| 1 |
/*------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* buffile.c |
| 4 |
* Management of large buffered temporary files. |
| 5 |
* |
| 6 |
* Portions Copyright (c) 1996-2025, 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 |
|
| 57 |
/* |
| 58 |
* We break BufFiles into gigabyte-sized segments, regardless of RELSEG_SIZE. |
| 59 |
* The reason is that we'd like large BufFiles to be spread across multiple |
| 60 |
* tablespaces when available. |
| 61 |
*/ |
| 62 |
#define MAX_PHYSICAL_FILESIZE 0x40000000 |
| 63 |
#define BUFFILE_SEG_SIZE (MAX_PHYSICAL_FILESIZE / BLCKSZ) |
| 64 |
|
| 65 |
/* |
| 66 |
* This data structure represents a buffered file that consists of one or |
| 67 |
* more physical files (each accessed through a virtual file descriptor |
| 68 |
* managed by fd.c). |
| 69 |
*/ |
| 70 |
struct BufFile |
| 71 |
{ |
| 72 |
int numFiles; /* number of physical files in set */ |
| 73 |
/* all files except the last have length exactly MAX_PHYSICAL_FILESIZE */ |
| 74 |
File *files; /* palloc'd array with numFiles entries */ |
| 75 |
|
| 76 |
bool isInterXact; /* keep open over transactions? */ |
| 77 |
bool dirty; /* does buffer need to be written? */ |
| 78 |
bool readOnly; /* has the file been set to read only? */ |
| 79 |
|
| 80 |
FileSet *fileset; /* space for fileset based segment files */ |
| 81 |
const char *name; /* name of fileset based BufFile */ |
| 82 |
|
| 83 |
/* |
| 84 |
* resowner is the ResourceOwner to use for underlying temp files. (We |
| 85 |
* don't need to remember the memory context we're using explicitly, |
| 86 |
* because after creation we only repalloc our arrays larger.) |
| 87 |
*/ |
| 88 |
ResourceOwner resowner; |
| 89 |
|
| 90 |
/* |
| 91 |
* "current pos" is position of start of buffer within the logical file. |
| 92 |
* Position as seen by user of BufFile is (curFile, curOffset + pos). |
| 93 |
*/ |
| 94 |
int curFile; /* file index (0..n) part of current pos */ |
| 95 |
off_t curOffset; /* offset part of current pos */ |
| 96 |
int pos; /* next read/write position in buffer */ |
| 97 |
int nbytes; /* total # of valid bytes in buffer */ |
| 98 |
|
| 99 |
/* |
| 100 |
* XXX Should ideally use PGIOAlignedBlock, but might need a way to avoid |
| 101 |
* wasting per-file alignment padding when some users create many files. |
| 102 |
*/ |
| 103 |
PGAlignedBlock buffer; |
| 104 |
}; |
| 105 |
|
| 106 |
static BufFile *makeBufFileCommon(int nfiles); |
| 107 |
static BufFile *makeBufFile(File firstfile); |
| 108 |
static void extendBufFile(BufFile *file); |
| 109 |
static void BufFileLoadBuffer(BufFile *file); |
| 110 |
static void BufFileDumpBuffer(BufFile *file); |
| 111 |
static void BufFileFlush(BufFile *file); |
| 112 |
static File MakeNewFileSetSegment(BufFile *buffile, int segment); |
| 113 |
|
| 114 |
/* |
| 115 |
* Create BufFile and perform the common initialization. |
| 116 |
*/ |
| 117 |
static BufFile * |
| 118 |
makeBufFileCommon(int nfiles) |
| 119 |
{ |
| 120 |
BufFile *file = (BufFile *) palloc(sizeof(BufFile)); |
| 121 |
|
| 122 |
file->numFiles = nfiles; |
| 123 |
file->isInterXact = false; |
| 124 |
file->dirty = false; |
| 125 |
file->resowner = CurrentResourceOwner; |
| 126 |
file->curFile = 0; |
| 127 |
file->curOffset = 0; |
| 128 |
file->pos = 0; |
| 129 |
file->nbytes = 0; |
| 130 |
|
| 131 |
return file; |
| 132 |
} |
| 133 |
|
| 134 |
/* |
| 135 |
* Create a BufFile given the first underlying physical file. |
| 136 |
* NOTE: caller must set isInterXact if appropriate. |
| 137 |
*/ |
| 138 |
static BufFile * |
| 139 |
makeBufFile(File firstfile) |
| 140 |
{ |
| 141 |
BufFile *file = makeBufFileCommon(1); |
| 142 |
|
| 143 |
file->files = (File *) palloc(sizeof(File)); |
| 144 |
file->files[0] = firstfile; |
| 145 |
file->readOnly = false; |
| 146 |
file->fileset = NULL; |
| 147 |
file->name = NULL; |
| 148 |
|
| 149 |
return file; |
| 150 |
} |
| 151 |
|
| 152 |
/* |
| 153 |
* Add another component temp file. |
| 154 |
*/ |
| 155 |
static void |
| 156 |
extendBufFile(BufFile *file) |
| 157 |
{ |
| 158 |
File pfile; |
| 159 |
ResourceOwner oldowner; |
| 160 |
|
| 161 |
/* Be sure to associate the file with the BufFile's resource owner */ |
| 162 |
oldowner = CurrentResourceOwner; |
| 163 |
CurrentResourceOwner = file->resowner; |
| 164 |
|
| 165 |
if (file->fileset == NULL) |
| 166 |
pfile = OpenTemporaryFile(file->isInterXact); |
| 167 |
else |
| 168 |
pfile = MakeNewFileSetSegment(file, file->numFiles); |
| 169 |
|
| 170 |
Assert(pfile >= 0); |
| 171 |
|
| 172 |
CurrentResourceOwner = oldowner; |
| 173 |
|
| 174 |
file->files = (File *) repalloc(file->files, |
| 175 |
(file->numFiles + 1) * sizeof(File)); |
| 176 |
file->files[file->numFiles] = pfile; |
| 177 |
file->numFiles++; |
| 178 |
} |
| 179 |
|
| 180 |
/* |
| 181 |
* Create a BufFile for a new temporary file (which will expand to become |
| 182 |
* multiple temporary files if more than MAX_PHYSICAL_FILESIZE bytes are |
| 183 |
* written to it). |
| 184 |
* |
| 185 |
* If interXact is true, the temp file will not be automatically deleted |
| 186 |
* at end of transaction. |
| 187 |
* |
| 188 |
* Note: if interXact is true, the caller had better be calling us in a |
| 189 |
* memory context, and with a resource owner, that will survive across |
| 190 |
* transaction boundaries. |
| 191 |
*/ |
| 192 |
BufFile * |
| 193 |
BufFileCreateTemp(bool interXact) |
| 194 |
{ |
| 195 |
BufFile *file; |
| 196 |
File pfile; |
| 197 |
|
| 198 |
/* |
| 199 |
* Ensure that temp tablespaces are set up for OpenTemporaryFile to use. |
| 200 |
* Possibly the caller will have done this already, but it seems useful to |
| 201 |
* double-check here. Failure to do this at all would result in the temp |
| 202 |
* files always getting placed in the default tablespace, which is a |
| 203 |
* pretty hard-to-detect bug. Callers may prefer to do it earlier if they |
| 204 |
* want to be sure that any required catalog access is done in some other |
| 205 |
* resource context. |
| 206 |
*/ |
| 207 |
PrepareTempTablespaces(); |
| 208 |
|
| 209 |
pfile = OpenTemporaryFile(interXact); |
| 210 |
Assert(pfile >= 0); |
| 211 |
|
| 212 |
file = makeBufFile(pfile); |
| 213 |
file->isInterXact = interXact; |
| 214 |
|
| 215 |
return file; |
| 216 |
} |
| 217 |
|
| 218 |
/* |
| 219 |
* Build the name for a given segment of a given BufFile. |
| 220 |
*/ |
| 221 |
static void |
| 222 |
FileSetSegmentName(char *name, const char *buffile_name, int segment) |
| 223 |
{ |
| 224 |
snprintf(name, MAXPGPATH, "%s.%d", buffile_name, segment); |
| 225 |
} |
| 226 |
|
| 227 |
/* |
| 228 |
* Create a new segment file backing a fileset based BufFile. |
| 229 |
*/ |
| 230 |
static File |
| 231 |
MakeNewFileSetSegment(BufFile *buffile, int segment) |
| 232 |
{ |
| 233 |
char name[MAXPGPATH]; |
| 234 |
File file; |
| 235 |
|
| 236 |
/* |
| 237 |
* It is possible that there are files left over from before a crash |
| 238 |
* restart with the same name. In order for BufFileOpenFileSet() not to |
| 239 |
* get confused about how many segments there are, we'll unlink the next |
| 240 |
* segment number if it already exists. |
| 241 |
*/ |
| 242 |
FileSetSegmentName(name, buffile->name, segment + 1); |
| 243 |
FileSetDelete(buffile->fileset, name, true); |
| 244 |
|
| 245 |
/* Create the new segment. */ |
| 246 |
FileSetSegmentName(name, buffile->name, segment); |
| 247 |
file = FileSetCreate(buffile->fileset, name); |
| 248 |
|
| 249 |
/* FileSetCreate would've errored out */ |
| 250 |
Assert(file > 0); |
| 251 |
|
| 252 |
return file; |
| 253 |
} |
| 254 |
|
| 255 |
/* |
| 256 |
* Create a BufFile that can be discovered and opened read-only by other |
| 257 |
* backends that are attached to the same SharedFileSet using the same name. |
| 258 |
* |
| 259 |
* The naming scheme for fileset based BufFiles is left up to the calling code. |
| 260 |
* The name will appear as part of one or more filenames on disk, and might |
| 261 |
* provide clues to administrators about which subsystem is generating |
| 262 |
* temporary file data. Since each SharedFileSet object is backed by one or |
| 263 |
* more uniquely named temporary directory, names don't conflict with |
| 264 |
* unrelated SharedFileSet objects. |
| 265 |
*/ |
| 266 |
BufFile * |
| 267 |
BufFileCreateFileSet(FileSet *fileset, const char *name) |
| 268 |
{ |
| 269 |
BufFile *file; |
| 270 |
|
| 271 |
file = makeBufFileCommon(1); |
| 272 |
file->fileset = fileset; |
| 273 |
file->name = pstrdup(name); |
| 274 |
file->files = (File *) palloc(sizeof(File)); |
| 275 |
file->files[0] = MakeNewFileSetSegment(file, 0); |
| 276 |
file->readOnly = false; |
| 277 |
|
| 278 |
return file; |
| 279 |
} |
| 280 |
|
| 281 |
/* |
| 282 |
* Open a file that was previously created in another backend (or this one) |
| 283 |
* with BufFileCreateFileSet in the same FileSet using the same name. |
| 284 |
* The backend that created the file must have called BufFileClose() or |
| 285 |
* BufFileExportFileSet() to make sure that it is ready to be opened by other |
| 286 |
* backends and render it read-only. If missing_ok is true, which indicates |
| 287 |
* that missing files can be safely ignored, then return NULL if the BufFile |
| 288 |
* with the given name is not found, otherwise, throw an error. |
| 289 |
*/ |
| 290 |
BufFile * |
| 291 |
BufFileOpenFileSet(FileSet *fileset, const char *name, int mode, |
| 292 |
bool missing_ok) |
| 293 |
{ |
| 294 |
BufFile *file; |
| 295 |
char segment_name[MAXPGPATH]; |
| 296 |
Size capacity = 16; |
| 297 |
File *files; |
| 298 |
int nfiles = 0; |
| 299 |
|
| 300 |
files = palloc(sizeof(File) * capacity); |
| 301 |
|
| 302 |
/* |
| 303 |
* We don't know how many segments there are, so we'll probe the |
| 304 |
* filesystem to find out. |
| 305 |
*/ |
| 306 |
for (;;) |
| 307 |
{ |
| 308 |
/* See if we need to expand our file segment array. */ |
| 309 |
if (nfiles + 1 > capacity) |
| 310 |
{ |
| 311 |
capacity *= 2; |
| 312 |
files = repalloc(files, sizeof(File) * capacity); |
| 313 |
} |
| 314 |
/* Try to load a segment. */ |
| 315 |
FileSetSegmentName(segment_name, name, nfiles); |
| 316 |
files[nfiles] = FileSetOpen(fileset, segment_name, mode); |
| 317 |
if (files[nfiles] <= 0) |
| 318 |
break; |
| 319 |
++nfiles; |
| 320 |
|
| 321 |
CHECK_FOR_INTERRUPTS(); |
| 322 |
} |
| 323 |
|
| 324 |
/* |
| 325 |
* If we didn't find any files at all, then no BufFile exists with this |
| 326 |
* name. |
| 327 |
*/ |
| 328 |
if (nfiles == 0) |
| 329 |
{ |
| 330 |
/* free the memory */ |
| 331 |
pfree(files); |
| 332 |
|
| 333 |
if (missing_ok) |
| 334 |
return NULL; |
| 335 |
|
| 336 |
ereport(ERROR, |
| 337 |
(errcode_for_file_access(), |
| 338 |
errmsg("could not open temporary file \"%s\" from BufFile \"%s\": %m", |
| 339 |
segment_name, name))); |
| 340 |
} |
| 341 |
|
| 342 |
file = makeBufFileCommon(nfiles); |
| 343 |
file->files = files; |
| 344 |
file->readOnly = (mode == O_RDONLY); |
| 345 |
file->fileset = fileset; |
| 346 |
file->name = pstrdup(name); |
| 347 |
|
| 348 |
return file; |
| 349 |
} |
| 350 |
|
| 351 |
/* |
| 352 |
* Delete a BufFile that was created by BufFileCreateFileSet in the given |
| 353 |
* FileSet using the given name. |
| 354 |
* |
| 355 |
* It is not necessary to delete files explicitly with this function. It is |
| 356 |
* provided only as a way to delete files proactively, rather than waiting for |
| 357 |
* the FileSet to be cleaned up. |
| 358 |
* |
| 359 |
* Only one backend should attempt to delete a given name, and should know |
| 360 |
* that it exists and has been exported or closed otherwise missing_ok should |
| 361 |
* be passed true. |
| 362 |
*/ |
| 363 |
void |
| 364 |
BufFileDeleteFileSet(FileSet *fileset, const char *name, bool missing_ok) |
| 365 |
{ |
| 366 |
char segment_name[MAXPGPATH]; |
| 367 |
int segment = 0; |
| 368 |
bool found = false; |
| 369 |
|
| 370 |
/* |
| 371 |
* We don't know how many segments the file has. We'll keep deleting |
| 372 |
* until we run out. If we don't manage to find even an initial segment, |
| 373 |
* raise an error. |
| 374 |
*/ |
| 375 |
for (;;) |
| 376 |
{ |
| 377 |
FileSetSegmentName(segment_name, name, segment); |
| 378 |
if (!FileSetDelete(fileset, segment_name, true)) |
| 379 |
break; |
| 380 |
found = true; |
| 381 |
++segment; |
| 382 |
|
| 383 |
CHECK_FOR_INTERRUPTS(); |
| 384 |
} |
| 385 |
|
| 386 |
if (!found && !missing_ok) |
| 387 |
elog(ERROR, "could not delete unknown BufFile \"%s\"", name); |
| 388 |
} |
| 389 |
|
| 390 |
/* |
| 391 |
* BufFileExportFileSet --- flush and make read-only, in preparation for sharing. |
| 392 |
*/ |
| 393 |
void |
| 394 |
BufFileExportFileSet(BufFile *file) |
| 395 |
{ |
| 396 |
/* Must be a file belonging to a FileSet. */ |
| 397 |
Assert(file->fileset != NULL); |
| 398 |
|
| 399 |
/* It's probably a bug if someone calls this twice. */ |
| 400 |
Assert(!file->readOnly); |
| 401 |
|
| 402 |
BufFileFlush(file); |
| 403 |
file->readOnly = true; |
| 404 |
} |
| 405 |
|
| 406 |
/* |
| 407 |
* Close a BufFile |
| 408 |
* |
| 409 |
* Like fclose(), this also implicitly FileCloses the underlying File. |
| 410 |
*/ |
| 411 |
void |
| 412 |
BufFileClose(BufFile *file) |
| 413 |
{ |
| 414 |
int i; |
| 415 |
|
| 416 |
/* flush any unwritten data */ |
| 417 |
BufFileFlush(file); |
| 418 |
/* close and delete the underlying file(s) */ |
| 419 |
for (i = 0; i < file->numFiles; i++) |
| 420 |
FileClose(file->files[i]); |
| 421 |
/* release the buffer space */ |
| 422 |
pfree(file->files); |
| 423 |
pfree(file); |
| 424 |
} |
| 425 |
|
| 426 |
/* |
| 427 |
* BufFileLoadBuffer |
| 428 |
* |
| 429 |
* Load some data into buffer, if possible, starting from curOffset. |
| 430 |
* At call, must have dirty = false, pos and nbytes = 0. |
| 431 |
* On exit, nbytes is number of bytes loaded. |
| 432 |
*/ |
| 433 |
static void |
| 434 |
BufFileLoadBuffer(BufFile *file) |
| 435 |
{ |
| 436 |
File thisfile; |
| 437 |
instr_time io_start; |
| 438 |
instr_time io_time; |
| 439 |
|
| 440 |
/* |
| 441 |
* Advance to next component file if necessary and possible. |
| 442 |
*/ |
| 443 |
if (file->curOffset >= MAX_PHYSICAL_FILESIZE && |
| 444 |
file->curFile + 1 < file->numFiles) |
| 445 |
{ |
| 446 |
file->curFile++; |
| 447 |
file->curOffset = 0; |
| 448 |
} |
| 449 |
|
| 450 |
thisfile = file->files[file->curFile]; |
| 451 |
|
| 452 |
if (track_io_timing) |
| 453 |
INSTR_TIME_SET_CURRENT(io_start); |
| 454 |
else |
| 455 |
INSTR_TIME_SET_ZERO(io_start); |
| 456 |
|
| 457 |
/* |
| 458 |
* Read whatever we can get, up to a full bufferload. |
| 459 |
*/ |
| 460 |
file->nbytes = FileRead(thisfile, |
| 461 |
file->buffer.data, |
| 462 |
sizeof(file->buffer.data), |
| 463 |
file->curOffset, |
| 464 |
WAIT_EVENT_BUFFILE_READ); |
| 465 |
if (file->nbytes < 0) |
| 466 |
{ |
| 467 |
file->nbytes = 0; |
| 468 |
ereport(ERROR, |
| 469 |
(errcode_for_file_access(), |
| 470 |
errmsg("could not read file \"%s\": %m", |
| 471 |
FilePathName(thisfile)))); |
| 472 |
} |
| 473 |
|
| 474 |
if (track_io_timing) |
| 475 |
{ |
| 476 |
INSTR_TIME_SET_CURRENT(io_time); |
| 477 |
INSTR_TIME_ACCUM_DIFF(pgBufferUsage.temp_blk_read_time, io_time, io_start); |
| 478 |
} |
| 479 |
|
| 480 |
/* we choose not to advance curOffset here */ |
| 481 |
|
| 482 |
if (file->nbytes > 0) |
| 483 |
pgBufferUsage.temp_blks_read++; |
| 484 |
} |
| 485 |
|
| 486 |
/* |
| 487 |
* BufFileDumpBuffer |
| 488 |
* |
| 489 |
* Dump buffer contents starting at curOffset. |
| 490 |
* At call, should have dirty = true, nbytes > 0. |
| 491 |
* On exit, dirty is cleared if successful write, and curOffset is advanced. |
| 492 |
*/ |
| 493 |
static void |
| 494 |
BufFileDumpBuffer(BufFile *file) |
| 495 |
{ |
| 496 |
int wpos = 0; |
| 497 |
int bytestowrite; |
| 498 |
File thisfile; |
| 499 |
|
| 500 |
/* |
| 501 |
* Unlike BufFileLoadBuffer, we must dump the whole buffer even if it |
| 502 |
* crosses a component-file boundary; so we need a loop. |
| 503 |
*/ |
| 504 |
while (wpos < file->nbytes) |
| 505 |
{ |
| 506 |
off_t availbytes; |
| 507 |
instr_time io_start; |
| 508 |
instr_time io_time; |
| 509 |
|
| 510 |
/* |
| 511 |
* Advance to next component file if necessary and possible. |
| 512 |
*/ |
| 513 |
if (file->curOffset >= MAX_PHYSICAL_FILESIZE) |
| 514 |
{ |
| 515 |
while (file->curFile + 1 >= file->numFiles) |
| 516 |
extendBufFile(file); |
| 517 |
file->curFile++; |
| 518 |
file->curOffset = 0; |
| 519 |
} |
| 520 |
|
| 521 |
/* |
| 522 |
* Determine how much we need to write into this file. |
| 523 |
*/ |
| 524 |
bytestowrite = file->nbytes - wpos; |
| 525 |
availbytes = MAX_PHYSICAL_FILESIZE - file->curOffset; |
| 526 |
|
| 527 |
if ((off_t) bytestowrite > availbytes) |
| 528 |
bytestowrite = (int) availbytes; |
| 529 |
|
| 530 |
thisfile = file->files[file->curFile]; |
| 531 |
|
| 532 |
if (track_io_timing) |
| 533 |
INSTR_TIME_SET_CURRENT(io_start); |
| 534 |
else |
| 535 |
INSTR_TIME_SET_ZERO(io_start); |
| 536 |
|
| 537 |
bytestowrite = FileWrite(thisfile, |
| 538 |
file->buffer.data + wpos, |
| 539 |
bytestowrite, |
| 540 |
file->curOffset, |
| 541 |
WAIT_EVENT_BUFFILE_WRITE); |
| 542 |
if (bytestowrite <= 0) |
| 543 |
ereport(ERROR, |
| 544 |
(errcode_for_file_access(), |
| 545 |
errmsg("could not write to file \"%s\": %m", |
| 546 |
FilePathName(thisfile)))); |
| 547 |
|
| 548 |
if (track_io_timing) |
| 549 |
{ |
| 550 |
INSTR_TIME_SET_CURRENT(io_time); |
| 551 |
INSTR_TIME_ACCUM_DIFF(pgBufferUsage.temp_blk_write_time, io_time, io_start); |
| 552 |
} |
| 553 |
|
| 554 |
file->curOffset += bytestowrite; |
| 555 |
wpos += bytestowrite; |
| 556 |
|
| 557 |
pgBufferUsage.temp_blks_written++; |
| 558 |
} |
| 559 |
file->dirty = false; |
| 560 |
|
| 561 |
/* |
| 562 |
* At this point, curOffset has been advanced to the end of the buffer, |
| 563 |
* ie, its original value + nbytes. We need to make it point to the |
| 564 |
* logical file position, ie, original value + pos, in case that is less |
| 565 |
* (as could happen due to a small backwards seek in a dirty buffer!) |
| 566 |
*/ |
| 567 |
file->curOffset -= (file->nbytes - file->pos); |
| 568 |
if (file->curOffset < 0) /* handle possible segment crossing */ |
| 569 |
{ |
| 570 |
file->curFile--; |
| 571 |
Assert(file->curFile >= 0); |
| 572 |
file->curOffset += MAX_PHYSICAL_FILESIZE; |
| 573 |
} |
| 574 |
|
| 575 |
/* |
| 576 |
* Now we can set the buffer empty without changing the logical position |
| 577 |
*/ |
| 578 |
file->pos = 0; |
| 579 |
file->nbytes = 0; |
| 580 |
} |
| 581 |
|
| 582 |
/* |
| 583 |
* BufFileRead variants |
| 584 |
* |
| 585 |
* Like fread() except we assume 1-byte element size and report I/O errors via |
| 586 |
* ereport(). |
| 587 |
* |
| 588 |
* If 'exact' is true, then an error is also raised if the number of bytes |
| 589 |
* read is not exactly 'size' (no short reads). If 'exact' and 'eofOK' are |
| 590 |
* true, then reading zero bytes is ok. |
| 591 |
*/ |
| 592 |
static size_t |
| 593 |
BufFileReadCommon(BufFile *file, void *ptr, size_t size, bool exact, bool eofOK) |
| 594 |
{ |
| 595 |
size_t start_size = size; |
| 596 |
size_t nread = 0; |
| 597 |
size_t nthistime; |
| 598 |
|
| 599 |
BufFileFlush(file); |
| 600 |
|
| 601 |
while (size > 0) |
| 602 |
{ |
| 603 |
if (file->pos >= file->nbytes) |
| 604 |
{ |
| 605 |
/* Try to load more data into buffer. */ |
| 606 |
file->curOffset += file->pos; |
| 607 |
file->pos = 0; |
| 608 |
file->nbytes = 0; |
| 609 |
BufFileLoadBuffer(file); |
| 610 |
if (file->nbytes <= 0) |
| 611 |
break; /* no more data available */ |
| 612 |
} |
| 613 |
|
| 614 |
nthistime = file->nbytes - file->pos; |
| 615 |
if (nthistime > size) |
| 616 |
nthistime = size; |
| 617 |
Assert(nthistime > 0); |
| 618 |
|
| 619 |
memcpy(ptr, file->buffer.data + file->pos, nthistime); |
| 620 |
|
| 621 |
file->pos += nthistime; |
| 622 |
ptr = (char *) ptr + nthistime; |
| 623 |
size -= nthistime; |
| 624 |
nread += nthistime; |
| 625 |
} |
| 626 |
|
| 627 |
if (exact && |
| 628 |
(nread != start_size && !(nread == 0 && eofOK))) |
| 629 |
ereport(ERROR, |
| 630 |
errcode_for_file_access(), |
| 631 |
file->name ? |
| 632 |
errmsg("could not read from file set \"%s\": read only %zu of %zu bytes", |
| 633 |
file->name, nread, start_size) : |
| 634 |
errmsg("could not read from temporary file: read only %zu of %zu bytes", |
| 635 |
nread, start_size)); |
| 636 |
|
| 637 |
return nread; |
| 638 |
} |
| 639 |
|
| 640 |
/* |
| 641 |
* Legacy interface where the caller needs to check for end of file or short |
| 642 |
* reads. |
| 643 |
*/ |
| 644 |
size_t |
| 645 |
BufFileRead(BufFile *file, void *ptr, size_t size) |
| 646 |
{ |
| 647 |
return BufFileReadCommon(file, ptr, size, false, false); |
| 648 |
} |
| 649 |
|
| 650 |
/* |
| 651 |
* Require read of exactly the specified size. |
| 652 |
*/ |
| 653 |
void |
| 654 |
BufFileReadExact(BufFile *file, void *ptr, size_t size) |
| 655 |
{ |
| 656 |
BufFileReadCommon(file, ptr, size, true, false); |
| 657 |
} |
| 658 |
|
| 659 |
/* |
| 660 |
* Require read of exactly the specified size, but optionally allow end of |
| 661 |
* file (in which case 0 is returned). |
| 662 |
*/ |
| 663 |
size_t |
| 664 |
BufFileReadMaybeEOF(BufFile *file, void *ptr, size_t size, bool eofOK) |
| 665 |
{ |
| 666 |
return BufFileReadCommon(file, ptr, size, true, eofOK); |
| 667 |
} |
| 668 |
|
| 669 |
/* |
| 670 |
* BufFileWrite |
| 671 |
* |
| 672 |
* Like fwrite() except we assume 1-byte element size and report errors via |
| 673 |
* ereport(). |
| 674 |
*/ |
| 675 |
void |
| 676 |
BufFileWrite(BufFile *file, const void *ptr, size_t size) |
| 677 |
{ |
| 678 |
size_t nthistime; |
| 679 |
|
| 680 |
Assert(!file->readOnly); |
| 681 |
|
| 682 |
while (size > 0) |
| 683 |
{ |
| 684 |
if (file->pos >= BLCKSZ) |
| 685 |
{ |
| 686 |
/* Buffer full, dump it out */ |
| 687 |
if (file->dirty) |
| 688 |
BufFileDumpBuffer(file); |
| 689 |
else |
| 690 |
{ |
| 691 |
/* Hmm, went directly from reading to writing? */ |
| 692 |
file->curOffset += file->pos; |
| 693 |
file->pos = 0; |
| 694 |
file->nbytes = 0; |
| 695 |
} |
| 696 |
} |
| 697 |
|
| 698 |
nthistime = BLCKSZ - file->pos; |
| 699 |
if (nthistime > size) |
| 700 |
nthistime = size; |
| 701 |
Assert(nthistime > 0); |
| 702 |
|
| 703 |
memcpy(file->buffer.data + file->pos, ptr, nthistime); |
| 704 |
|
| 705 |
file->dirty = true; |
| 706 |
file->pos += nthistime; |
| 707 |
if (file->nbytes < file->pos) |
| 708 |
file->nbytes = file->pos; |
| 709 |
ptr = (const char *) ptr + nthistime; |
| 710 |
size -= nthistime; |
| 711 |
} |
| 712 |
} |
| 713 |
|
| 714 |
/* |
| 715 |
* BufFileFlush |
| 716 |
* |
| 717 |
* Like fflush(), except that I/O errors are reported with ereport(). |
| 718 |
*/ |
| 719 |
static void |
| 720 |
BufFileFlush(BufFile *file) |
| 721 |
{ |
| 722 |
if (file->dirty) |
| 723 |
BufFileDumpBuffer(file); |
| 724 |
|
| 725 |
Assert(!file->dirty); |
| 726 |
} |
| 727 |
|
| 728 |
/* |
| 729 |
* BufFileSeek |
| 730 |
* |
| 731 |
* Like fseek(), except that target position needs two values in order to |
| 732 |
* work when logical filesize exceeds maximum value representable by off_t. |
| 733 |
* We do not support relative seeks across more than that, however. |
| 734 |
* I/O errors are reported by ereport(). |
| 735 |
* |
| 736 |
* Result is 0 if OK, EOF if not. Logical position is not moved if an |
| 737 |
* impossible seek is attempted. |
| 738 |
*/ |
| 739 |
int |
| 740 |
BufFileSeek(BufFile *file, int fileno, off_t offset, int whence) |
| 741 |
{ |
| 742 |
int newFile; |
| 743 |
off_t newOffset; |
| 744 |
|
| 745 |
switch (whence) |
| 746 |
{ |
| 747 |
case SEEK_SET: |
| 748 |
if (fileno < 0) |
| 749 |
return EOF; |
| 750 |
newFile = fileno; |
| 751 |
newOffset = offset; |
| 752 |
break; |
| 753 |
case SEEK_CUR: |
| 754 |
|
| 755 |
/* |
| 756 |
* Relative seek considers only the signed offset, ignoring |
| 757 |
* fileno. Note that large offsets (> 1 GB) risk overflow in this |
| 758 |
* add, unless we have 64-bit off_t. |
| 759 |
*/ |
| 760 |
newFile = file->curFile; |
| 761 |
newOffset = (file->curOffset + file->pos) + offset; |
| 762 |
break; |
| 763 |
case SEEK_END: |
| 764 |
|
| 765 |
/* |
| 766 |
* The file size of the last file gives us the end offset of that |
| 767 |
* file. |
| 768 |
*/ |
| 769 |
newFile = file->numFiles - 1; |
| 770 |
newOffset = FileSize(file->files[file->numFiles - 1]); |
| 771 |
if (newOffset < 0) |
| 772 |
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 |
break; |
| 778 |
default: |
| 779 |
elog(ERROR, "invalid whence: %d", whence); |
| 780 |
return EOF; |
| 781 |
} |
| 782 |
while (newOffset < 0) |
| 783 |
{ |
| 784 |
if (--newFile < 0) |
| 785 |
return EOF; |
| 786 |
newOffset += MAX_PHYSICAL_FILESIZE; |
| 787 |
} |
| 788 |
if (newFile == file->curFile && |
| 789 |
newOffset >= file->curOffset && |
| 790 |
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 |
file->pos = (int) (newOffset - file->curOffset); |
| 799 |
return 0; |
| 800 |
} |
| 801 |
/* Otherwise, must reposition buffer, so flush any dirty data */ |
| 802 |
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 |
if (newFile == file->numFiles && newOffset == 0) |
| 812 |
{ |
| 813 |
newFile--; |
| 814 |
newOffset = MAX_PHYSICAL_FILESIZE; |
| 815 |
} |
| 816 |
while (newOffset > MAX_PHYSICAL_FILESIZE) |
| 817 |
{ |
| 818 |
if (++newFile >= file->numFiles) |
| 819 |
return EOF; |
| 820 |
newOffset -= MAX_PHYSICAL_FILESIZE; |
| 821 |
} |
| 822 |
if (newFile >= file->numFiles) |
| 823 |
return EOF; |
| 824 |
/* Seek is OK! */ |
| 825 |
file->curFile = newFile; |
| 826 |
file->curOffset = newOffset; |
| 827 |
file->pos = 0; |
| 828 |
file->nbytes = 0; |
| 829 |
return 0; |
| 830 |
} |
| 831 |
|
| 832 |
void |
| 833 |
BufFileTell(BufFile *file, int *fileno, off_t *offset) |
| 834 |
{ |
| 835 |
*fileno = file->curFile; |
| 836 |
*offset = file->curOffset + file->pos; |
| 837 |
} |
| 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 |
BufFileSeekBlock(BufFile *file, int64 blknum) |
| 852 |
{ |
| 853 |
return BufFileSeek(file, |
| 854 |
(int) (blknum / BUFFILE_SEG_SIZE), |
| 855 |
(off_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 |
BufFileSize(BufFile *file) |
| 867 |
{ |
| 868 |
int64 lastFileSize; |
| 869 |
|
| 870 |
/* Get the size of the last physical file. */ |
| 871 |
lastFileSize = FileSize(file->files[file->numFiles - 1]); |
| 872 |
if (lastFileSize < 0) |
| 873 |
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 |
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 |
BufFileAppend(BufFile *target, BufFile *source) |
| 903 |
{ |
| 904 |
int64 startBlock = (int64) target->numFiles * BUFFILE_SEG_SIZE; |
| 905 |
int newNumFiles = target->numFiles + source->numFiles; |
| 906 |
int i; |
| 907 |
|
| 908 |
Assert(source->readOnly); |
| 909 |
Assert(!source->dirty); |
| 910 |
|
| 911 |
if (target->resowner != source->resowner) |
| 912 |
elog(ERROR, "could not append BufFile with non-matching resource owner"); |
| 913 |
|
| 914 |
target->files = (File *) |
| 915 |
repalloc(target->files, sizeof(File) * newNumFiles); |
| 916 |
for (i = target->numFiles; i < newNumFiles; i++) |
| 917 |
target->files[i] = source->files[i - target->numFiles]; |
| 918 |
target->numFiles = newNumFiles; |
| 919 |
|
| 920 |
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 |
BufFileTruncateFileSet(BufFile *file, int fileno, off_t offset) |
| 929 |
{ |
| 930 |
int numFiles = file->numFiles; |
| 931 |
int newFile = fileno; |
| 932 |
off_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 |
for (i = file->numFiles - 1; i >= fileno; i--) |
| 943 |
{ |
| 944 |
if ((i != fileno || offset == 0) && i != 0) |
| 945 |
{ |
| 946 |
FileSetSegmentName(segment_name, file->name, i); |
| 947 |
FileClose(file->files[i]); |
| 948 |
if (!FileSetDelete(file->fileset, segment_name, true)) |
| 949 |
ereport(ERROR, |
| 950 |
(errcode_for_file_access(), |
| 951 |
errmsg("could not delete fileset \"%s\": %m", |
| 952 |
segment_name))); |
| 953 |
numFiles--; |
| 954 |
newOffset = MAX_PHYSICAL_FILESIZE; |
| 955 |
|
| 956 |
/* |
| 957 |
* This is required to indicate that we have deleted the given |
| 958 |
* fileno. |
| 959 |
*/ |
| 960 |
if (i == fileno) |
| 961 |
newFile--; |
| 962 |
} |
| 963 |
else |
| 964 |
{ |
| 965 |
if (FileTruncate(file->files[i], offset, |
| 966 |
WAIT_EVENT_BUFFILE_TRUNCATE) < 0) |
| 967 |
ereport(ERROR, |
| 968 |
(errcode_for_file_access(), |
| 969 |
errmsg("could not truncate file \"%s\": %m", |
| 970 |
FilePathName(file->files[i])))); |
| 971 |
newOffset = offset; |
| 972 |
} |
| 973 |
} |
| 974 |
|
| 975 |
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 |
if (newFile == file->curFile && |
| 982 |
newOffset >= file->curOffset && |
| 983 |
newOffset <= file->curOffset + file->nbytes) |
| 984 |
{ |
| 985 |
/* No need to reset the current pos if the new pos is greater. */ |
| 986 |
if (newOffset <= file->curOffset + file->pos) |
| 987 |
file->pos = (int) (newOffset - file->curOffset); |
| 988 |
|
| 989 |
/* Adjust the nbytes for the current buffer. */ |
| 990 |
file->nbytes = (int) (newOffset - file->curOffset); |
| 991 |
} |
| 992 |
else if (newFile == file->curFile && |
| 993 |
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 |
file->curOffset = newOffset; |
| 1001 |
file->pos = 0; |
| 1002 |
file->nbytes = 0; |
| 1003 |
} |
| 1004 |
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 |
file->curFile = newFile; |
| 1011 |
file->curOffset = newOffset; |
| 1012 |
file->pos = 0; |
| 1013 |
file->nbytes = 0; |
| 1014 |
} |
| 1015 |
/* Nothing to do, if the truncate point is beyond current file. */ |
| 1016 |
} |