nonexistent files here and in SlruPhysicalReadPage.)
990 : *
991 : * Note: it is possible for more than one backend to be executing this
992 : * code simultaneously for different pages of the same file. Hence,
993 : * don't use O_EXCL or O_TRUNC or anything like that.
994 : */
995 7346355 : SlruFileName(ctl, path, segno);
996 7346355 : fd = OpenTransientFile(path, O_RDWR | O_CREAT | PG_BINARY);
997 7346355 : if (fd < 0)
998 : {
999 0 : slru_errcause = SLRU_OPEN_FAILED;
1000 0 : slru_errno = errno;
1001 0 : return false;
1002 : }
1003 :
1004 7346355 : if (fdata)
1005 : {
1006 2765 : if (fdata->num_files < MAX_WRITEALL_BUFFERS)
1007 : {
1008 2765 : fdata->fd[fdata->num_files] = fd;
1009 2765 : fdata->segno[fdata->num_files] = segno;
1010 2765 : fdata->num_files++;
1011 : }
1012 : else
1013 : {
1014 : /*
1015 : * In the unlikely event that we exceed MAX_WRITEALL_BUFFERS,
1016 : * fall back to treating it as a standalone write.
1017 : */
1018 0 : fdata = NULL;
1019 : }
1020 : }
1021 : }
1022 :
1023 7346594 : errno = 0;
1024 7346594 : pgstat_report_wait_start(WAIT_EVENT_SLRU_WRITE);
1025 7346594 : if (pg_pwrite(fd, shared->page_buffer[slotno], BLCKSZ, offset) != BLCKSZ)
1026 : {
1027 0 : pgstat_report_wait_end();
1028 : /* if write didn't set errno, assume problem is no disk space */
1029 0 : if (errno == 0)
1030 0 : errno = ENOSPC;
1031 0 : slru_errcause = SLRU_WRITE_FAILED;
1032 0 : slru_errno = errno;
1033 0 : if (!fdata)
1034 0 : CloseTransientFile(fd);
1035 0 : return false;
1036 : }
1037 7346594 : pgstat_report_wait_end();
1038 :
1039 : /* Queue up a sync request for the checkpointer. */
1040 7346594 : if (ctl->sync_handler != SYNC_HANDLER_NONE)
1041 : {
1042 : FileTag tag;
1043 :
1044 434279 : INIT_SLRUFILETAG(tag, ctl->sync_handler, segno);
1045 434279 : if (!RegisterSyncRequest(&tag, SYNC_REQUEST, false))
1046 : {
1047 : /* No space to enqueue sync request. Do it synchronously. */
1048 6 : pgstat_report_wait_start(WAIT_EVENT_SLRU_SYNC);
1049 6 : if (pg_fsync(fd) != 0)
1050 : {
1051 0 : pgstat_report_wait_end();
1052 0 : slru_errcause = SLRU_FSYNC_FAILED;
1053 0 : slru_errno = errno;
1054 0 : CloseTransientFile(fd);
1055 0 : return false;
1056 : }
1057 6 : pgstat_report_wait_end();
1058 : }
1059 : }
1060 :
1061 : /* Close file, unless part of flush request. */
1062 7346594 : if (!fdata)
1063 : {
1064 7343590 : if (CloseTransientFile(fd) != 0)
1065 : {
1066 0 : slru_errcause = SLRU_CLOSE_FAILED;
1067 0 : slru_errno = errno;
1068 0 : return false;
1069 : }
1070 : }
1071 :
1072 7346594 : return true;
1073 : }
1074 :
1075 : /*
1076 : * Issue the error message after failure of SlruPhysicalReadPage or
1077 : * SlruPhysicalWritePage. Call this after cleaning up shared-memory state.
1078 : */
1079 : static void
1080 1 : SlruReportIOError(SlruCtl ctl, int64 pageno, const void *opaque_data)
1081 : {
1082 1 : int64 segno = pageno / SLRU_PAGES_PER_SEGMENT;
1083 1 : int rpageno = pageno % SLRU_PAGES_PER_SEGMENT;
1084 1 : int offset = rpageno * BLCKSZ;
1085 : char path[MAXPGPATH];
1086 :
1087 1 : SlruFileName(ctl, path, segno);
1088 1 : errno = slru_errno;
1089 1 : switch (slru_errcause)
1090 : {
1091 1 : case SLRU_OPEN_FAILED:
1092 1 : ereport(ERROR,
1093 : (errcode_for_file_access(),
1094 : errmsg("could not open file \"%s\": %m", path),
1095 : opaque_data ? ctl->errdetail_for_io_error(opaque_data) : 0));
1096 : break;
1097 0 : case SLRU_SEEK_FAILED:
1098 0 : ereport(ERROR,
1099 : (errcode_for_file_access(),
1100 : errmsg("could not seek in file \"%s\" to offset %d: %m",
1101 : path, offset),
1102 : opaque_data ? ctl->errdetail_for_io_error(opaque_data) : 0));
1103 : break;
1104 0 : case SLRU_READ_FAILED:
1105 0 : if (errno)
1106 0 : ereport(ERROR,
1107 : (errcode_for_file_access(),
1108 : errmsg("could not read from file \"%s\" at offset %d: %m",
1109 : path, offset),
1110 : opaque_data ? ctl->errdetail_for_io_error(opaque_data) : 0));
1111 : else
1112 0 : ereport(ERROR,
1113 : (errmsg("could not read from file \"%s\" at offset %d: read too few bytes",
1114 : path, offset),
1115 : opaque_data ? ctl->errdetail_for_io_error(opaque_data) : 0));
1116 : break;
1117 0 : case SLRU_WRITE_FAILED:
1118 0 : if (errno)
1119 0 : ereport(ERROR,
1120 : (errcode_for_file_access(),
1121 : errmsg("Could not write to file \"%s\" at offset %d: %m",
1122 : path, offset),
1123 : opaque_data ? ctl->errdetail_for_io_error(opaque_data) : 0));
1124 : else
1125 0 : ereport(ERROR,
1126 : (errmsg("Could not write to file \"%s\" at offset %d: wrote too few bytes.",
1127 : path, offset),
1128 : opaque_data ? ctl->errdetail_for_io_error(opaque_data) : 0));
1129 : break;
1130 0 : case SLRU_FSYNC_FAILED:
1131 0 : ereport(data_sync_elevel(ERROR),
1132 : (errcode_for_file_access(),
1133 : errmsg("could not fsync file \"%s\": %m",
1134 : path),
1135 : opaque_data ? ctl->errdetail_for_io_error(opaque_data) : 0));
1136 0 : break;
1137 0 : case SLRU_CLOSE_FAILED:
1138 0 : ereport(ERROR,
1139 : (errcode_for_file_access(),
1140 : errmsg("could not close file \"%s\": %m",
1141 : path),
1142 : opaque_data ? ctl->errdetail_for_io_error(opaque_data) : 0));
1143 : break;
1144 0 : default:
1145 : /* can't get here, we trust */
1146 0 : elog(ERROR, "unrecognized SimpleLru error cause: %d",
1147 : (int) slru_errcause);
1148 : break;
1149 : }
1150 0 : }
1151 :
1152 : /*
1153 : * Mark a buffer slot "most recently used".
1154 : */
1155 : static inline void
1156 8604087 : SlruRecentlyUsed(SlruShared shared, int slotno)
1157 : {
1158 8604087 : int bankno = SlotGetBankNumber(slotno);
1159 8604087 : int new_lru_count = shared->bank_cur_lru_count[bankno];
1160 :
1161 : Assert(shared->page_status[slotno] != SLRU_PAGE_EMPTY);
1162 :
1163 : /*
1164 : * The reason for the if-test is that there are often many consecutive
1165 : * accesses to the same page (particularly the latest page). By
1166 : * suppressing useless increments of bank_cur_lru_count, we reduce the
1167 : * probability that old pages' counts will "wrap around" and make them
1168 : * appear recently used.
1169 : *
1170 : * We allow this code to be executed concurrently by multiple processes
1171 : * within SimpleLruReadPage_ReadOnly(). As long as int reads and writes
1172 : * are atomic, this should not cause any completely-bogus values to enter
1173 : * the computation. However, it is possible for either bank_cur_lru_count
1174 : * or individual page_lru_count entries to be "reset" to lower values than
1175 : * they should have, in case a process is delayed while it executes this
1176 : * function. With care in SlruSelectLRUPage(), this does little harm, and
1177 : * in any case the absolute worst possible consequence is a nonoptimal
1178 : * choice of page to evict. The gain from allowing concurrent reads of
1179 : * SLRU pages seems worth it.
1180 : */
1181 8604087 : if (new_lru_count != shared->page_lru_count[slotno])
1182 : {
1183 7484084 : shared->bank_cur_lru_count[bankno] = ++new_lru_count;
1184 7484084 : shared->page_lru_count[slotno] = new_lru_count;
1185 : }
1186 8604087 : }
1187 :
1188 : /*
1189 : * Select the slot to re-use when we need a free slot for the given page.
1190 : *
1191 : * The target page number is passed not only because we need to know the
1192 : * correct bank to use, but also because we need to consider the possibility
1193 : * that some other process reads in the target page while we are doing I/O to
1194 : * free a slot. Hence, check or recheck to see if any slot already holds the
1195 : * target page, and return that slot if so. Thus, the returned slot is
1196 : * *either* a slot already holding the pageno (could be any state except
1197 : * EMPTY), *or* a freeable slot (state EMPTY or CLEAN).
1198 : *
1199 : * The correct bank lock must be held at entry, and will be held at exit.
1200 : */
1201 : static int
1202 7734802 : SlruSelectLRUPage(SlruCtl ctl, int64 pageno)
1203 : {
1204 7734802 : SlruShared shared = ctl->shared;
1205 :
1206 : /* Outer loop handles restart after I/O */
1207 : for (;;)
1208 7343191 : {
1209 : int cur_count;
1210 15077993 : int bestvalidslot = 0; /* keep compiler quiet */
1211 15077993 : int best_valid_delta = -1;
1212 15077993 : int64 best_valid_page_number = 0; /* keep compiler quiet */
1213 15077993 : int bestinvalidslot = 0; /* keep compiler quiet */
1214 15077993 : int best_invalid_delta = -1;
1215 15077993 : int64 best_invalid_page_number = 0; /* keep compiler quiet */
1216 15077993 : int bankno = pageno % ctl->nbanks;
1217 15077993 : int bankstart = bankno * SLRU_BANK_SIZE;
1218 15077993 : int bankend = bankstart + SLRU_BANK_SIZE;
1219 :
1220 : Assert(LWLockHeldByMe(SimpleLruGetBankLock(ctl, pageno)));
1221 :
1222 : /* See if page already has a buffer assigned */
1223 251020458 : for (int slotno = bankstart; slotno < bankend; slotno++)
1224 : {
1225 236313978 : if (shared->page_status[slotno] != SLRU_PAGE_EMPTY &&
1226 236259506 : shared->page_number[slotno] == pageno)
1227 371513 : return slotno;
1228 : }
1229 :
1230 : /*
1231 : * If we find any EMPTY slot, just select that one. Else choose a
1232 : * victim page to replace. We normally take the least recently used
1233 : * valid page, but we will never take the slot containing
1234 : * latest_page_number, even if it appears least recently used. We
1235 : * will select a slot that is already I/O busy only if there is no
1236 : * other choice: a read-busy slot will not be least recently used once
1237 : * the read finishes, and waiting for an I/O on a write-busy slot is
1238 : * inferior to just picking some other slot. Testing shows the slot
1239 : * we pick instead will often be clean, allowing us to begin a read at
1240 : * once.
1241 : *
1242 : * Normally the page_lru_count values will all be different and so
1243 : * there will be a well-defined LRU page. But since we allow
1244 : * concurrent execution of SlruRecentlyUsed() within
1245 : * SimpleLruReadPage_ReadOnly(), it is possible that multiple pages
1246 : * acquire the same lru_count values. In that case we break ties by
1247 : * choosing the furthest-back page.
1248 : *
1249 : * Notice that this next line forcibly advances cur_lru_count to a
1250 : * value that is certainly beyond any value that will be in the
1251 : * page_lru_count array after the loop finishes. This ensures that
1252 : * the next execution of SlruRecentlyUsed will mark the page newly
1253 : * used, even if it's for a page that has the current counter value.
1254 : * That gets us back on the path to having good data when there are
1255 : * multiple pages with the same lru_count.
1256 : */
1257 14706480 : cur_count = (shared->bank_cur_lru_count[bankno])++;
1258 249955757 : for (int slotno = bankstart; slotno < bankend; slotno++)
1259 : {
1260 : int this_delta;
1261 : int64 this_page_number;
1262 :
1263 235252869 : if (shared->page_status[slotno] == SLRU_PAGE_EMPTY)
1264 3592 : return slotno;
1265 :
1266 235249277 : this_delta = cur_count - shared->page_lru_count[slotno];
1267 235249277 : if (this_delta < 0)
1268 : {
1269 : /*
1270 : * Clean up in case shared updates have caused cur_count
1271 : * increments to get "lost". We back off the page counts,
1272 : * rather than trying to increase cur_count, to avoid any
1273 : * question of infinite loops or failure in the presence of
1274 : * wrapped-around counts.
1275 : */
1276 0 : shared->page_lru_count[slotno] = cur_count;
1277 0 : this_delta = 0;
1278 : }
1279 :
1280 : /*
1281 : * If this page is the one most recently zeroed, don't consider it
1282 : * an eviction candidate. See comments in SimpleLruZeroPage for an
1283 : * explanation about the lack of a memory barrier here.
1284 : */
1285 235249277 : this_page_number = shared->page_number[slotno];
1286 235249277 : if (this_page_number ==
1287 235249277 : pg_atomic_read_u64(&shared->latest_page_number))
1288 8911 : continue;
1289 :
1290 235240366 : if (shared->page_status[slotno] == SLRU_PAGE_VALID)
1291 : {
1292 235239910 : if (this_delta > best_valid_delta ||
1293 0 : (this_delta == best_valid_delta &&
1294 0 : ctl->PagePrecedes(this_page_number,
1295 : best_valid_page_number)))
1296 : {
1297 33198436 : bestvalidslot = slotno;
1298 33198436 : best_valid_delta = this_delta;
1299 33198436 : best_valid_page_number = this_page_number;
1300 : }
1301 : }
1302 : else
1303 : {
1304 456 : if (this_delta > best_invalid_delta ||
1305 0 : (this_delta == best_invalid_delta &&
1306 0 : ctl->PagePrecedes(this_page_number,
1307 : best_invalid_page_number)))
1308 : {
1309 456 : bestinvalidslot = slotno;
1310 456 : best_invalid_delta = this_delta;
1311 456 : best_invalid_page_number = this_page_number;
1312 : }
1313 : }
1314 : }
1315 :
1316 : /*
1317 : * If all pages (except possibly the latest one) are I/O busy, we'll
1318 : * have to wait for an I/O to complete and then retry. In that
1319 : * unhappy case, we choose to wait for the I/O on the least recently
1320 : * used slot, on the assumption that it was likely initiated first of
1321 : * all the I/Os in progress and may therefore finish first.
1322 : */
1323 14702888 : if (best_valid_delta < 0)
1324 : {
1325 0 : SimpleLruWaitIO(ctl, bestinvalidslot);
1326 0 : continue;
1327 : }
1328 :
1329 : /*
1330 : * If the selected page is clean, we're set.
1331 : */
1332 14702888 : if (!shared->page_dirty[bestvalidslot])
1333 7359697 : return bestvalidslot;
1334 :
1335 : /*
1336 : * Write the page.
1337 : */
1338 7343191 : SlruInternalWritePage(ctl, bestvalidslot, NULL);
1339 :
1340 : /*
1341 : * Now loop back and try again. This is the easiest way of dealing
1342 : * with corner cases such as the victim page being re-dirtied while we
1343 : * wrote it.
1344 : */
1345 : }
1346 : }
1347 :
1348 : /*
1349 : * Write dirty pages to disk during checkpoint or database shutdown. Flushing
1350 : * is deferred until the next call to ProcessSyncRequests(), though we do fsync
1351 : * the containing directory here to make sure that newly created directory
1352 : * entries are on disk.
1353 : */
1354 : void
1355 9227 : SimpleLruWriteAll(SlruCtl ctl, bool allow_redirtied)
1356 : {
1357 9227 : SlruShared shared = ctl->shared;
1358 : SlruWriteAllData fdata;
1359 9227 : int64 pageno = 0;
1360 9227 : int prevbank = SlotGetBankNumber(0);
1361 : bool ok;
1362 :
1363 : /* update the stats counter of flushes */
1364 9227 : pgstat_count_slru_flush(shared->slru_stats_idx);
1365 :
1366 : /*
1367 : * Find and write dirty pages
1368 : */
1369 9227 : fdata.num_files = 0;
1370 :
1371 9227 : LWLockAcquire(&shared->bank_locks[prevbank].lock, LW_EXCLUSIVE);
1372 :
1373 224251 : for (int slotno = 0; slotno < shared->num_slots; slotno++)
1374 : {
1375 215024 : int curbank = SlotGetBankNumber(slotno);
1376 :
1377 : /*
1378 : * If the current bank lock is not same as the previous bank lock then
1379 : * release the previous lock and acquire the new lock.
1380 : */
1381 215024 : if (curbank != prevbank)
1382 : {
1383 4212 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1384 4212 : LWLockAcquire(&shared->bank_locks[curbank].lock, LW_EXCLUSIVE);
1385 4212 : prevbank = curbank;
1386 : }
1387 :
1388 : /* Do nothing if slot is unused */
1389 215024 : if (shared->page_status[slotno] == SLRU_PAGE_EMPTY)
1390 208698 : continue;
1391 :
1392 6326 : SlruInternalWritePage(ctl, slotno, &fdata);
1393 :
1394 : /*
1395 : * In some places (e.g. checkpoints), we cannot assert that the slot
1396 : * is clean now, since another process might have re-dirtied it
1397 : * already. That's okay.
1398 : */
1399 : Assert(allow_redirtied ||
1400 : shared->page_status[slotno] == SLRU_PAGE_EMPTY ||
1401 : (shared->page_status[slotno] == SLRU_PAGE_VALID &&
1402 : !shared->page_dirty[slotno]));
1403 : }
1404 :
1405 9227 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1406 :
1407 : /*
1408 : * Now close any files that were open
1409 : */
1410 9227 : ok = true;
1411 11992 : for (int i = 0; i < fdata.num_files; i++)
1412 : {
1413 2765 : if (CloseTransientFile(fdata.fd[i]) != 0)
1414 : {
1415 0 : slru_errcause = SLRU_CLOSE_FAILED;
1416 0 : slru_errno = errno;
1417 0 : pageno = fdata.segno[i] * SLRU_PAGES_PER_SEGMENT;
1418 0 : ok = false;
1419 : }
1420 : }
1421 9227 : if (!ok)
1422 0 : SlruReportIOError(ctl, pageno, NULL);
1423 :
1424 : /* Ensure that directory entries for new files are on disk. */
1425 9227 : if (ctl->sync_handler != SYNC_HANDLER_NONE)
1426 7388 : fsync_fname(ctl->Dir, true);
1427 9227 : }
1428 :
1429 : /*
1430 : * Remove all segments before the one holding the passed page number
1431 : *
1432 : * All SLRUs prevent concurrent calls to this function, either with an LWLock
1433 : * or by calling it only as part of a checkpoint. Mutual exclusion must begin
1434 : * before computing cutoffPage. Mutual exclusion must end after any limit
1435 : * update that would permit other backends to write fresh data into the
1436 : * segment immediately preceding the one containing cutoffPage. Otherwise,
1437 : * when the SLRU is quite full, SimpleLruTruncate() might delete that segment
1438 : * after it has accrued freshly-written data.
1439 : */
1440 : void
1441 1911 : SimpleLruTruncate(SlruCtl ctl, int64 cutoffPage)
1442 : {
1443 1911 : SlruShared shared = ctl->shared;
1444 : int prevbank;
1445 :
1446 : /* update the stats counter of truncates */
1447 1911 : pgstat_count_slru_truncate(shared->slru_stats_idx);
1448 :
1449 : /*
1450 : * Scan shared memory and remove any pages preceding the cutoff page, to
1451 : * ensure we won't rewrite them later. (Since this is normally called in
1452 : * or just after a checkpoint, any dirty pages should have been flushed
1453 : * already ... we're just being extra careful here.)
1454 : */
1455 1994 : restart:
1456 :
1457 : /*
1458 : * An important safety check: the current endpoint page must not be
1459 : * eligible for removal. This check is just a backstop against wraparound
1460 : * bugs elsewhere in SLRU handling, so we don't care if we read a slightly
1461 : * outdated value; therefore we don't add a memory barrier.
1462 : */
1463 1994 : if (ctl->PagePrecedes(pg_atomic_read_u64(&shared->latest_page_number),
1464 : cutoffPage))
1465 : {
1466 0 : ereport(LOG,
1467 : (errmsg("could not truncate directory \"%s\": apparent wraparound",
1468 : ctl->Dir)));
1469 0 : return;
1470 : }
1471 :
1472 1994 : prevbank = SlotGetBankNumber(0);
1473 1994 : LWLockAcquire(&shared->bank_locks[prevbank].lock, LW_EXCLUSIVE);
1474 47831 : for (int slotno = 0; slotno < shared->num_slots; slotno++)
1475 : {
1476 45920 : int curbank = SlotGetBankNumber(slotno);
1477 :
1478 : /*
1479 : * If the current bank lock is not same as the previous bank lock then
1480 : * release the previous lock and acquire the new lock.
1481 : */
1482 45920 : if (curbank != prevbank)
1483 : {
1484 917 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1485 917 : LWLockAcquire(&shared->bank_locks[curbank].lock, LW_EXCLUSIVE);
1486 917 : prevbank = curbank;
1487 : }
1488 :
1489 45920 : if (shared->page_status[slotno] == SLRU_PAGE_EMPTY)
1490 40497 : continue;
1491 5423 : if (!ctl->PagePrecedes(shared->page_number[slotno], cutoffPage))
1492 5116 : continue;
1493 :
1494 : /*
1495 : * If page is clean, just change state to EMPTY (expected case).
1496 : */
1497 307 : if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
1498 307 : !shared->page_dirty[slotno])
1499 : {
1500 224 : shared->page_status[slotno] = SLRU_PAGE_EMPTY;
1501 224 : continue;
1502 : }
1503 :
1504 : /*
1505 : * Hmm, we have (or may have) I/O operations acting on the page, so
1506 : * we've got to wait for them to finish and then start again. This is
1507 : * the same logic as in SlruSelectLRUPage. (XXX if page is dirty,
1508 : * wouldn't it be OK to just discard it without writing it?
1509 : * SlruMayDeleteSegment() uses a stricter qualification, so we might
1510 : * not delete this page in the end; even if we don't delete it, we
1511 : * won't have cause to read its data again. For now, keep the logic
1512 : * the same as it was.)
1513 : */
1514 83 : if (shared->page_status[slotno] == SLRU_PAGE_VALID)
1515 83 : SlruInternalWritePage(ctl, slotno, NULL);
1516 : else
1517 0 : SimpleLruWaitIO(ctl, slotno);
1518 :
1519 83 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1520 83 : goto restart;
1521 : }
1522 :
1523 1911 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1524 :
1525 : /* Now we can remove the old segment(s) */
1526 1911 : (void) SlruScanDirectory(ctl, SlruScanDirCbDeleteCutoff, &cutoffPage);
1527 : }
1528 :
1529 : /*
1530 : * Delete an individual SLRU segment.
1531 : *
1532 : * NB: This does not touch the SLRU buffers themselves, callers have to ensure
1533 : * they either can't yet contain anything, or have already been cleaned out.
1534 : */
1535 : static void
1536 142003 : SlruInternalDeleteSegment(SlruCtl ctl, int64 segno)
1537 : {
1538 : char path[MAXPGPATH];
1539 :
1540 : /* Forget any fsync requests queued for this segment. */
1541 142003 : if (ctl->sync_handler != SYNC_HANDLER_NONE)
1542 : {
1543 : FileTag tag;
1544 :
1545 13224 : INIT_SLRUFILETAG(tag, ctl->sync_handler, segno);
1546 13224 : RegisterSyncRequest(&tag, SYNC_FORGET_REQUEST, true);
1547 : }
1548 :
1549 : /* Unlink the file. */
1550 142003 : SlruFileName(ctl, path, segno);
1551 142003 : ereport(DEBUG2, (errmsg_internal("removing file \"%s\"", path)));
1552 142003 : unlink(path);
1553 142003 : }
1554 :
1555 : /*
1556 : * Delete an individual SLRU segment, identified by the segment number.
1557 : */
1558 : void
1559 2 : SlruDeleteSegment(SlruCtl ctl, int64 segno)
1560 : {
1561 2 : SlruShared shared = ctl->shared;
1562 2 : int prevbank = SlotGetBankNumber(0);
1563 : bool did_write;
1564 :
1565 : /* Clean out any possibly existing references to the segment. */
1566 2 : LWLockAcquire(&shared->bank_locks[prevbank].lock, LW_EXCLUSIVE);
1567 2 : restart:
1568 2 : did_write = false;
1569 34 : for (int slotno = 0; slotno < shared->num_slots; slotno++)
1570 : {
1571 : int64 pagesegno;
1572 32 : int curbank = SlotGetBankNumber(slotno);
1573 :
1574 : /*
1575 : * If the current bank lock is not same as the previous bank lock then
1576 : * release the previous lock and acquire the new lock.
1577 : */
1578 32 : if (curbank != prevbank)
1579 : {
1580 0 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1581 0 : LWLockAcquire(&shared->bank_locks[curbank].lock, LW_EXCLUSIVE);
1582 0 : prevbank = curbank;
1583 : }
1584 :
1585 32 : if (shared->page_status[slotno] == SLRU_PAGE_EMPTY)
1586 0 : continue;
1587 :
1588 32 : pagesegno = shared->page_number[slotno] / SLRU_PAGES_PER_SEGMENT;
1589 : /* not the segment we're looking for */
1590 32 : if (pagesegno != segno)
1591 7 : continue;
1592 :
1593 : /* If page is clean, just change state to EMPTY (expected case). */
1594 25 : if (shared->page_status[slotno] == SLRU_PAGE_VALID &&
1595 25 : !shared->page_dirty[slotno])
1596 : {
1597 25 : shared->page_status[slotno] = SLRU_PAGE_EMPTY;
1598 25 : continue;
1599 : }
1600 :
1601 : /* Same logic as SimpleLruTruncate() */
1602 0 : if (shared->page_status[slotno] == SLRU_PAGE_VALID)
1603 0 : SlruInternalWritePage(ctl, slotno, NULL);
1604 : else
1605 0 : SimpleLruWaitIO(ctl, slotno);
1606 :
1607 0 : did_write = true;
1608 : }
1609 :
1610 : /*
1611 : * Be extra careful and re-check. The IO functions release the control
1612 : * lock, so new pages could have been read in.
1613 : */
1614 2 : if (did_write)
1615 0 : goto restart;
1616 :
1617 2 : SlruInternalDeleteSegment(ctl, segno);
1618 :
1619 2 : LWLockRelease(&shared->bank_locks[prevbank].lock);
1620 2 : }
1621 :
1622 : /*
1623 : * Determine whether a segment is okay to delete.
1624 : *
1625 : * segpage is the first page of the segment, and cutoffPage is the oldest (in
1626 : * PagePrecedes order) page in the SLRU containing still-useful data. Since
1627 : * every core PagePrecedes callback implements "wrap around", check the
1628 : * segment's first and last pages:
1629 : *
1630 : * first<cutoff && last<cutoff: yes
1631 : * first<cutoff && last>=cutoff: no; cutoff falls inside this segment
1632 : * first>=cutoff && last<cutoff: no; wrap point falls inside this segment
1633 : * first>=cutoff && last>=cutoff: no; every page of this segment is too young
1634 : */
1635 : static bool
1636 1044096 : SlruMayDeleteSegment(SlruCtl ctl, int64 segpage, int64 cutoffPage)
1637 : {
1638 1044096 : int64 seg_last_page = segpage + SLRU_PAGES_PER_SEGMENT - 1;
1639 :
1640 : Assert(segpage % SLRU_PAGES_PER_SEGMENT == 0);
1641 :
1642 1186803 : return (ctl->PagePrecedes(segpage, cutoffPage) &&
1643 142707 : ctl->PagePrecedes(seg_last_page, cutoffPage));
1644 : }
1645 :
1646 : #ifdef USE_ASSERT_CHECKING
1647 : static void
1648 : SlruPagePrecedesTestOffset(SlruCtl ctl, int per_page, uint32 offset)
1649 : {
1650 : TransactionId lhs,
1651 : rhs;
1652 : int64 newestPage,
1653 : oldestPage;
1654 : TransactionId newestXact,
1655 : oldestXact;
1656 :
1657 : /*
1658 : * Compare an XID pair having undefined order (see RFC 1982), a pair at
1659 : * "opposite ends" of the XID space. TransactionIdPrecedes() treats each
1660 : * as preceding the other. If RHS is oldestXact, LHS is the first XID we
1661 : * must not assign.
1662 : */
1663 : lhs = per_page + offset; /* skip first page to avoid non-normal XIDs */
1664 : rhs = lhs + (1U << 31);
1665 : Assert(TransactionIdPrecedes(lhs, rhs));
1666 : Assert(TransactionIdPrecedes(rhs, lhs));
1667 : Assert(!TransactionIdPrecedes(lhs - 1, rhs));
1668 : Assert(TransactionIdPrecedes(rhs, lhs - 1));
1669 : Assert(TransactionIdPrecedes(lhs + 1, rhs));
1670 : Assert(!TransactionIdPrecedes(rhs, lhs + 1));
1671 : Assert(!TransactionIdFollowsOrEquals(lhs, rhs));
1672 : Assert(!TransactionIdFollowsOrEquals(rhs, lhs));
1673 : Assert(!ctl->PagePrecedes(lhs / per_page, lhs / per_page));
1674 : Assert(!ctl->PagePrecedes(lhs / per_page, rhs / per_page));
1675 : Assert(!ctl->PagePrecedes(rhs / per_page, lhs / per_page));
1676 : Assert(!ctl->PagePrecedes((lhs - per_page) / per_page, rhs / per_page));
1677 : Assert(ctl->PagePrecedes(rhs / per_page, (lhs - 3 * per_page) / per_page));
1678 : Assert(ctl->PagePrecedes(rhs / per_page, (lhs - 2 * per_page) / per_page));
1679 : Assert(ctl->PagePrecedes(rhs / per_page, (lhs - 1 * per_page) / per_page)
1680 : || (1U << 31) % per_page != 0); /* See CommitTsPagePrecedes() */
1681 : Assert(ctl->PagePrecedes((lhs + 1 * per_page) / per_page, rhs / per_page)
1682 : || (1U << 31) % per_page != 0);
1683 : Assert(ctl->PagePrecedes((lhs + 2 * per_page) / per_page, rhs / per_page));
1684 : Assert(ctl->PagePrecedes((lhs + 3 * per_page) / per_page, rhs / per_page));
1685 : Assert(!ctl->PagePrecedes(rhs / per_page, (lhs + per_page) / per_page));
1686 :
1687 : /*
1688 : * GetNewTransactionId() has assigned the last XID it can safely use, and
1689 : * that XID is in the *LAST* page of the second segment. We must not
1690 : * delete that segment.
1691 : */
1692 : newestPage = 2 * SLRU_PAGES_PER_SEGMENT - 1;
1693 : newestXact = newestPage * per_page + offset;
1694 : Assert(newestXact / per_page == newestPage);
1695 : oldestXact = newestXact + 1;
1696 : oldestXact -= 1U << 31;
1697 : oldestPage = oldestXact / per_page;
1698 : Assert(!SlruMayDeleteSegment(ctl,
1699 : (newestPage -
1700 : newestPage % SLRU_PAGES_PER_SEGMENT),
1701 : oldestPage));
1702 :
1703 : /*
1704 : * GetNewTransactionId() has assigned the last XID it can safely use, and
1705 : * that XID is in the *FIRST* page of the second segment. We must not
1706 : * delete that segment.
1707 : */
1708 : newestPage = SLRU_PAGES_PER_SEGMENT;
1709 : newestXact = newestPage * per_page + offset;
1710 : Assert(newestXact / per_page == newestPage);
1711 : oldestXact = newestXact + 1;
1712 : oldestXact -= 1U << 31;
1713 : oldestPage = oldestXact / per_page;
1714 : Assert(!SlruMayDeleteSegment(ctl,
1715 : (newestPage -
1716 : newestPage % SLRU_PAGES_PER_SEGMENT),
1717 : oldestPage));
1718 : }
1719 :
1720 : /*
1721 : * Unit-test a PagePrecedes function.
1722 : *
1723 : * This assumes every uint32 >= FirstNormalTransactionId is a valid key. It
1724 : * assumes each value occupies a contiguous, fixed-size region of SLRU bytes.
1725 : * (MultiXactMemberCtl separates flags from XIDs. NotifyCtl has
1726 : * variable-length entries, no keys, and no random access. These unit tests
1727 : * do not apply to them.)
1728 : */
1729 : void
1730 : SlruPagePrecedesUnitTests(SlruCtl ctl, int per_page)
1731 : {
1732 : /* Test first, middle and last entries of a page. */
1733 : SlruPagePrecedesTestOffset(ctl, per_page, 0);
1734 : SlruPagePrecedesTestOffset(ctl, per_page, per_page / 2);
1735 : SlruPagePrecedesTestOffset(ctl, per_page, per_page - 1);
1736 : }
1737 : #endif
1738 :
1739 : /*
1740 : * SlruScanDirectory callback
1741 : * This callback reports true if there's any segment wholly prior to the
1742 : * one containing the page passed as "data".
1743 : */
1744 : bool
1745 834300 : SlruScanDirCbReportPresence(SlruCtl ctl, char *filename, int64 segpage,
1746 : void *data)
1747 : {
1748 834300 : int64 cutoffPage = *(int64 *) data;
1749 :
1750 834300 : if (SlruMayDeleteSegment(ctl, segpage, cutoffPage))
1751 101 : return true; /* found one; don't iterate any more */
1752 :
1753 834199 : return false; /* keep going */
1754 : }
1755 :
1756 : /*
1757 : * SlruScanDirectory callback.
1758 : * This callback deletes segments prior to the one passed in as "data".
1759 : */
1760 : static bool
1761 209796 : SlruScanDirCbDeleteCutoff(SlruCtl ctl, char *filename, int64 segpage,
1762 : void *data)
1763 : {
1764 209796 : int64 cutoffPage = *(int64 *) data;
1765 :
1766 209796 : if (SlruMayDeleteSegment(ctl, segpage, cutoffPage))
1767 141993 : SlruInternalDeleteSegment(ctl, segpage / SLRU_PAGES_PER_SEGMENT);
1768 :
1769 209796 : return false; /* keep going */
1770 : }
1771 :
1772 : /*
1773 : * SlruScanDirectory callback.
1774 : * This callback deletes all segments.
1775 : */
1776 : bool
1777 8 : SlruScanDirCbDeleteAll(SlruCtl ctl, char *filename, int64 segpage, void *data)
1778 : {
1779 8 : SlruInternalDeleteSegment(ctl, segpage / SLRU_PAGES_PER_SEGMENT);
1780 :
1781 8 : return false; /* keep going */
1782 : }
1783 :
1784 : /*
1785 : * An internal function used by SlruScanDirectory().
1786 : *
1787 : * Returns true if a file with a name of a given length may be a correct
1788 : * SLRU segment.
1789 : */
1790 : static inline bool
1791 1056318 : SlruCorrectSegmentFilenameLength(SlruCtl ctl, size_t len)
1792 : {
1793 1056318 : if (ctl->long_segment_names)
1794 2375 : return (len == 15); /* see SlruFileName() */
1795 : else
1796 :
1797 : /*
1798 : * Commit 638cf09e76d allowed 5-character lengths. Later commit
1799 : * 73c986adde5 allowed 6-character length.
1800 : *
1801 : * Note: There is an ongoing plan to migrate all SLRUs to 64-bit page
1802 : * numbers, and the corresponding 15-character file names, which may
1803 : * eventually deprecate the support for 4, 5, and 6-character names.
1804 : */
1805 1053943 : return (len == 4 || len == 5 || len == 6);
1806 : }
1807 :
1808 : /*
1809 : * Scan the SimpleLru directory and apply a callback to each file found in it.
1810 : *
1811 : * If the callback returns true, the scan is stopped. The last return value
1812 : * from the callback is returned.
1813 : *
1814 : * The callback receives the following arguments: 1. the SlruCtl struct for the
1815 : * slru being truncated; 2. the filename being considered; 3. the page number
1816 : * for the first page of that file; 4. a pointer to the opaque data given to us
1817 : * by the caller.
1818 : *
1819 : * Note that the ordering in which the directory is scanned is not guaranteed.
1820 : *
1821 : * Note that no locking is applied.
1822 : */
1823 : bool
1824 6174 : SlruScanDirectory(SlruCtl ctl, SlruScanCallback callback, void *data)
1825 : {
1826 6174 : bool retval = false;
1827 : DIR *cldir;
1828 : struct dirent *clde;
1829 : int64 segno;
1830 : int64 segpage;
1831 :
1832 6174 : cldir = AllocateDir(ctl->Dir);
1833 1062391 : while ((clde = ReadDir(cldir, ctl->Dir)) != NULL)
1834 : {
1835 : size_t len;
1836 :
1837 1056318 : len = strlen(clde->d_name);
1838 :
1839 1056318 : if (SlruCorrectSegmentFilenameLength(ctl, len) &&
1840 1044104 : strspn(clde->d_name, "0123456789ABCDEF") == len)
1841 : {
1842 1044104 : segno = strtoi64(clde->d_name, NULL, 16);
1843 1044104 : segpage = segno * SLRU_PAGES_PER_SEGMENT;
1844 :
1845 1044104 : elog(DEBUG2, "SlruScanDirectory invoking callback on %s/%s",
1846 : ctl->Dir, clde->d_name);
1847 1044104 : retval = callback(ctl, clde->d_name, segpage, data);
1848 1044104 : if (retval)
1849 101 : break;
1850 : }
1851 : }
1852 6174 : FreeDir(cldir);
1853 :
1854 6174 : return retval;
1855 : }
1856 :
1857 : /*
1858 : * Individual SLRUs (clog, ...) have to provide a sync.c handler function so
1859 : * that they can provide the correct "SlruCtl" (otherwise we don't know how to
1860 : * build the path), but they just forward to this common implementation that
1861 : * performs the fsync.
1862 : */
1863 : int
1864 2 : SlruSyncFileTag(SlruCtl ctl, const FileTag *ftag, char *path)
1865 : {
1866 : int fd;
1867 : int save_errno;
1868 : int result;
1869 :
1870 2 : SlruFileName(ctl, path, ftag->segno);
1871 :
1872 2 : fd = OpenTransientFile(path, O_RDWR | PG_BINARY);
1873 2 : if (fd < 0)
1874 0 : return -1;
1875 :
1876 2 : pgstat_report_wait_start(WAIT_EVENT_SLRU_FLUSH_SYNC);
1877 2 : result = pg_fsync(fd);
1878 2 : pgstat_report_wait_end();
1879 2 : save_errno = errno;
1880 :
1881 2 : CloseTransientFile(fd);
1882 :
1883 2 : errno = save_errno;
1884 2 : return result;
1885 : }
|
