| 1 | /*------------------------------------------------------------------------- |
| 2 | * |
| 3 | * parsexlog.c |
| 4 | * Functions for reading Write-Ahead-Log |
| 5 | * |
| 6 | * Portions Copyright (c) 1996-2019, PostgreSQL Global Development Group |
| 7 | * Portions Copyright (c) 1994, Regents of the University of California |
| 8 | * |
| 9 | *------------------------------------------------------------------------- |
| 10 | */ |
| 11 | |
| 12 | #include "postgres_fe.h" |
| 13 | |
| 14 | #include <unistd.h> |
| 15 | |
| 16 | #include "pg_rewind.h" |
| 17 | #include "filemap.h" |
| 18 | |
| 19 | #include "access/rmgr.h" |
| 20 | #include "access/xlog_internal.h" |
| 21 | #include "access/xlogreader.h" |
| 22 | #include "catalog/pg_control.h" |
| 23 | #include "catalog/storage_xlog.h" |
| 24 | #include "commands/dbcommands_xlog.h" |
| 25 | |
| 26 | |
| 27 | /* |
| 28 | * RmgrNames is an array of resource manager names, to make error messages |
| 29 | * a bit nicer. |
| 30 | */ |
| 31 | #define PG_RMGR(symname,name,redo,desc,identify,startup,cleanup,mask) \ |
| 32 | name, |
| 33 | |
| 34 | static const char *RmgrNames[RM_MAX_ID + 1] = { |
| 35 | #include "access/rmgrlist.h" |
| 36 | }; |
| 37 | |
| 38 | static void extractPageInfo(XLogReaderState *record); |
| 39 | |
| 40 | static int xlogreadfd = -1; |
| 41 | static XLogSegNo xlogreadsegno = -1; |
| 42 | static char xlogfpath[MAXPGPATH]; |
| 43 | |
| 44 | typedef struct |
| 45 | { |
| 46 | const char *; |
| 47 | int ; |
| 48 | } ; |
| 49 | |
| 50 | static int SimpleXLogPageRead(XLogReaderState *xlogreader, |
| 51 | XLogRecPtr targetPagePtr, |
| 52 | int reqLen, XLogRecPtr targetRecPtr, char *readBuf, |
| 53 | TimeLineID *pageTLI); |
| 54 | |
| 55 | /* |
| 56 | * Read WAL from the datadir/pg_wal, starting from 'startpoint' on timeline |
| 57 | * index 'tliIndex' in target timeline history, until 'endpoint'. Make note of |
| 58 | * the data blocks touched by the WAL records, and return them in a page map. |
| 59 | */ |
| 60 | void |
| 61 | (const char *datadir, XLogRecPtr startpoint, int tliIndex, |
| 62 | XLogRecPtr endpoint) |
| 63 | { |
| 64 | XLogRecord *record; |
| 65 | XLogReaderState *xlogreader; |
| 66 | char *errormsg; |
| 67 | XLogPageReadPrivate private; |
| 68 | |
| 69 | private.datadir = datadir; |
| 70 | private.tliIndex = tliIndex; |
| 71 | xlogreader = XLogReaderAllocate(WalSegSz, &SimpleXLogPageRead, |
| 72 | &private); |
| 73 | if (xlogreader == NULL) |
| 74 | pg_fatal("out of memory" ); |
| 75 | |
| 76 | do |
| 77 | { |
| 78 | record = XLogReadRecord(xlogreader, startpoint, &errormsg); |
| 79 | |
| 80 | if (record == NULL) |
| 81 | { |
| 82 | XLogRecPtr errptr; |
| 83 | |
| 84 | errptr = startpoint ? startpoint : xlogreader->EndRecPtr; |
| 85 | |
| 86 | if (errormsg) |
| 87 | pg_fatal("could not read WAL record at %X/%X: %s" , |
| 88 | (uint32) (errptr >> 32), (uint32) (errptr), |
| 89 | errormsg); |
| 90 | else |
| 91 | pg_fatal("could not read WAL record at %X/%X" , |
| 92 | (uint32) (errptr >> 32), (uint32) (errptr)); |
| 93 | } |
| 94 | |
| 95 | extractPageInfo(xlogreader); |
| 96 | |
| 97 | startpoint = InvalidXLogRecPtr; /* continue reading at next record */ |
| 98 | |
| 99 | } while (xlogreader->ReadRecPtr != endpoint); |
| 100 | |
| 101 | XLogReaderFree(xlogreader); |
| 102 | if (xlogreadfd != -1) |
| 103 | { |
| 104 | close(xlogreadfd); |
| 105 | xlogreadfd = -1; |
| 106 | } |
| 107 | } |
| 108 | |
| 109 | /* |
| 110 | * Reads one WAL record. Returns the end position of the record, without |
| 111 | * doing anything with the record itself. |
| 112 | */ |
| 113 | XLogRecPtr |
| 114 | readOneRecord(const char *datadir, XLogRecPtr ptr, int tliIndex) |
| 115 | { |
| 116 | XLogRecord *record; |
| 117 | XLogReaderState *xlogreader; |
| 118 | char *errormsg; |
| 119 | XLogPageReadPrivate private; |
| 120 | XLogRecPtr endptr; |
| 121 | |
| 122 | private.datadir = datadir; |
| 123 | private.tliIndex = tliIndex; |
| 124 | xlogreader = XLogReaderAllocate(WalSegSz, &SimpleXLogPageRead, |
| 125 | &private); |
| 126 | if (xlogreader == NULL) |
| 127 | pg_fatal("out of memory" ); |
| 128 | |
| 129 | record = XLogReadRecord(xlogreader, ptr, &errormsg); |
| 130 | if (record == NULL) |
| 131 | { |
| 132 | if (errormsg) |
| 133 | pg_fatal("could not read WAL record at %X/%X: %s" , |
| 134 | (uint32) (ptr >> 32), (uint32) (ptr), errormsg); |
| 135 | else |
| 136 | pg_fatal("could not read WAL record at %X/%X" , |
| 137 | (uint32) (ptr >> 32), (uint32) (ptr)); |
| 138 | } |
| 139 | endptr = xlogreader->EndRecPtr; |
| 140 | |
| 141 | XLogReaderFree(xlogreader); |
| 142 | if (xlogreadfd != -1) |
| 143 | { |
| 144 | close(xlogreadfd); |
| 145 | xlogreadfd = -1; |
| 146 | } |
| 147 | |
| 148 | return endptr; |
| 149 | } |
| 150 | |
| 151 | /* |
| 152 | * Find the previous checkpoint preceding given WAL location. |
| 153 | */ |
| 154 | void |
| 155 | findLastCheckpoint(const char *datadir, XLogRecPtr forkptr, int tliIndex, |
| 156 | XLogRecPtr *lastchkptrec, TimeLineID *lastchkpttli, |
| 157 | XLogRecPtr *lastchkptredo) |
| 158 | { |
| 159 | /* Walk backwards, starting from the given record */ |
| 160 | XLogRecord *record; |
| 161 | XLogRecPtr searchptr; |
| 162 | XLogReaderState *xlogreader; |
| 163 | char *errormsg; |
| 164 | XLogPageReadPrivate private; |
| 165 | |
| 166 | /* |
| 167 | * The given fork pointer points to the end of the last common record, |
| 168 | * which is not necessarily the beginning of the next record, if the |
| 169 | * previous record happens to end at a page boundary. Skip over the page |
| 170 | * header in that case to find the next record. |
| 171 | */ |
| 172 | if (forkptr % XLOG_BLCKSZ == 0) |
| 173 | { |
| 174 | if (XLogSegmentOffset(forkptr, WalSegSz) == 0) |
| 175 | forkptr += SizeOfXLogLongPHD; |
| 176 | else |
| 177 | forkptr += SizeOfXLogShortPHD; |
| 178 | } |
| 179 | |
| 180 | private.datadir = datadir; |
| 181 | private.tliIndex = tliIndex; |
| 182 | xlogreader = XLogReaderAllocate(WalSegSz, &SimpleXLogPageRead, |
| 183 | &private); |
| 184 | if (xlogreader == NULL) |
| 185 | pg_fatal("out of memory" ); |
| 186 | |
| 187 | searchptr = forkptr; |
| 188 | for (;;) |
| 189 | { |
| 190 | uint8 info; |
| 191 | |
| 192 | record = XLogReadRecord(xlogreader, searchptr, &errormsg); |
| 193 | |
| 194 | if (record == NULL) |
| 195 | { |
| 196 | if (errormsg) |
| 197 | pg_fatal("could not find previous WAL record at %X/%X: %s" , |
| 198 | (uint32) (searchptr >> 32), (uint32) (searchptr), |
| 199 | errormsg); |
| 200 | else |
| 201 | pg_fatal("could not find previous WAL record at %X/%X" , |
| 202 | (uint32) (searchptr >> 32), (uint32) (searchptr)); |
| 203 | } |
| 204 | |
| 205 | /* |
| 206 | * Check if it is a checkpoint record. This checkpoint record needs to |
| 207 | * be the latest checkpoint before WAL forked and not the checkpoint |
| 208 | * where the master has been stopped to be rewinded. |
| 209 | */ |
| 210 | info = XLogRecGetInfo(xlogreader) & ~XLR_INFO_MASK; |
| 211 | if (searchptr < forkptr && |
| 212 | XLogRecGetRmid(xlogreader) == RM_XLOG_ID && |
| 213 | (info == XLOG_CHECKPOINT_SHUTDOWN || |
| 214 | info == XLOG_CHECKPOINT_ONLINE)) |
| 215 | { |
| 216 | CheckPoint checkPoint; |
| 217 | |
| 218 | memcpy(&checkPoint, XLogRecGetData(xlogreader), sizeof(CheckPoint)); |
| 219 | *lastchkptrec = searchptr; |
| 220 | *lastchkpttli = checkPoint.ThisTimeLineID; |
| 221 | *lastchkptredo = checkPoint.redo; |
| 222 | break; |
| 223 | } |
| 224 | |
| 225 | /* Walk backwards to previous record. */ |
| 226 | searchptr = record->xl_prev; |
| 227 | } |
| 228 | |
| 229 | XLogReaderFree(xlogreader); |
| 230 | if (xlogreadfd != -1) |
| 231 | { |
| 232 | close(xlogreadfd); |
| 233 | xlogreadfd = -1; |
| 234 | } |
| 235 | } |
| 236 | |
| 237 | /* XLogreader callback function, to read a WAL page */ |
| 238 | static int |
| 239 | (XLogReaderState *xlogreader, XLogRecPtr targetPagePtr, |
| 240 | int reqLen, XLogRecPtr targetRecPtr, char *readBuf, |
| 241 | TimeLineID *pageTLI) |
| 242 | { |
| 243 | XLogPageReadPrivate *private = (XLogPageReadPrivate *) xlogreader->private_data; |
| 244 | uint32 targetPageOff; |
| 245 | XLogRecPtr targetSegEnd; |
| 246 | XLogSegNo targetSegNo; |
| 247 | int r; |
| 248 | |
| 249 | XLByteToSeg(targetPagePtr, targetSegNo, WalSegSz); |
| 250 | XLogSegNoOffsetToRecPtr(targetSegNo + 1, 0, WalSegSz, targetSegEnd); |
| 251 | targetPageOff = XLogSegmentOffset(targetPagePtr, WalSegSz); |
| 252 | |
| 253 | /* |
| 254 | * See if we need to switch to a new segment because the requested record |
| 255 | * is not in the currently open one. |
| 256 | */ |
| 257 | if (xlogreadfd >= 0 && |
| 258 | !XLByteInSeg(targetPagePtr, xlogreadsegno, WalSegSz)) |
| 259 | { |
| 260 | close(xlogreadfd); |
| 261 | xlogreadfd = -1; |
| 262 | } |
| 263 | |
| 264 | XLByteToSeg(targetPagePtr, xlogreadsegno, WalSegSz); |
| 265 | |
| 266 | if (xlogreadfd < 0) |
| 267 | { |
| 268 | char xlogfname[MAXFNAMELEN]; |
| 269 | |
| 270 | /* |
| 271 | * Since incomplete segments are copied into next timelines, switch to |
| 272 | * the timeline holding the required segment. Assuming this scan can |
| 273 | * be done both forward and backward, consider also switching timeline |
| 274 | * accordingly. |
| 275 | */ |
| 276 | while (private->tliIndex < targetNentries - 1 && |
| 277 | targetHistory[private->tliIndex].end < targetSegEnd) |
| 278 | private->tliIndex++; |
| 279 | while (private->tliIndex > 0 && |
| 280 | targetHistory[private->tliIndex].begin >= targetSegEnd) |
| 281 | private->tliIndex--; |
| 282 | |
| 283 | XLogFileName(xlogfname, targetHistory[private->tliIndex].tli, |
| 284 | xlogreadsegno, WalSegSz); |
| 285 | |
| 286 | snprintf(xlogfpath, MAXPGPATH, "%s/" XLOGDIR "/%s" , private->datadir, xlogfname); |
| 287 | |
| 288 | xlogreadfd = open(xlogfpath, O_RDONLY | PG_BINARY, 0); |
| 289 | |
| 290 | if (xlogreadfd < 0) |
| 291 | { |
| 292 | pg_log_error("could not open file \"%s\": %m" , xlogfpath); |
| 293 | return -1; |
| 294 | } |
| 295 | } |
| 296 | |
| 297 | /* |
| 298 | * At this point, we have the right segment open. |
| 299 | */ |
| 300 | Assert(xlogreadfd != -1); |
| 301 | |
| 302 | /* Read the requested page */ |
| 303 | if (lseek(xlogreadfd, (off_t) targetPageOff, SEEK_SET) < 0) |
| 304 | { |
| 305 | pg_log_error("could not seek in file \"%s\": %m" , xlogfpath); |
| 306 | return -1; |
| 307 | } |
| 308 | |
| 309 | |
| 310 | r = read(xlogreadfd, readBuf, XLOG_BLCKSZ); |
| 311 | if (r != XLOG_BLCKSZ) |
| 312 | { |
| 313 | if (r < 0) |
| 314 | pg_log_error("could not read file \"%s\": %m" , xlogfpath); |
| 315 | else |
| 316 | pg_log_error("could not read file \"%s\": read %d of %zu" , |
| 317 | xlogfpath, r, (Size) XLOG_BLCKSZ); |
| 318 | |
| 319 | return -1; |
| 320 | } |
| 321 | |
| 322 | Assert(targetSegNo == xlogreadsegno); |
| 323 | |
| 324 | *pageTLI = targetHistory[private->tliIndex].tli; |
| 325 | return XLOG_BLCKSZ; |
| 326 | } |
| 327 | |
| 328 | /* |
| 329 | * Extract information on which blocks the current record modifies. |
| 330 | */ |
| 331 | static void |
| 332 | (XLogReaderState *record) |
| 333 | { |
| 334 | int block_id; |
| 335 | RmgrId rmid = XLogRecGetRmid(record); |
| 336 | uint8 info = XLogRecGetInfo(record); |
| 337 | uint8 rminfo = info & ~XLR_INFO_MASK; |
| 338 | |
| 339 | /* Is this a special record type that I recognize? */ |
| 340 | |
| 341 | if (rmid == RM_DBASE_ID && rminfo == XLOG_DBASE_CREATE) |
| 342 | { |
| 343 | /* |
| 344 | * New databases can be safely ignored. It won't be present in the |
| 345 | * source system, so it will be deleted. There's one corner-case, |
| 346 | * though: if a new, different, database is also created in the source |
| 347 | * system, we'll see that the files already exist and not copy them. |
| 348 | * That's OK, though; WAL replay of creating the new database, from |
| 349 | * the source systems's WAL, will re-copy the new database, |
| 350 | * overwriting the database created in the target system. |
| 351 | */ |
| 352 | } |
| 353 | else if (rmid == RM_DBASE_ID && rminfo == XLOG_DBASE_DROP) |
| 354 | { |
| 355 | /* |
| 356 | * An existing database was dropped. We'll see that the files don't |
| 357 | * exist in the target data dir, and copy them in toto from the source |
| 358 | * system. No need to do anything special here. |
| 359 | */ |
| 360 | } |
| 361 | else if (rmid == RM_SMGR_ID && rminfo == XLOG_SMGR_CREATE) |
| 362 | { |
| 363 | /* |
| 364 | * We can safely ignore these. The file will be removed from the |
| 365 | * target, if it doesn't exist in source system. If a file with same |
| 366 | * name is created in source system, too, there will be WAL records |
| 367 | * for all the blocks in it. |
| 368 | */ |
| 369 | } |
| 370 | else if (rmid == RM_SMGR_ID && rminfo == XLOG_SMGR_TRUNCATE) |
| 371 | { |
| 372 | /* |
| 373 | * We can safely ignore these. When we compare the sizes later on, |
| 374 | * we'll notice that they differ, and copy the missing tail from |
| 375 | * source system. |
| 376 | */ |
| 377 | } |
| 378 | else if (info & XLR_SPECIAL_REL_UPDATE) |
| 379 | { |
| 380 | /* |
| 381 | * This record type modifies a relation file in some special way, but |
| 382 | * we don't recognize the type. That's bad - we don't know how to |
| 383 | * track that change. |
| 384 | */ |
| 385 | pg_fatal("WAL record modifies a relation, but record type is not recognized: " |
| 386 | "lsn: %X/%X, rmgr: %s, info: %02X" , |
| 387 | (uint32) (record->ReadRecPtr >> 32), (uint32) (record->ReadRecPtr), |
| 388 | RmgrNames[rmid], info); |
| 389 | } |
| 390 | |
| 391 | for (block_id = 0; block_id <= record->max_block_id; block_id++) |
| 392 | { |
| 393 | RelFileNode rnode; |
| 394 | ForkNumber forknum; |
| 395 | BlockNumber blkno; |
| 396 | |
| 397 | if (!XLogRecGetBlockTag(record, block_id, &rnode, &forknum, &blkno)) |
| 398 | continue; |
| 399 | |
| 400 | /* We only care about the main fork; others are copied in toto */ |
| 401 | if (forknum != MAIN_FORKNUM) |
| 402 | continue; |
| 403 | |
| 404 | process_block_change(forknum, rnode, blkno); |
| 405 | } |
| 406 | } |
| 407 | |