| 1 |
/*------------------------------------------------------------------------- |
| 2 |
* |
| 3 |
* reinit.c |
| 4 |
* Reinitialization of unlogged relations |
| 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/reinit.c |
| 11 |
* |
| 12 |
*------------------------------------------------------------------------- |
| 13 |
*/ |
| 14 |
|
| 15 |
#include "postgres.h" |
| 16 |
|
| 17 |
#include <unistd.h> |
| 18 |
|
| 19 |
#include "common/relpath.h" |
| 20 |
#include "postmaster/startup.h" |
| 21 |
#include "storage/copydir.h" |
| 22 |
#include "storage/fd.h" |
| 23 |
#include "storage/reinit.h" |
| 24 |
#include "utils/hsearch.h" |
| 25 |
#include "utils/memutils.h" |
| 26 |
|
| 27 |
static void ResetUnloggedRelationsInTablespaceDir(const char *tsdirname, |
| 28 |
int op); |
| 29 |
static void ResetUnloggedRelationsInDbspaceDir(const char *dbspacedirname, |
| 30 |
int op); |
| 31 |
|
| 32 |
typedef struct |
| 33 |
{ |
| 34 |
RelFileNumber relnumber; /* hash key */ |
| 35 |
} unlogged_relation_entry; |
| 36 |
|
| 37 |
/* |
| 38 |
* Reset unlogged relations from before the last restart. |
| 39 |
* |
| 40 |
* If op includes UNLOGGED_RELATION_CLEANUP, we remove all forks of any |
| 41 |
* relation with an "init" fork, except for the "init" fork itself. |
| 42 |
* |
| 43 |
* If op includes UNLOGGED_RELATION_INIT, we copy the "init" fork to the main |
| 44 |
* fork. |
| 45 |
*/ |
| 46 |
void |
| 47 |
ResetUnloggedRelations(int op) |
| 48 |
{ |
| 49 |
char temp_path[MAXPGPATH + sizeof(PG_TBLSPC_DIR) + sizeof(TABLESPACE_VERSION_DIRECTORY)]; |
| 50 |
DIR *spc_dir; |
| 51 |
struct dirent *spc_de; |
| 52 |
MemoryContext tmpctx, |
| 53 |
oldctx; |
| 54 |
|
| 55 |
/* Log it. */ |
| 56 |
elog(DEBUG1, "resetting unlogged relations: cleanup %d init %d", |
| 57 |
(op & UNLOGGED_RELATION_CLEANUP) != 0, |
| 58 |
(op & UNLOGGED_RELATION_INIT) != 0); |
| 59 |
|
| 60 |
/* |
| 61 |
* Just to be sure we don't leak any memory, let's create a temporary |
| 62 |
* memory context for this operation. |
| 63 |
*/ |
| 64 |
tmpctx = AllocSetContextCreate(CurrentMemoryContext, |
| 65 |
"ResetUnloggedRelations", |
| 66 |
ALLOCSET_DEFAULT_SIZES); |
| 67 |
oldctx = MemoryContextSwitchTo(tmpctx); |
| 68 |
|
| 69 |
/* Prepare to report progress resetting unlogged relations. */ |
| 70 |
begin_startup_progress_phase(); |
| 71 |
|
| 72 |
/* |
| 73 |
* First process unlogged files in pg_default ($PGDATA/base) |
| 74 |
*/ |
| 75 |
ResetUnloggedRelationsInTablespaceDir("base", op); |
| 76 |
|
| 77 |
/* |
| 78 |
* Cycle through directories for all non-default tablespaces. |
| 79 |
*/ |
| 80 |
spc_dir = AllocateDir(PG_TBLSPC_DIR); |
| 81 |
|
| 82 |
while ((spc_de = ReadDir(spc_dir, PG_TBLSPC_DIR)) != NULL) |
| 83 |
{ |
| 84 |
if (strcmp(spc_de->d_name, ".") == 0 || |
| 85 |
strcmp(spc_de->d_name, "..") == 0) |
| 86 |
continue; |
| 87 |
|
| 88 |
snprintf(temp_path, sizeof(temp_path), "%s/%s/%s", |
| 89 |
PG_TBLSPC_DIR, spc_de->d_name, TABLESPACE_VERSION_DIRECTORY); |
| 90 |
ResetUnloggedRelationsInTablespaceDir(temp_path, op); |
| 91 |
} |
| 92 |
|
| 93 |
FreeDir(spc_dir); |
| 94 |
|
| 95 |
/* |
| 96 |
* Restore memory context. |
| 97 |
*/ |
| 98 |
MemoryContextSwitchTo(oldctx); |
| 99 |
MemoryContextDelete(tmpctx); |
| 100 |
} |
| 101 |
|
| 102 |
/* |
| 103 |
* Process one tablespace directory for ResetUnloggedRelations |
| 104 |
*/ |
| 105 |
static void |
| 106 |
ResetUnloggedRelationsInTablespaceDir(const char *tsdirname, int op) |
| 107 |
{ |
| 108 |
DIR *ts_dir; |
| 109 |
struct dirent *de; |
| 110 |
char dbspace_path[MAXPGPATH * 2]; |
| 111 |
|
| 112 |
ts_dir = AllocateDir(tsdirname); |
| 113 |
|
| 114 |
/* |
| 115 |
* If we get ENOENT on a tablespace directory, log it and return. This |
| 116 |
* can happen if a previous DROP TABLESPACE crashed between removing the |
| 117 |
* tablespace directory and removing the symlink in pg_tblspc. We don't |
| 118 |
* really want to prevent database startup in that scenario, so let it |
| 119 |
* pass instead. Any other type of error will be reported by ReadDir |
| 120 |
* (causing a startup failure). |
| 121 |
*/ |
| 122 |
if (ts_dir == NULL && errno == ENOENT) |
| 123 |
{ |
| 124 |
ereport(LOG, |
| 125 |
(errcode_for_file_access(), |
| 126 |
errmsg("could not open directory \"%s\": %m", |
| 127 |
tsdirname))); |
| 128 |
return; |
| 129 |
} |
| 130 |
|
| 131 |
while ((de = ReadDir(ts_dir, tsdirname)) != NULL) |
| 132 |
{ |
| 133 |
/* |
| 134 |
* We're only interested in the per-database directories, which have |
| 135 |
* numeric names. Note that this code will also (properly) ignore "." |
| 136 |
* and "..". |
| 137 |
*/ |
| 138 |
if (strspn(de->d_name, "0123456789") != strlen(de->d_name)) |
| 139 |
continue; |
| 140 |
|
| 141 |
snprintf(dbspace_path, sizeof(dbspace_path), "%s/%s", |
| 142 |
tsdirname, de->d_name); |
| 143 |
|
| 144 |
if (op & UNLOGGED_RELATION_INIT) |
| 145 |
ereport_startup_progress("resetting unlogged relations (init), elapsed time: %ld.%02d s, current path: %s", |
| 146 |
dbspace_path); |
| 147 |
else if (op & UNLOGGED_RELATION_CLEANUP) |
| 148 |
ereport_startup_progress("resetting unlogged relations (cleanup), elapsed time: %ld.%02d s, current path: %s", |
| 149 |
dbspace_path); |
| 150 |
|
| 151 |
ResetUnloggedRelationsInDbspaceDir(dbspace_path, op); |
| 152 |
} |
| 153 |
|
| 154 |
FreeDir(ts_dir); |
| 155 |
} |
| 156 |
|
| 157 |
/* |
| 158 |
* Process one per-dbspace directory for ResetUnloggedRelations |
| 159 |
*/ |
| 160 |
static void |
| 161 |
ResetUnloggedRelationsInDbspaceDir(const char *dbspacedirname, int op) |
| 162 |
{ |
| 163 |
DIR *dbspace_dir; |
| 164 |
struct dirent *de; |
| 165 |
char rm_path[MAXPGPATH * 2]; |
| 166 |
|
| 167 |
/* Caller must specify at least one operation. */ |
| 168 |
Assert((op & (UNLOGGED_RELATION_CLEANUP | UNLOGGED_RELATION_INIT)) != 0); |
| 169 |
|
| 170 |
/* |
| 171 |
* Cleanup is a two-pass operation. First, we go through and identify all |
| 172 |
* the files with init forks. Then, we go through again and nuke |
| 173 |
* everything with the same OID except the init fork. |
| 174 |
*/ |
| 175 |
if ((op & UNLOGGED_RELATION_CLEANUP) != 0) |
| 176 |
{ |
| 177 |
HTAB *hash; |
| 178 |
HASHCTL ctl; |
| 179 |
|
| 180 |
/* |
| 181 |
* It's possible that someone could create a ton of unlogged relations |
| 182 |
* in the same database & tablespace, so we'd better use a hash table |
| 183 |
* rather than an array or linked list to keep track of which files |
| 184 |
* need to be reset. Otherwise, this cleanup operation would be |
| 185 |
* O(n^2). |
| 186 |
*/ |
| 187 |
ctl.keysize = sizeof(Oid); |
| 188 |
ctl.entrysize = sizeof(unlogged_relation_entry); |
| 189 |
ctl.hcxt = CurrentMemoryContext; |
| 190 |
hash = hash_create("unlogged relation OIDs", 32, &ctl, |
| 191 |
HASH_ELEM | HASH_BLOBS | HASH_CONTEXT); |
| 192 |
|
| 193 |
/* Scan the directory. */ |
| 194 |
dbspace_dir = AllocateDir(dbspacedirname); |
| 195 |
while ((de = ReadDir(dbspace_dir, dbspacedirname)) != NULL) |
| 196 |
{ |
| 197 |
ForkNumber forkNum; |
| 198 |
unsigned segno; |
| 199 |
unlogged_relation_entry ent; |
| 200 |
|
| 201 |
/* Skip anything that doesn't look like a relation data file. */ |
| 202 |
if (!parse_filename_for_nontemp_relation(de->d_name, |
| 203 |
&ent.relnumber, |
| 204 |
&forkNum, &segno)) |
| 205 |
continue; |
| 206 |
|
| 207 |
/* Also skip it unless this is the init fork. */ |
| 208 |
if (forkNum != INIT_FORKNUM) |
| 209 |
continue; |
| 210 |
|
| 211 |
/* |
| 212 |
* Put the RelFileNumber into the hash table, if it isn't already. |
| 213 |
*/ |
| 214 |
(void) hash_search(hash, &ent, HASH_ENTER, NULL); |
| 215 |
} |
| 216 |
|
| 217 |
/* Done with the first pass. */ |
| 218 |
FreeDir(dbspace_dir); |
| 219 |
|
| 220 |
/* |
| 221 |
* If we didn't find any init forks, there's no point in continuing; |
| 222 |
* we can bail out now. |
| 223 |
*/ |
| 224 |
if (hash_get_num_entries(hash) == 0) |
| 225 |
{ |
| 226 |
hash_destroy(hash); |
| 227 |
return; |
| 228 |
} |
| 229 |
|
| 230 |
/* |
| 231 |
* Now, make a second pass and remove anything that matches. |
| 232 |
*/ |
| 233 |
dbspace_dir = AllocateDir(dbspacedirname); |
| 234 |
while ((de = ReadDir(dbspace_dir, dbspacedirname)) != NULL) |
| 235 |
{ |
| 236 |
ForkNumber forkNum; |
| 237 |
unsigned segno; |
| 238 |
unlogged_relation_entry ent; |
| 239 |
|
| 240 |
/* Skip anything that doesn't look like a relation data file. */ |
| 241 |
if (!parse_filename_for_nontemp_relation(de->d_name, |
| 242 |
&ent.relnumber, |
| 243 |
&forkNum, &segno)) |
| 244 |
continue; |
| 245 |
|
| 246 |
/* We never remove the init fork. */ |
| 247 |
if (forkNum == INIT_FORKNUM) |
| 248 |
continue; |
| 249 |
|
| 250 |
/* |
| 251 |
* See whether the OID portion of the name shows up in the hash |
| 252 |
* table. If so, nuke it! |
| 253 |
*/ |
| 254 |
if (hash_search(hash, &ent, HASH_FIND, NULL)) |
| 255 |
{ |
| 256 |
snprintf(rm_path, sizeof(rm_path), "%s/%s", |
| 257 |
dbspacedirname, de->d_name); |
| 258 |
if (unlink(rm_path) < 0) |
| 259 |
ereport(ERROR, |
| 260 |
(errcode_for_file_access(), |
| 261 |
errmsg("could not remove file \"%s\": %m", |
| 262 |
rm_path))); |
| 263 |
else |
| 264 |
elog(DEBUG2, "unlinked file \"%s\"", rm_path); |
| 265 |
} |
| 266 |
} |
| 267 |
|
| 268 |
/* Cleanup is complete. */ |
| 269 |
FreeDir(dbspace_dir); |
| 270 |
hash_destroy(hash); |
| 271 |
} |
| 272 |
|
| 273 |
/* |
| 274 |
* Initialization happens after cleanup is complete: we copy each init |
| 275 |
* fork file to the corresponding main fork file. Note that if we are |
| 276 |
* asked to do both cleanup and init, we may never get here: if the |
| 277 |
* cleanup code determines that there are no init forks in this dbspace, |
| 278 |
* it will return before we get to this point. |
| 279 |
*/ |
| 280 |
if ((op & UNLOGGED_RELATION_INIT) != 0) |
| 281 |
{ |
| 282 |
/* Scan the directory. */ |
| 283 |
dbspace_dir = AllocateDir(dbspacedirname); |
| 284 |
while ((de = ReadDir(dbspace_dir, dbspacedirname)) != NULL) |
| 285 |
{ |
| 286 |
ForkNumber forkNum; |
| 287 |
RelFileNumber relNumber; |
| 288 |
unsigned segno; |
| 289 |
char srcpath[MAXPGPATH * 2]; |
| 290 |
char dstpath[MAXPGPATH]; |
| 291 |
|
| 292 |
/* Skip anything that doesn't look like a relation data file. */ |
| 293 |
if (!parse_filename_for_nontemp_relation(de->d_name, &relNumber, |
| 294 |
&forkNum, &segno)) |
| 295 |
continue; |
| 296 |
|
| 297 |
/* Also skip it unless this is the init fork. */ |
| 298 |
if (forkNum != INIT_FORKNUM) |
| 299 |
continue; |
| 300 |
|
| 301 |
/* Construct source pathname. */ |
| 302 |
snprintf(srcpath, sizeof(srcpath), "%s/%s", |
| 303 |
dbspacedirname, de->d_name); |
| 304 |
|
| 305 |
/* Construct destination pathname. */ |
| 306 |
if (segno == 0) |
| 307 |
snprintf(dstpath, sizeof(dstpath), "%s/%u", |
| 308 |
dbspacedirname, relNumber); |
| 309 |
else |
| 310 |
snprintf(dstpath, sizeof(dstpath), "%s/%u.%u", |
| 311 |
dbspacedirname, relNumber, segno); |
| 312 |
|
| 313 |
/* OK, we're ready to perform the actual copy. */ |
| 314 |
elog(DEBUG2, "copying %s to %s", srcpath, dstpath); |
| 315 |
copy_file(srcpath, dstpath); |
| 316 |
} |
| 317 |
|
| 318 |
FreeDir(dbspace_dir); |
| 319 |
|
| 320 |
/* |
| 321 |
* copy_file() above has already called pg_flush_data() on the files |
| 322 |
* it created. Now we need to fsync those files, because a checkpoint |
| 323 |
* won't do it for us while we're in recovery. We do this in a |
| 324 |
* separate pass to allow the kernel to perform all the flushes |
| 325 |
* (especially the metadata ones) at once. |
| 326 |
*/ |
| 327 |
dbspace_dir = AllocateDir(dbspacedirname); |
| 328 |
while ((de = ReadDir(dbspace_dir, dbspacedirname)) != NULL) |
| 329 |
{ |
| 330 |
RelFileNumber relNumber; |
| 331 |
ForkNumber forkNum; |
| 332 |
unsigned segno; |
| 333 |
char mainpath[MAXPGPATH]; |
| 334 |
|
| 335 |
/* Skip anything that doesn't look like a relation data file. */ |
| 336 |
if (!parse_filename_for_nontemp_relation(de->d_name, &relNumber, |
| 337 |
&forkNum, &segno)) |
| 338 |
continue; |
| 339 |
|
| 340 |
/* Also skip it unless this is the init fork. */ |
| 341 |
if (forkNum != INIT_FORKNUM) |
| 342 |
continue; |
| 343 |
|
| 344 |
/* Construct main fork pathname. */ |
| 345 |
if (segno == 0) |
| 346 |
snprintf(mainpath, sizeof(mainpath), "%s/%u", |
| 347 |
dbspacedirname, relNumber); |
| 348 |
else |
| 349 |
snprintf(mainpath, sizeof(mainpath), "%s/%u.%u", |
| 350 |
dbspacedirname, relNumber, segno); |
| 351 |
|
| 352 |
fsync_fname(mainpath, false); |
| 353 |
} |
| 354 |
|
| 355 |
FreeDir(dbspace_dir); |
| 356 |
|
| 357 |
/* |
| 358 |
* Lastly, fsync the database directory itself, ensuring the |
| 359 |
* filesystem remembers the file creations and deletions we've done. |
| 360 |
* We don't bother with this during a call that does only |
| 361 |
* UNLOGGED_RELATION_CLEANUP, because if recovery crashes before we |
| 362 |
* get to doing UNLOGGED_RELATION_INIT, we'll redo the cleanup step |
| 363 |
* too at the next startup attempt. |
| 364 |
*/ |
| 365 |
fsync_fname(dbspacedirname, true); |
| 366 |
} |
| 367 |
} |
| 368 |
|
| 369 |
/* |
| 370 |
* Basic parsing of putative relation filenames. |
| 371 |
* |
| 372 |
* This function returns true if the file appears to be in the correct format |
| 373 |
* for a non-temporary relation and false otherwise. |
| 374 |
* |
| 375 |
* If it returns true, it sets *relnumber, *fork, and *segno to the values |
| 376 |
* extracted from the filename. If it returns false, these values are set to |
| 377 |
* InvalidRelFileNumber, InvalidForkNumber, and 0, respectively. |
| 378 |
*/ |
| 379 |
bool |
| 380 |
parse_filename_for_nontemp_relation(const char *name, RelFileNumber *relnumber, |
| 381 |
ForkNumber *fork, unsigned *segno) |
| 382 |
{ |
| 383 |
unsigned long n, |
| 384 |
s; |
| 385 |
ForkNumber f; |
| 386 |
char *endp; |
| 387 |
|
| 388 |
*relnumber = InvalidRelFileNumber; |
| 389 |
*fork = InvalidForkNumber; |
| 390 |
*segno = 0; |
| 391 |
|
| 392 |
/* |
| 393 |
* Relation filenames should begin with a digit that is not a zero. By |
| 394 |
* rejecting cases involving leading zeroes, the caller can assume that |
| 395 |
* there's only one possible string of characters that could have produced |
| 396 |
* any given value for *relnumber. |
| 397 |
* |
| 398 |
* (To be clear, we don't expect files with names like 0017.3 to exist at |
| 399 |
* all -- but if 0017.3 does exist, it's a non-relation file, not part of |
| 400 |
* the main fork for relfilenode 17.) |
| 401 |
*/ |
| 402 |
if (name[0] < '1' || name[0] > '9') |
| 403 |
return false; |
| 404 |
|
| 405 |
/* |
| 406 |
* Parse the leading digit string. If the value is out of range, we |
| 407 |
* conclude that this isn't a relation file at all. |
| 408 |
*/ |
| 409 |
errno = 0; |
| 410 |
n = strtoul(name, &endp, 10); |
| 411 |
if (errno || name == endp || n <= 0 || n > PG_UINT32_MAX) |
| 412 |
return false; |
| 413 |
name = endp; |
| 414 |
|
| 415 |
/* Check for a fork name. */ |
| 416 |
if (*name != '_') |
| 417 |
f = MAIN_FORKNUM; |
| 418 |
else |
| 419 |
{ |
| 420 |
int forkchar; |
| 421 |
|
| 422 |
forkchar = forkname_chars(name + 1, &f); |
| 423 |
if (forkchar <= 0) |
| 424 |
return false; |
| 425 |
name += forkchar + 1; |
| 426 |
} |
| 427 |
|
| 428 |
/* Check for a segment number. */ |
| 429 |
if (*name != '.') |
| 430 |
s = 0; |
| 431 |
else |
| 432 |
{ |
| 433 |
/* Reject leading zeroes, just like we do for RelFileNumber. */ |
| 434 |
if (name[1] < '1' || name[1] > '9') |
| 435 |
return false; |
| 436 |
|
| 437 |
errno = 0; |
| 438 |
s = strtoul(name + 1, &endp, 10); |
| 439 |
if (errno || name + 1 == endp || s <= 0 || s > PG_UINT32_MAX) |
| 440 |
return false; |
| 441 |
name = endp; |
| 442 |
} |
| 443 |
|
| 444 |
/* Now we should be at the end. */ |
| 445 |
if (*name != '\0') |
| 446 |
return false; |
| 447 |
|
| 448 |
/* Set out parameters and return. */ |
| 449 |
*relnumber = (RelFileNumber) n; |
| 450 |
*fork = f; |
| 451 |
*segno = (unsigned) s; |
| 452 |
return true; |
| 453 |
} |