Line data Source code
1 : /*-------------------------------------------------------------------------
2 : *
3 : * buffile.c
4 : * Management of large buffered temporary files.
5 : *
6 : * Portions Copyright (c) 1996-2026, PostgreSQL Global Development Group
7 : * Portions Copyright (c) 1994, Regents of the University of California
8 : *
9 : * IDENTIFICATION
10 : * src/backend/storage/file/buffile.c
11 : *
12 : * NOTES:
13 : *
14 : * BufFiles provide a very incomplete emulation of stdio atop virtual Files
15 : * (as managed by fd.c). Currently, we only support the buffered-I/O
16 : * aspect of stdio: a read or write of the low-level File occurs only
17 : * when the buffer is filled or emptied. This is an even bigger win
18 : * for virtual Files than for ordinary kernel files, since reducing the
19 : * frequency with which a virtual File is touched reduces "thrashing"
20 : * of opening/closing file descriptors.
21 : *
22 : * Note that BufFile structs are allocated with palloc(), and therefore
23 : * will go away automatically at query/transaction end. Since the underlying
24 : * virtual Files are made with OpenTemporaryFile, all resources for
25 : * the file are certain to be cleaned up even if processing is aborted
26 : * by ereport(ERROR). The data structures required are made in the
27 : * palloc context that was current when the BufFile was created, and
28 : * any external resources such as temp files are owned by the ResourceOwner
29 : * that was current at that time.
30 : *
31 : * BufFile also supports temporary files that exceed the OS file size limit
32 : * (by opening multiple fd.c temporary files). This is an essential feature
33 : * for sorts and hashjoins on large amounts of data.
34 : *
35 : * BufFile supports temporary files that can be shared with other backends, as
36 : * infrastructure for parallel execution. Such files need to be created as a
37 : * member of a SharedFileSet that all participants are attached to.
38 : *
39 : * BufFile also supports temporary files that can be used by the single backend
40 : * when the corresponding files need to be survived across the transaction and
41 : * need to be opened and closed multiple times. Such files need to be created
42 : * as a member of a FileSet.
43 : *-------------------------------------------------------------------------
44 : */
45 :
46 : #include "postgres.h"
47 :
48 : #include "commands/tablespace.h"
49 : #include "executor/instrument.h"
50 : #include "miscadmin.h"
51 : #include "pgstat.h"
52 : #include "storage/buffile.h"
53 : #include "storage/bufmgr.h"
54 : #include "storage/fd.h"
55 : #include "utils/resowner.h"
56 : #include "utils/wait_event.h"
57 :
58 : /*
59 : * We break BufFiles into gigabyte-sized segments, regardless of RELSEG_SIZE.
60 : * The reason is that we'd like large BufFiles to be spread across multiple
61 : * tablespaces when available.
62 : */
63 : #define MAX_PHYSICAL_FILESIZE 0x40000000
64 : #define BUFFILE_SEG_SIZE (MAX_PHYSICAL_FILESIZE / BLCKSZ)
65 :
66 : /*
67 : * This data structure represents a buffered file that consists of one or
68 : * more physical files (each accessed through a virtual file descriptor
69 : * managed by fd.c).
70 : */
71 : struct BufFile
72 : {
73 : int numFiles; /* number of physical files in set */
74 : /* all files except the last have length exactly MAX_PHYSICAL_FILESIZE */
75 : File *files; /* palloc'd array with numFiles entries */
76 :
77 : bool isInterXact; /* keep open over transactions? */
78 : bool dirty; /* does buffer need to be written? */
79 : bool readOnly; /* has the file been set to read only? */
80 :
81 : FileSet *fileset; /* space for fileset based segment files */
82 : const char *name; /* name of fileset based BufFile */
83 :
84 : /*
85 : * resowner is the ResourceOwner to use for underlying temp files. (We
86 : * don't need to remember the memory context we're using explicitly,
87 : * because after creation we only repalloc our arrays larger.)
88 : */
89 : ResourceOwner resowner;
90 :
91 : /*
92 : * "current pos" is position of start of buffer within the logical file.
93 : * Position as seen by user of BufFile is (curFile, curOffset + pos).
94 : */
95 : int curFile; /* file index (0..n) part of current pos */
96 : pgoff_t curOffset; /* offset part of current pos */
97 : int64 pos; /* next read/write position in buffer */
98 : int64 nbytes; /* total # of valid bytes in buffer */
99 :
100 : /*
101 : * XXX Should ideally use PGIOAlignedBlock, but might need a way to avoid
102 : * wasting per-file alignment padding when some users create many files.
103 : */
104 : PGAlignedBlock buffer;
105 : };
106 :
107 : static BufFile *makeBufFileCommon(int nfiles);
108 : static BufFile *makeBufFile(File firstfile);
109 : static void extendBufFile(BufFile *file);
110 : static void BufFileLoadBuffer(BufFile *file);
111 : static void BufFileDumpBuffer(BufFile *file);
112 : static void BufFileFlush(BufFile *file);
113 : static File MakeNewFileSetSegment(BufFile *buffile, int segment);
114 :
115 : /*
116 : * Create BufFile and perform the common initialization.
117 : */
118 : static BufFile *
119 6275 : makeBufFileCommon(int nfiles)
120 : {
121 6275 : BufFile *file = palloc_object(BufFile);
122 :
123 6275 : file->numFiles = nfiles;
124 6275 : file->isInterXact = false;
125 6275 : file->dirty = false;
126 6275 : file->resowner = CurrentResourceOwner;
127 6275 : file->curFile = 0;
128 6275 : file->curOffset = 0;
129 6275 : file->pos = 0;
130 6275 : file->nbytes = 0;
131 :
132 6275 : return file;
133 : }
134 :
135 : /*
136 : * Create a BufFile given the first underlying physical file.
137 : * NOTE: caller must set isInterXact if appropriate.
138 : */
139 : static BufFile *
140 1962 : makeBufFile(File firstfile)
141 : {
142 1962 : BufFile *file = makeBufFileCommon(1);
143 :
144 1962 : file->files = palloc_object(File);
145 1962 : file->files[0] = firstfile;
146 1962 : file->readOnly = false;
147 1962 : file->fileset = NULL;
148 1962 : file->name = NULL;
149 :
150 1962 : return file;
151 : }
152 :
153 : /*
154 : * Add another component temp file.
155 : */
156 : static void
157 0 : extendBufFile(BufFile *file)
158 : {
159 : File pfile;
160 : ResourceOwner oldowner;
161 :
162 : /* Be sure to associate the file with the BufFile's resource owner */
163 0 : oldowner = CurrentResourceOwner;
164 0 : CurrentResourceOwner = file->resowner;
165 :
166 0 : if (file->fileset == NULL)
167 0 : pfile = OpenTemporaryFile(file->isInterXact);
168 : else
169 0 : pfile = MakeNewFileSetSegment(file, file->numFiles);
170 :
171 : Assert(pfile >= 0);
172 :
173 0 : CurrentResourceOwner = oldowner;
174 :
175 0 : file->files = (File *) repalloc(file->files,
176 0 : (file->numFiles + 1) * sizeof(File));
177 0 : file->files[file->numFiles] = pfile;
178 0 : file->numFiles++;
179 0 : }
180 :
181 : /*
182 : * Create a BufFile for a new temporary file (which will expand to become
183 : * multiple temporary files if more than MAX_PHYSICAL_FILESIZE bytes are
184 : * written to it).
185 : *
186 : * If interXact is true, the temp file will not be automatically deleted
187 : * at end of transaction.
188 : *
189 : * Note: if interXact is true, the caller had better be calling us in a
190 : * memory context, and with a resource owner, that will survive across
191 : * transaction boundaries.
192 : */
193 : BufFile *
194 1962 : BufFileCreateTemp(bool interXact)
195 : {
196 : BufFile *file;
197 : File pfile;
198 :
199 : /*
200 : * Ensure that temp tablespaces are set up for OpenTemporaryFile to use.
201 : * Possibly the caller will have done this already, but it seems useful to
202 : * double-check here. Failure to do this at all would result in the temp
203 : * files always getting placed in the default tablespace, which is a
204 : * pretty hard-to-detect bug. Callers may prefer to do it earlier if they
205 : * want to be sure that any required catalog access is done in some other
206 : * resource context.
207 : */
208 1962 : PrepareTempTablespaces();
209 :
210 1962 : pfile = OpenTemporaryFile(interXact);
211 : Assert(pfile >= 0);
212 :
213 1962 : file = makeBufFile(pfile);
214 1962 : file->isInterXact = interXact;
215 :
216 1962 : return file;
217 : }
218 :
219 : /*
220 : * Build the name for a given segment of a given BufFile.
221 : */
222 : static void
223 9290 : FileSetSegmentName(char *name, const char *buffile_name, int segment)
224 : {
225 9290 : snprintf(name, MAXPGPATH, "%s.%d", buffile_name, segment);
226 9290 : }
227 :
228 : /*
229 : * Create a new segment file backing a fileset based BufFile.
230 : */
231 : static File
232 1943 : MakeNewFileSetSegment(BufFile *buffile, int segment)
233 : {
234 : char name[MAXPGPATH];
235 : File file;
236 :
237 : /*
238 : * It is possible that there are files left over from before a crash
239 : * restart with the same name. In order for BufFileOpenFileSet() not to
240 : * get confused about how many segments there are, we'll unlink the next
241 : * segment number if it already exists.
242 : */
243 1943 : FileSetSegmentName(name, buffile->name, segment + 1);
244 1943 : FileSetDelete(buffile->fileset, name, true);
245 :
246 : /* Create the new segment. */
247 1943 : FileSetSegmentName(name, buffile->name, segment);
248 1943 : file = FileSetCreate(buffile->fileset, name);
249 :
250 : /* FileSetCreate would've errored out */
251 : Assert(file > 0);
252 :
253 1943 : return file;
254 : }
255 :
256 : /*
257 : * Create a BufFile that can be discovered and opened read-only by other
258 : * backends that are attached to the same SharedFileSet using the same name.
259 : *
260 : * The naming scheme for fileset based BufFiles is left up to the calling code.
261 : * The name will appear as part of one or more filenames on disk, and might
262 : * provide clues to administrators about which subsystem is generating
263 : * temporary file data. Since each SharedFileSet object is backed by one or
264 : * more uniquely named temporary directory, names don't conflict with
265 : * unrelated SharedFileSet objects.
266 : */
267 : BufFile *
268 1943 : BufFileCreateFileSet(FileSet *fileset, const char *name)
269 : {
270 : BufFile *file;
271 :
272 1943 : file = makeBufFileCommon(1);
273 1943 : file->fileset = fileset;
274 1943 : file->name = pstrdup(name);
275 1943 : file->files = palloc_object(File);
276 1943 : file->files[0] = MakeNewFileSetSegment(file, 0);
277 1943 : file->readOnly = false;
278 :
279 1943 : return file;
280 : }
281 :
282 : /*
283 : * Open a file that was previously created in another backend (or this one)
284 : * with BufFileCreateFileSet in the same FileSet using the same name.
285 : * The backend that created the file must have called BufFileClose() or
286 : * BufFileExportFileSet() to make sure that it is ready to be opened by other
287 : * backends and render it read-only. If missing_ok is true, which indicates
288 : * that missing files can be safely ignored, then return NULL if the BufFile
289 : * with the given name is not found, otherwise, throw an error.
290 : */
291 : BufFile *
292 2643 : BufFileOpenFileSet(FileSet *fileset, const char *name, int mode,
293 : bool missing_ok)
294 : {
295 : BufFile *file;
296 : char segment_name[MAXPGPATH];
297 2643 : Size capacity = 16;
298 : File *files;
299 2643 : int nfiles = 0;
300 :
301 2643 : files = palloc_array(File, capacity);
302 :
303 : /*
304 : * We don't know how many segments there are, so we'll probe the
305 : * filesystem to find out.
306 : */
307 : for (;;)
308 : {
309 : /* See if we need to expand our file segment array. */
310 5013 : if (nfiles + 1 > capacity)
311 : {
312 0 : capacity *= 2;
313 0 : files = repalloc_array(files, File, capacity);
314 : }
315 : /* Try to load a segment. */
316 5013 : FileSetSegmentName(segment_name, name, nfiles);
317 5013 : files[nfiles] = FileSetOpen(fileset, segment_name, mode);
318 5013 : if (files[nfiles] <= 0)
319 2643 : break;
320 2370 : ++nfiles;
321 :
322 2370 : CHECK_FOR_INTERRUPTS();
323 : }
324 :
325 : /*
326 : * If we didn't find any files at all, then no BufFile exists with this
327 : * name.
328 : */
329 2643 : if (nfiles == 0)
330 : {
331 : /* free the memory */
332 273 : pfree(files);
333 :
334 273 : if (missing_ok)
335 273 : return NULL;
336 :
337 0 : ereport(ERROR,
338 : (errcode_for_file_access(),
339 : errmsg("could not open temporary file \"%s\" from BufFile \"%s\": %m",
340 : segment_name, name)));
341 : }
342 :
343 2370 : file = makeBufFileCommon(nfiles);
344 2370 : file->files = files;
345 2370 : file->readOnly = (mode == O_RDONLY);
346 2370 : file->fileset = fileset;
347 2370 : file->name = pstrdup(name);
348 :
349 2370 : return file;
350 : }
351 :
352 : /*
353 : * Delete a BufFile that was created by BufFileCreateFileSet in the given
354 : * FileSet using the given name.
355 : *
356 : * It is not necessary to delete files explicitly with this function. It is
357 : * provided only as a way to delete files proactively, rather than waiting for
358 : * the FileSet to be cleaned up.
359 : *
360 : * Only one backend should attempt to delete a given name, and should know
361 : * that it exists and has been exported or closed otherwise missing_ok should
362 : * be passed true.
363 : */
364 : void
365 352 : BufFileDeleteFileSet(FileSet *fileset, const char *name, bool missing_ok)
366 : {
367 : char segment_name[MAXPGPATH];
368 352 : int segment = 0;
369 352 : bool found = false;
370 :
371 : /*
372 : * We don't know how many segments the file has. We'll keep deleting
373 : * until we run out. If we don't manage to find even an initial segment,
374 : * raise an error.
375 : */
376 : for (;;)
377 : {
378 391 : FileSetSegmentName(segment_name, name, segment);
379 391 : if (!FileSetDelete(fileset, segment_name, true))
380 352 : break;
381 39 : found = true;
382 39 : ++segment;
383 :
384 39 : CHECK_FOR_INTERRUPTS();
385 : }
386 :
387 352 : if (!found && !missing_ok)
388 0 : elog(ERROR, "could not delete unknown BufFile \"%s\"", name);
389 352 : }
390 :
391 : /*
392 : * BufFileExportFileSet --- flush and make read-only, in preparation for sharing.
393 : */
394 : void
395 379 : BufFileExportFileSet(BufFile *file)
396 : {
397 : /* Must be a file belonging to a FileSet. */
398 : Assert(file->fileset != NULL);
399 :
400 : /* It's probably a bug if someone calls this twice. */
401 : Assert(!file->readOnly);
402 :
403 379 : BufFileFlush(file);
404 379 : file->readOnly = true;
405 379 : }
406 :
407 : /*
408 : * Close a BufFile
409 : *
410 : * Like fclose(), this also implicitly FileCloses the underlying File.
411 : */
412 : void
413 6113 : BufFileClose(BufFile *file)
414 : {
415 : int i;
416 :
417 : /* flush any unwritten data */
418 6113 : BufFileFlush(file);
419 : /* close and delete the underlying file(s) */
420 12379 : for (i = 0; i < file->numFiles; i++)
421 6266 : FileClose(file->files[i]);
422 : /* release the buffer space */
423 6113 : pfree(file->files);
424 6113 : pfree(file);
425 6113 : }
426 :
427 : /*
428 : * BufFileLoadBuffer
429 : *
430 : * Load some data into buffer, if possible, starting from curOffset.
431 : * At call, must have dirty = false, pos and nbytes = 0.
432 : * On exit, nbytes is number of bytes loaded.
433 : */
434 : static void
435 63075 : BufFileLoadBuffer(BufFile *file)
436 : {
437 : File thisfile;
438 : instr_time io_start;
439 : instr_time io_time;
440 :
441 : /*
442 : * Advance to next component file if necessary and possible.
443 : */
444 63075 : if (file->curOffset >= MAX_PHYSICAL_FILESIZE &&
445 0 : file->curFile + 1 < file->numFiles)
446 : {
447 0 : file->curFile++;
448 0 : file->curOffset = 0;
449 : }
450 :
451 63075 : thisfile = file->files[file->curFile];
452 :
453 63075 : if (track_io_timing)
454 0 : INSTR_TIME_SET_CURRENT(io_start);
455 : else
456 63075 : INSTR_TIME_SET_ZERO(io_start);
457 :
458 : /*
459 : * Read whatever we can get, up to a full bufferload.
460 : */
461 126150 : file->nbytes = FileRead(thisfile,
462 63075 : file->buffer.data,
463 : sizeof(file->buffer.data),
464 : file->curOffset,
465 : WAIT_EVENT_BUFFILE_READ);
466 63075 : if (file->nbytes < 0)
467 : {
468 0 : file->nbytes = 0;
469 0 : ereport(ERROR,
470 : (errcode_for_file_access(),
471 : errmsg("could not read file \"%s\": %m",
472 : FilePathName(thisfile))));
473 : }
474 :
475 63075 : if (track_io_timing)
476 : {
477 0 : INSTR_TIME_SET_CURRENT(io_time);
478 0 : INSTR_TIME_ACCUM_DIFF(pgBufferUsage.temp_blk_read_time, io_time, io_start);
479 : }
480 :
481 : /* we choose not to advance curOffset here */
482 :
483 63075 : if (file->nbytes > 0)
484 61388 : pgBufferUsage.temp_blks_read++;
485 63075 : }
486 :
487 : /*
488 : * BufFileDumpBuffer
489 : *
490 : * Dump buffer contents starting at curOffset.
491 : * At call, should have dirty = true, nbytes > 0.
492 : * On exit, dirty is cleared if successful write, and curOffset is advanced.
493 : */
494 : static void
495 72521 : BufFileDumpBuffer(BufFile *file)
496 : {
497 72521 : int64 wpos = 0;
498 : int64 bytestowrite;
499 : File thisfile;
500 :
501 : /*
502 : * Unlike BufFileLoadBuffer, we must dump the whole buffer even if it
503 : * crosses a component-file boundary; so we need a loop.
504 : */
505 145042 : while (wpos < file->nbytes)
506 : {
507 : int64 availbytes;
508 : instr_time io_start;
509 : instr_time io_time;
510 :
511 : /*
512 : * Advance to next component file if necessary and possible.
513 : */
514 72521 : if (file->curOffset >= MAX_PHYSICAL_FILESIZE)
515 : {
516 0 : while (file->curFile + 1 >= file->numFiles)
517 0 : extendBufFile(file);
518 0 : file->curFile++;
519 0 : file->curOffset = 0;
520 : }
521 :
522 : /*
523 : * Determine how much we need to write into this file.
524 : */
525 72521 : bytestowrite = file->nbytes - wpos;
526 72521 : availbytes = MAX_PHYSICAL_FILESIZE - file->curOffset;
527 :
528 72521 : if (bytestowrite > availbytes)
529 0 : bytestowrite = availbytes;
530 :
531 72521 : thisfile = file->files[file->curFile];
532 :
533 72521 : if (track_io_timing)
534 0 : INSTR_TIME_SET_CURRENT(io_start);
535 : else
536 72521 : INSTR_TIME_SET_ZERO(io_start);
537 :
538 72521 : bytestowrite = FileWrite(thisfile,
539 72521 : file->buffer.data + wpos,
540 : bytestowrite,
541 : file->curOffset,
542 : WAIT_EVENT_BUFFILE_WRITE);
543 72521 : if (bytestowrite <= 0)
544 0 : ereport(ERROR,
545 : (errcode_for_file_access(),
546 : errmsg("could not write to file \"%s\": %m",
547 : FilePathName(thisfile))));
548 :
549 72521 : if (track_io_timing)
550 : {
551 0 : INSTR_TIME_SET_CURRENT(io_time);
552 0 : INSTR_TIME_ACCUM_DIFF(pgBufferUsage.temp_blk_write_time, io_time, io_start);
553 : }
554 :
555 72521 : file->curOffset += bytestowrite;
556 72521 : wpos += bytestowrite;
557 :
558 72521 : pgBufferUsage.temp_blks_written++;
559 : }
560 72521 : file->dirty = false;
561 :
562 : /*
563 : * At this point, curOffset has been advanced to the end of the buffer,
564 : * ie, its original value + nbytes. We need to make it point to the
565 : * logical file position, ie, original value + pos, in case that is less
566 : * (as could happen due to a small backwards seek in a dirty buffer!)
567 : */
568 72521 : file->curOffset -= (file->nbytes - file->pos);
569 72521 : if (file->curOffset < 0) /* handle possible segment crossing */
570 : {
571 0 : file->curFile--;
572 : Assert(file->curFile >= 0);
573 0 : file->curOffset += MAX_PHYSICAL_FILESIZE;
574 : }
575 :
576 : /*
577 : * Now we can set the buffer empty without changing the logical position
578 : */
579 72521 : file->pos = 0;
580 72521 : file->nbytes = 0;
581 72521 : }
582 :
583 : /*
584 : * BufFileRead variants
585 : *
586 : * Like fread() except we assume 1-byte element size and report I/O errors via
587 : * ereport().
588 : *
589 : * If 'exact' is true, then an error is also raised if the number of bytes
590 : * read is not exactly 'size' (no short reads). If 'exact' and 'eofOK' are
591 : * true, then reading zero bytes is ok.
592 : */
593 : static size_t
594 15785618 : BufFileReadCommon(BufFile *file, void *ptr, size_t size, bool exact, bool eofOK)
595 : {
596 15785618 : size_t start_size = size;
597 15785618 : size_t nread = 0;
598 : size_t nthistime;
599 :
600 15785618 : BufFileFlush(file);
601 :
602 31585555 : while (size > 0)
603 : {
604 15801624 : if (file->pos >= file->nbytes)
605 : {
606 : /* Try to load more data into buffer. */
607 63075 : file->curOffset += file->pos;
608 63075 : file->pos = 0;
609 63075 : file->nbytes = 0;
610 63075 : BufFileLoadBuffer(file);
611 63075 : if (file->nbytes <= 0)
612 1687 : break; /* no more data available */
613 : }
614 :
615 15799937 : nthistime = file->nbytes - file->pos;
616 15799937 : if (nthistime > size)
617 15740563 : nthistime = size;
618 : Assert(nthistime > 0);
619 :
620 15799937 : memcpy(ptr, file->buffer.data + file->pos, nthistime);
621 :
622 15799937 : file->pos += nthistime;
623 15799937 : ptr = (char *) ptr + nthistime;
624 15799937 : size -= nthistime;
625 15799937 : nread += nthistime;
626 : }
627 :
628 15785618 : if (exact &&
629 1687 : (nread != start_size && !(nread == 0 && eofOK)))
630 0 : ereport(ERROR,
631 : errcode_for_file_access(),
632 : file->name ?
633 : errmsg("could not read from file set \"%s\": read only %zu of %zu bytes",
634 : file->name, nread, start_size) :
635 : errmsg("could not read from temporary file: read only %zu of %zu bytes",
636 : nread, start_size));
637 :
638 15785618 : return nread;
639 : }
640 :
641 : /*
642 : * Legacy interface where the caller needs to check for end of file or short
643 : * reads.
644 : */
645 : size_t
646 0 : BufFileRead(BufFile *file, void *ptr, size_t size)
647 : {
648 0 : return BufFileReadCommon(file, ptr, size, false, false);
649 : }
650 :
651 : /*
652 : * Require read of exactly the specified size.
653 : */
654 : void
655 10350039 : BufFileReadExact(BufFile *file, void *ptr, size_t size)
656 : {
657 10350039 : BufFileReadCommon(file, ptr, size, true, false);
658 10350039 : }
659 :
660 : /*
661 : * Require read of exactly the specified size, but optionally allow end of
662 : * file (in which case 0 is returned).
663 : */
664 : size_t
665 5435579 : BufFileReadMaybeEOF(BufFile *file, void *ptr, size_t size, bool eofOK)
666 : {
667 5435579 : return BufFileReadCommon(file, ptr, size, true, eofOK);
668 : }
669 :
670 : /*
671 : * BufFileWrite
672 : *
673 : * Like fwrite() except we assume 1-byte element size and report errors via
674 : * ereport().
675 : */
676 : void
677 15431677 : BufFileWrite(BufFile *file, const void *ptr, size_t size)
678 : {
679 : size_t nthistime;
680 :
681 : Assert(!file->readOnly);
682 :
683 30891247 : while (size > 0)
684 : {
685 15459570 : if (file->pos >= BLCKSZ)
686 : {
687 : /* Buffer full, dump it out */
688 45892 : if (file->dirty)
689 45576 : BufFileDumpBuffer(file);
690 : else
691 : {
692 : /* Hmm, went directly from reading to writing? */
693 316 : file->curOffset += file->pos;
694 316 : file->pos = 0;
695 316 : file->nbytes = 0;
696 : }
697 : }
698 :
699 15459570 : nthistime = BLCKSZ - file->pos;
700 15459570 : if (nthistime > size)
701 15389313 : nthistime = size;
702 : Assert(nthistime > 0);
703 :
704 15459570 : memcpy(file->buffer.data + file->pos, ptr, nthistime);
705 :
706 15459570 : file->dirty = true;
707 15459570 : file->pos += nthistime;
708 15459570 : if (file->nbytes < file->pos)
709 15456994 : file->nbytes = file->pos;
710 15459570 : ptr = (const char *) ptr + nthistime;
711 15459570 : size -= nthistime;
712 : }
713 15431677 : }
714 :
715 : /*
716 : * BufFileFlush
717 : *
718 : * Like fflush(), except that I/O errors are reported with ereport().
719 : */
720 : static void
721 15827283 : BufFileFlush(BufFile *file)
722 : {
723 15827283 : if (file->dirty)
724 26945 : BufFileDumpBuffer(file);
725 :
726 : Assert(!file->dirty);
727 15827283 : }
728 :
729 : /*
730 : * BufFileSeek
731 : *
732 : * Like fseek(), except that target position needs two values in order to
733 : * work when logical filesize exceeds maximum value representable by pgoff_t.
734 : * We do not support relative seeks across more than that, however.
735 : * I/O errors are reported by ereport().
736 : *
737 : * Result is 0 if OK, EOF if not. Logical position is not moved if an
738 : * impossible seek is attempted.
739 : */
740 : int
741 71592 : BufFileSeek(BufFile *file, int fileno, pgoff_t offset, int whence)
742 : {
743 : int newFile;
744 : pgoff_t newOffset;
745 :
746 71592 : switch (whence)
747 : {
748 71257 : case SEEK_SET:
749 71257 : if (fileno < 0)
750 0 : return EOF;
751 71257 : newFile = fileno;
752 71257 : newOffset = offset;
753 71257 : break;
754 0 : case SEEK_CUR:
755 :
756 : /*
757 : * Relative seek considers only the signed offset, ignoring
758 : * fileno.
759 : */
760 0 : newFile = file->curFile;
761 0 : newOffset = (file->curOffset + file->pos) + offset;
762 0 : break;
763 335 : case SEEK_END:
764 :
765 : /*
766 : * The file size of the last file gives us the end offset of that
767 : * file.
768 : */
769 335 : newFile = file->numFiles - 1;
770 335 : newOffset = FileSize(file->files[file->numFiles - 1]);
771 335 : if (newOffset < 0)
772 0 : ereport(ERROR,
773 : (errcode_for_file_access(),
774 : errmsg("could not determine size of temporary file \"%s\" from BufFile \"%s\": %m",
775 : FilePathName(file->files[file->numFiles - 1]),
776 : file->name)));
777 335 : break;
778 0 : default:
779 0 : elog(ERROR, "invalid whence: %d", whence);
780 : return EOF;
781 : }
782 71592 : while (newOffset < 0)
783 : {
784 0 : if (--newFile < 0)
785 0 : return EOF;
786 0 : newOffset += MAX_PHYSICAL_FILESIZE;
787 : }
788 71592 : if (newFile == file->curFile &&
789 71439 : newOffset >= file->curOffset &&
790 52239 : newOffset <= file->curOffset + file->nbytes)
791 : {
792 : /*
793 : * Seek is to a point within existing buffer; we can just adjust
794 : * pos-within-buffer, without flushing buffer. Note this is OK
795 : * whether reading or writing, but buffer remains dirty if we were
796 : * writing.
797 : */
798 36419 : file->pos = (int64) (newOffset - file->curOffset);
799 36419 : return 0;
800 : }
801 : /* Otherwise, must reposition buffer, so flush any dirty data */
802 35173 : BufFileFlush(file);
803 :
804 : /*
805 : * At this point and no sooner, check for seek past last segment. The
806 : * above flush could have created a new segment, so checking sooner would
807 : * not work (at least not with this code).
808 : */
809 :
810 : /* convert seek to "start of next seg" to "end of last seg" */
811 35173 : if (newFile == file->numFiles && newOffset == 0)
812 : {
813 0 : newFile--;
814 0 : newOffset = MAX_PHYSICAL_FILESIZE;
815 : }
816 35173 : while (newOffset > MAX_PHYSICAL_FILESIZE)
817 : {
818 0 : if (++newFile >= file->numFiles)
819 0 : return EOF;
820 0 : newOffset -= MAX_PHYSICAL_FILESIZE;
821 : }
822 35173 : if (newFile >= file->numFiles)
823 0 : return EOF;
824 : /* Seek is OK! */
825 35173 : file->curFile = newFile;
826 35173 : file->curOffset = newOffset;
827 35173 : file->pos = 0;
828 35173 : file->nbytes = 0;
829 35173 : return 0;
830 : }
831 :
832 : void
833 88672 : BufFileTell(BufFile *file, int *fileno, pgoff_t *offset)
834 : {
835 88672 : *fileno = file->curFile;
836 88672 : *offset = file->curOffset + file->pos;
837 88672 : }
838 :
839 : /*
840 : * BufFileSeekBlock --- block-oriented seek
841 : *
842 : * Performs absolute seek to the start of the n'th BLCKSZ-sized block of
843 : * the file. Note that users of this interface will fail if their files
844 : * exceed BLCKSZ * PG_INT64_MAX bytes, but that is quite a lot; we don't
845 : * work with tables bigger than that, either...
846 : *
847 : * Result is 0 if OK, EOF if not. Logical position is not moved if an
848 : * impossible seek is attempted.
849 : */
850 : int
851 69412 : BufFileSeekBlock(BufFile *file, int64 blknum)
852 : {
853 138824 : return BufFileSeek(file,
854 69412 : (int) (blknum / BUFFILE_SEG_SIZE),
855 69412 : (pgoff_t) (blknum % BUFFILE_SEG_SIZE) * BLCKSZ,
856 : SEEK_SET);
857 : }
858 :
859 : /*
860 : * Returns the amount of data in the given BufFile, in bytes.
861 : *
862 : * Returned value includes the size of any holes left behind by BufFileAppend.
863 : * ereport()s on failure.
864 : */
865 : int64
866 291 : BufFileSize(BufFile *file)
867 : {
868 : int64 lastFileSize;
869 :
870 : /* Get the size of the last physical file. */
871 291 : lastFileSize = FileSize(file->files[file->numFiles - 1]);
872 291 : if (lastFileSize < 0)
873 0 : ereport(ERROR,
874 : (errcode_for_file_access(),
875 : errmsg("could not determine size of temporary file \"%s\" from BufFile \"%s\": %m",
876 : FilePathName(file->files[file->numFiles - 1]),
877 : file->name)));
878 :
879 291 : return ((file->numFiles - 1) * (int64) MAX_PHYSICAL_FILESIZE) +
880 : lastFileSize;
881 : }
882 :
883 : /*
884 : * Append the contents of the source file to the end of the target file.
885 : *
886 : * Note that operation subsumes ownership of underlying resources from
887 : * "source". Caller should never call BufFileClose against source having
888 : * called here first. Resource owners for source and target must match,
889 : * too.
890 : *
891 : * This operation works by manipulating lists of segment files, so the
892 : * file content is always appended at a MAX_PHYSICAL_FILESIZE-aligned
893 : * boundary, typically creating empty holes before the boundary. These
894 : * areas do not contain any interesting data, and cannot be read from by
895 : * caller.
896 : *
897 : * Returns the block number within target where the contents of source
898 : * begins. Caller should apply this as an offset when working off block
899 : * positions that are in terms of the original BufFile space.
900 : */
901 : int64
902 153 : BufFileAppend(BufFile *target, BufFile *source)
903 : {
904 153 : int64 startBlock = (int64) target->numFiles * BUFFILE_SEG_SIZE;
905 153 : int newNumFiles = target->numFiles + source->numFiles;
906 : int i;
907 :
908 : Assert(source->readOnly);
909 : Assert(!source->dirty);
910 :
911 153 : if (target->resowner != source->resowner)
912 0 : elog(ERROR, "could not append BufFile with non-matching resource owner");
913 :
914 153 : target->files = (File *)
915 153 : repalloc(target->files, sizeof(File) * newNumFiles);
916 306 : for (i = target->numFiles; i < newNumFiles; i++)
917 153 : target->files[i] = source->files[i - target->numFiles];
918 153 : target->numFiles = newNumFiles;
919 :
920 153 : return startBlock;
921 : }
922 :
923 : /*
924 : * Truncate a BufFile created by BufFileCreateFileSet up to the given fileno
925 : * and the offset.
926 : */
927 : void
928 9 : BufFileTruncateFileSet(BufFile *file, int fileno, pgoff_t offset)
929 : {
930 9 : int numFiles = file->numFiles;
931 9 : int newFile = fileno;
932 9 : pgoff_t newOffset = file->curOffset;
933 : char segment_name[MAXPGPATH];
934 : int i;
935 :
936 : /*
937 : * Loop over all the files up to the given fileno and remove the files
938 : * that are greater than the fileno and truncate the given file up to the
939 : * offset. Note that we also remove the given fileno if the offset is 0
940 : * provided it is not the first file in which we truncate it.
941 : */
942 18 : for (i = file->numFiles - 1; i >= fileno; i--)
943 : {
944 9 : if ((i != fileno || offset == 0) && i != 0)
945 : {
946 0 : FileSetSegmentName(segment_name, file->name, i);
947 0 : FileClose(file->files[i]);
948 0 : if (!FileSetDelete(file->fileset, segment_name, true))
949 0 : ereport(ERROR,
950 : (errcode_for_file_access(),
951 : errmsg("could not delete fileset \"%s\": %m",
952 : segment_name)));
953 0 : numFiles--;
954 0 : newOffset = MAX_PHYSICAL_FILESIZE;
955 :
956 : /*
957 : * This is required to indicate that we have deleted the given
958 : * fileno.
959 : */
960 0 : if (i == fileno)
961 0 : newFile--;
962 : }
963 : else
964 : {
965 9 : if (FileTruncate(file->files[i], offset,
966 : WAIT_EVENT_BUFFILE_TRUNCATE) < 0)
967 0 : ereport(ERROR,
968 : (errcode_for_file_access(),
969 : errmsg("could not truncate file \"%s\": %m",
970 : FilePathName(file->files[i]))));
971 9 : newOffset = offset;
972 : }
973 : }
974 :
975 9 : file->numFiles = numFiles;
976 :
977 : /*
978 : * If the truncate point is within existing buffer then we can just adjust
979 : * pos within buffer.
980 : */
981 9 : if (newFile == file->curFile &&
982 9 : newOffset >= file->curOffset &&
983 9 : newOffset <= file->curOffset + file->nbytes)
984 : {
985 : /* No need to reset the current pos if the new pos is greater. */
986 0 : if (newOffset <= file->curOffset + file->pos)
987 0 : file->pos = (int64) newOffset - file->curOffset;
988 :
989 : /* Adjust the nbytes for the current buffer. */
990 0 : file->nbytes = (int64) newOffset - file->curOffset;
991 : }
992 9 : else if (newFile == file->curFile &&
993 9 : newOffset < file->curOffset)
994 : {
995 : /*
996 : * The truncate point is within the existing file but prior to the
997 : * current position, so we can forget the current buffer and reset the
998 : * current position.
999 : */
1000 0 : file->curOffset = newOffset;
1001 0 : file->pos = 0;
1002 0 : file->nbytes = 0;
1003 : }
1004 9 : else if (newFile < file->curFile)
1005 : {
1006 : /*
1007 : * The truncate point is prior to the current file, so need to reset
1008 : * the current position accordingly.
1009 : */
1010 0 : file->curFile = newFile;
1011 0 : file->curOffset = newOffset;
1012 0 : file->pos = 0;
1013 0 : file->nbytes = 0;
1014 : }
1015 : /* Nothing to do, if the truncate point is beyond current file. */
1016 9 : }
|
